Академический Документы
Профессиональный Документы
Культура Документы
For photocopying of material in this volume, please pay a copying fee through the Copyright
Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to
photocopy is not required from the publisher.
Printed in Singapore.
PREFACE
Asset allocation investigates the optimal division of a portfolio among different asset
classes. Standard theory involves the optimal mix of risky stocks, bonds, and cash
together with various subdivisions of these asset classes. Underlying this is the insight
that diversification allows for achieving a balance between risk and return: by using
different types of investment, losses may be limited and returns are made less volatile
without losing too much potential gain.
These insights are made precise using the benchmark theory of mathematical
finance, the Black-Scholes-Merton theory, based on Brownian motion as the driving
noise process for risky asset prices. Here, the distributions of financial returns of the
risky assets in a portfolio are multivariate normal, thus relating to the standard mean-
variance portfolio theory of Markowitz with its risk-return paradigm as above.
Recent years have seen many empirical studies shedding doubt on the Black-
Scholes-Merton model, and motivating various alternative modeling approaches,
which were able to reproduce the stylized facts of asset returns (such as heavy tails and
volatility clustering) much better. Also, various new asset classes and specific financial
tools for achieving better diversification have been created and entered the investment
universe.
This book combines academic research and practical expertise on these new (often
called alternative) assets and trading strategies in a unique way. We include the prac-
titioners’ viewpoint on new asset classes as well as academic research on modeling
approaches, for new asset classes. In particular, alternative asset classes such as power
forward contracts, forward freight agreements, and investment in photovoltaic facil-
ities are discussed in detail, both on a stand-alone basis and with a view to their
effects on diversification in combination with classical asset. We also analyse credit-
related portfolio instruments and their effect in achieving an optimal asset allocation.
In this context, we highlight aspects of financial structures which may sometimes be
neglected, such as default risk of issuer in case of certificates or the role that model
v
May 13, 2010 10:6 WSPC/SPI-B913 b913-fm FA
vi Preface
risk plays within asset allocation problems. This leads naturally to the use of robust
asset allocation strategies.
Extending the classical mean-variance portfolio setting, we include dynamic port-
folio strategies and illustrate different portfolio protection strategies. In particular, we
compare the benefits of such strategies and investigate conditions under which Con-
stant Proportion Portfolio Insurance (CPPI) may be prefered to Option-Based Portfolio
Insurance (OBPI) and vice versa. We also contribute to the understanding of gap risk
by analyzing this risk for CPPI and Constant Proportion Debt Obligations (CPDO) in
a sophisticated modeling framework. Such analyses are supplemented and extended
by an investigation of the optimality of hedging approaches such as variance-optimal
hedging and semistatic variants of classical hedging strategies.
Many of the articles can serve as guides for the implementation of various models.
In addition, we also present state-of-the-art models and explain modern tools from
financial mathematics, such as Markov-Switching models, time-changed Lévy models,
variants of lognormal approximations, and copula structures.
This books combines a unique mix of authors. Also many of our students improved
the outcome of the project with critical and insightful comments. Particular thanks goes
to Georg Grüll, Peter Hieber, Julia Kraus, Matthias Lutz, Jan-Frederik Mai, Kathrin
Maul, Kevin Metka, Daniela Neykova, Johannes Rauch, Andreas Rupp, Daniela Selch,
and Christofer Vogt.
CONTENTS
Preface v
vii
May 13, 2010 10:6 WSPC/SPI-B913 b913-fm FA
viii Contents
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.2 Overview on Alternative Real Assets . . . . . . . . . . . . . . . . . . . . 52
3.3 Modeling Photovoltaic Investments . . . . . . . . . . . . . . . . . . . . 53
3.3.1 General Approach . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.3.2 Definition of the Investment Project . . . . . . . . . . . . . . . . 54
May 13, 2010 10:6 WSPC/SPI-B913 b913-fm FA
Contents ix
x Contents
Contents xi
xii Contents
Chapter 11. Portfolio Insurances, CPPI and CPDO, Truth or Illusion? 259
Elisabeth Joossens and Wim Schoutens
11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
11.2 Credit Risk and Credit Default Swaps . . . . . . . . . . . . . . . . . . . 261
11.2.1 Credit Risk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
11.2.2 Credit Default Swaps (CDS) . . . . . . . . . . . . . . . . . . . . 265
11.3 Portfolio Insurances . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
May 13, 2010 10:6 WSPC/SPI-B913 b913-fm FA
Contents xiii
Chapter 12. On the Benefits of Robust Asset Allocation for CPPI Strategies 295
Katrin Schöttle and Ralf Werner
12.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296
12.2 The Financial Market . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296
12.2.1 The Basic Financial Market . . . . . . . . . . . . . . . . . . . . 297
12.2.2 The Riskless Asset . . . . . . . . . . . . . . . . . . . . . . . . . 298
12.2.3 The Risky Asset . . . . . . . . . . . . . . . . . . . . . . . . . . 298
12.2.4 Classical Mean–Variance Analysis . . . . . . . . . . . . . . . . 300
12.2.5 The Trading Strategy . . . . . . . . . . . . . . . . . . . . . . . . 302
12.3 The Standard CPPI Strategy . . . . . . . . . . . . . . . . . . . . . . . . 302
12.3.1 The Simple Case . . . . . . . . . . . . . . . . . . . . . . . . . . 303
12.3.2 The General Case . . . . . . . . . . . . . . . . . . . . . . . . . 305
12.3.3 Shortfall Probability of CPPI Strategies . . . . . . . . . . . . . . 308
12.3.4 Improving CPPI Strategies . . . . . . . . . . . . . . . . . . . . . 310
12.3.5 CPPI Strategies Under Estimation Risk . . . . . . . . . . . . . . 313
12.4 Robust Mean–Variance Optimization and Improved CPPI Strategies . . . 316
12.4.1 Robust Mean–Variance Analysis . . . . . . . . . . . . . . . . . . 317
12.4.2 Uncertainty Sets Via Expert Opinions or Related Estimators . . . 317
May 13, 2010 10:6 WSPC/SPI-B913 b913-fm FA
xiv Contents
Contents xv
Index 395
This page intentionally left blank
May 12, 2010 17:46 WSPC/SPI-B913 b913-ch01 FA
Part I
Alternative Investments
May 12, 2010 17:46 WSPC/SPI-B913 b913-ch01 FA
SOCIALLY RESPONSIBLE
INVESTMENTS
1
SVEN HROß∗ , CHRISTOFER VOGT† and RUDI ZAGST‡
HVB-Stiftungsinstitut für Finanzmathematik,
Technische Universität München, Boltzmannstr. 3,
85747 München, Germany
∗
shross@gmx.de
†
vogtchri@aol.com
‡
zagst@tum.de
Within the last two decades, the market of socially responsible investing (SRI) has seen
unprecedented growth and has become more and more important, not only because of
the current financial crisis. This chapter gives a survey of the asset class SRI in general,
i.e., market development and investment possibilities. Moreover, the question “How
sustainable is sustainability?” is addressed by analyzing SAM Group sustainability
rankings of the years 2001–2007. Furthermore, the ability of SRI to contribute to
diversification within a portfolio is scrutinized. The analysis is based on simulated
returns generated by an autoregressive Markov-Switching model and accounts for
different levels of investors’ risk aversion. Optimal portfolios consisting of stocks,
bonds, and the respective SRI index show that risk–averse investors mix SRI to an
established portfolio consisting of bonds and stocks to reduce the risk and increase
the performance. Additionally, the asset class SRI is found to be a substitute for the
asset class stocks.
3
May 12, 2010 17:46 WSPC/SPI-B913 b913-ch01 FA
4 Hroß et al.
1.1. INTRODUCTION
There are different ways to describe socially responsible investing (SRI). Reference 1
defines SRI as the integration of environmental, social, and corporate governance
(ESG) considerations into investment management processes and ownership practices
hoping that these factors can have an impact on financial performance. Responsible
investment can be practiced across all asset classes.
Several reasons can be stated, why the field of SRI has gained great public interest
as well as rising economic importance in recent years. Simultaneously to the on-going
climate change debate, public scrutiny and political attention have put pressure on
businesses to consider both social and environmental issues in their activities.
Accompanied by these developments, the SRI market grew strongly during the last
decade. SRI does no longer represent a negligible economical niche, but as stated in [2]
it might play a crucial financial role in the future. The current size of the worldwide
SRI market is according to [3] approximately 5 trillion. With 53% market share, the
greatest part of the SRI market is based in Europe followed by the United States with
39%. The rest of the world represents only 8% of the SRI market.
According to [4], the size of the SRI market in the United States was $639 bil-
lion in 1995 and then grew up to $2159 billion in 1999, which means an average
annual growth rate of 36%. From 1999 to 2005, SRI investment volumes only slightly
grew up to $2290 billion, but then growth accelerated again resulting in $2711 billion
in 2007.
The European SRI market experienced an average growth rate of 51% since 2002
from an absolute investment volume of 336 billion in 2002 up to 2665 billion in
2007. Reference 3 estimates that the share of SRI in the total European fund market is
about 17.6% in 2008 and largely driven by institutional investors.
There are several possibilities to invest into SRI. For example, the SAM Group
(www.sam-group.com) offers a wide range of funds covering the total SRI market and
also special funds, e.g., on Islamic sustainability. There are also sustainably managed
fixed-income funds available. Another possibility is the direct investment into non-
listed companies or projects. In this context, projects like wind farms or solar parks can
be mentioned as suitable investment possibilities. Moreover, certificates are available
on the market which allow the investor to participate in the SRI market, e.g., index
certificates on the European Renewable Energy Index (ERIX Index Certificate, Societe
Generale, ISIN: DE000SG1ERX7).
The structure of this chapter is as follows. Section 1.2 gives an overview on recent
research on SRI. Section 1.3 answers the question “How sustainable is sustainability?”
by using Markov transition matrices. Section 1.4 then analyzes SRI in a portfolio con-
text by generating optimal portfolios for different investors using a Markov-Switching
model and different optimization frameworks. Finally, Sec. 1.5 concludes.
May 12, 2010 17:46 WSPC/SPI-B913 b913-ch01 FA
During the last years, several empirical studies analyzed whether SRI produces or
destroys shareholder wealth. Many early studies on the performance of SRI use regres-
sion models with one or two factors and try to measure Jensen’s alpha. Reference 5
compares 32 SRI funds to 320 non-SRI funds in the United States between 1981 and
1990 and finds no significant average alphas with respect to a value-weighted NYSE
index. More advanced studies apply a matching approach to compare SRI and non-SRI
funds with similar characteristics, e.g., fund universe and size. Within this approach,
management and transaction costs can be included into the analysis, see, e.g., [6]
or [7]. As a result, no significant performance differences between SRI and non-SRI
could be observed. One problem is that important characteristics might not be taken
into consideration. Reference 8 applies a four factor model according to [9] using as
regression factors the excess market return, SMB (“Small-minus-Big”: The difference
between the return of a small- and of a large-cap portfolio), HML (“High-minus-Low”:
The return difference between a value- and a growth-portfolio, i.e., a portfolio con-
taining firms that dispose of a high book-to-market ratio versus firms with a low value
relating to this ratio), and MOM (“Momentum”: The return difference betweeen two
portfolios, one consisting of last year’s best performers and the other of the worst
performers) in order to analyze the performance of United States, German, and British
SRI funds. The authors build two portfolios for each country, one containing all SRI
funds, the other the conventional funds, and find under — as well as outperformance of
SRI, but none of the differences are significant. Furthermore, SRI funds seem to have
an investment bias toward growth stocks (low book-to-market value) and small caps
(lower market-capitalization). Reference 10 uses eco-efficiency rankings of Innovest
to evaluate two equity portfolios that differ in eco-efficiency. The high-ranked portfo-
lio shows significantly higher returns than its low-ranked counterpart over the period
1995–2003. In contrast, [11] finds that SRI investors have to pay for their constrained
investment style. Another approach is to look at SRI equity indices to avoid usual prob-
lems of mutual funds during a performance analysis, e.g., transaction costs of funds
or effects of management skills. Reference 12 analyzes 29 SRI indices and applies
different settings to test for differences in risk-adjusted performance compared to a
suitable benchmark. The study concludes that SRI screens do not lead to significant
performance difference of SRI indices. Yet, no final answer to the question whether
SRI produces or destroys shareholder wealth can be given. Independent of these find-
ings, SRI market growth might simply come from the non-financial utility gained by
SRI investors. To the authors’ best knowledge, there is yet no such study scrutinizing
this effect. Therefore, the focus of Sec. 1.4 lies on the benefits of SRI in a portfolio
context. For this, optimal portfolios of bonds, stocks, and SRI will be constructed for
different investor types and in different optimization frameworks.
May 12, 2010 17:46 WSPC/SPI-B913 b913-ch01 FA
6 Hroß et al.
counter-intuitive, due to the fact that sustainability is a long-term affair and thus should
not be subject to large-sized jumps, unless extraordinary events occur, e.g., an envi-
ronmental disaster on an oil producer’s platform. For the following analysis, data of
those companies are used for which the sustainability scores are available for two
consecutive years. For the entire six-year time period, this leads to a total dataset of
2125 observations. The calculation of the transition matrices is performed as follows:
For every single year, companies are ranked by their sustainability score, whereby for
every year the 25% best rated companies are assigned to the 1st quartile, the next 25%
to the 2nd quartile, and so on. Based on this allocation, empirical transition probabil-
ities from one of the four quartiles to any of the four quartiles after one year can be
calculated.
8 Hroß et al.
Altogether, the results provide evidence to the assumption that sustainability rank-
ings do not have a high degree of short-term variation.
After having analyzed the sustainability of sustainability in the preceding section, this
section will scrutinize how SRI can be evaluated with regard to the portfolio context.
The main questions to be answered are whether investors shall add SRI investments
to their portfolio, and if so, with which weighting.
In the conducted portfolio case study, the SRI market is represented by the
Advanced Sustainable Performance Index (ASPI). The ASPI is a European index con-
sisting of 120 companies and is published by Vigeo Group, an extra-financial supplier
and rating agency in the field of sustainable development and social responsibility
(for further information see [13]). In order to include dividend payments to the anal-
ysis, total return indices are used, i.e., dividends are reinvested. This approach has
two main advantages. First, the index already represents a selected basket of the asset
category SRI and the time series are readily available. Second, the predefined index is
widespread and thus has the advantage that the companies’ specific risks are already
eliminated by diversification. As a result, only the diversification effect of the asset
class SRI itself is observed.
By comparing mean and standard deviation of bonds and stocks, it becomes evi-
dent that most of the risk–averse investors would invest the bulk of their wealth in
bonds. This is due to the extremely high mean for bonds (5.92% per annum) combined
with a low standard deviation. Additionally, bonds display the highest skewness and
lowest excess kurtosis, which is generally preferred by risk–averse investors. As it is
a debatable point whether past returns indicate the future in a sufficient way, experts’
forecasts about expected returns are often used to solve this shortcoming. By using the
Black–Litterman approach to adjust the empirical returns, the empirical mean µemp
itself as well as absolute and relative forecasts are taken into account (see, e.g., [14]).
This approach can be interpreted as a linear combination of these two components at
a given confidence level τ regarding the forecasts. The Black–Litterman expectations
µBL can be expressed by (given that L is invertible)
µBL = τ · L−1 q + (1 − τ) · µemp , (1.1)
where L represents the linear transformation Lµ of the asset classes expected return
vector µ and for each forecast, whose actual value is specified in q. The assumptions
about the forecasts are taken from [15]. This means that the annual return of bonds is
expected to be 3.96% and the equity risk premium amounts to 3.5%. The additional
assumption that the difference of the means of stocks and SRI does not change leads to
1 0 0 0.0396/250
L = −1 1 0 and q = 0.0350/250 .
0 −1 1 0.0303/250
For τ = 0, the Black–Litterman expectations are equal to the empirical means,
while τ = 1 leads to expectations which are completely driven by the forecasts.
Table 1.5 provides the Black–Litterman expectations for a confidence level of τ = 0.75.
May 12, 2010 17:46 WSPC/SPI-B913 b913-ch01 FA
10 Hroß et al.
The empirical returns are adjusted for the Black–Litterman expectations by apply-
ing a linear shift to the whole dataset. Hence, all other empirical statistics (except for
the autocorrelation in squared returns) are unaffected.
Skewness and excess kurtosis in Table 1.4 lead to the presumption that the returns
of all three asset classes are non-normally distributed. To test for non-normality and
autocorrelation, a Jarque–Bera test and a Ljung–Box-Q test (see [16] and [17]) were
applied. The p-values of both tests are given in Table 1.6. Italicised values are those
smaller than 0.05, for which the null hypothesis can be rejected at a significance level
of 5%.
As correlations are essential for diversification in a portfolio context, the corre-
lations of the empirical daily log-returns are listed in Table 1.7. The high correlation
between stocks and SRI is in line with the findings of [12]. One reason for this fact
may be that SRI stocks are simply a subset of the whole stock universe. Nevertheless,
each correlation coefficient smaller than one allows to benefit from diversification.
12 Hroß et al.
The parameters allow for the derivation of crucial information about the two
possible meta states. State 1 characterizes a bull market for all three assets. The expected
return of stocks and SRI is almost the same (15.4% p.a. for stocks and 16% p.a. for
SRI) whereas the standard deviation is much higher for SRI. In state 2, the expected
return of bonds only suffers a small decline compared to state 1 (from 4.6% p.a. to
3.8% p.a.) but the volatility nearly doubles in this regime. For stocks and SRI state 2
resembles a bearish market with huge losses for both asset classes (−26.9% p.a. for
stocks and −13.9% p.a. for SRI) and high standard deviations. The observation of an
increasing volatility in falling markets can be found in empirical studies like [20]. The
transition probabilities p12 and p21 of the Markov chain imply a realistic stability of
the two possible states. If the market is in a bullish or a bearish scenario respectively,
the market remains in the current state on the following day with a high probability.
As the fitting is based on minimizing the sum of squared deviations between the
empirical and theoretical statistics, these are outlined in Table 1.9.
The close match is emphasized by the theoretical correlations of the model, which
fit the empirical correlations very well (compare Tables 1.7 and 1.10).
It is worthwhile to have a look at the correlation structures of the error terms t ,
which are depicted in Table 1.11. The correlation of the error terms between stocks and
SRI is about 0.9 in the “bad” state (state 2) and thus much higher than in the “good”
state (state 1) with about 0.47.
State 1
State 2
14 Hroß et al.
0.11
Bonds
0.10 Stocks
SRI
0.09 Bonds, stocks
Expected return
0.07
0.06
0.05
0.04
0.00 0.02 0.04 0.06 0.08 0.10 0.12 0.14 0.16 0.18 0.20
Standard deviation
The traditional mean–variance framework based on [21] only takes the first two
moments of the return distribution into account. It is defined by
λ T
x x, max xT µ − (1.3)
2 x
where µ denotes the vector of expected asset returns, the return covariance–matrix,
and λ the risk–aversion parameter of the investor.
Figure 1.1 shows the diversification effect which can be obtained by including the
asset SRI into an existing portfolio of bonds and stocks.
Due to the importance of higher moments within the portfolio context, the power-
utility model and frameworks which maximize different performance measures are
introduced in the following.
The power-utility model is defined by the optimization problem
max E[γ −1 (1 + R(x))γ ], γ < 1 ∧ γ = 0
max E[U(R(x))] = x
, (1.4)
x max E[ln(1 + R(x))], γ=0
x
where γ describes the risk aversion of the investor and R(x) is the portfolio return
depending on the portfolio weights x.
The third optimization model maximizes the performance measure
introduced
in [22]. In this framework, each investor can set a threshold to classify the returns
into gains (returns above the threshold) and losses (returns below the threshold).
is
May 12, 2010 17:46 WSPC/SPI-B913 b913-ch01 FA
defined as the ratio of probability-weighted gains to losses and thus equal to the upside
potential divided by the downside potential. The corresponding optimization problem
is given by
E[R(x) − τ]+
max
τ (R(x)) = max , (1.5)
x x E[τ − R(x)]−
where τ is the loss threshold and R(x) the portfolio return for portfolio weights x.
The performance measure Score-value considers the difference of upside and
downside potential. The risk-free rate r is used as threshold and the downside potential
is weighted with a risk–aversion parameter λSc . This leads to an optimization problem
defined by
max ScoreλSc (R(x)) = max E[R(x) − r]+ − λSc · E[r − R(x)]− . (1.6)
x x
The Mean-Conditional Value at Risk (MCVaR) is a risk measure referring to the tail
of a distribution. It is based on the Conditional Value at Risk (CVaR) defined by
CVaR(R(x)) = E[R(x)|R(x) < VaR(R(x))], (1.7)
where VaR(R(x)) is the Value at Risk (see, e.g., [23]) of the portfolio return R(x) at
a given confidence level α (in the case under consideration α = 99.5%). In order to
consider both risk and return, the optimization problem is given by
max MCVaRλMCVaR (R(x)) = max E[R(x)] − λMCVaR CVaR(R(x)) (1.8)
x x
16 Hroß et al.
Table 1.13 Weights and Performance Measures of Optimal Portfolios (One Year).
Portfolio weights
are presented in detail here. Table 1.13 shows the optimal portfolio weights as well
as the computed performance measures
, Score-value, MCVaR, and Sharpe ratio.
The first three were already introduced in Sec. 1.4.5. The well-known Sharpe ratio
measures the expected excess return over the riskless investment in units of standard
May 12, 2010 17:46 WSPC/SPI-B913 b913-ch01 FA
deviation (see [24]). The low risk–averse Investor C invests largely into SRI. In the
mean-variance (MV), the power-utility (PU), and the Omega framework, he allocates
around 90% into SRI and in the Score-value and the MCVaR framework he is com-
pletely invested in SRI. As Investor C does not invest in stocks at all, the remainder
of his wealth is invested in bonds. The strategy of Investor B is very similar with the
only difference being that the portfolio weights of the asset class SRI are smaller in the
MV, PU, and Omega framework. The best diversified portfolios are held by the highly
risk–averse Investor A. Most of his wealth — between 61% and 68% — is invested
into bonds. The majority of the remainder is allocated to SRI, leading to a fraction of
25–35% and only a small amount is invested into stocks (between 3–8%). Overall, the
results emphasize that SRI can be interpreted as a substitute for stocks and that only
the most risk–averse investor allocates money to the substituted asset class in order to
benefit from the diversification effect.
Figures 1.2 (a)–(e) illustrate the portfolio allocations in dependence of the respec-
tive risk–aversion parameter. The vertical lines indicate the three considered investor
types. The solid line displays the risk of the portfolios measured by the standard
deviation of the portfolios’ returns. As mentioned above, the proportion invested into
stocks is due to the substitution effect always very small or even zero. A further inter-
esting issue is the development of the proportion invested into bonds. In the MV,
PU, and Omega framework, the fraction of bonds increases slowly with increasing
risk–aversion whereas it ascends late but steeply in the Score-value and the MCVaR
framework what results in an enormous reduction of risk from that point on.
In order to get more information about the substitution effect, a sensitivity analysis
is performed for the mean–variance framework with a one-year time horizon. As only
the most risk–averse investor allocates parts of his wealth to stocks, the question
arises under which circumstances the less risk–averse Investor C mixes stocks into
his portfolio. For this purpose, the Black–Litterman expectation for the annual return
of SRI is gradually decreased, starting from a value of 10.22% (see Table 1.5). With
the help of the overall probability π1 (see Sec. 1.4.4) the amount of reduction can
be split up for the two regimes. For state 1 the reduction is π1 · and for state 2 it
equals (1 − π1 ) · . While facing the same risk exposure, the upper hurdle rate for the
expected return is determined by a value of 9.27%. This value represents the highest
expected return of SRI for which Investor C still invests a small amount into stocks.
A further decrease of the expected return leads to a partial replacement of SRI by
stocks. The lower hurdle rate is given by an annual SRI return of 7.01%, i.e., if the
expected return is below the lower hurdle rate Investor C will not invest in SRI anymore
and therefore create the appropriate benchmark portfolio of 70% stocks and 30%
bonds.
May 12, 2010 17:46 WSPC/SPI-B913 b913-ch01 FA
18 Hroß et al.
1.5. CONCLUSION
SRI is a growing asset class. By analyzing SAM sustainability scores, it was shown
that an SRI portfolio has a high degree of consistency, i.e., sustainable companies are
likely to stay sustainable in the future. For the best ranked companies, the probability
of being in the 1st quartile again in the next year is 70% on average. Moreover, optimal
May 12, 2010 17:46 WSPC/SPI-B913 b913-ch01 FA
portfolios for different investor types are constructed. The parameter estimates of the
underlying Markov-Switching model are based on a time series ranging from 1992
to 2008. The main finding of the conducted case study is that SRI turns out to be a
substitute for stocks and that the less risk–averse an investor is the more he invests in
SRI, respectively, does a full investment into this asset class. More risk–averse investors
use all three asset classes in order to gain from diversification effects. Nevertheless,
only a very small fraction (<8%) is invested into stocks.
References
[1] Mercer (2008). The language of socially responsible investing, an industry guide to key
terms and organizations (January, 2008).
[2] Lydenberg, S (2006). Envisioning socially responsible investing — a model for 2006.
Journal of Corporate Citizenship, 7, 57–77.
[3] European Social Investment Forum. European SRI study (2008).
[4] Renneboog, L, J Horst and C Zhang (2008). Socially responsible investments: Institutional
aspects, performance, and investor behaviour. Journal of Banking and Finance, 32(9),
1723–1742.
[5] Hamilton, S, H Jo and M Statman (1993). Doing well while doing good? The invest-
ment performance of socially responsible mutual funds. Financial Anaylsts Journal, 49(6),
62–66.
[6] Mallin, C, B Saadouni and R Briston (1995). The financial performance of ethical invest-
ment trusts. Journal of Business Finance & Accounting, 22, 483–496.
[7] Statman, M (2000). Socially responsible mutual funds. Financial Analysts Journal, 56,
30–39.
[8] Bauer, R, K Koedijk and R Otten (2005). International evidence on ethical mutual fund
performance and investment style. Journal of Banking and Finance, 29, 1751–1767.
[9] Carhart, M (1997). On the persistence in mutual fund performance. Journal of Finance,
52, 57–82.
[10] Derwall, J, N Guenster, R Bauer and K Koedijk (2005). The eco-efficiency premium puzzle.
Financial Analyst Journal, 61, 51–63.
[11] Geczy, C, R Stambaugh and D Levin (2005). Investing in socially responsible mutual
funds. Wharton School Working Paper.
[12] Schröder, M (2007). Is there a difference? The performance characteristics of SRI equity
indexes. Journal of Business Finance and Accounting, 34(1&2), 331–348.
[13] Vigeo (2008). Methodology and composition of the ASPI Eurozone. URL http://
www.vigeo.com/csr-rating-agency/en/nos-produits-isr/indice-aspi/indice-aspi.html.
[14] Black, F and R Litterman (1990). Asset allocation: Combining investor views with market
equilibrium. Goldman Sachs Fixed Income Research.
[15] Aigner, P, G Beyschlag, T Friederich, M Kalepky and R Zagst (2008). Optimal risk-return
profiles for portfolios including stocks, bonds, and listed private equity. Submitted for
publication.
[16] Jarque, C and A Bera (1980). Efficient tests for normality, homoskedasticity and serial
independence of regression residuals. Economic Letters 6, 6(3), 255–259.
May 12, 2010 17:46 WSPC/SPI-B913 b913-ch01 FA
20 Hroß et al.
[17] Ljung, G and G Box (1978). On a measure of lack of fit intime series models. Biometrika,
65(2), 297–303.
[18] Timmermann, A (2000). Moments of Markov switching models.
[19] Höcht, S, K-H Ng, J Wolf and R Zagst (2008). Optimal portfolio allocation with Asian
hedge funds and Asian REITs. International Journal of Services Sciences, 1(1), 36–68.
[20] French, K, W Schwert and R Stambaugh (1987). Expected stocks returns and volatility.
Journal of Financial Economics, 19, 3–29.
[21] Markowitz, H (1952). Portfolio selection. Journal of Finance, 7(1), 77–91.
[22] Shadwick, W and C Keating (2002). A universal performance measure. Journal of Perfor-
mance Measurement, 6(2), 59–84.
[23] Jorion, P (2000). Value at Risk: The New Benchmark for Managing Financial Risk, 2nd
Ed., McGraw-Hill.
[24] Sharpe, W (1964). Capital asset prices, a theory of market equilibrium under conditions
of risk. Journal of Finance, 19, 77–91.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch02 FA
21
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch02 FA
22 Aigner et al.
2.1. INTRODUCTION
Private equity investments have probably been ventured since men try to augment
their personal wealth. The voyage of Christopher Columbus, who received funding
from various noblemen for his supposed expedition to Asia, is often referred to as a
well-known example of the historical roots of this asset class. By the middle of the
last century, individual financiers, affectionately dubbed business angels, crossed the
threshold to the present-day notion of private equity investment: They granted money
along with management support to high-tech start-up companies and actively sup-
ported young entrepreneurs in order to maximize the value of their own investment. In
the 1970s, investment constraints previously imposed on institutional investors were
relaxed by providing for vast additional capital channels. Thus, the venture capital
market appeared significantly altered by the late 1980s: Not only high-tech start-up
companies engrossed the investors’ attention, yet more established enterprises and
firms not having technology-related products in their portfolio came into financial
support as well, and debt financing grew more and more important. A profuse array
of various funding practices such as restructuring, buyout, or mezzanine emerged.
Together with the traditional venture capital, they evolved to an asset class known
as private equity, which enjoyed enormous growth and returns over the last five
years.
With the credit crisis followed by an overall economic downturn becoming the
prominent topic in 2008, one may raise the question whether private equity is still an
asset class worthwhile considering. And indeed, both columnists and scientists forecast
great changes which the industry is about to experience: Since the credit markets have
basically come to halt, procuring debt for leveraged buyout transactions was virtually
impossible, resulting deal volumes to plummet by 75% in 2008. Default probabilities
for portfolio companies may be as high as 50% and consequently, most institutional
investors have been reducing the exposure to private equity, see [16].
However, private equity still displays characteristics, which make it an interesting
part of a well-diversified investment strategy: Since private equity tends to display
only a moderate correlation to the public equity market, it may extend the portfolio’s
efficient frontier, and thus, yield a better risk–return profile. Furthermore, the persis-
tence in performance of top-quartile private equity firms has also proven valid in past
crises and is likely to do so in the future, see [1, 5, 21]. Private equity funds might
even play a key role in the economic recovery process. Currently, the private equity
industry holds $450 billion waiting to be invested [16]; thus, besides governments
and sovereign wealth funds, they are the only source for liquidity to an ailing econ-
omy. Since private equity funds tend to generate even greater returns when acquiring
portfolio companies during grave economic periods, now might be the right timing to
increase one’s exposure to private equity vehicles, see [1].
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch02 FA
The remainder of this article is structured as follows: First the basic ideas and
different categories concerning private equity are delineated, then various investment
vehicles are expounded, and finally means of determining the optimal fraction of listed
private equity entities in a portfolio are developed.
Delineating the notion of “private equity” is certainly not facile. In general, private
equity is a broad term that refers to any type of equity investment in an asset in
which the equity is not freely tradable on a public stock market. Furthermore, a private
equity investment usually entails a rather long engagement period, although it is always
temporary with an exit strategy sometimes already existing at the time of the investment
being ventured. Before one can probe into the various classes of private equity (see
Sec. 2.3) it is pertinent to scrutinize the criteria which distinguish them.
Early Stage Financing. Early stage financing comprises seed financing, start-up
financing, and first stage financing. During the seed phase solely a conception of the
business and the product exist, with the founder himself or some acquaintances of
his raising the money required. The start-up period is dedicated to the composition
of a business plan while the financial burden is still primarily imposed on the previ-
ously mentioned persons. With the first stage commencing, the entrepreneur launches
production and selling of his product. These first revenues, however, are easily out-
numbered by the increasing expenditures.
Expansion Stage Financing. During the expansion stage, the business strives for
continuous growth and attempts to break even; then, an expansion to new markets
and a diversification of the array of products is aimed at, and the sales and distribution
process is adapted to this development. Compared with the early stage, the risk involved
is lowered significantly.
Late Stage Financing. With the late stage commencing, the business is usually fairly
well established. According to its situation and its prospects, there are various possi-
bilities to be proceeded with. One might be the alignment of the enterprise to a future
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch02 FA
24 Aigner et al.
IPO, i.e., initial public offering of its share at a stock market; doing this, bridge financ-
ing provides for an ameliorated equity ratio. A financing aimed at changing the capital
structure of the portfolio company is also called “recapitalization” or “consolidation.”
Another possible strategy might be a separation of the firm from its former owners.
This can be achieved by a management-buyout (MBO) with the former executives
becoming the new shareholders or by a management-buyin (MBI), where previously
unengaged investors assume both stakes and operational control. With the new owner
usually not being able to raise the money required for the purchase on their own,
buyout financing, which can be composed of equity, debt, and mezzanine products,
is required. If the debt ratio of the deal exceeds 50%, the expression of a leveraged
buyout applies, see [13].
Turnaround financing comes in handy if the enterprise struggles and feels inclined
toward augmenting the equity ratio to become attractive to debt capital providers again.
Funds specializing in this practice usually offer also consultant services to increase
the odds of a successful turnaround. This type of investment strategy is sometimes
referred to as “distressed financing.”
Trade Sale. Selling the investment to another corporation is called “trade sale.” As
the purchaser predominately operates in the same industry like the portfolio company,
he may benefit from strategic advantages when incorporating the venture to be sold into
his corporate structure and thus pays even a higher price than an IPO might generate.
Additionally, transaction fees are usually lower than in an IPO sale. Although the
portfolio company mostly thrives upon such an arrangement, its management might
be hesitant, as their own position and independence could be affected.
due to lofty yield expectations of the purchaser, the capital gains obtainable with a
secondary sale are rather limited while transaction costs tend to be fairly moderate.
Equity Financing. Equity is, as the expression “private equity” suggests, the most
common capital source of the private equity industry. An equity investor enjoys a
proportional claim to a company’s assets, which usually depends on his share of the
company’s book value. He is mostly entitled to vote at the shareholders’ meeting and
thus, wields influence on the company’s business strategy and its management. In
case of an insolvency, equity is characterized by its junior treatment in the liquidation
process. Equity investors only recover the invested capital after all claims resulting
from debt financing are satisfied. Therefore, they face the peril of not only losing the
invested principal, yet also might be obliged to inject fresh capital into the firm, if it has
a corporate structure rendering the stakeholder personally liable, see [4]. Evidently an
equity investment is deemed to be riskier (than debt financing), and thus the expected
return is higher due to the risk premium, which an equity investor expects. Owing
to these characteristics, equity is the preferred type of financing for rather young
companies, as it does not require a collateral which these companies mostly could not
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch02 FA
26 Aigner et al.
provide for. Equity is provided to the portfolio company either by purchasing shares
from the previous stakeholder, or raising “fresh” equity capital by the private equity
investor, or granting equity capital in the form of a dormant partnership.
Debt Financing. Opposed to equity, debt is provided to a company only for a pre-
determined and limited time. The debt investor’s claim is senior to claims of equity
providers in case of an insolvency. Therefore, he is not liable, and he only risks the
loss of the committed principal in case of a default. The debt investor’s return, i.e.,
the interest payments, are due at predefined, fixed dates regardless of the portfolio
company’s performance, which accounts for the name “fixed income investment.” As
the cash flows — apart from default situations — are predictable, the risk and, conse-
quently, also the return associated with debt financing is lower than in the equity case.
Debt is predominately a prevalent means of financing late-stage portfolio companies.
Furthermore, debt does not affect the ownership situation, which is often desired by
family-owned portfolio companies, and might grant tax-related benefits to the invested
firms, as it can deduct the interest payments from the taxable revenues.
28 Aigner et al.
company’s balance sheet and the acquired company’s free cash flow will be used to
repay it, see [4].
Although the sources of value generation for buyout deals are manifold, they can
be basically cut down to the following three aspects:
Bondholders. Investors holding bonds of the target company are the obvious losers
of a leveraged buyout deal: As previously mentioned, the debt of a buyout is mostly
collateralized with the target company’s assets; therefore, the existing bond holders
would have to “share” the target company with the upcoming bondholders in case
of a default. Thus, bond prices tend to plummet, as soon as a leveraged buyout is
announced.
Tax shields. Debt financing displays the great advantage that the interest the company
pays to creditors is a tax-deductible expense, while dividends, the “cost” of equity
capital, are not. The amount of tax payments which the company can avoid by incurring
debt is called “tax-shield.” Provided that the value of a company is unaffected by the
way it is financed (Modigliani–Miller Theorem [20]), the present value of the tax-
shield equals the increase in value of a firm when a private equity fund raises the
amount of debt in its target’s capital structure.
Incentives. Greater incentives for the company’s management and its employees
can account for yet another source of value generation as people need to increase
their efforts to come up with the cash needed for debt service. Furthermore, senior
managers are usually obliged to place a considerable amount of their personal wealth
in the company they lead.
When contemplating an investment in private equity, one should bear in mind that
there are manifold possibilities of obtaining exposure to this asset class each of which
are visualized in Fig. 2.1.
30 Aigner et al.
Limited Partner. The limited partners provide for the bulk of the committed capital
(∼99%) and are accordingly called “investor” or “sponsor.” The investment activity
of large institutional investors carries today’s industry. Therefore, the main players in
the market are pension funds, insurers, funds-of-funds, banks, government agencies,
corporate investors, foundations, and family offices.
Banks. Banks usually play an important role in leveraged buyout transactions as they
provide for the necessary amount of debt capital. The target company predominantly
serves as collateral for these loans (see Fig. 2.2).
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch02 FA
Management Fees. Management fees provide for a base compensation so the fund
manager can support ongoing activities of the fund. On the other hand, the management
team should not make any profits on those fees, as compensation reasonably ought
to be performance driven. Therefore, annual fees range between 1.5% and 2.5% of
the committed capital. Levying fees on a committed capital basis prevents the fund
managers from going “for volume instead of quality,” i.e., drawing down and investing
more capital than he can find remunerative investment opportunities, see [19].
Gross/Net Cash Flows. As illustrated in Fig. 2.3 one can differ between two types of
cash flows: (1) Deal cash flows occur between the private equity fund and its portfolio
companies; as carried interest and management fees are not yet deducted, they are
also referred to as “gross cash flows.” (2) Fund cash flows occur between the private
equity fund and the investor (limited partner). Carried interest and management fees
had already been deducted at that point; thus, they are also referred to as “net cash
flow” or net distributions.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch02 FA
32 Aigner et al.
2.3.4. Fund-of-Funds
2.3.4.1. Structure of a private equity fund-of-funds
However, the process of selecting a successful fund is not trivial either, as choosing
an apt general partner is crucial to the fund performance which differs immensely.
Therefore, the investor may want to outsource this activity to a professional, together
with all the other chores entailed by being a limited partner. Hence, a fund-of-funds can
be deemed a vehicle that pools a group of investors and uses the capital to assemble a
diversified portfolio of funds. The organization of funds-of-funds basically resembles
the structure of regular private equity funds (see Fig. 2.4).
2.3.4.2. Advantages
Diversification. The possibility of greater diversification is definitely one of the
greatest advantages of funds-of-funds, as various studies have shown that they perform
similarly to individual funds but with less pronounced extremes. With funds-of-funds
participating in up to 100 individual private equity funds averaging about 20 portfolio
companies, the investor can engage in about 2000 target enterprises. Furthermore, very
specific expertise is required when investing in new industries, emerging regions, or
financing stages. As an individual fund cannot specialize in every field, it is certainly
pertinent to get exposure to different private equity markets via a fund-of-funds.
Resources. Usually both institutional investors and individuals lack the expertise of
research, due diligence, ongoing monitoring, reporting, and administration of a private
equity fund investment.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch02 FA
Liquidity Management. The vast majority of private equity funds do not know
either the timing or the size of the distribution cash flows. However, they do follow the
pattern with cash out-flows during the first half and in-flows during the second half of
a fund’s lifespan. With a fund-of-funds having many individual funds in its portfolio,
the aggregate cash flows become more predictable and thus they can agree upon pre-
defined payout schedules, which especially favor the needs of retail customers.
2.3.4.3. Disadvantages
The greatest drawback associated with a fund-of-funds is the additional layer of fees
the investor is charged. Funds-of-funds charge a management fee customarily ranging
between 1.0% and 1.5% of the committed assets and additionally receive carried
interest once the hurdle rate is attained. However, funds-of-funds predominantly deduct
only 5% of the gains. Furthermore, investors typically have fewer possibilities of
wielding influence on the development of their money compared to a single-fund
investment. Beyond, funds-of-funds do not offer a tailor-made investment focus which
some investors may need in order to offset their underexposure to some countries or
industries.
34 Aigner et al.
to this asset class while maintaining a decent level of liquidity. As private equity
companies primarily hold the corporate structure of a limited liability partnership,
they are not publicly quoted. However, there are some previously or currently listed
companies whose core business is private equity. Furthermore, investment funds which
invest a predetermined share in private equity, and specially structured investment
vehicles which invest in private equity directly are to be found. Of course, one may
also purchase bonds issued by private equity companies or resort to a warrant with a
private equity company’s share as underlying value. Moreover, there are certificates
with a private equity index or a specific basket comprising private equity portfolios
with different investment focuses as underlying values.
LPX GmbH is a company based in Switzerland that specializes in constructing
indices based on listed private equity (LPE). It provides the global LPE index LPX50,
LPX Major Market, and LPX Composite; LPX GmbH also calculates and publishes
[17] regional indices such as the LPX Europe as well as style indices such as the LPX
Buyout, LPX Indirect, and the LPXVenture. For the portfolio calculations subsequently
conducted the LPX 50 is considered. The LPX 50 is a global index that consists of the
50 largest liquid LPE companies covered by LPX. The total return index (as opposed
to a price index) is applied here, which means that dividend payments are also included
by reinvesting them in stocks.
After a detailed description of the different types of private equity investments, the
benefits of private equity for asset allocation purposes are analyzed in the following.
Like every other alternative asset class, private equity is supposed to provide diversifi-
cation benefits, i.e., reducing the risk of a portfolio while keeping the expected return
at a high level, or, respectively, increasing the expected return while not bearing a
higher risk. Unfortunately, it is not easy to measure the performance of private equity
investments. This is due to the fact that — except for listed private equity vehicles —
there is usually only a very limited secondary market. Hence, although it is possible
to calculate the return of a private equity investment after a fund is liquidated, it is
difficult to find a fair value during the fund’s lifetime. However, these fair values are
necessary to identify the risk profile of private equity investments. Therefore, the sub-
sequent study is only based on listed private equity as it provides the only private equity
investments, which are liquid enough to have realistic price time series. The following
methodology and the results of this section are based on the study which is presented in
detail in [2].
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch02 FA
36 Aigner et al.
38 Aigner et al.
(e.g., in Brunner/Hafner [6]). Here, the Markov–Switching model allowing for state-
independent (first lag) autoregressive dynamics is used.
This Markov–Switching process is defined as follows:
with st denoting the state of markets at time t, µst ,a denoting the mean return of asset
a ∈ {1, 2, 3} (where 1 =ˆ bonds, 2 =ˆ stocks, 3 = ˆ LPE) in state st , εt = (εt,1 , εt,2 , εt,3 )T
the innovation at time t, εt ∼ N(0, st ), and the autocorrelation parameter φ satisfying
|φ| < 1.
There are two possible states: If the markets are in state 1 then st = 1, if they are
in state 2 then st = 2. The changes of the state st over time are modeled by a Markov
chain. The transition probabilities of changing from state i at time t − 1 into state j at
time t are given by the matrix:
p11 p12
P= , where P(st = j|st−1 = i) = pij .
p21 p22
with µ1 , µ2 ∈ R3 , φ ∈ (−1, 1)3 , p12 , p21 ∈ [0, 1] and 1 , 2 ∈ R3×3 the covariance
matrices for the two states 1 and 2.
This model is described in detail in [2]. Another possible way is to apply univariate
processes for each asset and then include the correlation structure between the asset
returns by generating multivariate normally distributed residuals. This would allow
for all possible combinations of different individual market states, where each asset
class can be in one of two states, and therefore has one drawback: As the correlation
would only be accounted for by fitting the residuals, the second big part which is able
to influence the correlation is neglected, which is the correlation occurring from the
state combination. Therefore, instead of modeling independent Markov chains, the
ones resorted to in this study are related to each other in the way that only two meta-
states are allowed for: st = 1 and st = 2, i.e., all assets are in the same state at each
point of time t. Of course, the more market states are taken into account, the better the
empirical moments and autocorrelations may be matched. However, the (quadratic)
increase of the parameter vector itself causes the fitting method to be less efficient.
Furthermore, overfitting problems are more likely to occur. Therefore, as few states as
possible should be used. In this case, two market states already enable the model to
match the empirical parameters in a very good way.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch02 FA
0.4512 0.5488
P=
0.4944 0.5056
st = 1 st = 2
40 Aigner et al.
show some realistic characteristics, e.g., the correlation between the error terms of LPE
and stocks is much higher (0.92412) for the bad market state 1 than for the “better”
state 2 (0.41966). The parameters are fitted in such a way that the deviation between
the empirical statistics of the time series and the theoretical statistics of the model
is minimized. The (relative) deviation of the theoretical statistics (determined by the
parameters) and the empirical statistics is always smaller than 1% (only the deviation
for the kurtosis of bonds is higher (3.5%)).
Sharpe Ratio. The Sharpe ratio, see [23, 24], was first introduced by William F.
Sharpe and is one of the most common and intuitive measures. It is defined as follows:
µR − r
SR(R) = (2.3)
σR
with portfolio return R, its mean µR , standard deviation σR , and risk-free rate r.
Despite its intuitive formula (expected excess return per unit of risk in terms of
standard deviation), it has one major shortcoming: Since risk is only measured in
terms of standard deviation, the Sharpe ratio solely uses the first two moments of
the return distribution and ignores characteristics of non-normally distributed returns.
Therefore, it is not appropriate (or at least not sufficient) for return distributions, which
show non-normal characteristics, such as it is the present case.
Omega Measure. The Omega measure, see [22], is based on the idea that an investor
partition returns into gain and loss, i.e., in returns above and below a return threshold
τ. Thus, the Omega measure represents the probability weighted ratio of returns above
and below the threshold or, in other words, upside per downside:
E[R − τ]+
τ (R) = . (2.4)
E[τ − R]+
This definition can be transformed into τ (R) = (µR − τ)/(E[τ − R]+ ) + 1. This
notation reminds of the Sharpe ratio. But unlike the standard deviation for the Sharpe
ratio, the downside potential (as a measure of risk) exhibits the crucial advantage of
considering the whole return distribution of the portfolio return R where R is smaller
than τ (i.e., “left” of the threshold τ). However, the Omega measure still does not allow
to account for risk aversion (as it is not the same as choosing the threshold τ).
Score Value. The score value, see [10, 14], corrects the upside by the weighted
downside and is defined as follows:
where λSc defines the risk aversion of the investor. A high λSc stands for a high risk
aversion and vice versa. This performance measure encompasses the complete return
distribution of a portfolio as well as an investor-specific risk aversion parameter λSc . It
is obvious that the higher the Score value the better the portfolio under this measure.
In this study, the parameter τSc for the Score value is set to be equal to the risk-free
rate, i.e., τSc = r. However, the Score value also allows to adjust τSc according to an
investor’s individual preferences.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch02 FA
42 Aigner et al.
with R = R(w) = i wi Ri denoting the portfolio return (subject to w), Ri the return
of asset i, and γ capturing the investor’s attitude toward the risk. The aim is to find
those weights w, which maximize the expected utility:
max E[U(R(w))]. (2.9)
w∈Z
Omega Measure and Score Value Frameworks. As the motivation for the Omega
measure and the Score value is to capture more than the first two moments of the
return distributions of a portfolio, the portfolios can be optimized by maximizing
each of these performance measures, and thereby taking advantage of their respective
characteristics. The restrictions are the same as in the other models, i.e., w ∈ Z.
Investor: A B C
44 Aigner et al.
Portfolio weights
Performance measures
a time horizon of one year. There are several conclusions to be drawn from this
tables: There is almost always the same order of how much weight the performance
measures assign to the LPX 50 Index: The fractions increase as follows: mean–variance
framework < power–utility framework < Omega measure framework. The highest
fraction for the LPX 50 is always recommended by the Score value framework, whereas
the mean–variance framework allocates the smallest portions to listed private equity
among all optimization frameworks.
As theory suggests, the higher the risk aversion of the investor the smaller the
investment in listed private equity, because this asset is riskier than bonds and stocks.
In fact, the most risk–averse investor A would not consider a private equity investment
at all. Investor B puts lower fractions into the LPX 50 (7.5–11.8%) than investor C
(16.9–27.9%).
Another interesting issue to examine is whether stocks or bonds lose weight when
the LPX 50 index is included in the portfolio. Overall, it can be observed that bonds
hardly change their weight, and thus almost the whole weight of the LPX 50 is drawn
from the stocks. It is rather the case that the weight of bonds slightly increases when
taking listed private equity into account. The least risky asset class, bonds, is regarded
as a counterpart for the highly risky listed private equity for two reasons: its low risk
itself (measured in any risk measure) and the low correlation with both stocks and
private equity. Figures 2.5–2.8 show the portfolio allocation and the risk–aversion
parameters in different optimization frameworks subject to the standard deviation.
46 Aigner et al.
2.5. CONCLUSION
48 Aigner et al.
equity fund (or fund-of-funds). For this reason, the conducted study about the benefits
of private equity vehicles is limited to listed private equity only.
Monthly log returns of the LPX50 total return index showed a significant non-
normality and autocorrelation. Capturing these characteristics is not possible with stan-
dard approaches. Therefore, the concept of Markov–Switching models is introduced
to model more realistic return time series in order to find optimal asset allocations.
Depending on the degree of risk aversion of an investor and the optimization method
used, listed private equity should have a weight between 0% and 27.9% in an optimal
portfolio, where only the most risk–averse investor type is supposed to ignore listed
private equity for a one-year investment horizon.
References
[1] Aigner, P, S Albrecht, G Beyschlag, T Friederich, M Kalepky and R Zagst (2008). What
drives PE? Analyses of success factors for private equity funds. The Journal of Private
Equity, (Fall), pp. 63–85.
[2] Aigner, P, G Beyschlag, T Friederich, M Kalepky and R Zagst (2009). Optimal risk-return
profiles for portfolios including stocks, bonds, and listed private equity (submitted).
[3] Black, F and R Litterman (1990). Asset allocation: Combining investor views with market
equilibrium. Goldman Sachs Fixed Income Research.
[4] Brealey, R and S Myers (1991). Principles of Corporate Finance. NewYork: McGraw-Hill.
[5] Brigl, M, A Herrera, H Liechtenstein, H Meerkatt, J Prats and J Rose (2008). The advantage
of persistence, how the best private-equity firms ‘beat the fade.’ The Boston Consulting
Group, Inc., and the IESE Business School of the University of Navarra.
[6] Brunner, B and R Hafner (2006). Modelling Hedge Fund Returns: An Asset Allocation
Perspective, pp. 429–452. Risk Books.
[7] Damaschke, D and P Züchner. Private Equity Secondary Market, Grundlagen und Praxis
für Investoren, Consultants und Wissenschaftler. Saarbrücken: VDM Verlag Dr. Müller.
[8] Dimson, E, P Marsh and M Staunton (2006). The worldwide equity premium: A smaller
puzzle. London Business School, Working Paper.
[9] Donaldson, G, M Kamstra and L Kramer (2007). Estimating the ex ante equity premium.
Rotman School of Management Working Paper No. 07/02.
[10] Gong, X, M Huber, S Lanzinner and R Zagst (2006). Zertifikate. Mehrwert für Privatan-
leger? Zeitschrift für das gesamte Kreditwesen, 22, 1235–1239.
[11] Grabenwarter, U and T Weidig (2005). Exposed to the J-Curve, Understanding and Man-
aging Private Equity Fund Investments. London: Euromoney Institutional Investor Plc.,
[12] Grauer, R and N Hakansson (1986). A half century of returns on levered and unlevered
portfolios of stocks, bonds and bills. Journal of Business, 59, 287–318.
[13] Grunert, C (2005). Private Equity im Asset Management Institutioneller Investoren. Frank-
furt: Bankakademie-Verlag GmbH.
[14] Höcht, S, K.-H. Ng, J Wolf and R Zagst (2008). Optimal portfolio allocation with Asian
hedge funds and Asian reits. Journal of Services Sciences, 1(1), 36–68.
[15] Jesch, Th (2004). Private-Equity-Beteiligungen, Wirtschaftliche, rechtliche und steuerliche
Rahmenbedingungen aus Investorensicht. Wiesbaden: Verlag Dr. Th. Gabler.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch02 FA
[16] Liechtenstein, H and H Meerkatt (2008). Get ready for the private-equity shakeout. The
Boston Consulting Group, Inc., and the IESE Business School of the University of Navarra.
[17] LPX GmbH. Guide to the LPX Equity Indices.
[18] Markowitz, H (1952). Portfolio selection. Journal of Finance, 7(1), 77–91.
[19] Meyer, Th and P-Y Mathonet (2005). Beyond the J-Curve, Managing a Portfolio of Venture
Capital and Private Equity Funds. West Sussex: John Wiley & Sons.
[20] Modigliani, F and M Miller (1958). The cost of capital, corporation finance and the theory
of investment. American Economic Review, pp. 261–297.
[21] Rouvinez, C (2006). Top quartile persistence in private equity. Private Equity International,
pp. 76–79.
[22] Shadwick, W and C Keating (2002). A universal performance measure. Journal of Perfor-
mance Measurement, 6(3), 59–84.
[23] Sharpe, WF (1964). Capital asset prices —A theory of market equilibrium under conditions
of risk. Journal of Finance, 19, 77–91.
[24] Sharpe, WF (1994). The Sharpe ratio. Journal of Portfolio Management, pp. 49–58.
[25] Timmermann, A (2000). Moments of markov switching models. Journal of Econometrics,
96, 75–111.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch02 FA
In this chapter, we are introducing the asset class “Alternative Real Assets”. This asset class
provides access to “real” investment opportunities, real meaning both partially inflation protected
and solid, via limited partnership fund structures. This (sub-)asset class comprises fund vehicles
that invest, for example, in infrastructure, shipping, or renewable energy projects. For some of
these alternative real assets, financial incentives are provided, such as preferential feed-in tariffs
for solar-generated electricity.
We discuss alternative real assets, including a comparison to other asset classes and a descrip-
tion of distinctive features, with a focus on photovoltaic investments. The bottom-up modeling
approach of individual photovoltaic projects and the approach to incorporate these investments
in portfolio analytics are presented. Based on a comprehensive framework, we are finally able
to evaluate the benefits of photovoltaic investments in a multi-asset portfolio.
Investing in photovoltaic facilities is environmentally supportive. We are able, as well, to
show that under reasonable assumptions, financing solar plants is an attractive asset class, both
stand-alone and in a portfolio context.
51
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch03 FA
52 Mader et al.
3.1. INTRODUCTION
This chapter introduces the asset class “Alternative Real Assets.” Alternative real assets
provide access to investment opportunities in special forms of project finance, usually
via limited partnership structures. This asset class comprises fund vehicles that invest,
for example, in infrastructure, shipping, or renewable energy projects. For some of
these alternative real assets, financial incentives are in place, such as preferential feed-
in tariffs for solar-generated electricity.
The analysis of photovoltaic investments is the focus of this article. A bottom-up
modeling of investments in photovoltaic facilities, as well as the approach to incorpo-
rate these investments in portfolio analytics, is presented. Based on a comprehensive
framework, we are able to evaluate the benefits of photovoltaic investments in a multi-
asset portfolio. Empirical data and realistic modeling assumptions are the result of
a joint research project with a leading provider of closed-end funds in the area of
photovoltaic investments.
This chapter is organized as follows. Section 3.2 provides an introduction to alter-
native real assets. In Sec. 3.3, we present our bottom-up approach to model alternative
real asset investments. Section 3.4 describes how to incorporate these investments to a
portfolio. It shows the results of integrating photovoltaic investments into a traditional
portfolio, including changes in portfolio statistics and the corresponding substitution
effects. A conclusion is given in the last section.
Broadly speaking, asset classes contain investments with similar features including
their risk and return characteristics and their sensitivity toward major market fac-
tors. The investment universe is usually divided into traditional asset classes and
non-traditional or alternative asset classes that emerged over the last few decades.
Traditional asset classes include equities, nominal government, or corporate bonds.
Examples for non-traditional or alternative asset classes are private equity or hedge
fund investments.
The rationale for investing in the different asset classes varies broadly, which is in
turn the reason why investors should diversify their asset allocation. For equity invest-
ments, a main investment objective is to participate in the growth potential and payout
of listed companies. The rationale for bond investments includes income generation,
deflation protection, and reduction of portfolio volatility.
Traditional real assets include commodities, inflation-linked bonds, and real estate
investments (direct or indirect, but not listed real estate investments). The categoriza-
tion to this asset class subset is usually based on either the inflation protection feature
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch03 FA
such as real return, or the tangibility of the assets (see [3], p. 20, for a definition of real
assets). Alternative real assets are represented by investments in shipping, aviation,
infrastructure, or renewable energy facilities like photovoltaic or wind power plants.
The real assets that provide the necessary cash flows may be roads, ships, or solar
plants. Therefore, alternative real assets provide direct exposure to specific projects
and the corresponding “real facilities” but no exposure to a listed company’s capital
structure. Due to the project finance character, alternative real assets are not very liquid
investments, even though secondary markets are already emerging. As revenues of the
financed projects are often linked to inflation, most alternative real assets provide some
inflation protection.
In the following, we will focus on investments in solar power plants. This kind
of alternative real assets is characterized by cash flows that are mainly driven by
solar radiation and preferential feed-in tariffs. These cash flows are often linked to
inflation and show a sensitivity to interest rates due to the partial debt financing of the
photovoltaic projects. More details on the modeling of photovoltaic investments are
given in the next section.
54 Mader et al.
exposure to a specific project with a well-defined revenue and cost structure, on the
other hand we have exposure to a broad mix of risk factors that drive the stock returns
of a listed company:
• Capital structure of the company
• Balance sheet effects
• General economic environment
• Change in exchange rates
• Differing expectations of market participants
• Analyst appraisals of profit expectations
Therefore, we believe that a different methodology than the top-down CAPM approach
is required, when dealing with this kind of project finance. However, in relevant aca-
demic and practitioner publications, frequently top-down approaches are used for eval-
uating the alternative real assets.
Based on expert information from a well-known German initiator of photovoltaic
projects (we would like to thank KGAL [KG Allgemeine Leasing GmbH & Co.] for
valuable insights to the financial modeling of solar plants), we analyzed the revenue
and cost structure reflecting the return profile of individual solar power plants. Thus,
we were able to identify crucial and cash flow relevant aspects of photovoltaic projects
that should be covered by an appropriate model:
• Revenues of the sale of electricity
• Cost of acquisition
• Operating expenses
• Interest on debt capital
• Tax payments and interest subsidy
Modeling these building blocks properly is at the heart of our bottom-up approach.
Reference [7] follows a similar bottom-up approach to derive the cash flows resulting
from shipping investments. However, they combine this procedure with a CAPM-
approach to finally be able to price a shipping investment. Recently, at the stock
market exchanges of Hamburg and Hannover, a ship fund index is calculated using a
bottom-up methodology. The index calculation is described in [14].
for feeding “clean” electricity into the grid. Such feed-in tariffs are guaranteed by
local law and paid by the government, securing cash inflow. In some countries, the
compensation is even linked to local consumer price indices and accordingly provides
partial inflation protection to potential investors.
Due to generous (inflation-linked) government compensations and a high level
of global irradiance, Spain is a favored country for building up photovoltaic facilities
and feeding the produced electricity into the local grid. For further comprehension,
global irradiance is the total incoming solar radiation on a horizontal plane at the
Earth’s surface, i.e., the incoming solar radiation after scattering effects in the Earth’s
atmosphere. Therefore, our representative photovoltaic project is located in Spain and
we have chosen a corresponding parametrization. This includes a project duration of
25 years, which corresponds to the maximum duration of the attractive feed-in tariffs.
When following a bottom-up approach, it is absolutely necessary to model the
project’s properties and the corresponding dependencies (especially the dependencies
on stochastic risk factors) with a high level of detail. To be able to model the cash
flow structure in an appropriate way, we set up a comprehensive bottom-up framework
covering all revenue and cost-related features of the sample solar plant. An overview
of the framework components and properties can be found in Table 3.1.
A special focus of our analysis is on the simulation of stochastic risk factors that
drive the revenue and the cost side of the photovoltaic project. The key risk factors
and their dependencies to various components of the bottom-up framework are listed
in Table 3.2.
These risk factors have to be modeled and simulated in an integrated approach.
This allows us to conduct a multi-scenario analysis of the inherent risk factors, the
Revenue Modeling of the generation of electricity out of solar radiation, and the
corresponding feed-in tariffs, including inflation-indexed tariff adjustments
Costs Considering acquisition and operating costs, as well as inflation-driven increase
over time
Financing Modeling of equity, debt capital, and shareholder debt capital, as well as
required liquidity reserves
Tax Considering tax payments on company level, tax deductibility of interest
payments on debt and shareholder debt, and tax loss carry-forwards to future
periods
Facility Considering annual power loss of the solar panels due to degradation effects,
liquidation proceeds are on par with deconstruction costs at project
termination
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch03 FA
56 Mader et al.
sample photovoltaic project and a portfolio consisting of traditional assets and the
photovoltaic project.
with the positive real numbers brw , âr , and σr , the time-dependent deterministic func-
tion θr and the standard Brownian motion WrQ , independent of the Brownian motions
mentioned before. The nominal short rate {rN (t)}t∈[0,T ∗ ] is defined as the sum of real
short rate and inflation, i.e.,
r N (t) = r(t) + i(t).
The term structure of interest rates can then be derived by the zero-coupon bond prices
obtained from this setting. As shown in [16], p. 4254f, the price of a zero-coupon bond
with maturity t < T ≤ T ∗ is given by
P(t, T ) = exp(A(t, T ) − B(t, T )r(t) − C(t, T )i(t) − D(t, T )w(t)),
where
1
B(t, T ) = (1 − exp(−âr (T − t)),
âr
1
C(t, T ) = (1 − exp(−âr (T − t)),
âi
brw 1 − exp(−âr (T − t)) exp(−âw (T − t)) − exp(−âr (T − t))
D(t, T ) = · + ,
âr âw âw − âr
T
1 2
A(t, T ) = (σ B(l, T )2 + σi2 C(l, T )2 + σw2 D(l, T )2 ) − θr (l)B(l, T )
t 2 r
− θi C(l, T ) − θw D(l, T ) dl.
Using Girsanov’s Theorem, the model equations can be derived under the real measure
IP instead of the equivalent martingale measure Q by replacing WiQ , WwQ , WrQ with
the independent standard Brownian motions Wi , Ww , Wr and using the parameters
ai = âi − λi σi2 , aw = âw − λw σw2 , and ar = âr − λr σr2 . λi , λw , and λr are obtained by
the change of measure, as shown in [10]. As can be seen, the term structure of interest
rates is driven by the real short rate process, as well as the underlying macroeconomic
factors of economic growth and inflation.
The third cascade contains equity assets. The equity prices {StE }t∈[0,T ∗ ] are driven
by the following dynamics:
dS E (t) = [αE + bEr r(t) − bEi i(t) + bEw w(t)]S E (t)dt + σE S E (t)dWE (t),
where bEr , bEi , bEw , and σE are positive real numbers, αE ∈ R, and WE (t) is a standard
Brownian motion, independent of those mentioned above.
58 Mader et al.
starting point for the generation of electricity in a solar plant. An appropriate starting
point for the modeling approach is average monthly global irradiance. Having a higher
frequency will, apart from data availability issues, increase the noise component of
the model, but not have a different impact on at most monthly simulated cash flows
generated by the photovoltaic plant. As the irradiance model component independently
adds to the basic framework, and the frequency of our data is also monthly, we develop
a discrete-time model. The following sections now present the building blocks for the
derivation of a discrete-time model for yearly global irradiance.
3000
2500
Annualized monthly solar irradiance
2000
1500
1000
500
0
0 10 20 30 40 50 60 70 80
Time in months
−3
x 10
4
3.5
2.5
Estimated Density
1.5
0.5
0
−400 −300 −200 −100 0 100 200 300 400 500
Monthly global solar irradiance
Figure 3.1 Left: Original (continuous line) and fitted (dashed line) annualized monthly global
radiation series. Right: Estimated density of residuals.
the location Toledo is shown in Fig. 3.1. The density was estimated using Gaussian
kernel smoothing, i.e., the density estimator
T
1 1 − (Y −Ŷ 2(t))2
fˆ(Y ) = √ e 2h
Th t=1 2π
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch03 FA
60 Mader et al.
1
was applied to the observed data series, where the bandwidth used was h = 3T4 5 σ,
and σ was obtained using the median absolute deviation estimator. (More informa-
tion on kernel smoothing can be found in [4]). To capture the skewness in the fluc-
tuation around the average seasonal global irradiance, we specify the noise process
{m }m∈{1,...,12·T ∗ } to follow a Markov switching model, i.e.,
(m) = µM
1 + X (m)(µ2 − µ1 )
M M M
9000
8000
7000
6000
Frequency
5000
4000
3000
2000
1000
0
−800 −600 −400 −200 0 200 400 600 800 1000
Annualized monthly global irradiance results
3500
3000
Annulized monthly global irradiance
2500
2000
1500
1000
500
0
0 5 10 15 20
Time
Figure 3.3 Original series (dots) and annualized monthly global irradiance quantiles (barred
lines).
We now use the model for Y M (m) to obtain the distribution of Y(t) by simulating
1000 trajectories of monthly global irradiance, each over a time span of 25 years.
By calculating yearly radiation on the simulated processes according to (3.3), we
generate 1000 paths of yearly global irradiance with 25 observations per path. The
resulting distribution can be seen in Fig. 3.4 for Toledo. Compared to monthly global
irradiance the skewness is reduced, but retained. The results we obtain for yearly
global irradiance are similar to the published information on average yearly global
irradiance, as available for example through radiation maps shown in [11] or offered
on the website www.meteonorm.com, which provides a comprehensive meteorological
reference. Table 3.3 below shows the means and relative standard deviations, i.e., the
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch03 FA
62 Mader et al.
standard deviation divided by the mean, resulting from the simulation of yearly global
irradiance for several sites, as well as the R2 -values of the estimated model.
Comparing means and standard deviations with the information in [11] shows them
to be in line with their data. The skewness of the distributions for the different locations
is very similar; the distributions vary mainly in location and scale. For the following
analysis of photovoltaic investments, we choose Toledo as a reference for yearly global
irradiance on a horizontal surface, because it shows average characteristics of global
irradiance in Spain, and is located close to a large number of existing solar plants.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch03 FA
64 Mader et al.
n
→n =
total
rp,1 rp,i .
i=1
Given these assumptions, it is easy to show that (for this result compare also [5],
p. 49.)
n−1
j
total
Var(rp,1 →n ) = n · 1 + 2 1− · kp (j) · σp2 , (3.4)
j=1
n
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch03 FA
with
total n
Var rp,1 →n = Var(rp,i ) + Cov(rp,i , rp,j )
i=1 i =j
n
i−1
= Var(rp,i ) + 2 · Cov(rp,i , rp,j )
i=1 j=1
n
i−1
= σp2 + 2 · (σp2 · kp (i − j))
i=1 j=1
n
i−1
= σp2 n + 2 · kp (i − j) .
i=1 j=1
n−1
n
= kp (k)
k=1 i=k+1
n−1
= (n − k) · kp (k),
k=1
Thus, knowing the degree of autocorrelation and having for instance yearly periods
allow us to calculate the annualized volatility from total volatility by dividing the
total volatility by the square root of the number of years adjusted for the degree
of autocorrelation by the factor 1/α. Assuming the absence of autocorrelation, only
the observable initial portfolio value and the observable portfolio value at the end
of the investment horizon are needed to calculate single-period portfolio log-return
volatility. The dependence of cash flows generated by a solar power plant over time
should mainly be driven by their dependence on inflation and interest rates, which
both have a negative impact on the cash flows, and both show positive autocorrelation.
Thus, one would expect positive autocorrelation, if any, in the single-period returns
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch03 FA
66 Mader et al.
3.4.3. Results
Tables 3.6–3.8 present the results of our portfolio optimizations for the three initial
allocations “10% Equity,” “20% Equity,” and “30% Equity.” Every table shows the
resulting optimized portfolios for similar expected risk or similar expected return
Substitution
Allocations Statistics w.r.t. “10% Equity”
Portfolio PV (%) E (%) B (%) Return (%) Risk (%) PV (%) E (%) B (%)
Substitution
Allocations Statistics w.r.t. “20% Equity”
Portfolio PV (%) E (%) B (%) Return (%) Risk (%) PV (%) E (%) B (%)
Substitution
Allocations Statistics w.r.t. “30% Equity”
Portfolio PV (%) E (%) B (%) Return (%) Risk (%) PV (%) E (%) B (%)
statistics when we allow the allocation of a maximum weight of 5%, 15%, or 25% to
the photovoltaic investment opportunity.
The “higher return” or “lower risk” portfolios, denoted by “HR X% PV”, and “LR
X% PV,” respectively, are given in the three tables together with the corresponding
expected risk and expected return figures. (X denotes the photovoltaic investment
proportion in the portfolio.) These efficient portfolios (with weights for photovoltaic
(PV), equity (E) and bond (B) investments), compared to the initial allocation, yield
the same expected return with less risk, or the same risk with more expected return.
The last three columns of Tables 3.6–3.8 present the weights differences between the
initial allocations and the new efficient weights, i.e., the substitutional effects due to
the introduction of photovoltaic investments.
The dominance of portfolios including the photovoltaic investment opportunity
is striking. In Table 3.6, we see the results for an initial equity allocation of 10%.
The higher return portfolios show an increase in expected return from 20 bp up to
70 bp depending on the maximum quota of photovoltaic investments in the portfolio.
For example, the higher return portfolio with a limit of 5% in photovoltaics (“HR
5% PV”) allocates the maximum of 5% to photovoltaic investments, 13% to equities
and 82% to bonds, increasing the expected return from 4.9% to 5.1% at the same
level of expected risk (3.5%). The lower risk portfolios (“LR X% PV”) decrease the
expected risk by up to 60 bp (or approximately 17% in relative terms) at the same level
of expected return. With higher photovoltaic limits, the risk and return characteristics
cannot even be brought down to the initial levels when selecting portfolios from the
efficient set. The minimum expected return for “LR 15% PV” is 5.1%, for “LR 25%
PV” it is 25%, respectively.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch03 FA
68 Mader et al.
In Table 3.7 we see higher return and lower risk portfolios for an initial allocation
of 20% in equites. The improvement on the return side ranges from 10 bp to 60 bp and
up to 110 bp (or approximately 27% in relative terms) on the risk side. The comparison
of the optimized portfolios to a 30% equity portfolio in Table 3.8 reveals up to 60 bp
more return and up to 160 bp (or approximately 31% in relative terms) less risk.
Interestingly, the optimized portfolio allocations in the case of the “10% Equity”
and “20% Equity” show a slight increase in the equity weight for the higher return
portfolios, as the photovoltaic investment seems to reduce the portfolio risk substan-
tially. Therefore, the bond allocation is reduced by up to 30% (“HR 25% PV” with
“10% Equity” as initial allocation). For the “30% Equity” portfolio, the optimized
higher return portfolios show a stable allocation in equities — bonds are substituted
for by the allocation to photovoltaic investments. When we optimize the lower risk
portfolios, the allocation to photovoltaic investments consistently substitutes equity
and bonds. For the “20% Equity” portfolio both substituted at similar amounts, for the
“30% Equity” portfolio compared to bonds more equity is substituted, and for “10%
Equity” portfolio more bonds are substituted compared to equity.
3.5. CONCLUSION
The chosen bottom-up framework allows for the inclusion of alternative real assets
when evaluating and optimizing the portfolio allocations. We can examine the port-
folio contribution of alternative real assets in general and we are able to customize
the modeling for specific investment assets in great detail. Beside the analysis in a
portfolio context, this bottom-up modeling also allows ongoing monitoring of specific
alternative real assets.
References
[1] Akinoglu, BG (2008). Recent advances in the relations between bright sunshine hours
and solar irradiation. In V Badescu (ed.), pp. 115–144. Modeling Solar Radiation at the
Earth’s Surface Springer Verlag.
[2] Alsaad, MA (1991). Correlations for predicting average global solar radiation incident on
a horizontal surface in amman. Journal of King Saud University. Engineering Sciences,
3(1) 121–134.
[3] Ankrim, EM and CR Hensel (1993). Commodities in asset allocation: A real-asset alter-
native to real estate?” Financial Analysts Journal, 49(3) 20–29.
[4] Bowman, AW and A Azzalini (1997). Applied smoothing techniques for data analysis.
Oxford Science Publications.
[5] Campbell, J, A Lo, AC MacKinlay (1997). The Econometrics of Financial Markets. Prince-
ton University Press.
[6] Conover, WJ (1998). Practical Nonparametric Statistics. Wiley.
[7] Drobetz, W, L Tegtmeier and M Topalov (2006). Bewertung von Kommanditanteilen
geschlossener Schiffsfonds mit dem Ertragswertverfahren. Finanz Betrieb, 6, 399–412.
[8] Hamilton, JD (1994). Time Series Analysis. Princeton University Press.
[9] Intriligator, M, R Bodkin and C Hsiao (1996). Econometric Models, Techniques, and
Applications. Prentice Hall.
[10] Schmidt, B and R Zagst (2000). A three-factor defaultable term structure model. Journal
of Fixed Income, 10(2), 63–79.
[11] Suri, M et al. (2007). Uncertainty in photovoltaic electricity yield prediction from fluctu-
ation of solar radiation. In Proceedings of the 22nd European Photovoltaic Solar Energy
Conference, Milano (preprint).
[12] Tegtmeier, L and M Topalov (2006). Die Bedeutung von Schiffsbeteiligungen im Rahmen
der Asset Allocation von Privatanlegern. Finanz Betrieb, 7–8, 506–509.
[13] Timmerman, A (2000). Moments of Markov switching models. Journal of Econometrics,
96, 75–111.
[14] TKL Fonds (2009). Leitfaden zum SFI TKL.Ship Fund Index.
[15] Vasicek, O (1977). An equillibrium characterization of the term structure. Journal of Finan-
cial Economics, 5, 177–188.
[16] Zagst, R, Th Meyer and H Hagedorn (2007). Integrated modelling of stock and bond
markets. Journal of International Finance, 19(1), 4252–4277.
[17] Zagst, R (2002). Interest Rate Management. Springer Verlag.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch03 FA
Despite the fact that the freight market is one of the oldest markets, it found only limited
attention in financial (mathematics) studies. One possible reason may be that this
market is very intransparent and only active market participants such as shipowners,
dockyards, and commodity traders were fully informed on market activities. With
the launch of freight market derivatives, however, many other financial institutions
started to actively trade on the freight market, for instance by taking positions on future
global trade activities via Forward FreightAgreements and Freight Futures. This article
provides an introduction to the freight market and its most commonly used financial
assets. We investigate the relationship between the freight market, freight derivatives,
and macro-economic variables. Further, we show how these relationships can be used
to better explain and predict the future price movements of freight derivatives. To
achieve this, we propose to use classical statistical methods, such as Vector Auto
71
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch04 FA
Regression and Vector Error Correction Models, to incorporate the information from
spot markets and related explanatory variables into the prediction process.
The oversea shipping market has been established in the 17th century and thus is
one of the oldest physical markets. In 2007, the volume of international seaborne trade
reached 8.02 billion tons [1]. In 2008, over 80 per cent of the worlds merchandise trade
volume was carried by sea, making the maritime transport the backbone of international
trade and globalization. Despite its size, it found only limited attention in financial
(mathematics) studies and for a long time it failed to attract the focus of investors
from outside the traditional shipping market environment. Since the shipping market
is exposed to various risks, it is highly volatile. Historically, these risks were mainly
due to unsuitable technical equipment, lack of reliable weather forecasts, piracy, etc.,
which in many cases led to the loss of freight, the ship, and the lives of the crew. While
natural risks are still present nowadays, financial risk for shipowners, commodity
traders, charterers, and other investors and market participants have become more
important.
4.1.1. Vessels
The modern fleet of vessels used for overseas shipping is well diversified. We see a
large range of modern cargo vessels, ranging from small coasters to very large tankers,
but generally we can classify them in terms of the groups given in Table 4.1 below.
Not reported in the table are liquefied natural gas (LNG) tankers, general cargo (Gen
Cargo), and Roll On-Roll Off (Ro-Ro) carriers, as well as vessels that are able to carry
a combination of chemicals and oil products. Furthermore, passenger cruise ships
are excluded in the analysis of this paper. In the 2008 report of the United Nations
Conference on Trade and Development (UNCTAD [1]), it is stated that the world vessel
fleet expanded by 7.2 per cent during 2007 to 1.12 billion deadweight tonnage (DWT)
at the beginning of 2008. Deadweight tonnage is a measure of how much weight of
cargo a ship can carry. Just to give an impression on prices, a 170,000 DWT dry bulk
carrier was priced at $97 million in December 2007 about 39 per cent more than a
year before, and 2.4 times the price paid in 2000. Greece, Japan, and Germany are the
world’s top three shipowning nations.
4.1.2. Cargo
The main cargos of tankers are crude oil products (“dirty oil”) and refined oil products
(“clean oil”), chemicals as well as liquefied natural and propane gas (LNG resp. LPG),
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch04 FA
whereas dry cargo such as ore, coal, bauxite, phosphates, cement, scrap, steel, sugar,
grain and rice, cars, containers are shipped via bulk carriers or container ships. Half
of the cargos are energy related while container traffic is just over 10% by weight, but
much higher in terms of value. Ship sizes are measured in DWT and include Capesize
>100,000 DWT, Panamax 60–100,000 DWT (capable to cross the Panama Canal),
Supramax 40–60,000 DWT as well as Handysize 10–40,000 DWT.
4.1.3. Routes
The biggest trading routes by vessel type and size are given in Table 4.2. In our analysis,
we restrict our focus to the three routes given in bold. These are the most frequently
served routes, and we were able to obtain a data set of high quality for them.
TD3 The “Tanker Dirty Route 3” route from Ras Tanura, Saudi Arabia to Chiba,
Japan is the most common trade route into Asia for very large crude carriers.
The standard 260,000 ton of non-heat crude carried from Saudi Arabia to Japan
on TD3 is approximately two million barrels of oil. As spot price (USD/Ton),
we make use of the Bloomberg price series “D27TAGJP Index.”
TC2 Being the freight element of the cross Atlantic gasoline arbitrage trade, the
“Tanker Clean Route 2” route is the most frequently traded clean product tanker
contract. The TC2 route is typically served by modern Panamax or Supramax
tankers carrying clean product (most often gasoline and other light distillates)
from Rotterdam to New York Harbor. For spot prices, we use the Baltic Clean
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch04 FA
C3 Tubaro to Beilun
C4 Richards Bay, South Africa to Rotterdam Coil
C5 Western OZ to Beilun
C7 Puerto Bolivar (Ecuador) to Rotterdam
Panamax
Tanker Index “BITY Index,” which is an average over all clean tanker routes for
the year 2005, and the “CMRTUKUA Index” for the remaining time, because
its price history starts in 2006 and all other prices are available from 2005.
C4 The most common dry cargo routes are the big coal and iron ore transportation
routes from Africa to Europe and from South America and Australia to Asian
destinations. We consider the “Capesize Route 4” route from Richards Bay,
South Africa to Rotterdam with spot prices represented by the SSY price history
“SSYWRBRT Index.”
As it can be seen from the spot prices of the three routes given in Fig. 4.1, volatility is
high in the freight market. Typically, the price of shipping one ton of cargo on one of
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch04 FA
40
20
0
2005 2006 2007 2008 2009
TC2 freight rates
60
40
20
0
2005 2006 2007 2008 2009
C4 freight rates
60
40
20
0
2005 2006 2007 2008 2009
Figure 4.1 Spot price histories in USD/Ton for the period 01/01/2005 to 20/2/2009.
the three routes trades between 15 and 35 USD, but strong deviations from this band
have occurred even before 2008. We further observe that the freight market experi-
enced many pronounced cycles of high as well as low rates. Also the financial crisis
of 2008/2009 had its effects on the freight market; in line with the stock markets, rates
dropped in late 2008 and remained at a low level until spring 2009 when they slightly
recovered. The effects of the credit crunch on the freight rates as well as on the entire
shipping industry is further discussed in [2]. The authors note that the economic slow-
down affected the shipping market more than most other markets, for example causing
an 94% drop of the Baltic Exchange Dry Index from May to December 2008. The con-
sequences of this variability are twofold: the typical market participants are in need
of hedging instruments to lay off some of the risk of adverse price movements, while
new investors and speculators are drawn to the market because large price movements
can imply large profits.
In order to understand this high variability of the freight rates, we first have to
identify the price drivers of the market, i.e., what variables govern the demand and the
supply of freight capacity, and therefore the freight rates.
at the typical cargos (oil products, chemicals, gas, ore, coal, bauxite, phosphates,
cement, scrap, steel, sugar, grain, rice), we expect this relationship to be rather strong.
What is the level of industrial production? Has the grain harvest been successful? Are
power stations importing more coal? How is the steel industry performing? Commodity
demand directly implies demand of shipping capacity. Stock indices and futures on
indices can be seen as a proxy for the physical industrial activities. Futures have the
additional feature of a time horizon for which expectations of the underlying prices
are captured. Further, we include various commodity prices as well as prices of fuel
and energy products.
Due to the globalization, many goods and commodities are imported from over-
seas. This is beneficial as a basic principal from economic theory, but additionally on
a short-to-medium horizon, exchange rates further influence the benefit. Thus, in the
analysis below we include the main FX rates as explanatory variables.
Note that the demand of shipping capacity can change rapidly. Extreme economi-
cal or environmental events, even exceptional weather conditions, can imply significant
and fast changes in the need of energy, building materials, goods, and thus of shipping
capacity.
4.2.3. Costs
The main components of a vessel’s costs are typically divided into voyage costs and
daily running costs and may include, among other factors, crew wages, overtime, pen-
sion contributions, insurance, traveling costs, victualing, insurance, deck and engine
room stores and spares, lubricating oil, periodic costs such as dry-docking, special
surveys, running repairs, office and/or management costs, port disbursement, canal
dues, commissions, taxes, etc. The main component, however, are bunker prices. With
bunker fuel accounting for between one-quarter and one-third of the cost of running a
vessel, oil price movements directly affect shipowners.
We further include interest rate products in the analysis below because the costs
of financing ships, financing freight for the duration of the journey, etc. are likely to
affect freight rates.
A freight derivative is a financial contract which sets an agreed future price for car-
rying commodities at sea. Freight derivatives were introduced to give freight market
participants an instrument to reduce their various risks. Given the current volatility of
the freight markets, managing freight market risk is a significant issue for the shipping
industry. For example, by locking in deals well in advance, shipping companies as
well as commodity traders achieve to be less exposed to spontaneous adverse price
movements on the spot markets.
The first contracts in this respect were time charter agreements. Here, the vessel
is chartered to a counter party over a certain time span, typically a few years. We now
see efforts to open the market to more participants, making it possible to trade not only
cash settled (the supertanker “AbQaiq” will not show up at your front door), but also
in standard contracts traded on an exchange and without counter party risk.
The two main institutions in the field are the Baltic Exchange in London [4] as well
as Imarex in Oslo [5], both providing further information on trading freight derivatvies
as well as background reports.
To obtain more standardized products, the contract details can be based on the
terms and conditions of the Forward Freight Agreement Brokers Association [6]. The
main terms of an agreement cover the following specifications:
• Contract route: the agreed physical trading route.
• Contract period: the period for which the prices are fixed.
• Contract quantity (Q): the amount of the cargo for which prices are fixed.
• Contract rate (CR): the rate at which differences will be settled.
• Settlement rate (SR): calculated as the average spot price during the contract period.
On settlement, typically within five days after the contract period, the two parties will
settle the contract according to the following simple calculation: the seller of the FFA
contract is required to pay the buyer an amount equal to (CR − SR) · Q.
Again, due to the inhomogeneity of the freight market, there is no spot price readily
available. Instead, daily freight rate assessments of the physical routes are provided by
independent organizations such as the Baltic Exchange. Their daily route assessments
are based on input from participating ship brokers according to fixtures concluded on
the day and market sentiment. Baltic spot rates aim to be “Fair Value” assessments of
the market on any day.
the front month contract and carefully reading the product specifications of our “spot”
prices, we believe to have a good proxy for freight spot prices at hand.
We now try to identify explanatory variables for the observed price movements in
freight spot and freight futures rates. Instead of manually setting up an economic
model for the relations between variables, we extract the information by means of
standard statistical Vector Auto Regression (VAR) and Vector Error Correction Models
(VECM).
As potential explanatory variables for the spot and future prices of our three routes,
we take variables from the groups identified in Sec. 4.2. In more detail, these are
reported in Tables 4.4 and 4.5. We consider the indices and futures as forward looking
indicators of economic activity. The exchange rates are motivated by the geographic
location of the main routes (linking Europe, Africa, Japan, and the United States) and
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch04 FA
the commodities are part of the freight. We expect that the fuel- and energy-related
commodities are particularly helpful since their prices are closely monitored economic
variables and their dependence on economic activity is high.
We introduce the following notation: since it will be clear from the context if spots
or futures prices are considered, both the freight spot and freight futures time series
are named F = (Ft ), and by ft = log(Ft+1 /Ft ) we refer to their log returns. Similarly,
by et = log(Et+1 /Et ), we refer to the log return series of the explanatory variable
E = (Et ) under consideration.
We work with the following specification of the VAR model.
20 20
µ + aj ft−j + bj et−j
f f
ft j=1 j=1 t
= + (4.1)
et 20 20 et
µ + ce + df
e j t−j j t−j
j=1 j=1
with an N0, distributed white noise error term , and all coefficients being real
numbers.
Thus, in the model framework used here, we assume that log returns of freight
rates and explanatory variables are given by a constant (µf resp. µe ) plus a weighted
sum of lagged “old” log returns, plus an error term which for simplicity we assume to
be normally distributed.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch04 FA
We use 20 lags because working with daily data, this number corresponds to
one month worth of data. It is desirable to include even more lags, to capture rela-
tions which hold over several months, but we found this yields unreliable, unsta-
ble results in most cases and thus we do not increase the maximal lag sizes at the
moment.
One application of the VAR model is that it allows for a Granger Causality test.
TD3 The tankers of this route carry crude oil from Saudi Arabia to Japan. Indeed, as
expected the spot prices of oil, priced in Saudi Arabia and Japan, can also be
found on the list, as well as the Nikkei 225, the major stock index for Japan.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch04 FA
Table 4.6 Granger Causality Test Results. For Each of the Three Routes We Report
the 15 Most Significant Causality Test P-Values, Averaged over FFA Spot and Futures.
TD3 TC2 C4
0.00 Dirty Tanker Ind. 0.00 Clean Tanker Ind. 0.00 Baltic Dry Ind.
0.02 LIBOR USD 1M 0.01 LIBOR USD 1M 0.00 Capesize Index
0.04 Oil Spot Arab. Gulf 0.01 FX EUR vs. USD 0.02 Dow Jones Indust.
0.05 Natural Gas 0.05 GSCI Energy 0.09 CRB Metal Index
0.07 Oil Spot UK 0.05 Dirty Tanker Ind. 0.10 CRB Comm. Index
0.09 Clean Tanker Ind. 0.07 GSCI 0.10 DAX
0.09 GSCI Energy 0.07 Oil Spot USA 0.12 FTSE 100
0.10 Crude Oil 0.07 Oil Future USA 2M 0.15 Oil Spot Asia
0.11 Oil Future USA 1M 0.09 CRB Comm. Ind. 0.15 EURO STOXX 50
0.11 Oil Future USA 4M 0.09 Natural Gas 0.19 S&P ASX Index
0.11 Oil Spot Asia 0.09 FX GP vs. USD 0.21 Oil Future USA 6M
0.12 Oil Future USA 2M 0.10 Oil Future USA 3M 0.21 Oil Future USA 3M
0.12 FX GP vs. USD 0.10 Oil Future USA 1M 0.21 Coil Rotterdam 2M
0.13 Oil Future USA 3M 0.11 Oil Future USA 5M 0.22 Crude Oil
0.14 Brent Crude 0.11 Oil Spot Asia 0.23 Aluminum
0.19 Nikkei 225 0.14 Oil Spot UK 0.25 Oil Future USA 2M
0.19 Oil Spot USA 0.46 Tanker DWT 0.26 Coil Rotterdam 1M
0.43 Dow Jones Indust. 0.49 Dow Jones Indust. 0.29 Coil Rich. Bay 2M
0.45 FX JPN vs USD 0.50 EURO STOXX 50 0.30 Coil Rotterdam 3M
0.58 Tanker DWT 0.55 DAX 0.65 Bulk Carrier DWT
Further, both the clean and dirty tanker indices from the Baltic Exchange rank
among the top variables.
TC2 This route’s vessels take oil products from Rotterdam to the United States. Again,
oil-related variables are on the list of the top 15 related variables, together with
exchange rates and oil futures prices.
C4 The capesize bulk carries of the C4 route mainly carry coil from Richards Bay,
South Africa to Rotterdam. Besides the energy related-variables, the metal and
commodity indices from CRB as well as the Coil Prices are among the top 15
out of the total 71 variables.
extend the bivariate model from Eq. (4.1) to a multivariate model and take those five
variables that minimize the sum of squared residuals of the freight rate log return
series. The reason for this additional selection algorithm is that five carefully selected
variables typically have a higher combined explanatory power than the five top individ-
ual explanatory variables. For example the Baltic Dry Index is very similar to the Baltic
Capesize Index, and therefore will not deliver substantial additional information.
Instead of trying out all 71 · 70 · 69 · 68 · 67 combinations of explanatory variables,
we apply a Greedy scheme which subsequently adds the variable that minimizes the
sum SSRFull in the two, three, four and five dimensional model.
The resulting top variables for the spot and first month contracts are given in
Table 4.7, where we applied the procedure on the entire data set. It is interesting to
see that besides the variables with the obvious relations other variables have a high
explanatory power as well, for example the LIBOR rates.
4.4.4. Cointegration
Another concept of dependence between two or more time series is that of cointegra-
tion. In the bivariate case, we say that two time series f and e are cointegrated if each
of the series itself is integrated of order one, while some nontrivial linear combination
(f, e) · a of the two series is stationary. a is then called cointegration vector. The
interest in cointegration theory among the economists is due to the fact that it allows
to model stable long-term relations in between non-stationary variables. We can use
cointegration analysis in three ways:
• Find an equilibrium relation in a system of price series.
Table 4.7 Top Five Selected Variables Through the Greedy Scheme for
the Entire Data Set. The First Set of the Five Variables Corresponds to
FFA Spot Prices and the Second Set to Front Month FFA Futures Prices.
TD3 TC2 C4
Top 1 Dirty Tanker Ind. Clean Tanker Ind. Baltic Dry Index
Top 2 Natural Gas LIBOR USD 1M LIBOR JPY 1M
Top 3 FX EUR vs. USD LIBOR JPY 3M Gas Oil
Top 4 LIBOR USD 1M FX EUR vs. USD Dow Jones Indust.
Top 5 LIBOR JPY 1M Copper GCC 200 (Africa)
The models used in this context are of the following form, where F and Ei , i =
1, . . . , 5 denote the freight and explanatory variables respectively, denotes first
order difference. Let yt be given by
then the Vector Error Correction Model (VECM) is given by the following represen-
tation:
20
yt = µ + j · yt−j + · yt−1 + t .
j=1
Thus, in this framework, we model the FFA log returns (which equal the first differ-
ences of log prices) to be given by the following composition: a constant µf , given by
the first entry of the real vector µ, plus a weighted sum of lagged log returns of the
FFA series and the lagged log returns of the explanatory variables. The weights of the
sum corresponding to the j -lagged log returns are given by the first row of the real–
matrix j . The third summand is given by the first entry of the matrix–vector product of
the cointegration matrix with the lagged log price series. The rows of the cointegra-
tion matrix describe the cointegration relations and the rank of the matrix corresponds
to the number of cointegration relationships. The last summand is given by the first
entry of a six-dimensional white noise error term , which we again assume to be
normally distributed. The explanatory variables are modeled using the same principle.
For the estimation of the models we utilize the R package “URCA” by Bernhard
Pfaff, where we find a fast estimation procedure based on the seminal papers by
Johansen in [8]– [10]. Further tests in the VECM framework are implemented as well.
These are typically employed to estimate the cointegration rank and to test if the
cointegration vector equals a certain exogeneously given vector, all of which we do
not need for our application of the model. The tests are rather involved and the test
statistics depend either on the rank or eigenvalues of the cointegration matrix or
on the residual vector . However, we checked on the rank and always found at least
one cointegration relation with the “top five” variables obtained through the selection
algorithm above which suffices for our task.
As an illustrative example, we find the cointegration relationship between the
logarithms of spot prices of the tanker route TD3 and five explanatory variables. In
order to be able to check the results, we manually selected five explanatory variables,
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch04 FA
We have made two models available and we will see in this section that both allow for
out of sample predictions of future prices. The extension of the VECM over the VAR
is the additional balance term which models the cointegration relationship. Since we
work with log returns in the VAR model and apply the VECM with its first differences
to log prices, the two models become well comparable. For a given calibration day T ,
we apply the following procedure to draw out of sample forecasts of freight spot and
freight futures prices F over the period [T, T + τ]. In the following, we set τ = 100,
and therefore make forecasts for the next 100 working days.
1. Preparation: prepare the historical data used for the estimation: for the freight rate
F under consideration and the potential explanatory variables:
1.1 Specify the estimation window [t, T ].
1.2 Exclude the explanatory variables which are too sparse in the estimation
window.
1.3 Fill the minor data gaps using Brownian Bridges.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch04 FA
1.4 Transform the prices to log prices for the VECM model.
1.5 Calculate the log returns for the VAR model.
2. Top Five: out of the remaining pool of cleaned explanatory variables, select the five
variables with the biggest combined explanatory power in the estimation window
via the algorithm explained in Sec. 4.4.3.
3. Estimate models: estimate the parameters of the VAR and VECM model from
Sec. 4.4 in the six-dimensional setup using the data in the estimation window.
4. Draw out of sample forecasts for the period [T, T + τ] and transform these to actual
FFA rates, i.e., transform predicted log returns to prices in the VAR model, and the
forecasted differenced log prices to actual prices in the VECM model.
The predictions occurring in the fourth step above are obtained in the natural way: In
two dimensions, the VAR model is set to the specification:
20 20
µ + aj ft−j + bj et−j
f f
ft j=1 j=1 t
= + ,
et 20
20 et
µ + ce + df
e j t−j j t−j
j=1 j=1
20
20
f̂T +1 = µf + aj fT +1−j + bj eT +1−j .
j=1 j=1
Further estimated log returns fˆT +2 , . . . , fˆT +τ are then received recursively, treating
the estimated forecasted values needed for the next day forecast as if they were true
log returns.
The step in the VECM model is carried out similarly: The model is given as
20
yt = µ + j · yt−j + · yt−1 + t ,
j=1
and increasing the time by one day as well as setting the random term to its mean
yields the estimate
20
ŷT +1 = µ + j · yT +1−j + · yT .
j=1
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch04 FA
90
80
70
60
50
40
30
20
10
Figure 4.2 Forecast carried out on 2/1/2007 for 100 working days and compared to actual
prices.
That is, in two dimensions only we get for the FFA term
20
20
log F̂T +1 = µf + aj log FT +1−j + bj log ET +1−j
j=1 j=1
Working with all the uncertainties of the freight market, like its sudden price movements
in spot as well as futures rates, we cannot expect that our forecasts will yield the correct
prices without error. We can however examine if this error is in the range of a very
conservative forecast model.
We examine the forecast error of our model with the following back test. As an error
quantity we define for a τ days ahead prediction carried out on day T the prediction
error PE as following:
F
T +τ − F̂T +τ
PET,τ =
,
FT
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch04 FA
i.e., we compare the absolute value of the relative error of actual rates to model implied
predicted rates. Note that PET,0 = 0. For the actual test, we apply an increasing window
with the initial window covering the period from 1/1/2005 to 1/1/2007. Starting with
this window, we forecast 100 working days and then increase the window by one
working day and repeat the forecast. We advocate the use of an increasing window
over the use of a moving window because the length of the data set allows robust
estimations of the model. For the results described below, we used 20 lags. Increasing
the window on a daily basis allows for 550 calibrations until reaching the end of the
time series on 20/2/2009.
As an additional reference, we include a third very conservative forecast model,
where we simply assume that the prices remain constant.
In order to demonstrate the developed methods, we perform the following backtest.
After preparing the data as described above, we run the following algorithm for the
spot and futures rates of the three routes under consideration.
1. Initializing:
2. Loop windows:
2.1. In each of the 550 possible steps, increase the window by one working day to
get the new window [1, T i+1 ].
2.2. Estimate the VAR and VECM model parameters using data only available
in the window, and draw forecasts as described in Sec. 4.5 for the next 100
working days beyond T i+1 .
2.3. Calculate and save the prediction error PET i+1 ,τ , for τ = 1, . . . , 100 for the
VAR and VECM model as well as for constant forecasts.
To visualize the results, we average the prediction error PET i ,τ for τ ∈ {1, 2, . . . , 100}
over the 550 out-of-sample forecasts drawn from the VAR and VECM model as well
as the constant forecasts. The resulting Fig. 4.3 shows the relative error for the TD3
spot prices. Starting with PET,0 = 0 for the “0-days”-ahead prediction, it follows the
error of the conservative constant out-of-sample forecast for the first “10-days”-ahead
predictions but then the error increases slightly.
This shows that the models are able to yield stable forecasts over periods as long
as 100 days ahead. Additional benefits of the models over the constant predictions are
for example readily available confidence regions and the possibility to check on trends
rather than just having constant predictions.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch04 FA
TD3 Spot
0.7
const.
0.6 VAR
VECM
0.5
0.4
0.3
0.2
0.1
0
0 10 20 30 40 50 60 70 80 90 100
4.7. CONCLUSION
After an introduction to the freight market and its two main derivatives, we examined
the relation of the highly volatile freight rates with explanatory variables. Incorporating
such physical and economical relations in a model is desirable, as it leads to a better
understanding of the freight market and may improve out-of-sample forecasts. We
proposed the use of two statistical methods to jointly model freight rates and related
explanatory variables.
After two very favorable years, the freight market currently suffers from the world
financial crisis. Large investments over the last years have made shipping capacities
available, but freight rates fell to a level which drove many participants out of the
physical market as well as its derivatives. However, with a revival of the world’s
economy, it is likely to see an upturn of the freight market again.
References
Power forward contracts deliver electricity over a specified period, and can be viewed
as a portfolio of forwards with maturity at each time instant in the delivery period. We
investigate the implied power forward dynamics from a geometric Brownian motion
specification of the forward price, which turns out to have a very complicated struc-
ture. Lognormal approximations are argued for, and we demonstrate that they work
excellently in many situations. In particular, we focus on the approximation suggested
by Bjerksund et al. [8], where the volatility of the power forward is simply the aver-
age of the fixed-maturity forward volatility. Although giving a superior model to the
moment matched dynamics, it fails to estimate the tails of the power forward distribu-
tion in some cases with extreme volatility and mean-reversion. We provide analytical
bounds in terms of geometric Brownian motions for the power forward dynamics, and
also compare the covariance structure with those implied by a geometric Brownian
motion.
93
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch05 FA
94 Benth
5.1. INTRODUCTION
The power markets have in the recent two decades been liberalized in Europe, America,
Asia, and Australia. This freeing of trade in power is a process still going on, and
gradually attracting investors not having any physical interest in power production or
consumption. Large investors like investment banks, hedge funds, and pension funds
see opportunities for portfolio diversification and speculation with this new asset class.
The power markets are naturally divided into spot and forward products, where the
former are usually linked to physical production and consumption, while the latter are
financial.
Naturally, investors without physical capacity are interested in the forward markets
for power. We shall concentrate on the markets for electricity, however, coal, gas, and
oil are also an integrated part of the power market being important fuels for power
production. Taking the Nordic market for electricity, NordPool, as an example, the
power forward contracts are financially settled against the system price in the market.
The system price is the “spot” of this market, an auction-based hourly price for next-
day delivery of electricity. The forward contracts are both of futures and forward type,
and deliver the money-equivalent to receiving the spot over a pre-determined delivery
period. Typically, the delivery period can be a week, month, quarter or year ahead, and
the forward price is given in terms of Euro/MWh.
In this paper, we will be concerned with the analysis of models for the forward
price dynamics. The basis for our studies is taken from Benth and Koekebakker [4],
Bjerksund et al. [8] and Kiesel et al. [17] (see also Benth et al. [6]), where the Heath–
Jarrow–Morton (HJM) technique taken from fixed-income theory is applied to the
power forward market. The main obstacle in using the HJM approach to power markets
is the fact that the forward contracts do not deliver the underlying spot at a fixed
maturity time in the future, but over a delivery period. This creates difficulties in stating
reasonable models, which are analytically and empirically tractable. The standard
approach is to suppose a dynamics for a (non-existing) fixed maturity contract, and
derive the power forward price by averaging over the delivery period. We investigate the
properties of such models and approximations, to understand the stochastic dynamics
of power forward prices.
The motivation for analyzing the price dynamics of power forwards is twofold.
First, by investing in such power products one is naturally interested in the risk asso-
ciated with the strategy chosen. To get hold of this, a proper understanding of the
stochastics of the dynamical behavior of prices is required. This entails in an analysis
of the volatility of the price. Secondly, there are markets for options on power forwards,
both organized and OTC. The arbitrage-free prices of such options are naturally con-
nected to the volatility of the underlying forward as well.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch05 FA
Our analysis and results are presented as follows: In the next section, we describe
the HJM approach to power forwards, recalling some existing results from Benth
et al. [6]. In Sec. 5.3, we discuss the stochastic dynamics of power forwards, and link
it to approximative geometric Brownian motion processes. In particular, we study the
Bjerksund, Rasmussen, and Stensland approximation to the power forward dynamics,
discuss the covariance structure of power forwards with different delivery periods
and their distributional properties. Several empirical examples are studied. Option
pricing is the topic in Sec. 5.4, where we compare the approximation by Bjerksund,
Rasmussen, and Stensland with the true prices, both for plain vanilla call options and
calendar spreads. We conclude in the final section.
Suppose St is the spot price of power at time t, being a stochastic process defined
on some filtered probability space (, F, F, Q) with time t running in the finite time
interval [0, Tmax ]. We assume that the filtration F satisfies the usual conditions (see
page 10 in Karatzas and Shreve [16]), and supports a Wiener process W = {Wt }t≥0 .
Throughout the paper, we work exclusively under a risk–neutral probability Q.
A long position entered at time t in a power forward contract settled continuously
against the spot price over a period [T1 , T2 ] will yield a profit/loss
T2
e−r(u−t) (Su − Ft (T1 , T2 ))du. (5.1)
T1
with
re−ru
w(u, T1 , T2 ) = . (5.3)
e−rT1 − e−rT2
If the forward is settled at the end of the delivery period T2 , we use w(u, T1 , T2 ) =
1/(T2 − T1 ) instead (see Sec. 4.1 in Benth et al. [6] for details). Obviously, we
implicitly assume that w(u, T1 , T2 )Su is integrable with respect to Q × du on
× [T1 , T2 ].
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch05 FA
96 Benth
for all 0 ≤ T1 < T2 . This leads us back to a model for Ft (u, u), which is equivalent to
a forward price for a contract with fixed maturity time u. Hence, from this view, it is
natural to use the HJM approach to state models for ft (u), and next derive the power
forward price Ft (T1 , T2 ) from the relation (5.4). In this paper, we shall investigate the
dynamics of the resulting power forward price, and investigate approximations of it in
view of option pricing.
Since ft (u) has to be a martingale, the dynamics will have the form
dft (u) = σ̃t (u)dWt ,
where t → σ̃t (u) is a stochastic process on [0, u] which is supposed to be integrable
with respect to W . The model is defined for general choices of maturity times u, with
the only restriction that u ≤ Tmax . To simplify matters, we shall concentrate on a
geometric Brownian motion structure, and assume that
dft (u) = σt (u)ft (u)dWt . (5.6)
Here, we let σt (u) be a deterministic volatility function with 0 ≤ t ≤ u ≤ Tmax ,
and t → σt (u) is square-integrable on [0, u] for every u ≤ Tmax . We have a model
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch05 FA
specifying the full behavior of the forward curve dynamics driven by a single Brownian
motion Wt . We remark in passing that there is no difficulty in assuming a multi-factor
model, but to stay focused we refrain from doing so.
Note that the volatility σt (u) depends on the maturity time u. This allows us to
incorporate the Samuelson effect into the model. The Samuelson effect tells us that the
volatility of the forward price converges to the spot volatility as time to maturity tends
to zero. This effect is heavily based on a mean-reverting property of the underlying
spot price dynamics. In fact, assume that the (risk–neutral) spot price follows the
exponential of an Ornstein–Uhlenbeck process Xt ,
dXt = α(µ − Xt )dt + σdWt ,
with α, σ, and µ being constants, and α, σ > 0. Then, we easily derive that
dft (u) = σe−α(u−t) ft (u)dWt .
Thus, σt (u) = σ exp(−α(u − t)), and letting u − t → 0, we find that the forward
volatility tends to σ at an exponential speed given in terms of the mean–reversion
coefficient α. There is both economical and empirical evidence for an explicit depen-
dence on the maturity time in the volatility function (see Geman [13] and Eydeland and
Wolynieck [10]). Other specifications of a term structure of volatility σt (u) for energy
and commodity markets have been suggested (see Clewlow and Strickland [9]). For
instance, since forwards in the long end of the curve indeed show price volatility con-
trary to the mean–reversion model discussed above, it has been suggested that the
volatility looks like
σt (u) = σ0 + σ1 e−α(u−t) , (5.7)
for two positive constants σ0 and σ1 . The first term σ0 is coming from a non-stationary
factor in the spot prices, like for instance a non-stationary long-term stochastic mean.
One may also introduce seasonality in the volatility term structure. See Benth and
Koekebakker [4] for a discussion of the various models with references to related
work. The volatility model in (5.7) will be our choice in many examples to follow, and
serve as a canonical choice of volatility in power markets.
The degenerate case when σt (u) = σt gives an explicit geometric Brownian motion
dynamics for Ft (T1 , T2 ). Since in that case
t
1 t 2
ft (u) = f0 (u) exp − σs ds + σs dWs
2 0 0
we find
T2 t
1 t 2
Ft (T1 , T2 ) = w(u, T1 , T2 )f0 (u)du exp − σ ds + σs dWs
T1 2 0 s 0
t
1 t 2
= F0 (T1 , T2 ) exp − σ ds + σs dWs .
2 0 s 0
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch05 FA
98 Benth
Hence, we have
dFt (T1 , T2 ) = σt Ft (T1 , T2 )dWt . (5.8)
When σt (u) explicitly depends on u, the simple geometric Brownian motion structure
is no longer true.
Before we proceed with analyzing the dynamics of Ft (T1 , T2 ), let us discuss how
one could infer estimates on the parameters of ft (u) when data for such contracts are
not available. One frequently used method is to smoothen the observed power forward
curve, that is, to derive a continuous curve of forward prices fˆt (u) from the discretely
observed power forwards Fˆ t (T1 , T2 ). A way to construct such a curve is proposed in
Benth et al. [5]. From a given a priori seasonal function, a fourth order polynomial
spline is fitted using a maximum smoothness criterion so that the implied power forward
prices are within the bid–ask spread of the market prices. Another similar approach is
described in Fleten and Lemming [11]. For each day, one can construct a time series
of such curves fˆt (u), t = 0, 1, . . . , and these can be used to infer estimates on the
volatility term structure σt (u) of the process ft (u). Obviously, one also gets estimates
on the drift of ft (u) from these data, where this drift is known as the risk premium in
the market.
A different way to approach the inference problem is to use implied volatilities
from quoted option prices in the market. This way to estimate is problematic since we
need to derive theoretical option pricing formulas for call and put options, which is not
straightforward at all due to the delivery period of the power forwards. Thus, we end
up with first having to perform a Monte Carlo simulation of prices in combination with
a numerical integration, a procedure which is slow when the goal is to derive implied
estimates on the volatility term structure σt (u). To speed up this process, one may
use some approximations of the dynamics of Ft (T1 , T2 ) where analytical (Black-76)
option pricing formulas become available. The validity of such approximations is the
target of the study in this paper.
One may leave completely the specification of ft (u), and directly specify the
dynamics of the power forwards Ft (T1 , T2 ), only for those being actually traded in
the market. This kind of LIBOR model for the power forward market is analyzed in
Benth and Koekebakker [4] and Kiesel et al. [17]. In the latter paper, they use it to find
implied estimates for the volatility term structure of power forwards in the German
EEX market.
In this Section, we analyze the time dynamics of power forwards resulting from the
dynamics of ft (u). Our standing assumption is the geometric Brownian motion struc-
ture of ft (u) defined in (5.6). We further analyze approximations of the power forward
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch05 FA
where we have used the stochastic Fubini Theorem (see pages 159, 160 in Protter [20]).
Next, applying a straightforward integration by parts yields
T2
w(u, T1 , T2 )σs (u)fs (u)du
T1
T2 u
∂σs (u)
= σs (T2 )Fs (T1 , T2 ) − w(v, T1 , T2 )fs (v)dvdu.
T1 ∂u T1
Dividing and multiplying with w(v, T1 , u) in the dv-integral on the right-hand side,
and observing that w(v, T1 , T2 )/w(v, T1 , u) is independent of v, gives us the following
expression:
T2
w(u, T1 , T2 )σs (u)fs (u)du
T1
T2
∂σs (u) w(v, T1 , T2 )
= σs (T2 )Fs (T1 , T2 ) − Fs (T1 , u)du.
T1 ∂u w(v, T1 , u)
Collecting terms together, we have proven:
Proposition 5.1. Given the dynamics (5.6) for ft (u), the dynamics of the power for-
ward Ft (T1 , T2 ) is
T2
dFt (T1 , T2 ) ∂σt (u) w(v, T1 , T2 ) Ft (T1 , u)
= σt (T2 ) − du dWt , (5.9)
Ft (T1 , T2 ) T1 ∂u w(v, T1 , u) Ft (T1 , T2 )
where w(v, T1 , T2 )/w(v, T1 , u) is independent of v.
Although ft (u) has a simple geometric Brownian motion dynamics with maturity-
dependent volatility, the power forward has a complex dynamics which involves not
only the power forward Ft (T1 , T2 ) itself, but the states of all power forwards Ft (T1 , u)
for u ∈ (T1 , T2 ]. The geometric Brownian motion structure is thus not inherited from
ft (u). We also note that a volatility σt (u), which is not depending on u will have
∂σt (u)/∂u = 0, and we obtain a geometric Brownian motion dynamics as we already
saw above in the derivation of (5.8). In the general case, this is a rather complex model,
especially from the point of view of simulation and pricing of options since in principle
it involves an infinite number of power forwards.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch05 FA
100 Benth
We see that the dynamics of Ft (T1 , T2 ) has a stochastic volatility described as a ratio
of two weighted averages of ft (u). The stochasticity comes in through fluctuations
of ft (u) in time t. If σt (u) is rather flat as a function of maturity u, we expect the
stochastic volatility to be close to σt . This happens, say, when we have a volatility
function for which limu→∞ σt (u) = σ̂t . Then, for delivery periods in the long end of
the curve, that is, T1 − t large, we will approximately have
T2 T2
w(u, T1 , T2 )σt (u)ft (u)du ≈ σ̂t w(u, T1 , T2 )ft (u)du,
T1 T1
and hence,
I.e., in the long end of the power forward curve, the prices move approximately accord-
ing to a geometric Brownian motion with maturity-independent volatility given as the
limit of σt (u) with respect to u. In the discussion leading to this result, we have implic-
itly assumed that ft (u) is well-behaved as a function of u ∈ [T1 , T2 ], for instance, that
it is continuous. As an example, consider σt (u) given as in (5.7). Then
lim σt (u) = σ0
u→∞
for all t ≥ 0. Thus, in this case, the power forward will for large values of T1 − t
behave approximately as a geometric Brownian motion with volatility given by the
long-term level σ0 . We note that closer to maturity the power forward will naturally
have increasing volatility due to the Samuelson effect of ft (u), and thus its dynamics
will depart from the constant volatility geometric Brownian motion.
However, one may ask the question whether there exist reasonable geometric
Brownian motion models, which approximate the power forward dynamics. Looking
at the mean-value theorem in calculus, we have that
T2
∗
ft (u (t)) = w(u, T1 , T2 )ft (u)du,
T1
∗
for some u (t) ∈ (T1 , T2 ). Hence,
which gives an explicit relation between a power forward price dynamics and a forward
price ft (u) being a geometric Brownian motion. However, the latter has a parameter
u, which in this case will depend on time t. The connection motivates to look for
approximative models, which indeed are lognormal processes.
Following the argument in Benth and Henriksen [3], we use a Taylor expansion
of the logarithmic function to find
1
2 ln(ft (u) + ft (v)) = ln ft (u) + ln ft (v) + (cosh(ln(ft (u)/ft (v))) − 1)
2
+ 2 ln(2) + O((ln(ft (u)/ft (v)))4 ).
Now, for two nearby delivery times u and v, ft (u) ≈ ft (v) as long as σ(t, u) is
sufficiently regular in u. Hence, we see that the logarithm of ft (u) + ft (v) is close to
the sum of logarithms of ft (u) and ft (v) and the constant 2 ln(2), which is following
a drifted Brownian motion. The crucial observation here is that there is the same
Brownian motion Wt driving both ft (u) and ft (v) making them perfectly correlated,
and it is only the regularity of σt (u) as a function of maturity that decides how good
the approximation is. Indeed, in Henriksen [14], it is demonstrated empirically that
a sum of positively correlated lognormal variables are very well approximated by a
lognormal variable. Of course, in our case, we have an integral of lognormal variables,
but still this discussion points toward the reasonability of geometric Brownian motion
approximations for power forwards.
We remark that a geometric Brownian motion dynamics for Ft (T1 , T2 ) in general
will violate the no-arbitrage condition (5.5) (see Lemma 6.1, page 168 in Benth et al.
[6]), and thus cannot be a valid dynamics for the whole power forward curve. However,
as dynamics for a contract with a specified delivery period [T1 , T2 ], it is arbitrage-free.
Since analytical pricing formulas for plain vanilla options are available for geometric
Bronwian motion dynamics, such approximations are particularly interesting.
102 Benth
where
w(u, T1 , T2 )σt (u)
wσ (t, u, T1 , T2 ) = T2 .
T1 w(v, T1 , T2 )σt (v)dv
T
Since wσ (t, u, T1 , T2 ) is non-negative and T12 wσ (t, u, T1 , T2 )du = 1, it is a weight-
ing function. Hence, the fraction on the right-hand side of the dynamics (5.11) is
the ratio between two weighted averages of ft (u) over u ∈ [T1 , T2 ]. As long as
wσ (t, u, T1 , T2 ) ≈ w(u, T1 , T2 ), this is close to one. This is equivalent to σt (u) being
approximately equal to its weighted average, i.e., being close to flat over the delivery
period. We consider an example, where this is true.
Consider σt (u) = σ exp(−α(u − t)) and w(u, T1 , T2 ) = 1/(T2 − T1 ). We find
αe−αu
wσ (t, u, T1 , T2 ) = .
e−αT1
− e−αT2
Note that ∂wσ (t, u, T1 , T2 )/∂u = −αwσ (t, u, T1 , T2 ) < 0, and the weighting function
is decreasing as a function of u. Its maximal value becomes
wσ (t, T1 , T1 , T2 ) = α(1 − e−α(T2 −T1 ) )−1 ,
and minimal value
wσ (t, T2 , T1 , T2 ) = α(eα(T2 −T1 ) − 1)−1 .
Moreover,
wσ (t, T1 , T1 , T2 ) − wσ (t, T2 , T1 , T2 ) = α.
If α is small, then the variation in wσ is correspondingly small. Furthermore, as long
as α(T2 − T1 ) is small, we have
eα(T2 −T1 ) ≈ 1 + α(T2 − T1 ).
Hence,
1
wσ (t, u, T1 , T2 ) ≈ = w(u, T1 , T2 ).
T2 − T 1
Thus, we have in this case that
T2
dFt (T1 , T2 ) ≈ w(u, T1 , T2 )σt (u)duFt (T1 , T2 )dWt .
T1
This is the approximation of the power forward price dynamics suggested by Bjerksund
et al. [8], and we introduce the BRS-approximation as
dFtBRS (T1 , T2 ) = BRS
t (T1 , T2 )FtBRS (T1 , T2 )dWt (5.12)
with the BRS-volatility function given by
T2
BRS
t (T1 , T2 ) = w(u, T1 , T2 )σt (u)du. (5.13)
T1
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch05 FA
The BRS-approximation will play a central role in the analysis to follow, since this
is the benchmark model for many applications and seems to be frequently used in
practice.
The approximation in (5.12) can also be argued for by looking at the returns of
ft (u). Thinking of Ft (T1 , T2 ) as a portfolio of ft (u)’s, it may seem natural that the
return from Ft (T1 , T2 ) is the sum of returns from the “assets” in the portfolio, namely
ft (u) for u ∈ [T1 , T2 ]. Hence,
T2 T2
dFt (T1 , T2 ) dft (u)
≈ w(u, T1 , T2 ) du = w(u, T1 , T2 )σt (u)du dWt .
Ft (T1 , T2 ) T1 ft (u) T1
In effect, in the above approximation, we are boldly assuming that the self-financing
property holds. Thus, we have again reached the approximative model in (5.12). It is
of course a question whether the return of a power forward approximately is equal to
the weighted average of the return of ft (u) in general.
dFtm (T1 , T2 ) = m m
t (T1 , T2 ) Ft (T1 , T2 )dWt , (5.14)
where the subscript “m” indicates that we are looking for a dynamic matching the first
two moments of F . The deterministic volatility function t → m t (T1 , T2 ) is assumed
to be square integrable for t ∈ [0, T1 ] to ensure that Ftm (T1 , T2 ) is a martingale for
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch05 FA
104 Benth
First, by the martingale properties of Ftm (T1 , T2 ) and ft (u) we find that
T2
E[Ftm (T1 , T2 )] = F0m (T1 , T2 ) = w(u, T1 , T2 )f0 (u)du = F0 (T1 , T2 ).
T1
The expectation inside the double integral on the right-hand side can be calculated by
appealing to well-known properties of geometric Brownian motion (note that we have
u ≤ v here),
1 t 2
E[ft (u)ft (v)] = f0 (u)f0 (v) exp − σs (u) + σs2 (v)ds
2 0
t
× E exp σs (u) + σs (v)dWs
0
1 t 2
= f0 (u)f0 (v) exp − σ (u) + σs (v)ds
2
2 0 s
t
1
exp (σs (u) + σs (v))2 ds
2 0
t
= f0 (u)f0 (v) exp σs (u)σs (v)ds .
0
t
F02 (T1 , T2 ) exp (m 2
s ) (T1 , T2 )ds
0
T2 v
=2 w(u, T1 , T2 )w(v, T1 , T2 )f0 (u)f0 (v)
T1 T1
t
× exp σs (u)σs (v)ds dudv.
0
T2 v
ξ(t, T1 , T2 ; η) = w(u, T1 , T2 )w(v, T1 , T2 )f0 (u)f0 (v)
T1 T1
t
×e 0 σs (u)σs (v)ds
η(u)η(v)dudv. (5.15)
ξ(t, T1 , T2 ; σt (·))
t ) (T1 , T2 ) =
2
(m . (5.16)
ξ(t, T1 , T2 ; 1)
For a given volatility function σt (u), we can calculate m t (T1 , T2 ) from (5.16).
From the definition of (5.15), we see that m
t (T 1 , T 2 ) depends on the initial forward
curve f0 (u) as well as the volatility σt (u) and the weighting function w(u, T1 , T2 ).
The integrand in the double integral forming ξ(t, T1 , T2 ; η) is rather complex due
t
to the exponentiating of 0 σs (u)σs (v)ds, and will not in general admit any simple
closed-form expression as is mostly the case for the BRS-volatility BRS t (T1 , T2 ). For
example, choosing the standard volatility σt (u) as in (5.7) leads to a double integral
of the exponential of an exponential. Hence, in order to calculate m t (T1 , T2 ), we
must perform a numerical integration. Even worse, we need to do such a numerical
integration at each time t we want to have the value of Ftm (T1 , T2 ) since the expression
inside the integrals is t-dependent. This may become very time consuming in practice.
However, when pricing a plain vanilla call option, we get away with one such numerical
integration, as we see from the Black-76 Formula (see Sec. 5.4). In any case, the
expression for the volatility resulting from moment matching does not seem to be as
operable in practice as for instance the BRS-volatility.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch05 FA
106 Benth
Let us compare m t (T1 , T2 ) with the BRS-volatility in (5.13). Note that we may
write
T2 2
ξ(t, T1 , T2 ; σt (·)) = w(u, T1 , T2 )σt (u)du ξ(t, T1 , T2 ; σ t (·, T1 , T2 )),
T1
with
σt (u)
σ̄t (u, T1 , T2 ) = T2 .
T1 w(v, T1 , T2 )σt (v)dv
If
w(u, T1 , T2 )σt (u)
wσ (t, T1 , T2 ) = T2 ≈ w(u, T1 , T2 ),
T1 w(v, T1 , T2 )σt (v)dv
we find
T2 2
ξ(t, T1 , T2 ; σt (·)) ≈ w(u, T1 , T2 )σt (u)du ξ(t, T1 , T2 ; 1).
T1
Recall that if σt (u) is given by (5.7) and w(u, T1 , T2 ) = 1/(T2 − T1 ), this property
holds as long as α, the speed of mean–reversion, is small. Hence,
t (T1 , T2 ) ≈ t
BRS
m (T1 , T2 ).
The two ways of finding a “reasonable” geometric Brownian motion specification will
therefore in some relevant cases be approximately equal. However, the expression for
BRS
t (T1 , T2 ) is considerably simpler to find, since it allows for analytical expressions
in most interesting cases, which is not the case for m t (T1 , T2 ). We shall therefore
concentrate our analysis on the BRS-approximation.
and
Let us compare the two expressions for the covariance. Introduce the weight
function
w(u, T1 , T2 )f0 (u) w(u, T1 , T2 )f0 (u)
wf (u, T1 , T2 ) = T2 = ,
F0 (T1 , T2 )
T1 w(v, T1 , T2 )f0 (v)dv
to find
T2 T4
= F0 (T1 , T2 )F0 (T3 , T4 ) wf (u, T1 , T2 )wf (v, T3 , T4 )
T1 T3
s
× (e 0 σr (u)σr (v)dr
− 1)dudv,
108 Benth
Proof. To prove the lower bound, use Jensen’s Inequality for the concave function
ln x to find
T2 T2
ln w(u, T1 , T2 )ft (u)du ≥ w(u, T1 , T2 ) ln ft (u)du.
T1 T1
and variance
t T2 2
w(u, T1 , T2 )σs (u)du ds.
0 T1
The latter is the variance of the logreturn from the BRS-approximation (5.12), where
we recall from (5.13) that
T2
BRS
t (T1 , T2 ) = w(u, T1 , T2 )σt (u)du.
T1
By the definition of FtBRS (T1 , T2 ) and the deterministic function c(t, T1 , T2 ) in the
Proposition, we find after taking exponentials
T2
w(u, T1 , T2 )ft (u)du
T1
T2 t
1 t 2
= w(u, T1 , T2 )f0 (u) exp − σ (u)ds + σs (u)dWs du
T1 2 0 s 0
T2 t
1 t 2
= F0 (T1 , T2 ) wf (u, T1 , T2 ) · exp − σs (u)ds + σs (u)dWs du.
T1 2 0 0
Again appealing to Jensen’s Inequality, but now using the convex function exp(x), we
find
T2 t
1 t 2
wf (u, T1 , T2 ) exp − σs (u)ds + σs (u)dWs du
T1 2 0 0
1 t T2
≤ exp − wf (u, T1 , T2 )σs2 (u)duds
2 0 T1
t T2
+ wf (u, T1 , T2 )σs (u)dudWs .
0 T1
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch05 FA
110 Benth
f
Thus, by the definition of Ft (T1 , T2 ),
T2 2
f 1 t T2
w(u, T1 , T2 )ft (u)du ≤ Ft (T1 , T2 ) exp wf (u, T1 , T2 )σs (u)du
T1 2 0 T1
T2
− 2
wf (u, T1 , T2 )σs (u)du ds
T1
f
≤ Ft (T1 , T2 ).
In the last inequality, we applied Jensen’s Inequality to the convex function x2 . This
shows the upper bound, and the Proposition is proved.
In the lower bound for Ft (T1 , T2 ), the deterministic function c(t, T1 , T2 ) is less that
one. This we see by using Jensen’s Inequality on the two convex functions exp(x)
and x2 . This means that the lower bound of Ft (T1 , T2 ) is less than FtBRS (T1 , T2 ), and
in theory the power forward price can be below the BRS-approximation. However,
there is a geometric Brownian motion bounding the power forward from above. This
geometric Brownian motion may even bound the BRS-approximation.
Figure 5.1 The distributions from simulations of a weekly (left column) and quarterly (right
column) power forward contract at time t = 1 with row 1 showing the “moderate” market
scenario, whereas rows 2 shows the “spiky” market scenarios. The BRS approximation is shown
in dashed lines, all distributions are plotted with a logarithmic frequency axis.
depicted in the right column. The BRS approximation is included with a dashed line,
using a logarithmic scale on the frequency axis. The first row in each figure shows the
“moderate” scenario, while the second row depicts the “spiky” case.
We observe that the distributions become wider when approaching T1 , in line
with the fact that σt (u) is increasing when approaching time to maturity. We notice
that the BRS approximation has close to an excellent fit for all sampling times t, the
two contract types, and the different market scenarios. This indicates that the BRS
approximation works excellently for short- and long-delivery periods, no matter how
far or close to start of delivery of the contract we are. However, one may spot a
slight underestimation of the true distribution for the weekly contract in the “spiky”
market scenario with t = 50. It seems that the BRS approximation for this situation is
underestimating in both tails.
To investigate this closer, we looked at the distribution of a weekly and monthly
power forward, both starting at T1 = 50, and maturing at T2 = 55 and T2 = 70
respectively. Further, since the effect seems to come from the exponential increase
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch05 FA
112 Benth
Figure 5.2 The distributions from simulations of a weekly (left column) and quarterly (right
column) power forward contract at time t = 25 with row 1 showing the “moderate” market
scenario, whereas row 2 shows the “spiky” market scenarios. The BRS approximation is shown
in dashed lines, all distributions are plotted with a logarithmic frequency axis.
in volatlity, we let σ0 = 0 in the “spiky” market scenario. In Fig. 5.4, we plot the
distributions for the two contracts in the two market scenarios at time t = 50. We
clearly see that the “spiky” market scenario produces a distribution for the power
forward, which has heavier tails than the BRS approximation. The weekly contract is
most extreme, whereas the monthly only has a heavier right tail. The effect is very small
in the “moderate” market scenario, perhaps most pronounced for the weekly contract.
It thus seems like that the BRS approximation becomes better when the delivery period
increases, and the farther away we are from T1 , the start of delivery.
To further analyze the distributional differences between the BRS approximation
and the power forward, we include in Fig. 5.5 plots of the 1%, 5%, 95%, and 99%
quantiles of the two for all times t = 40 up to t = 50, the start of delivery. We present
the weekly and monthly contracts, separating the moderate and spiky market scenar-
ios. The “moderate” market does not show any difference between the quantiles of
the power forward and its BRS approximation. Here, the tails are captured excellently
by the BRS approximation. The BRS approximation fails seriously, however, on the
95% and 99% quantiles of both, weekly and monthly forward contracts in the “spiky”
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch05 FA
Figure 5.3 The distributions from simulations of a weekly (left column) and quarterly (right
column) power forward contract at times t = 50 with row 1 showing the “moderate” market
scenario, whereas row 2 shows the “spiky” market scenarios. The BRS approximation is shown
in dashed lines, all distributions are plotted with a logarithmic frequency axis.
scenario. In fact, the quantiles are close to 5% underestimated by the BRS aproxi-
mation. There is also an underestimation of the lower quantiles, in particular for the
weekly contract. From the figure, we also see how the interplay between volatility
and mean–reversion makes the “moderate” market scenario being more risky far from
delivery, while the “spiky” regime becomes gradually more riskier toward the start of
delivery, and eventually, for times t close to 50, being significantly more volatile than
the “moderate” scenario. If we would include a contribution from σ0 in the “spiky”
scenario, the effects would be reduced. Furthermore, we clearly see that the longer the
delivery period, the less volatile are the prices. There is a significant smoothing effect
in the length of delivery period.
From these examples, we conclude that the BRS approximation works excellently
for cases, where the market is not mean–reverting strongly and where the volatility
is not high (in relative terms). When we use the approximation for markets with very
fast mean–reversion and high volatility, the tails are not well explained by the BRS
approximation. It seems to underestimate significantly the extreme quantiles, an effect
which will lead to an underestimation of option prices if considering for instance far
out-of-the-money calls.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch05 FA
114 Benth
Figure 5.4 The distributions of a weekly (left column) and monthly (right column) power
forward contracts at time t = 50 with row 1 showing the “moderate” market scenario, whereas
row 2 shows the “spiky” market scenarios. The BRS approximation is shown in dashed lines,
all distributions are plotted with a logarithmic frequency axis.
Proposition 5.4. Let r be the risk-free interest rate, K the strike, and 0 ≤ T ≤ T1
the exercise time for a call option written on a forward contract with price process
FtBRS (T1 , T2 ) defined as in (5.12). Then, the arbitrage-free option price at time t ≤ T is
18 18
16 16
14 14
12 12
10 10
8 8
6 6
4 4
40 45 50 40 45 50
13 13
12.5 12.5
12 12
11.5 11.5
11 11
10.5 10.5
10 10
9.5 9.5
9 9
8.5 8.5
8 8
40 45 50 40 45 50
Figure 5.5 The 1%, 5%, 95%, and 99% quantiles of a weekly (top row) and monthly (bottom
row) power forward contracts at all times t up to ten days prior to start of delivery. Left column
shows the case of “moderate” market, while the right row depicts the results for the “spiky”
market. The BRS approximation is shown with slashed lines.
where
T
d1 = d2 + (BRS
s )2 (T1 , T2 )ds
t
T
ln(FtBRS (T1 , T2 )/K) − 0.5 t (BRS s )2 (T1 , T2 )ds
d2 =
T BRS )2 (T , T )ds
t (s 1 2
116 Benth
For many relevant cases like the various choices of σt (u) considered in this paper,
the integral of (BRS
s )2 (T1 , T2 ) in the expression for the Black-76 price can be ana-
lytically calculated, and hence the formula provides a very rapid way to price call
options. Note that the moment matching volatility m t (T1 , T2 ) defined in (5.16) can
be substituted with the BRS volatility in the option pricing formula. However, the
definition of mt (T1 , T2 ) is complicated, and for the interesting situations it does not
have any analytical expression, and numerical integration must be used to price options.
In fact, one must already use a numerical approach to integrate σt (u) and the initial
forward curve f0 (u) to obtain the moment-matching volatility m t (T1 , T2 ). This may
be time consuming and not practical, in particular if we want to price dynamically or
to derive a hedge. In such a situation, we must update the volatility m at each time
step, which would be significantly more efficient to do with the BRS approximation.
As a numerical example, consider a call option on a power forward with weekly
delivery. The fixed-maturity volatility σt (u) is again given by (5.7), and as usual we
separate between a “moderate” and a “spiky” market scenario. The latter has parameters
specified as σ0 = 0, σ1 = 300%, and α = 0.2, whereas the “moderate” market scenario
has σ0 = 20%, σ1 = 50%, and α = 0.05. We know from the previous section that in
the “spiky” scenario the BRS approximation failed to capture the tails of the power
forward distribution. In this example, we investigate the consequences on call option
pricing, and in particular we focus on options which are far-out-of the money. The
initial forward curve is f0 (u) = F0 (T1 , T2 ) = 10, where we have used the weight
function w(u, T1 , T2 ) = 1/(T2 −T1 ). The options are based on a weekly power forward
contract, starting delivery at time T1 = 50. Further, since the deviation of the BRS
approximation from the power forward is greatest at start of delivery, we consider
options with exercise time T = 50. The risk-free interest rate is set equal to r = 5%.
In Table 5.1, we report the results from the numerical example. The power for-
ward option prices are based on 10,000 Monte Carlo simulations, while the BRS
Table 5.1 Call Option Prices on Power Forward and Its BRS Approxi-
mation. The Power Forward Has a Weekly Delivery from Time T1 = 50,
and Exercise Time Is T = 50.
Strike “Moderate” power BRS “Spiky” power BRS
approximation is resulting from the Black-76 formula. We observe that there is hardly
any difference in the “moderate” scenario, in line with the fact that the BRS approxima-
tion is fitting the power forward distribution extremely well in this situation. However,
when we let the mean–reversion increase and the spot volatility σ1 be very high,
there is a significant effect on option prices. We know that the BRS approximation
underestimates the tails of the power forward distribution. This results in an under-
estimation of the option prices when using the BRS approximation and the Black-76
formula. From Table 5.1, we observe that options prices are indeed lower for the BRS
approximation. Already for an in-the-money option with strike K = 9, the price is
about 9% understimated by the BRS approximation. Looking at the extreme K = 13,
we have close to 50% underestimation of the true price. Hence, the BRS approximation
may be seriously flawed in situations with a very steep volatility curve.
We also know from the analysis of the power forward and its BRS approximation
that the covariance structure between different contracts are not matching. In fact,
the BRS approximation is also underestimating this. We consider an example of a
calendar spread option, where we price options based on the difference between two
power forwards with different delivery periods. For the BRS approximation, we have
a Margrabe Formula for the price (see Margrabe [19]):
Proposition 5.5. The price at time t of a call option on the difference FtBRS (T1 , T2 ) −
FtBRS (T3 , T4 ) with exercise time T, where min(T1 , T2 ) ≥ T ≥ t, strike 0, and interest
rate r is given by
where
ln(FtBRS (T1 , T2 )/FtBRS (T3 , T4 )) − 0.52t (T )
d= ,
t (T )
T
BRS
2
t (T ) =
2
s (T1 , T2 ) − BRS s (T3 , T4 ) ds,
t
Proof. We include the proof for convenience. Let for simplicity t = 0, and factorize
out by FTBRS (T3 , T4 ) from the price to get
118 Benth
with
d Q̃ T
1 T
= exp BRS
s (T3 , T4 )dWs − (BRS
s )2
(T3 , T4 )ds .
dQ 0 2 0
In our numerical example, we consider two call option contracts. The first is a call
on the spread of two weekly contracts, one delivering from T1 = 50 to T2 = 55, and
the other over the time interval T1 = 70 to T2 = 75. The second is a call on the spread
of a weekly contract and a monthly, where the weekly is as the first power forward
above, and the monthly start delivering at T3 = 50, and ends at T4 = 70. The two
options are supposed to mimic a spread on power forward delivering far apart, and a
spread of two overlapping power forwards.
We consider an exercise time T = 50, the beginning of delivery of the first weekly
contract. Option prices are calculated in both the “moderate” and “spiky” scenarios
described above. The BRS approximation yields analytical prices by the Margrabe
formula in Proposition 5.5, while the “true” prices based on power forwards must be
simulated by Monte Carlo. We used 100,000 outcomes, a figure that gave sufficient
accuracy for our numerical example. The results are reported in Table 5.2. It is hard to
detect any significant difference between the real prices and the approximative ones
in any of the two markets scenarios. Checking various other exercise times T < 50
did not change the conclusions. Hence, although there is a difference in the covariance
structure between the BRS approximation and the power forward, this does not seem
to have any significant effect on calendar spreads. Due to the analyticity of the calendar
spread option price based on the BRS approximation, this is much more attractive to
use in practice since Monte Carlo is very slow.
There may be other specifications of the volatility structure that indeed gives
significant mispricing of calendar spread options when using the BRS approximation.
Inspecting the numbers in Table 5.2 closer, we observe that in all cases except the
“spiky” market scenario with two weekly contracts, the BRS approximation gives
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch05 FA
5.5. CONCLUSION
In this paper, we have analyzed the distributional properties of power forward contracts.
Power forward contracts deliver electricity over a specified period, and can be viewed as
a portfolio of forwards with maturity at each time instant in the delivery period. We have
investigated the implied power forward dynamics from a geometric Brownian motion
specification of the forward price. The implied dynamics is not in general a geometric
Brownian motion, however, as it turns out, can be reasonably well approximated by
one. We show that there exist upper and lower bounds in terms of geometric Brownian
motions for the power forwards.
We have spent most of our efforts in this paper to consider the BRS approximation,
where the power forward dynamics is approximated by a geometric Brownian motion,
with a volatility specified as the average of the fixed-delivery forward volatility. Both,
analytically and emprically, the approximation works very well. However, if the market
has very strong mean–reversion and excessively high volatility (as is the case for
many electricity markets), the BRS approximation may underestimate significantly
the tails of the power forward distribution. We show that this may have consequences
in risk management and option pricing, where the BRS approximation will seriously
understimate the true risk and price. On the other hand, it works excellently in more
moderate situations.
There are also signs of mis-representation of the true covariance structure between
power forward contracts by the BRS approximation. By looking at various calendar
spread options, it seems that this is not very significant in practice. However, the BRS
approximation allows for an analytical Margrabe formula to price calendar spread
options, which is far superior in efficiency compared to the Monte Carlo method
which must be used to price spread options on power forwards.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch05 FA
120 Benth
For basket option pricing, one is frequently using moment matching techniques
to fit a lognormal variable to a sum of lognormals. This approach can also be used for
power forwards, and provides an explicit representation of the approximating volatility.
However, this moment-matched volatility is in general very complicated and requires
numerical integration to be evaluated, whereas the BRS volatility is analytically avail-
able. Also, in many situations, the BRS volatility is a reasonable approximation of the
moment-matched volatility. Hence, the BRS approximation is to be preferred among
the two.
In this paper, we have only treated the case of one driving Brownian motion for
the dynamics. In practice, the dependence structure among different power forwards
is far more complicated (see Benth et al. [6]), and more factors are called for. Many
of the calculations above can be easily extended to cover such models. However,
another feature that should be investigated is the non-normality of returns of power
forwards. Empirical analysis of power forward prices has shown that the log returns
are significantly heavy tailed (see Benth and Koekebakker [4]). Indeed, as shown in
Frestad et al. [12], the normal inverse Gaussian (NIG) distribution fits the log returns
on power forwards with delivery periods ranging from weekly to yearly extremely
well. The four-parameter NIG distribution was introduced in finance by Barndorff-
Nielsen [2], and has later proved to be a suitable distribution class to model the log
returns of financial data. It has the normal distribution as a limiting case, and Frestad
et al. [12] demonstrate statistically that the estimated NIG distributions of the power
forward contracts are far from normal. To model dynamics, matching these statistical
observations requires the use of Lévy processes.
To indicate how such an extension could be achieved under the HJM paradigm
in our context, one could state an exponential Lévy dynamics for the forward process
ft (u) as follows:
Here, Lt is a Lévy process with NIG marginals, say. The maturity dependence of ft (u) is
modeled using the deterministic scaling function γt (u), which plays much the same role
as the volatility function σt (u). To have a risk–neutral dynamics, we must require that
E[exp(γt (u)Lt )] = 1,
for all t ≥ 0 and u ≥ t. The naive BRS approximation in this case would be to consider
with
T2
BRS
t (T1 , T2 ) = w(u, T1 , T2 )γt (u)du.
T1
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch05 FA
References
122 Benth
[19] Margrabe, W (1978). The value of an option to exchange one asset for another. Journal of
Finance, 33, 177–187.
[20] Protter, Ph (1990). Stochastic Integration and Differential Equations. Springer Verlag.
[21] Zacks, S and CP Tsokos (1978). The distribution of sums of dependent log-normal vari-
ables. Technical Report No. 31, Defense Technical Information Center.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch06 FA
Certificates have become very popular in Germany, Austria, and Switzerland in the
last few years. From a technical and legal point of view they are bonds. Thus, their
value actually also depends on the rating and creditworthiness of the issuing company.
This aspect is in general neglected in the pricing of these products. In the following,
we present a model which overcomes this lack and incorporates the default risk of the
issuing company in the pricing. We derive closed-form expressions for index, basket,
and bonus certificates under issuer risk in a Black–Scholes model framework. The
results are analyzed for different scenarios and compared with valuations in a model
which neglects issuer risk.
123
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch06 FA
6.1. INTRODUCTION
Certificates have become very popular with retail customers in Germany, Austria, and
Switzerland in the last few years. They are referred to as structured products. The
simplest structures are index certificates, which allow a retail customer to invest in
an index like the German stock index Deutscher Aktien Index (DAX) or the Dow
Jones Euro Stoxx 50. From a legal point of view, those products are bonds and the
investors are, thus, creditors of the respective issuer. This rather technical aspect has
been disregarded in the investment decision by many retail investors. Even the value
at risk figures which are sometimes indicated by banks to show the investors the risk
involved in single certificates are based on the assumption of a non-defaultable issuer.
However, in the case of an insolvency of the issuer, the investor may lose his total
investment regardless of the performance of the underlyings of the certificate. The
case of Lehman Brothers shows that this risk can materialize. Certificates differ in this
feature from an investment in funds. The investment in a fund is a so-called special
property and is not affected at all by the rating and creditworthiness of the issuing
company.
Issuer risk is the risk of loss on securities and other tradeable obligations because
the issuer does not fulfil his contractual obligations due to his insolvency. Up to today
this kind of risk has been hardly addressed in the pricing of exotic securities and
especially not from a retailer’s perspective but only in connection with regulatory
capital requirements like in Basel II. More details about modeling and evaluating
counterparty/issuer risk under an economic or regulatory perspective can be found
in [1–3].
Pricing securities under counterparty risk can be traced back to Merton [4]. John-
son and Stulz [5] analyzed the counterparty risk in option pricing. They used a firm
value model and assumed that the vulnerable option presents the single debt of the
company. A huge increase in the derivative’s value, thus, rises the risk of default of
the company. This approach is only appropriate when the derivative is the only or
the predominant source of funding of the counterparty. Hull and White [6] as well as
Jarrow and Turnbull [7] value so-called vulnerable options, options on a bond written
by a defaultable party, in a reduced-form model assuming independence between the
credit risk of the counterparty and the asset underlying the derivative. Cherubini and
Luciano, see [8,9], suggest to use a copula approach to value the counterparty risk in an
investment and to allow for a dependence structure between term-structure movements
and a default of one of the parties.
Klein [10] as well as Klein and Inglis [11] choose a firm-value model to account
for the issuer risk and to model the dependencies between the issuing firm and the
underlying. We follow their approach in that regard, and condition the payoff of the
certificate on the survival of the issuer: the certificate only pays back the total investment
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch06 FA
and gain as long as the issuer has not defaulted, i.e., its asset value has not fallen under
a certain barrier. Like Klein [10] and Klein and Inglis [11], we model the correlation
between the assets of the issuer and the asset underlying the derivative explicitly.
The barrier is exponentially increasing in time, and the issuer can default, anytime
before maturity. In the case of default, the investor recovers a constant fraction of
the market value of his investment. As we deal with retail products, we furthermore
assume that the exotic structures are fully hedged, i.e., all debt owed to the investor has
to be seen alongside assets which the company owns. In the case of default, however,
these assets do not cover the investors but are part of the insolvency estate. This
allows us to assume the boundary as deterministically increasing rather than stochastic,
see [11].
This implies that default occurs whenever the stock price S1,t falls to zero. As soon as
S1,t reaches zero, it remains there. In this framework, an European option is seen as
the corresponding down-and-out barrier option with an absorbing barrier for S1,t set
at 0. Hence, an equivalent time to default to τ is
ς = inf{t ∈ (t0 , T ] : S1,t ≤ 0}. (6.5)
The dynamics for S1,t are, prior to default, found by applying Itô to (6.4) and are
given by
dS1,t = S1,t (rt − d1,t )dt + (D1,t + S1,t )σ1 dW1 . (6.6)
The stock price of the issuer follows — prior to default — a shifted log-normal distribu-
tion. This distribution implies negative stock prices with positive probability, see [13]
for more details. The higher the leverage (debt-to-equity ratio), the higher is the prob-
ability of default and vice versa.
The dynamics of the underlying assets of a certificate are modeled by geometric
Brownian motions. The following system is considered to describe the assets of the
issuer and the underlying assets of the derivatives:
dS1,t = S1,t (rt − d1,t )dt + (D1,t + S1,t )σ1 dW1 (6.7)
dSi,t = Si,t (rt − di,t )dt + Si,t σi dWi for i = 2, . . . , n (6.8)
dWi , dWj = ρij dt, for i, j ∈ 1, . . . , n, i = j, −1 ≤ ρij ≤ 1,
where σi and ρij are constants, see [14].
Proposition 6.1. In the case when the investment in the equity is guaranteed by an
D
issuer who is defaultable with recovery rate zero, the price S2,t of this investment at
time t is given by
D
S2,t = S2D (S1 , S2 , t)
T
= e− t rs ds+ατ+β1 y1 +β2 y2 − 12 χ(δ21 +2ρδ1 δ2 +δ22 )
√
× (e(δ1 ψ1 +δ2 ψ2 ) χ
N2 (ψ1 , z2 , ρ)
√
(δ1 ξ1 +δ2 ξ2 ) χ
−e N2 (ξ1 , z2 , ρ)), (6.9)
where N2 (x, y, ρ) is the standard bivariate normal distribution function with correla-
tion ρ,
S1,t + D1,t T
y1 = ln , y2 = ln(S2,t e t (rs −d2,s )ds ),
D1,t
τ = T − t, χ = σ1 σ2 τ,
σ2 σ1
z1 = y1 , z2 = y2 ,
σ1 σ2
σ1 − ρσ2 σ2 − ρσ1
β1 = , β2 = ,
2σ1 (1 − ρ2 ) 2σ2 (1 − ρ2 )
σ1 σ2
δ1 = − β1 , δ2 = (1 − β2 ),
σ2 σ1
1 σ12 − 2σ1 σ2 ρ + σ22
α=− ,
4 2(1 − ρ2 )
z1 √ z2 √
ψ1 = √ + χ(δ1 + ρδ2 ), ψ2 = √ + χ(δ2 + ρδ1 ),
χ χ
z1 √ 2ρz1 − z2 √
ξ1 = − √ + χ(δ1 + ρδ2 ), ξ2 = − √ + χ(δ2 + ρδ1 ).
χ χ
For a proof see A.1. Products, which ensure a repayment of an investment with
notional 1, are internally stripped into a zero-coupon bond and several derivatives.
T
As known, a non-defaultable zero-coupon bond is priced by Zt = exp(− t rs ds).
Whereas when we assume that the issuer can default and — if he defaults — with no
recovery, the value of the zero-coupon bond can be computed by, see [13]
T
ZtD = e− t rs ds
(N (d3 ) − ex1 N (d4 )), (6.10)
with N (.) being the standard cumulative normal distribution function,
S1,t + D1,t x1 1√ x1 1√
x1 = ln , d3 = √ − χ, d4 = − √ − χ.
D1,t χ 2 χ 2
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch06 FA
When the issuer additionally promises a fixed interest rI on the investment this is val-
T
ued, assuming no default, by ZI,t = (1 + rI ) exp(− t rs ds) and accordingly by ZI,tD
=
T
exp(− t rs ds)(1+rI )(N (d3 )−e N (d4 )) in a world with defaults and zero recovery.
x1
Call options are introduced in the structured products to leverage the investment
and to grant the investor a high participation rate in the value of the underlying. If no
default of the issuer is assumed the call option with strike K is priced by the well-
T
known Black–Scholes formula Ct (S2 , K) = S2,t N (d1 ) − K exp(− t rs ds)N (d2 ),
√ √
where d1 = (ln(S2,t /K) + (r + 1/2σ22 )(T − t))/(σ T − t) and d2 = d1 − σ T − t.
Proposition 6.2. The price CtD at time t of a defaultable call option, when the payment
of the call option is guaranteed by an issuer who is defaultable with recovery rate zero,
is given by
T √
CtD (S1 , S2 , K) = Ke− (e− 2 χ(δ1 +2ρδ1 δ2 +δ2 ) (e(δ1 η1 +δ2 η2 )
1 2 2
t rs ds+ατ+β·X χ
N2 (η1 , η2 , ρ)
√
(δ1 ν1 +δ2 ν2 ) χ
−e N2 (ν1 , ν2 , ρ))
√
− 12 χ(δ− − − −2 − − − −
2
−e 1 +2ρδ1 δ2 +δ2 ) (e(δ1 η1 +δ2 η2 ) χ
N2 (η− −
1 , η2 , ρ)
− − √
(δ− −
−e 1 ν1 +δ2 ν2 ) χ N2 (ν1− , ν2− , ρ))), (6.11)
where
S1,t + D1,t S2,t T (rs −d2,s )ds
x1 = ln , x2 = ln et ,
D1,t K
σ2 σ1
z1 = x1 , z2 = x2 ,
σ1 σ2
− σ1 − σ2
δ1 = − β1 , δ2 = − β2 ,
σ2 σ1
z1 √ z2 √
η1 = √ + χ(δ1 + ρδ2 ), η2 = √ + χ(δ2 + ρδ1 ),
χ χ
z1 √ 2ρz1 − z2 √
ν1 = − √ + χ(δ1 + ρδ2 ), ν2 = − √ + χ(δ2 + ρδ1 ),
χ χ
z1 √ − z2 √ −
η−
1 = √ + χ(δ1 + ρδ−
2 ), η−
2 = √ + χ(δ2 + ρδ−
1 ),
χ χ
z1 √ 2ρz1 − z2 √ −
ν1− = − √ + χ(δ− − −
1 + ρδ2 ), ν2 = − √ + χ(δ2 + ρδ−
1 ),
χ χ
For a proof see A.2. Similarly, the non-defaultable digital option with strike K is priced
T
by Cd,t (S2 , K) = exp(− t rs ds)N (d2 ).
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch06 FA
D
Proposition 6.3. The price Cd,t at time t of a defaultable digital call option where the
payment of one is guaranteed by an issuer who is defaultable with recovery rate zero
is given by
D
Cd,t (S1 , S2 , K)
T √
rs ds+ατ+β1 x1 +β2 x2 − 12 χ(δ− − − − − − − −
2 2
= e− t 1 +2ρδ1 δ2 +δ2 ) (e(δ1 η1 +δ2 η2 ) χ
N2 (η− −
1 , η2 , ρ)
− − − − √
− e(δ1 ν1 +δ2 ν2 ) χ
N2 (ν1− , ν2− , ρ)), (6.12)
with parameters explained in Propositions 6.1 and 6.2.
For a proof see A.3. The bonus certificate includes, beside other derivatives, a knock-
out put option. The price of a non-defaultable knock-out put option with barrier B and
strike K is given by, see also [16],
Pk,t (S1 , S2 , K, B)
2β B S2
−
T B 2 ln − ln
= Pt (S2 , K) − Ke t rs ds
√ − β σ2 (T − t)
N K K
σ2 T − t K
2(β−1) B
T
− t d2 ds B 2 ln K − ln SK2
+ Se N √ + (1 − β )σ2 (T − t) ,
K σ2 T − t
(6.13)
where
T
Pt (S2 , K) = Ke− t rs ds N (−d2 ) − S2,t N (−d1 ),
∗ S2 B
x2 = ln , b = ln ,
K K
1 r − d2 1
β = α = − σ22 β ,
2
− ,
2 σ22 2
χ = σ22 τ
and d1 , d2 as in Proposition 6.2.
D
Proposition 6.4. The price Pk,t at time t of a defaultable knock-out put option, when
the payoff is guaranteed by an issuer who is defaultable with recovery rate zero, is
given by
T ∗ ∗
D
Pk,t (S1 , S2 , K, B) = Ke(− t rs ds+α τ+β1 x1 +β2 x2 )
√ 2 ∗
∗ ∗
× h(τ, z∗1 , z∗2 , y1∗ , y2∗ ) max(e−β1 (σ1 1−ρ y1 +σ1 ρy2 )−β2 (y2 σ2 +b)
√
−β1 (σ1 1−ρ2 y1∗ +σ1 ρy2∗ )+(1−β2 )(y2∗ σ2 +b)
−e , 0)dy2∗ dy1∗ , (6.14)
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch06 FA
where
1 x1 x∗ − b x∗ − b
z∗1 = −ρ 2 , z∗2 = 2 ,
1 − ρ2 σ1 σ2 σ2
1
σ σ − ρ 12 σ22 − (r − d2 )
2 1 2 1 σ2 − ρσ1 r − d2
β1 = , β2 = − ,
(1 − ρ )σ1 σ2
2 2 (1 − ρ )σ2
2 (1 − ρ2 )σ22
1 1 2 1 1
α∗ = − σ12 β1 − σ2 − (r − d2 ) β2 + σ12 β1 + σ22 β2 + ρσ1 σ2 β1 β2 .
2 2
2 2 2 2
h(τ, z∗1 , z∗2 , y1∗ , y2∗ ) is given in Appendix A.4.
where ς is given in (6.5). The index certificate can be valued by using the building
blocks of Proposition 6.1. Thus, the certificate is priced by the following formula
The number of parameters rises considerably if the notion of issuer risk is incorporated
in the pricing of derivatives. In the following, we show the results of the valuation of
the index certificate in different scenarios. The results are compared to prices which
neglect issuer risk. For all the scenario computations in this chapter, we have chosen
the following instrument parameterizations if not stated differently in the respective
examples:
Index certificate
110
Defaultable
Non-defaultable
100
90
80
Value
70
60
50
40
0 200 400 600 800 1000 1200
Debt level
Figure 6.1 shows that for increasing debt of the issuing company the value of the index
certificate declines and finally approaches RS2,t . The price of the non-defaultable cer-
tificate stays at par regardless of the debt level. This shows that for a retail investor a
simple comparison of the prices of index certificates of different issuers is not appropri-
ate in order to find the security with the best price-performance ratio. For his investment
decision, the investor has to take the rating and equity to debt ratio into consideration.
In Fig. 6.2, the impact of the volatility of the issuer’s asset (the volatility of
the underlying is kept to 0.4) and the correlation between the issuer’s asset and the
underlying is analyzed. For that reason, we choose a relatively low debt scenario. The
negative relationship between the volatility of the issuer’s assets and the price is clearly
visible. The probability of a default considerably increases in high volatility scenarios
and thus the price decreases. It strikes that the impact of the correlation on the price
grows with the volatility: for a very high volatility, a correlation near 1 leads to a
considerably higher value than a correlation near −1. In a low volatility and low debt
scenario with D0 = 50, the price is not much affected by the level of the correlation
as the probability of default is relatively low. However, in a high volatility and low
debt scenario with, thus, higher probability of default of the issuer, both assets tend
to move in the same direction if the correlation rises. Thus, in contrast to a negative
correlation, the probability that the issuer survives and the underlying asset features
a positive return goes up. A negative correlation results in a different performance of
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch06 FA
Index certificate
100
90
80
Value
70
60
50
1
40
0.7 0.5
0.6 0
0.5
0.4 −0.5
0.3
0.2 −1
Volatility S1 Correlation
the issuer’s assets and the underlying. Thus, the probability goes up that the certificate
matures at a low value or worthless.
point the investor does not profit from any increase in the stock price. A defaultable
participation guarantee certificate of an issuer with recovery rate R is, thus, priced by
For the participation guarantee certificate, we compute some scenario values. For these
computations, we choose a notional of 88, a start of the participation of K = 100, and
a participation rate of p = 50%.
In Fig. 6.3, we find the impact of the debt level on the participation guarantee
certificate similar to that of the index certificate: as before the price falls especially
steeply for smaller D0 levels.
In Fig. 6.4, the joint impact of the issuer’s volatility and the correlation is shown.
The form of the graphs resembles the graph of Fig. 6.2 and the same explanations
apply to explain the form. The effect of the correlation on the price of the participation
guarantee certificate is, however, clearly less distinct because the total price of the
certificate considerably depends on the value of the zero-coupon bond for which the
correlation is irrelevant.
80
Value
70
60
50
40
0 200 400 600 800 1000 1200
Debt level
95
90
85
80
Value
75
70
65
60 1
0.7 0.5
0.6 0
0.5
0.4 −0.5
0.3
0.2 −1
Volatility S1 Correlation
70
65
60
Value
55
50
45
40
35
30
0 200 400 600 800 1000 1200
Debt level
For the bonus guarantee certificate, we observe the same typical feature of the
graph, which shows prices for different debt levels, see Fig. 6.5, as for the other
examples: the graph decreases sharply for lower debt levels and approaches R times
the price of the analogous certificate of a non-defaultable issuer.
In Fig. 6.6 the relationship between price and volatility is of nearly linear kind:
the higher the volatility, the lower the price the investor has to pay. The impact of the
correlation is less distinct as in the case of index certificates due to the fact that a major
part of the value of the bonus guarantee depends on the zero-coupon bond.
75
70
Value
65
60 1
0.5
55 0
0.7
0.6 −0.5
0.5
0.4
0.3 −1 Correlation
0.2
Volatility S1
where ς is as in (6.5). The discount certificate can be valued by the following formula
Discount certificate
75
Defaultable
70 Non-defaultable
65
60
55
Value
50
45
40
35
30
0 200 400 600 800 1000 1200
Debt level
Discount certificate
75
70
65
60
Value
55
50
45
40
1
35
0.7 0.5
0.6 0
0.5
0.4 −0.5
0.3
0.2 −1
Volatility S1 Correlation
investment in the underlying and by buying a knock-out put option. The respective
barrier level is the protection barrier. When the issuer risk is incorporated in the pricing
model, the price is specified by
BC(t, S1 , S2 )
T
= S2,t + e− t rs ds
EQ [max[K − S2,T , 0]1{τ2 >T } | Ft ]
T
− (1 − R)e− t rs ds
EQ [(S2,T + max[K − S2,T , 0]1{τ2 >T } )1{ς≤T } | Ft ], (6.23)
where
τ2 = inf(t ∈ (t0 , T ] : S2,t ≤ B), (6.24)
where ς is as in (6.5). The following formula evaluates the payoff:
BC(t, S1 , S2 ) = R(S2,t + Pk,t (S2 , K, B))
D
+ (1 − R) S2,t (S1 , S2 ) + Pk,t
D
(S1 , S2 , K, B) . (6.25)
Finally, we show some exemplary computations of the bonus certificate. We assumed
a protection barrier B = 80 and a bonus barrier K = 120 for our computations.
Not surprisingly, the graph, i.e., Fig. 6.9, plotting the debt level of the issuer against
the value of the bonus certificates shows the same features as in the examples above.
Bonus certificate
110
Defaultable
Non-defaultable
100
90
80
Value
70
60
50
40
0 200 400 600 800 1000 1200
Debt level
Bonus certificate
100
90
80
70
Value
60
50
40 1
0.5
30 0
2.5
2 1.5 −0.5
1
0.5
0 −1
Correlation
Volatility S1
In Fig. 6.10, we see a strong impact of correlation in high volatility scenarios and
small influence in lower volatility scenarios. These characteristics can be explained by
the fact that correlation between issuer and underlying has a considerable impact on
the price of the building blocks of the bonus certificate.
6.4. CONCLUSION
D
Proof. S2,t fulfils the following partial differential equation and boundary conditions
1 2 ∂2 S2,t
D
1 ∂2 S2,t
D
∂2 S2,t
D
σ1 (S1 + D1,t )2 + σ22 S22 + ρσ1 σ2 (S1 + D1,t )S2
2 2
2 ∂S1 2 ∂S2 ∂S1 ∂S2
D
∂S2,t D
∂S2,t D
∂S2,t
+ (rt − d1,t )S1 + (rt − d2,t )S2 + − rt S2,t
D
= 0,
∂S 1 ∂S2 ∂t
S2 (t, 0, S2 ) = 0,
D
D
S2 (T, S1 , S2 ) = S2,T .
2 1 2
see [15], page 493. We apply the method of images, see [18], page 476ff, to restrict the
fundamental solution to the area the problem is defined for. This method is appropriate
when the region to which the solution should be bounded is highly symmetric: a solu-
tion for the free space is first derived (pF (χ, z1 , z2 , z1 , z2 )) and then restricted to the
defined region via symmetry, i.e., the principle of reflection. In this case, the image
point (z1 , z2 ) of the source point (z1 , z2 ) is given by
(z1 , z2 ) = (−z1 , −2ρz1 + z2 ).
Thus, the solution is found by
p(χ, z1 , z2 , z1 , z2 ) = pF (χ, z1 − z1 , z2 − z2 )
− pF (χ, z1 + z1 , z2 + 2ρz1 − z2 ).
In this case the solution for the free space is
2
(z +z −2ρz1 z2 ) 2
1 − 1 2
pF (χ, z1 , z2 , z1 , z2 ) = e 2(1−ρ2 )χ .
2π 1 − ρ2 χ
The solution of the original Dirichlet problem can be found by computing, see [17],
page 364ff,
S2D∗ (χ, z1 , z2 )
∞ ∞
σ1
σ2
−β1 z +(1−β2 ) z
= p (χ, z1 −
F
z1 , z2 − z2 )e σ2 1 σ1 2
dz1 dz2
−∞ 0
∞ ∞
σ1
σ2
−β1 z +(1−β2 ) z
− pF (χ, z1 + z1 , z2 + 2ρz1 − z2 )e σ2 1 σ1 2
dz1 dz2 .
−∞ 0
With
∞ ∞
ϕ= pF (χ, z1 , z2 )ek1 z1 +k2 z2 dz1 dz2
−∞ 0
Proof. The proof is similar to the one for the investment in the underlying. By the
means of the Feyman Kac Theorem, see [19], page 143ff, the following PDE can be
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch06 FA
derived:
2 D
1 2 2∂ C 1 2 2 ∂2 CD ∂2 C D
σ (S + D ) + σ S + ρσ σ (S + D )S
2 1 1 1,t
∂S12 2 2 2
∂S22
1 2 1 1,t 2
∂S1 ∂S2
∂CD ∂CD ∂CD
+ (rt − D1,t )S1 + (rt − d2,t )S2 + − rt C D = 0,
∂S1 ∂S 2 ∂t
C (t, 0, S2 ) = 0,
D
D
C (T, S1 , S2 ) = max(S2 − K, 0).
By applying the transformations, which have been introduced before the PDE can be
reduced to
D∗ ∗ ∗ ∗
∂C 1 ∂2 C D 1 ∂2 CD ∂2 C D
− − − ρ = 0,
∂χ 2 ∂z21 2 ∂z22 ∂z1 ∂z2
∗
CD (χ, 0, z2 ) = 0,
D∗ σ σ
−β1 σ1 z1 +(1−β2 ) σ2 z2
σ σ
−β1 σ1 z1 −β2 σ2 z2
C (0, z1 , z2 ) = max(e 2 1 −e 2 1 , 0).
Using the method of images on the probability density function as before the solution
is given by
∗
CD (χ, z1 , z2 )
∞ ∞
σ1
σ2
−β1 z +(1−β2 ) z
= pF (χ, z1 − z1 , z2 − z2 )e σ2 1 σ1 2
dz1 dz2
0 0
∞ ∞
σ1
σ2
−β1 z −β2 z
− p (χ, z1 −
F
z1 , z2 − z2 )e σ2 1 σ1 2
dz1 dz2
0 0
∞ ∞
σ1
σ2
−β1 z +(1−β2 ) z
− p (χ, z1 +
F
z1 , z2 + 2ρz1 − z2 )e σ2 1 σ1 2
dz1 dz2
0 0
∞ ∞
σ1
σ2
−β1 σ2 z1 −β2 σ1 z2
− pF (χ, z1 + z1 , z2 + 2ρz1 − z2 )e dz1 dz2
0 0
with
∞ ∞
ϕ= pF (χ, z1 , z2 )ek1 z1 +k2 z2 dz1 dz2
0 0
∞ ∞
σ1
σ2
−β1 z −β2 z
− p (χ, z1 +
F
z1 , z2 + 2ρz1 − z2 )e σ2 1 σ1 2
dz1 dz2 .
0 0
1 1 1 1
α∗ = − σ12 β1 − σ22 − (r − d2 ) β2 + σ12 β1 + σ22 β2 + ρσ1 σ2 β1 β2 ,
2 2
2 2 2 2
S1 + D1,t S2
x1 = ln x2∗ = ln ,
D1,t K
B
τ = T − t, b = ln ,
K
1 x1 x∗ − b x2∗ − b
z∗1 = −ρ 2 , z∗2 = ,
1 − ρ2 σ1 σ2 σ2
1
σ σ − ρ 12 σ22 − (r − d2 )
2 1 2 1 σ2 − ρσ1 r − d2
β1 = , β2 = − 2 ,
(1 − ρ2 )σ1 σ2 2 (1 − ρ )σ2
2 σ2 (1 − ρ2 )
the PDE can be reduced to
∗ ∗ ∗
1 ∂2 PkD 1 ∂2 PkD ∂PkD
−
2 ∂z2 − + = 0,
2 ∂z22 ∂τ
1
D∗ ∗
Pk (τ, z1 , 0) = 0,
D∗ ∗ 1 − ρ2 ∗
Pk τ, z1 , − z1 = 0,
ρ
√
P D∗
(0, z ∗ ∗
, z ) = max(e −β1 (σ1 1−ρ2 z∗1 +σ1 ρz∗2 )−β2 (z∗2 σ2 +b)
k 1 2
√ 2∗
∗ ∗
− e−β1 (σ1 1−ρ z1 +σ1 ρz2 )+(1−β2 )(z2 σ2 +b) , 0).
This PDE can be solved by integrating the payoff over the probability density function,
see [16]
√
∗ 2 ∗ ∗ ∗
PkD (τ, z∗1 , z∗2 ) = h(τ, z∗1 , z∗2 , y1∗ , y2∗ ) max(e−β1 (σ1 1−ρ y1 +σ1 ρy2 )−β2 (y2 σ2 +b)
√
1−ρ2 y1∗ +σ1 ρy2∗ )+(1−β2 )(y2∗ σ2 +b)
− e−β1 (σ1 , 0)dy2∗ dy1∗ ,
where
∞
2 − (q2 +qt2 ) nπθt nπθ q q
h(τ, z∗1 , z∗2 , y1∗ , y2∗ ) =
t
e 2τ sin sin I nπϕ ,
ϕτ n=1 ϕ ϕ τ
1 − ρ2
tan ϕ = − , ϕ ∈ [0, π],
ρ
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch06 FA
y2∗
tan θ = , θ ∈ [0, ϕ],
y1∗
z∗2
tan θt = , θt ∈ [0, ϕ],
z∗1
q = y∗1 2 + y∗2 2 ,
qt = z∗1 2 + z∗2 2 .
While the region is given as follow:
1 − ρ2 ∗
(−∞, 0) × −∞, − y1 ∪ (0, +∞) if {ρ > 0} ∩ {y1∗ > 0}
ρ
1 − ρ2 ∗
(−∞, 0) × − y1 , +∞ ∪ (−∞, 0) if {ρ > 0} ∩ {y1∗ < 0}
ρ
=
1 − ρ2 ∗
(0, +∞) × 0, − if {ρ < 0} ∩ {y1∗ > 0}
y1
ρ
(−∞, 0) × − 1 − ρ 2
y1∗ , 0 if {ρ < 0} ∩ y1∗ < 0 ,
ρ
References
[1] Canabarro, E and D Duffie (2003). Measuring and marking counterparty risk. In:
Asset/Liability Management for Financial Institutions, L Tilman (ed.). Institutional
Investor Books.
[2] Pykhtin, M and S Zhu (2006). Measuring counterparty credit risk for trading products
under Basel II, Working Paper.
[3] Pykhtin, M and S Zhu (2007). A guide to modeling counterparty credit risk, GARP Risk
Review.
[4] Merton, RC (1974). On the pricing of corporate debt: The risk structure of interest rates.
Journal of Finance, 29, 449–470.
[5] Johnson, H and R Stulz (1987). The pricing of options with default risk. Journal of Finance,
42, 267–280.
[6] Hull, JC and AD White (1995). The impact on default risk on the prices of options and
other derivative securities. Journal of Banking and Finance, 19, 299–322.
[7] Jarrow, R and S Turnbull (1995). Pricing derivatives on financial securities subject to credit
risk. Journal of Finance, 50, 53–85.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch06 FA
[8] Cherubini, U and E Luciano (2002). Copula vulnerability. Risk, 15, 83–86.
[9] Cherubini, U and E Luciano (2004). Pricing swap credit risk with copulas, Working Paper.
[10] Klein, P (1996). Pricing Black–Scholes options with correlated credit risk. Journal of Bank-
ing and Finance, 20, 1211–1229.
[11] Klein, P and M Inglis (2001). Pricing vulnerable European options when the option payoff
can increase the risk of financial distress. Journal of Banking and Finance, 25, 993–1012.
[12] Stamicar, R and C Finger (2005). Incorporating equity derivatives into the CreditGrades
model, RiskMetrics Group.
[13] Sepp,A (2006). Extended CreditGrades model with stochastic volatility and jumps. Wilmott
Magazine, 54, 50–62.
[14] Hull, JC (2002). Options, Futures, and Other Derivatives, 5th edn. Prentice Hall.
[15] Lipton, A (2001). Mathematical Methods for Foreign Exchange: A Financial Engineer’s
Approach. Singapore: World Scientific Publishing Co. Pte. Ltd.
[16] He, H et al. (1998). Double lookbacks. Mathematical Finance, 8, 201–228.
[17] Karatzas, I and SE Shreve (1991). Brownian Motion and Stochastic Calculus, 2nd edn.
Berlin: Springer-Verlag.
[18] Zauderer, E (2006). Partial Differential Equations of Applied Mathematics. John Wiley &
Sons, Inc.
[19] Øksendal, B (2003). Stochastic Differential Equations, 6th edn. Berlin: Springer-Verlag.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch07 FA
147
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch07 FA
7.1. INTRODUCTION
The credit market experienced an enormous growth over the last years. As part of
the credit market, the credit derivatives market was the worldwide fastest-growing
derivatives market (see [1]). The market size expanded from USD 180 billion in
1996 via USD 5,021 billion in 2004 to USD 20,207 billion in 2006 and to esti-
mated USD 33,120 billion in 2008 (see [2]). Due to the financial crisis, the growth
was rapidly stopped in 2008 and turned into a decline of the market size. Still, the
actual notional amount outstanding of the credit default swaps (CDS), which are the
most important instruments in this market, exceeded the estimation of the British
Bankers’ Association by far with a notional amount outstanding of USD 41,868 bil-
lion (see [3]). After a standardization of credit derivative contracts and the introduction
of CDS indices, a revolution in terms of liquidity has taken place, which is only one
reason why credit instruments are very attractive to investors. In addition, credit instru-
ments such as corporate bonds, credit derivatives, and securitizations often have an
appealing risk–return profile allowing to enhance the portfolio return. Furthermore,
due to the correlation structure of their returns to those of traditional asset classes
such as stocks and government bonds, they offer high potential for diversification.
Finally, they allow to manage the credit risk exposure. Even knowing the potential
benefits of different credit instruments, investors still have to know how to com-
bine them optimally with traditional asset classes, i.e., they have to decide on the
optimal proportion of credit-related products, especially for their individual level of
risk–aversion.
A reasonable asset allocation including credit instruments needs to account for
the distinct return characteristics of these instruments. We exemplarily analyze the
return properties of daily log-returns of U.S. Lehman aggregates over the period from
7 November, 2002 to 29 September, 2006. The descriptive statistics of the returns
are summarized in Table 7.1, where it becomes obvious that the risk–return profile of
credit instruments cannot be sufficiently described by mean and variance alone.
The simulation framework consists of a model to simulate risk factors on the one hand
and correlated default times on the other hand.
First, we introduce a model to simulate risk factors. We simulate the short rate,
the growth rate of the Gross Domestic Product (GDP), and the short-rate spread
with the extended model of Schmid and Zagst (see [4]). The short inflation and
the returns of an equity index are simulated with the integrated market model sug-
gested by [5]. The latter can be embedded into the extended model of Schmid and
Zagst.
In the following, we assume that markets are frictionless and perfectly competi-
tive, that trading takes place continuously, that there are neither taxes nor transaction
costs or informational asymmetries, and that investors act as price takers. We fix a
terminal time horizon T ∗ . Uncertainty in the financial market is modeled by a com-
plete probability space (, G, P) and all random variables and stochastic processes
introduced below are defined on this probability space. We assume that (, G, P) is
equipped with three filtrations H, F, and G, i.e., three increasing families of sub-
σ-fields of G. The default time τ of an obligor is an arbitrary non-negative random
variable on (, G, P). For the sake of convenience, we assume that P[τ = 0] = 0
and P[τ > t] > 0 for every t ∈ (0, T ∗ ]. For a given default time τ, we introduce
the associated default indicator or hazard function H(t) = 1{τ≤t} , t ∈ (0, T ∗ ]. Let
H = (Ht )0≤t≤T ∗ be the filtration generated by the process H. In addition, we define
the filtration F = (Ft )0≤t≤T ∗ as the filtration generated by the multi-dimensional
standard Brownian motion Wt = (Wr,t , Wω,t , Wu,t , Ws,t ) and G = (Gt )0≤t≤T ∗ as
the enlarged filtration G = H ∨ F, i.e., for every t we set Gt = Ht ∨Ft . All
filtrations are assumed to satisfy the usual conditions of completeness and right-
continuity. For the sake of simplicity, we furthermore assume that F0 is trivial. It
should be emphasized that τ is not necessarily a stopping time with respect to the
filtration F but of course with respect to the filtration G. If we assumed that τ
was a stopping time with respect to F, then it would be necessarily a predictable
stopping time.
We assume that there exists a measure Q ∼ P such that all discounted price
processes of the financial instruments are martingales
t
relative to (Q, G).As numéraires,
we choose the money-market account Bt = e 0 rl dl with rt denoting the non-defaultable
short rate. We will assume throughout the chapter that for any t ∈ (0, T ∗ ] the σ-fields
FT ∗ and Ht are conditionally independent (under Q) given Ft . Following [8], p. 167
and p. 242, this is equivalent to the assumption that for any t ∈ (0, T ∗ ] and any Q-
integrable FT ∗ -measurable random variable X, we have EQ [X | Gt ] = EQ [X | Ft ].
We start by defining all processes under the real-world measure P and will then work
under the equivalent martingale measure Q.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch07 FA
processes under this measure. The relationship between the parameters is given by
âk = ak + λk σk2 , where k = r, ω, s, u. We obtain λr , λω , λs , λu when changing measure
from P to Q by applying Girsanov’s Theorem (see, for example, [4]).
The processes introduced earlier imply analytical formulas for the zero rates at
time t to maturity T of non-defaultable and defaultable bonds, R(t, T ) and Rd (t, T ),
as well as for the credit spreads, S(t, T ). So, we can determine the complete term
structure. The explicit formulas can be found in the Appendix.
To simulate correlated default times, we use Li’s approach, who presents an effi-
cient algorithm in [16]. We use it with a one-factor copula as applied for example
by [17]. The one-factor copula is a simple, but powerful way to quickly define a corre-
lation structure between several variables. Therefore, it has become very popular and
the standard approach in practice to simulate correlated default times. The underlying
idea is the following: In reality, more firms default during a recession than during a
booming period. This implies that each firm is subject to the same set of macroeco-
nomic environment and that there exists a dependence among the firms. The Gaussian
one-factor copula is often applied due to an easy implementation and the appealing
properties of a standard normal distribution, such as the stability under convolution.
We rather use the NIG copula which is able to overcome some modeling deficiencies
of the Gaussian copula, e.g., the lack of tail dependence. The NIG distribution is a
mixture of normal and inverse Gaussian distributions. It is a four parameter distribu-
tion with very interesting properties. It can produce fat tails and skewness, it is stable
under convolution (under certain conditions), and the density function, the distribution
function and the inverse distribution function can be computed sufficiently fast (see [6]
and [18]). A definition of the NIG distribution and its most important properties can
be found in [6].
Before presenting the copula model, we explain the idea of the Large Homoge-
neous Portfolio (LHP) approach, introduced by [19]. It assumes a constant default
correlation structure over the reference credit portfolio, with the same default prob-
abilities and the same recovery rate in case of default, and it models default using
a one-factor Gaussian copula. Reference [6] modified the LHP model by replacing
the Gaussian distribution with the NIG distribution. Their one-factor copula model is
briefly introduced in the following.
Consider a homogeneous portfolio of m credit instruments. The standardized asset
return of the i-th issuer in the portfolio, Ai , is assumed to be of the form
Ai = aM + 1 − a2 Xi (7.7)
and
√ √ √ √
1 − a2 1 − a2 1 − a2 βγ 2 1 − a2 γ 3
Xi ∼ N IG α, β, − , , (7.9)
a a a α2 a α2
where γ = α2 − β2 .
Then, the asset returns Ai also follow an NIG distribution with the parameters
α β βγ 2 γ 3
Ai ∼ N IG , , − 2 , 2 =: N IG A (7.10)
a a aα aα
and Ai has zero mean and unit variance. The factor M represents the systematic com-
mon market factor and Xi represents firm-specific factors. Equation (7.7) defines a
correlation structure between the random variables Ai . Then, the correlation between
the asset returns of two issuers is given by a2 , in the case of a homogeneous portfolio.
Conditional on M the asset returns of different issuers are independent.
Let us assume that default at time t occurs when the asset return of obligor i
falls below the threshold C(t), i.e., Ai ≤ C(t). Using this copula model, the variable
Ai is then mapped to default time τi of the i-th issuer with a percentile-to-percentile
transformation as described for example by [17] or [20]:
p(t) = P[τi ≤ t] = P[Ai ≤ C(t)] = N IG A (C(t)). (7.11)
p(t) = P[τi ≤ t] is the real-world distribution of default times estimated with a
migration matrix following [20]. According to Eq. (7.11), we conclude that C(t) =
N IG −1
A (p(t)) and thus
P[τi ≤ t] = P[Ai ≤ N IG −1 −1
A (p(t))] = P[p (N IG A (Ai )) ≤ t].
Now, we can simulate Ai according to Eq. (7.7) and determine the default times via
τi = p−1 (N IG(Ai )).
Based on the simulations according to the previous section, we price the following
financial instruments: Government and corporate coupon bonds, funded CDS and
funded CDS indices. The pricing for all instruments is done under the following con-
ditions. We price the instruments at every simulated time step tk , k ∈ {0, . . . , p}, with
t0 < t1 < · · · < tp = T sim , where T sim denotes the end date of the simulation.
We assume that the cash flows occurring during the lifetime of the instrument are
reinvested in the respective instrument. For simplicity reasons, we assume a constant
recovery rate REC of 40% for all credit instruments in the case of a default rather than
using a stochastic recovery model. The constant recovery rate of 40% is in accordance
with the recovery rates for the iTraxx Europe (see www.indexco.com). The cash flows
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch07 FA
at default time τ are invested in a risk-free cash account and they are compounded
at every simulated time step with the default risk-free government rate. We always
assume an initial investment of one unit.
The pricing and total return calculation of the bonds and the equity index is
straightforward. The pricing of the funded CDS and the funded CDS index, however,
is not trivial and is therefore derived in the following.
A CDS can be issued in a funded version. Then, the investor (protection seller)
buys a floating rate note (FRN), which pays a coupon of the three-month LIBOR plus
the fixed CDS spread on a quarterly basis. If no credit event occurs, the coupon is paid
until maturity. In this case, the investor receives the notional amount of the FRN. If,
however, a credit event of the reference entity takes place, the investor receives the
recovery value and the contract terminates. The functionality of a funded CDS index
is similar to funded CDS. It can be considered as a portfolio of funded CDS. If a credit
event of a reference entity takes place, the investor receives the recovery value, but
the notional amount of the FRN is reduced by the weight of the reference entity. If we
assume, for example, the first default in the reference portfolio containing 125 equally
weighted names, like the broadest and most actively traded investment-grade indices
CDX.NA.IG and the iTraxx Europe for North America and Europe, respectively, the
notional amout is reduced by 1/125 from 100% to 99.2%. The future coupon payments
are based on the new notional amount while the coupon rate remains unchanged.
We consider CDS and the CDS index from an investor’s perspective. Hence, we
view these instruments as an investment rather than as a means of hedging. Further-
more, we use the funded version of CDS and the CDS index for two reasons. First,
we can exclude counterparty risk from our considerations. Second, there are investors
who are for regulatory restrictions or due to internal investment policies not allowed
to enter into unfunded credit-derivative contracts.
Before explaining the total return components of these instruments, we briefly
describe the relationship between the relevant time steps. At every simulated time step,
tk ∈ {0, . . . , T sim }, we price the funded CDS/CDS index. Furthermore, we assume that
t1c < · · · < tnc = T denote the spread payment dates, where t1c denotes the next spread
payment date following the current pricing day, tk , and T denotes the maturity of the
CDS/CDS index. We denote the previous spread payment date or the settlement date,
if the first spread payment has not yet been made, by t0c . Spread payments are made in
arrear — at time tic for the payment period from ti−1c
to tic .
A funded CDS/CDS index has the following total return components: Present value
of the pure, default risk-free FRN at tk paying LIBOR (PVFRN ), coupon payments on the
notional amount at t0c , comprising LIBOR and the fixed spread S CDS , compounded to the
following simulated date tk (PVS ), recovery payments at default time τ, compounded to
the following simulated date tk (PVREC ), and the present value of the CDS/CDS index
(PVCDS ). PVCDS can be calculated for every pricing day tk . Its value can be either
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch07 FA
where
tic = tic −ti−1
c
, S CDS is the fixed annual spread of the CDS/CDS index, P(tk , tic )
is the discount factor and 1 − EL(tk , tic , λk ) denotes the probability of no default up
to time tic .
The protection leg is the present value of all expected protection payments made
by the protection seller. As we assume a constant recovery rate REC, we can calculate
the protection leg by
T
Prot Leg(tk , T, λk ) = (1 − REC) P(tk , l) dEL(tk , l, λk )
tk
n
≈ (1 − REC) (EL(tk , tic , λk ) − EL(tk , ti−1
c
, λk ))P(tk , tic ).
i=1
At issuance of the funded CDS, the fixed annual spread S CDS is determined so that the
value of the premium leg equals the value of the protection leg:
(1 − REC) ni=1 (EL(tk , tic , λk ) − EL(tk , ti−1
c
, λk ))P(tk , tic )
S CDS = n . (7.12)
i=1
ti (1 − EL(tk , ti , λk ))P(tk , ti )
c c c
The equality of premium leg and protection leg as shown in Eq. (7.12) must also hold
for every day, when S CDS is substituted by the quoted par yield spread. For the sake
of simplicity, we use the simulated zero spread as the quoted par yield spread. The
absolute difference between the spreads, however, is very low. We analyzed historical
credit spreads from 31 December, 1991 to 29 September, 2006. The average difference
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch07 FA
in this period was at most 1.5 bps for rating classes AA, A2 and BBB and maturities of
up to five years. The worst case on a single day was 4.9 bps. After this substitution, we
can solve the equation for λk and determine the constant (implied) default intensity at tk .
Let N(tk , τ) denote the notional amount of a funded CDS/CDS index at time step
tk . Then, the present value of the CDS/CDS index is given by
PVCDS (tk , T, S CDS , λk )
= (Prem Leg(tk , T, S CDS , λk ) − Prot Leg(tk , T, λk ))N(tk , τ). (7.13)
The calculation of the present values of the other total return components is straight-
forward. Then, we determine the total return of the investment, RCDS , for every tk > 0
by
PVCDS (tk , T, S CDS , λk ) + PVFRN (tk ) + PVREC (tk ) + PVS (tk )
RCDS (tk−1 , tk ) = − 1.
PVCDS (tk−1 , T, S CDS , λk−1 ) + PVFRN (tk−1 ) + PVREC (tk−1 )
For all optimization criteria, we make the following assumptions. There are neither
transaction costs nor taxes; all securities can be divided arbitrarily; the portfolios
remain unchanged over time; there are n given assets to invest in with returns Ri ,
i = 1, . . . , n; the expected return of asset i is given by µi := EP [Ri ] and µ :=
(µ1 , . . . , µn )T ; xi is the portfolio weight of asset i with ni=1 xi = 1, and the portfolio
is denoted by x := (x1 , . . . , xn )T .
as seen earlier. Therefore, this criterion is only of limited use for the optimization of
portfolios including credit instruments.
where CR (α) denotes the α-quantile of the portfolio’s return distribution R(x). From
the formulas above, it becomes evident that the CVaR provides information on the
negative tail of a return distribution since it is not only focussed on the α-quantile but
also takes into account the shape of its tail. This is of great importance if instruments
may default and so produce fat tails. The corresponding CVaR optimization is given by
We fit our model to market data as of 30 September, 2006 (simulation start date). To
calibrate the simulation model, we use parameters estimated by [21] and [22]. The
parameters for the short rate and the GDP growth rates are given by ar = 0.37867,
âr = 0.24782, br = 0.13315, σr = 0.01496, aω = 1.18532, âω = 0.26847, θω =
0.01583, σω = 0.00601. The credit spread parameters are displayed in Table 7.2.
We adjust the estimated parameters âs , âu , and bsu to better meet historical data in
terms of average spreads and spread ranges. For the inflation process and the process
for the equity-index returns, we use the following parameters (see [5] and [22]): ai =
0.64073, âi = 0.50319, θi = 1.04790, σi = 0.01447. αE = −3.07791, bER =
3.94000, bEi = 9.32436, bEω = 5.10643, σE = 0.16000. For the equity index, we
adjust the level of the returns and the standard deviation to better reflect the updated
historical data.
We want the U.S. CDS index to represent the current composition of the
CDX.NA.IG in terms of proportion of different rating classes. However, we assume the
U.S. CDS index (short CDX) only to be composed of the rating classes AA, A, and
BBB with the weights 12.28%, 39.47%, and 48.25%. We added the AAA rating class
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch07 FA
AA A2 BBB1
to rating class AA and BB to BBB, as there are only parameters for the rating classes
AA, A, and BBB available. Furthermore, we normalized the weights to sum up to one.
For the simulation of correlated default times, we use the average one-year migra-
tion matrix for the United States, from 1981 to 2005, provided by [23]. The parameters
for the one-factor copula as of simulation start date are estimated in accordance with the
model introduced by [6]. We use the following parameters a = 0.37029, α = 0.70138,
β = 0. So, the correlation a of a single reference entity to the common market factor
is equivalent to a correlation between two reference entities of a2 = 0.13711.
Having simulated 5,000 scenarios on a quarterly basis with government interest
rates, credit spreads and correlated default times, we can now price various financial
instruments. All bonds, CDS and the CDX, are assumed to have an initial term of
approximately five years.
Table 7.3 reports the return characteristics for investment horizons one year and
three years. Here, we make some general observations. We see higher expected returns
and higher standard deviations or levels of CVaR with decreasing credit quality of
the bonds. Besides, we observe that — as soon as defaults have occurred — the
distributions of credit instruments become non-normal. This can be shown by some
statistics: The skewness of the credit instruments is negative, the excess kurtosis is
significantly higher than 0 and minimum values always indicate defaults. The non-
normality can also be verified easily by Jarque–Bera tests. For the one year investment
horizon, the AA and A-rated CDS and the CDX have considerably less volatility and
a lower CVaR than government and corporate bonds. At the same time, CDS and the
CDX have a similar expected return to government bonds. Though, comparing the one-
year and the three-year investment horizon, we make an interesting observation. For the
three-year horizon, the CDS have a similar or higher risk than the corporate bonds and a
May 12, 2010
Asset Allocation with Credit Instruments
17:47
Table 7.3 Key Statistics of Total Returns of Financial Instruments.
WSPC/SPI-B913
1 Year
Mean 5.55% 10.43% 6.03% 6.19% 6.37% 5.49% 5.75% 5.92% 6.10%
Median 5.50% 8.25% 6.01% 6.18% 6.51% 5.52% 5.77% 5.95% 6.27%
σ 3.16% 19.30% 3.35% 3.57% 4.73% 0.74% 1.13% 1.70% 3.46%
CVaR(1%) 2.55% 32.14% 3.24% 5.34% 17.71% −2.60% −2.64% −0.39% 13.24%
Min −4.35% −43.08% −59.17% −58.22% −58.17% −13.27% −57.44% −57.72% −57.28%
Max 17.01% 88.10% 17.18% 18.12% 19.63% 7.59% 8.46% 8.54% 10.00%
Skewness 0.0900 0.5568 −1.3820 −3.3200 −6.7629 −5.1385 −34.6157 −30.6476 −16.6226
Excess kurtosis 0.0276 0.4571 28.2614 59.8199 90.1579 109.6775 1927.6047 1135.5933 298.5225
3 Years
Mean 16.57% 33.44% 18.04% 18.55% 19.08% 16.43% 17.20% 17.73% 18.41%
Median 16.58% 24.29% 18.16% 18.86% 19.99% 16.46% 17.25% 17.93% 19.10%
b913-ch07
σ 2.67% 52.74% 3.86% 5.34% 8.77% 3.33% 4.18% 5.43% 8.06%
CVaR(1%) −9.41% 54.45% −1.53% 14.83% 53.12% −5.77% 0.65% 15.09% 51.90%
Min 7.32% −67.75% −55.18% −55.41% −55.20% −30.08% −54.21% −54.58% −54.77%
Max 25.13% 439.46% 27.26% 28.12% 30.60% 27.25% 29.06% 30.19% 32.05%
FA
Skewness −0.0188 1.3790 −9.2575 −9.8637 −7.1960 −0.9070 −7.3939 −8.4513 −7.1697
Excess kurtosis −0.0459 3.8584 172.4228 131.3360 56.2670 10.8976 123.0516 107.1262 59.9592
159
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch07 FA
lower return. This phenomenon can be easily explained by the nature of the instruments.
Fixed-income instruments are exposed to market risk which is the higher, the longer
the time to maturity. CDS and the CDX are very similar to an FRN and only have a
very small market risk. They pay the short-term interest rate, which is lower than the
medium-term interest rates if the curve is normal. In addition, the outcome of interest
rates is more volatile over a longer investment horizon. In Table 7.4, we show the linear
correlations between the returns of the financial instruments for the investment horizons
one year and three years. We make the following main observations: Government-bond
and corporate-bond returns have a high positive correlation. This is reasonable as the
main driver of the bond return is the government rate, which is the same for all bonds.
Differences in bond returns come from different credit spreads and different default
times. Bond and equity-index returns are negatively correlated, which is also known
from historical time series. There is a small positive or negative correlation between
returns of corporate bonds and funded CDS/CDX. This may seem counterintuitive at
first sight. A closer look at the nature of these instruments proves otherwise. We need
to differentiate between two opposing effects, which can be attributed to single return
components of the funded CDS/CDX. On the one hand, the present value of the CDS
PVCDS (see Eq. (7.13)) strongly behaves like the price of the corporate bond. Zero rates
being equal, PVCDS and the price of the corporate bond decrease if spreads increase
resulting in a high positive correlation. On the other hand, bonds are fixed income
instruments with prices that strongly depend on the level and the structure of zero
rates. If zero rates are falling, the bond returns increase as bond prices increase. For
the FRN linked to the CDS/CDX, it is the other way round. Since its return strongly
depends on the floating LIBOR, the return of the FRN tends to decrease with falling
interest rates resulting in a high negative correlation between bond returns and returns
of funded CDS/CDX. All in all, the two opposing effects cancel each other out to a
certain degree resulting in a small positive or negative correlation between corporate
bonds and funded CDS/CDX. The returns of the equity index and CDS or the CDX
are positively correlated. This is reasonable since both equity-index returns and the
three-month government rate are to a large proportion driven by the short rate, which
was empirically confirmed by [5].
To conclude, we identify appealing risk–return profiles and low correlations
between bonds, the equity index and CDS/CDX. Therefore, investors should bene-
fit from holding a portfolio consisting of traditional financial instruments and credit
instruments.
After having analyzed the return characteristics, we can turn to portfolio optimiza-
tion. For this purpose, we consider a one-year and a three-year investment horizon and
the following three investment universes: Initial investment universe consisting of
government bonds and an equity index; extension by corporate bonds; extension by
CDS and the CDX.
May 12, 2010
Asset Allocation with Credit Instruments
17:47
Table 7.4 Correlation of Total Returns of Financial Instruments.
Govt. Equity Bond Bond Bond CDS CDS CDS CDS
bond index AA A BBB index AA A BBB
WSPC/SPI-B913
1 Year
Govt. bond 1.0000 −0.1184 0.9508 0.8824 0.6828 −0.2338 −0.1569 −0.1013 −0.0510
Equity index −0.1184 1.0000 −0.1182 −0.1110 −0.0835 0.1124 0.0751 0.0639 0.0104
Bond AA 0.9508 −0.1182 1.0000 0.8391 0.6511 −0.2227 −0.0781 −0.0986 −0.0465
Bond A 0.8824 −0.1110 0.8391 1.0000 0.6002 −0.1934 −0.1385 −0.0417 −0.0482
Bond BBB 0.6828 −0.0835 0.6511 0.6002 1.0000 −0.0104 −0.1061 0.0265 0.0560
CDS index −0.2338 0.1124 −0.2227 −0.1934 −0.0104 1.0000 0.2652 0.2802 0.1830
CDS AA −0.1569 0.0751 −0.0781 −0.1385 −0.1061 0.2652 1.0000 0.1107 0.0632
CDS A −0.1013 0.0639 −0.0986 −0.0417 0.0265 0.2802 0.1107 1.0000 0.0395
CDS BBB −0.0510 0.0104 −0.0465 −0.0482 0.0560 0.1830 0.0632 0.0395 1.0000
3 Years
Govt. bond 1.0000 −0.1629 0.7001 0.5086 0.3271 −0.4689 −0.3812 −0.2919 −0.1814
Equity index −0.1629 1.0000 −0.1068 −0.1014 −0.0593 0.2651 0.2201 0.1764 0.1045
b913-ch07
Bond AA 0.7001 −0.1068 1.0000 0.3543 0.2511 −0.2249 −0.2387 −0.1934 −0.0579
Bond A 0.5086 −0.1014 0.3543 1.0000 0.2027 −0.1772 −0.1930 −0.0712 −0.0726
Bond BBB 0.3271 −0.0593 0.2511 0.2027 1.0000 −0.0816 −0.1103 −0.0901 −0.0209
CDS index −0.4689 0.2651 −0.2249 −0.1772 −0.0816 1.0000 0.6809 0.5481 0.4063
CDS AA −0.3812 0.2201 −0.2387 −0.1930 −0.1103 0.6809 1.0000 0.4113 0.2732
FA
CDS A −0.2919 0.1764 −0.1934 −0.0712 −0.0901 0.5481 0.4113 1.0000 0.2070
CDS BBB −0.1814 0.1045 −0.0579 −0.0726 −0.0209 0.4063 0.2732 0.2070 1.0000
161
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch07 FA
11
10
Expected return (%)
4
0 5 10 15 20 25
Standard deviation (%)
100
90
Portfolio weights (%)
80
70
60
50
40
30
20
10
0
3.30
4.17
0.61
0.76
1.14
1.75
2.48
5.08
6.04
7.03
8.05
9.09
10.16
11.26
12.38
13.52
14.67
15.83
16.99
18.17
government bonds are substituted to a large extent with corporate bonds, CDS, and
the CDX. As already seen from Table 7.3, AA and A-rated CDS and the CDX have
a similar expected return compared to government bonds but a considerably lower
standard deviation for a one-year horizon. Therefore, particularly for lower levels of
standard deviation, optimal portfolios contain a considerable proportion of CDS and
the CDX. For levels of standard deviation higher than 7.64% (one-year horizon) and
10.38% (three-year horizon), an optimal portfolio only consists of corporate bonds and
an equity index. For lower levels of standard deviation, corporate bonds are partially
replaced by CDS and the CDX. There is only a rather small proportion of government
bonds for a low level of risk, due to the risk–return profile of government bonds
and CDS.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch07 FA
35
33
31
Expected return (%)
29
27
25
23
21
19
17
15
0 10 20 30 40 50 60
Standard deviation (%)
100
90
Portfolio weights (%)
80
70
60
50
40
30
20
10
0
8.21
10.81
1.50
1.84
2.49
3.86
5.86
13.55
16.38
19.26
22.17
25.13
28.13
31.17
34.23
37.31
40.40
43.50
46.62
49.74
the CVaR optimization at first. Then, we describe and explain some differences of
mean-variance and CVaR optimization. We begin with the efficient frontiers. Adding
credit instruments to a portfolio only consisting of the initial investment universe leads
to an upward shift and a shift to the left of the efficient frontier. This indicates that
there is an enormous potential for return enhancement or risk reduction. The potential
depends on the investment horizon. It is exemplarily illustrated for the minimum-CVaR
portfolio as a reference point, in Table 7.6.
Table 7.6 reveals that for a one-year investment horizon there is a high potential to
either reduce the risk for a given level of return (from a CVaR of 2.10% to a CVaR of
only −3.12%), or to enhance the portfolio return for a given level of risk (from 5.79%
to 6.78%). Also for a three-year investment horizon there is a high potential to reduce
risk and/or enhance return.
Next, we analyze the optimal asset allocations for an investor using the CVaR-
criterion. In an optimal portfolio, government bonds are partially substituted with
corporate bonds, CDS and the CDX compared to the initial investment universe. As
explained earlier, AA- and A-rated CDS and the CDX have a similar expected return
to government bonds, but a lower CVaR for a one-year horizon. Particularly for low
levels of CVaR, corporate bonds are partially substituted with CDS and the CDX.
For levels of CVaR higher than 18.29% (one-year horizon) and 14.47% (three-year
horizon), an optimal portfolio only consists of corporate bonds and an equity index.
If, however, lower levels of CVaR are of interest, optimal allocations are composed of
a considerable proportion of credit derivatives.
Having a closer look at the resulting optimal allocations of mean-variance and
CVaR optimizations, we can identify some differences. For a one-year investment
horizon, optimal allocations contain CDS even for high levels of expected returns.
While mean-variance optimal portfolios are composed of CDS up to an expected
portfolio return of 7.91%, CVaR optimal portfolios hold CDS up to a return level of
8.94%. In contrast, CVaR optimal portfolios contain significantly less BBB-rated, but
significantly more AA-rated corporate bonds. For a three-year investment horizon,
CVaR optimal portfolios hold CDS up to an expected portfolio return of 25.30%,
while mean-variance optimal portfolios only consist of CDS up to returns of 21.53%.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch07 FA
11
10
Expected return (%)
4
−10 −5 0 5 10 15 20 25 30 35
CVaR (%)
100
90
Portfolio weights (%)
80
70
60
50
40
30
20
10
0
9.62
11.57
13.55
15.58
17.62
19.77
22.03
24.32
26.64
28.97
31.36
−3.30
−2.78
−1.60
−0.20
1.27
2.77
4.27
5.92
7.74
CVaR (%)
35
33
31
29
Expected return (%)
27
25
23
21
19
17
15
−20 −10 0 10 20 30 40 50 60
CVaR (%)
100
90
Portfolio weights (%)
80
70
60
50
40
30
20
10
0
−12.00
−11.02
−9.34
−7.18
−4.67
−1.94
0.84
3.78
6.98
10.51
14.17
18.02
21.92
25.93
30.07
34.31
47.12
38.55
42.79
51.57
CVaR (%)
Equity Index Government Bonds CDX
with either the mean-variance or the CVaR criterion, can add performance to his
portfolio for a given level of risk, or he can reduce risk for a given target return level.
This can be realized due to a low correlation between the different instruments and
their attractive risk–return profile.
not mean that this investor is seeking risk. He is rather willing to bear more risk than
the risk–averse investor in compensation for a higher risk premium. Their respective
benchmark portfolio is defined on the initial investment universe and is exhibited in
Table 7.7.
The procedure for selecting the optimal portfolios to compare with the benchmark
portfolios is the following. At first, we determine the risk and return characteristics
of the benchmark portfolios, denoted as “riskB ” and “returnB ”, where “riskB ” refers
to the relevant risk measure: If optimal portfolios are selected using mean-variance
approach (MV), the relevant risk measure is standard deviation, in the case of CVaR
optimization it is the CVaR. We use these risk values as reference points to find the
optimal portfolios with identical risk values on the efficient frontiers of the most
extended investment universe based on the results of mean-variance and the CVaR-
optimization presented in the previous sections (see Figs. 7.1–7.4). The results are
shown in Tables 7.8 and 7.9 for the one-year and the three-year investment horizon,
respectively.
In the upper parts of the Tables 7.8 and 7.9 the risk–return profiles of the benchmark
portfolios and the optimal portfolios (denoted as “risk∗ ”, “return∗ ”) can be found. In
addition, the corresponding optimal asset allocations are displayed in the lower part
of these tables.
Tables 7.8 and 7.9 reveal some interesting characteristics of mean-variance opti-
mal and CVaR optimal portfolios for the risk–averse and risk–affine investor defined
in Table 7.7. At first, we focus on some similarities. Mean-variance and CVaR opti-
mization lead to similar expected portfolio returns. Extending the investment universe
by credit instruments an investor always can enhance the portfolio return. A risk–
averse investor with a one-year investment horizon can increase his portfolio return
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch07 FA
from 7.01% to 7.50% applying the mean-variance optimization and to 7.47% applying
the CVaR optimization. A risk–affine investor with a three-year investment horizon
increases his portfolio return from 28.38% to 29.12% (MV) or to 29.18% (CVaR).
The proportion of the equity index in an optimal portfolio is always rather close to the
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch07 FA
proportion of the benchmark portfolio, i.e., it is approximately 30% for the risk–averse
investor and 70% for the risk–affine investor. However, the equity proportion is always
slightly higher for investors using the CVaR criterion. The two investors should never
allocate funds in government bonds to build an optimal portfolio. The proportion of the
government bonds in the benchmark portfolios is rather reallocated to the credit instru-
ments in the optimized portfolios, particularly to corporate bonds and CDS. For the
considered levels of risk, the CDS Index does not play a role. For the one-year hori-
zon both optimization approaches lead to a proportion of approximately 11–12% of
CDS for the risk–averse investor. Independent of optimization criterion or investment
horizon, the risk–affine investor should not invest in CDS. Having a closer look at the
proportions of the credit instruments, we can figure out some interesting differences.
A risk–averse investor applying the CVaR criterion and having three-year investment
horizon should still allocate approximately 12% to CDS, while this investor using
the mean-variance criterion should not hold CDS in his portfolio. This indicates that
mean-variance optimization underestimates the benefit of credit derivatives for longer
investment horizons. For both investment horizons, mean-variance optimal portfo-
lios contain a considerably higher proportion of BBB-rated corporate bonds than the
CVaR-optimal portfolios. This means that mean-variance optimization overestimates
the benefits of BBB-rated corporate bonds by only taking the expected return and
volatility into account but mainly ignoring the tail events. These effects were already
explained in the previous section.
To sum up, comparing the resulting optimal asset allocations for a representative
risk–averse and risk–affine investor, we see that, independent of investment horizon and
optimization criterion, an investor always benefits from substituting government bonds
by corporate bonds, CDS, and the CDX. The resulting optimal allocations, however,
strongly depend on the investor type, the optimization criterion and the investment
horizon.
model to simulate correlated default times with an NIG one-factor copula, which
is an extension of the Gaussian one-factor copula. The NIG version is appealing since
it allows to generate more realistic properties of default times such as a higher prob-
ability of joint defaults of different issuers. After the introduction of general con-
ditions for the total return calculations, we have explained how to price CDS. For
portfolio optimization we have applied two criteria — the mean-variance and the
CVaR optimization. The former can be viewed as traditional portfolio optimization.
Its main disadvantage is that it only takes the first two moments of a return dis-
tribution into account, which is obviously not appropriate when also allowing for
credit instruments. The latter is able to overcome this drawback since it considers
the left tail of a distribution. So, it particularly takes defaults into account. Finally,
we have presented the model parameters and we have applied our asset allocation
framework to the U.S. market for a one-year and a three-year investment horizon.
We have found that credit instruments have an appealing risk–return profile and a
correlation structure providing a considerable potential for diversification. Moreover,
we have found that mean-variance optimization overestimates the benefits of low-
rated bonds by only focussing on mean and variance, and mainly ignoring the tail
events.
Realistic modeling of return characteristics is very important when instruments
with specific return characteristics are included in an asset allocation. Otherwise, opti-
mization results are not more than an approximation. Our model provides realistic
return distributions. Furthermore, it can be fitted to market data and it can be easily
extended by other credit instruments. Therefore, it is appropriate for an asset allocation
with credit instruments.
APPENDIX
The zero rates R(t, T ) and Rd (t, T ) at time t to maturity T of non-defaultable and
defaultable bonds as well as the credit spreads S(t, T ) are given by
1
R(t, T ) = − [A(t, T ) − B(t, T )rt − D(t, T )ωt ],
T −t
1
Rd (t, T ) = − [Ad (t, T ) − B(t, T )rt − Dd (t, T )ωt − Ed (t, T )ut − F d (t, T )st ],
T −t
1
S(t, T ) = − [Ad (t, T ) − A(t, T ) − (Dd (t, T ) − D(t, T ))ωt
T −t
− Ed (t, T )ut − F d (t, T )st ],
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch07 FA
with
T
1 2
A(t, T ) = (σ B(l, T ) + σω D(l, T ) ) − θr (l)B(l, T ) − θω D(l, T ) dl,
2 2 2
t 2 r
T
1
Ad (t, T ) = [σr2 B(l, T )2 + σω2 Dd (l, T )2 + σu2 Ed (l, T )2 + σs2 F d (l, T )2 ]dl
2 t
T
− [θr (l)B(l, T ) + θω Dd (l, T ) + θu Ed (l, T ) + θs F d (l, T )]dl,
t
1
B(t, T ) = (1 − e−âr (T −t) ),
âr
br 1 − e−âω (T −t) e−âω (T −t) − e−âr (T −t)
D(t, T ) = + ,
âr âω âω − âr
bsω 1 − e−âω (T −t) e−âω (T −t) − e−âs (T −t)
D (t, T ) = D(t, T ) −
d
+ ,
âs âω âω − âs
bsu 1 − e−âu (T −t) e−âu (T −t) − e−âs (T −t)
E (t, T ) =
d
+ ,
âs âu âu − âs
1
F d (t, T ) = (1 − e−âs (T −t) ).
âs
For an explicit derivation of the zero rates, we refer to [4].
References
[1] FitchRatings (2005). Global credit derivatives survey: Risk dispersion accelerates.
[2] British Bankers’ Association (2006). BBA credit derivatives report 2006 — executive
summary.
[3] Bank for International Settlements (2009). BIS quarterly review June 2009: International
banking and financial market developments.
[4] Antes, S., M Ilg, B Schmid and R Zagst (2008). Empirical evaluation of hybrid defaultable
bond pricing models. Journal Applied Mathematical Finance, 15(3), 219–249.
[5] Zagst, R., T Meyer and H Hagedorn (2007). Integrated modelling of stock and bond
markets. International Journal of Finance, 19(1), 4252–4277.
[6] Kalemanova,A., B Schmid and R Werner (2007). The Normal Inverse Gaussian distribution
for sythetic CDO pricing. Journal of Derivatives, Spring.
[7] Markowitz, H. (1952). Portfolio selection. Journal of Finance, 7, 77–91.
[8] Bielecki, T and M Rutkowski (2004). Credit Risk: Modeling, Valuation and Hedging.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch07 FA
[9] Hull, J and A White (1994). Numerical procedures for implementing term structure models
ii: Two-factor models. Journal of Derivatives.
[10] Nelson, CR and A Siegel (1987). Parsimonious modelling of yield curves. The Journal of
Business, 60(4), 473–489.
[11] Arin, KP and A Mamun (2005). Is it inflation or inflation variability? A note on the stock
return-inflation puzzle. Finance Letters, 3(6), 19–23.
[12] Sellin, P (2001). Monetary policy and the stock market. Theory and empirical evidence.
Journal of Economic Surveys, 15(4).
[13] Feldstein, M (1980). Inflation and the stock market. American Economic Review, 70,
839–847.
[14] Fama, EF (1981). Stock returns, real activity, inflation and money. American Economic
Review, 71, 545–565.
[15] Friedman, M (1977). Nobel lecture: Inflation and unemployment. Journal of Political
Economy, 85, 451–472.
[16] Li, DX (2000). On default correlation: A copula approach. Journal of Fixed Income, 9,
43–54.
[17] Hull, J and A White (2004). Valuation of a CDO and an n-th to default CDS without a
Monte Carlo simulation. Journal of Derivatives.
[18] Kalemanova, A and R Werner (2006). A short note on the efficient implementation of the
Normal Inverse Gaussian distribution. Working Paper.
[19] Vasicek, O (1987). Probability of loss on loan portflio. Memo, KMV Corporation.
[20] Bluhm, C (2003). CDO modeling: Techniques, examples and applications. Working Paper.
[21] Schmid, B., R Zagst and S Antes (2006). Pricing of credit derivatives. Working Paper.
[22] Zagst, R (2006). Integrated risk management — distinguished lecure series 2006. Lecture
notes.
[23] Standard & Poor’s (2006) Annual 2005 global corporate default study and rating transi-
tions.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch07 FA
The dependence of extreme financial events among different asset classes is taken
under consideration on a portfolio level. For this, a new product group, called cross
asset portfolio derivatives, is introduced and explained in the light of related existing
products and pricing methods. A classification is presented and features of these prod-
ucts are described. Finally, two modeling and pricing frameworks using multivariate
stochastic processes and (hierarchical) copulas, respectively, are suggested.
This chapter focuses on the dependence of tail events, i.e., the risk of rare market events
occurring interdependently. A current trend in product development is to apply credit
175
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch08 FA
and the total loss is realized with the 20th, the result is an option-like payoff. This
option-type behavior can be used to introduce the leverage into the investment.
Due to the so-called waterfall structure of the tranched CTSO, the investor can
choose different subordination levels. This choice determines the real trigger event
exposure to the portfolio. A less senior tranche bears a higher trigger risk, but bearing
this risk is compensated with higher spread on the outstanding notional. In case of a
CDO, one often classifies the subordination in equity, mezzanine, and senior tranches.
These names intend to reflect the different risk exposures.
Example 2 (General structure of an nth-to-trigger basket). Consider a portfolio
of 10 TS, with a notional of one million US$ each, and an investor (insurance buyer)
who wants to protect herself against the first loss in the portfolio by means of a first-to-
trigger swap. For this, she pays periodic premium payments on a pre-specified sched-
ule. As soon as one of the TS is triggered, premium payments stop and she receives
a contingent payment in the amount of reimbursement rate times the notional. For
n > 1, nth-to-trigger derivatives are defined similarly, with the first trigger replaced
by the nth trigger.
Figure 8.1 shows the classification of cross asset portfolio derivatives within the
product group of trigger derivatives. The specific products mentioned denote examples
of the respective group.
At first glance, CFO and CTSO share many similarities, but the structural features as
well as the involved risks differ quite significantly. The following structural features
are presented in [3].
new investment technologies based on asset classes, they are familiar with. Traditional
CFO, using equity and commodity indices as underlying, did not become as popular
as their counterparts: cash flow CO. In 2000–2001 CFO experienced a strong growth,
with hedge fund investments being included. The idea was to use a technology, which
generates benefits from a combination of highly volatile and uncorrelated trading
strategies (long-short equity, convertible arbitrage, event-driven, and fixed income
arbitrage). On the negative side, these CFO investments included multiple fees due to
the investment in tradable funds. According to [4], securitized hedge fund investments
can provide the following benefits:
In 2003–2004, according to [3], hedge fund-linked CFO activity slowed down due to
accounting issues. Still, market interest remained. The only innovative feature in these
transactions, compared to usual CO, was the use of an uncommon underlying.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch08 FA
systematic risk factors. The risk/return profile of CTSO tranches differs significantly
from the one of the underlying assets, shifting risk away from volatility toward tail
events. The impact of changes in correlation depends on the specific tranche as the
following two examples illustrate:
(1) A long position in an equity piece of a CCO, where the trigger levels are set
below and above the current spot prices, is long in joint probabilities of large price
swings in the commodity market, i.e., the buyer of that piece expects rising joint
probabilities of large price movements.
(2) A short position in a senior piece of a cross asset CTSO is long in joint probabilities
of distortions in the underlying markets, i.e., the short seller expects the likelihood
of a market crash to rise.
Considering inter-market correlation is not new, but the securitization in terms of the
aforementioned products is. However, in order to derive the correlation of trigger events
in different financial markets, analyzing historical data is not sufficient. Mathematical
models, such as structural and/or copula-based models known from credit markets,
have to be implemented in order to get an intuition for the traded tail event correlations
in CTSO.
N · s −rt tk
n
= e k Q(τ > tk ). (8.1)
4 k=1
n
= RIR · N e−rtk tk Q(tk−1 < τ ≤ tk ). (8.2)
k=1
Hence, from the perspective of the protection buyer, the expected value of a TS at
initiation of the contract is given by
n
N · s −rk tk
n
TS(0, T) = RIR · N e−rk tk Q(tk−1 < τ ≤ tk ) − e Q(τ > tk ).
4
k=1 k=1
The par spread sf , computed to allow both parties to enter the TS at zero cost, is found
by solving TS(0, T) = 0 for s, i.e.,
n −rtk tk
RIR k=1 e Q(tk−1 < τ ≤ tk )
sf = n −rtk tk
. (8.3)
1
4 k=1 e Q(τ > tk )
Equations (8.1) and (8.3) do not account for accrued interest, which might additionally
be included as part of the terms of contract. If accrued interest is stipulated, an additional
premium payment is added to the premium leg, which is proportional to the time
between the trigger event τ and the last payment date tk−1 and is conditional on the
event that the trigger falls into the respective period (tk−1 , tk ].
N ·s (n)
n
= e−rtk tk Q(τ (n) > tk ), (8.4)
4 k=1
n
−rtk tk
EDTL (n)
= EQ RIR · N e 11{tk−1 <τ (n) ≤tk }
k=1
n
= RIR · N e−rtk tk Q(tk−1 < τ (n) ≤ tk ). (8.5)
k=1
The fair spread sf(n) is again obtained from equating both legs and solving for sf(n) . This
yields
(n) RIR nk=1 e−rtk tk Q(tk−1 < τ (n) ≤ tk )
sf = n −rtk tk
. (8.6)
1
4 k=1 e Q(τ (n) > tk )
The total nominal value of the CTSO is given by M = I · N. This nominal value is
segmented using the attachment and detachment point of each tranche. More precisely,
we define the partition 0 = l1 < u1 = l2 < · · · < uJ−1 = lJ < uJ = M. Based on
the overall portfolio loss, the loss affecting tranche j results in
j
Lt = min{max{0, Lt − lj }, uj − lj }, t ∈ [0, T ], j ∈ {1, . . . , J}. (8.8)
Given the payment schedule T = {0 = t0 < t1 < · · · < tn = T } and the pricing
measure Q, the expected discounted premium and trigger legs of tranche j are given by
n
−rtk tk j
EDPL = EQ
(j)
s · tk · e
(j)
Mtk , j ∈ {1, . . . , J}, (8.10)
k=1
n
−rtk tk j j
EDTL (j)
= EQ e (Ltk − Ltk−1 ) , j ∈ {1, . . . , J}. (8.11)
k=1
a proxy variable. An example for the latter is the firm-value interpretation used
in structural-default models. Typically, the trigger time τi is then defined as some
first-passage time of the respective univariate marginal below or above some trig-
ger level. Depending on the choice of univariate model, one has different options to
couple the marginals. For instance, if the univariate marginals are driven by Brow-
nian motions, a natural approach is to assume these Brownian components as being
correlated. If more general Lévy processes are used, common jumps might also be
used to introduce dependence. References for the modeling of stocks and credit prod-
ucts are [2] and [1]. References for the modeling of interest rates and energy, as
examples for other underlyings than stocks and credit, are the books [5] and [6],
respectively.
The major advantage of the structural approach is that it is fully dynamic, since all
relevant objects are modeled as marginals of one high-dimensional stochastic process.
Besides the static pricing of all univariate and portfolio derivatives at a time, this
additionally induces a dynamic which can be useful for the pricing of options on these
derivatives.
However, the structural approach has some drawbacks. Firstly, it is not obvious
how dependence shall be introduced to univariate processes from different model
classes, e.g., classical diffusions and pure jump processes. Secondly, analytically solv-
ing a non-trivial multi-dimensional first-passage time problem is virtually impossible.
This becomes evident when the example of [7] is considered, who succeeds in present-
ing a (highly non-trivial) formula for the joint probability of two correlated diffusions
to remain above some threshold level for a given time. Hence, pricing portfolio trigger
swaps requires a Monte Carlo engine, which is slow and biased if correlated processes
are sampled on a discrete grid and monitored for some threshold level. Thirdly, for
dependent stochastic processes, it is typically not possible to separate the parameters
of the marginals from those specifying the dependence structure; an exception are pro-
cesses purely driven by correlated Brownian motions. This complicates the calibration
of the model. Finally, the dependence structure of the resulting vector (τ1 , . . . , τI ) is
unknown for all non-trivial examples.
canonical construction
τi = inf{t ≥ 0 : pi (t) ≤ Ui }, Ui ∼ Uniform(0, 1). (8.13)
This representation is especially convenient for simulations, since sampling τi boils
down to sampling the univariate trigger variable Ui and solving, assuming a continuous
and strictly decreasing pi , pi (t) = Ui for t. Considering all random times τ1 , . . . , τt
at once it is possible to introduce dependence to the model by assuming the vector of
trigger times (U1 , . . . , UI ) to be distributed according to some I-dimensional copula
C. Note that, by the defining properties of a copula, it is guaranteed that the univariate
marginal distributions (of all τi ) are preserved, which is highly convenient in a cali-
bration. If sampling strategies for C are known, it is straightforward to sample from
the vector of random times (τ1 , . . . , τI ): sample (U1 , . . . , UI ) ∼ C and locate each
univariate trigger time τi .
The major advantage of the copula approach is the separation of marginals from
the dependence structure. Firstly, this allows the use of different model classes for
the univariate marginals. For instance, one might use a classical geometric Brownian
motion for one asset class, a pure-jump Lévy process for a second asset class, and
an intensity model for the asset class credit. For each of these classes, one can derive
univariate distributions for the respective τi ’s, which are then coupled by means of
the copula C. Using copulas with tail-dependence and singularities allows to include
Armageddon scenarios (with multiple trigger events at a time) and trigger clusters.
Secondly, the approach is well-suited for simulations, as long as sampling routines for
the copula are known. Thirdly, the calibration of the multivariate model is simplified
when the univariate marginals can be calibrated individually, followed by a calibration
of the dependence structure in a second step.
However, there are some shortfalls of the copula approach that need to be
addressed. Firstly, the model is static in the sense that a multivariate distribution of
(τ1 , . . . , τI ) is specified, but no dynamic model leading to it. While this is unprob-
lematic for pure pricing problems, it rules out applications such as the simulation of
the time evolution of resulting model spreads. Secondly, it is not clear which class of
copulas is best suited for modeling the dependence structure of the respective market.
As long as there is no liquid market for tail dependence, it is not clear which copula
should be preferred and how this copula should be parametrized.
Let us briefly give some references for the required sampling of copulas. In high
dimensions, sampling strategies for some elliptical copulas (including the Gauss-
and t-copula) are known, see e.g., [10]. For exchangeable (meaning that all pair-
wise correlations are identical) Archimedean copulas the standard reference is [13], an
important contribution is [14]. Recently, some authors considered hierarchical (nested)
Archimedean copulas and sampling strategies for these, see e.g., [15]– [17]. Sampling
Marshall–Olkin type copulas in high dimensions is considered in [18].
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch08 FA
Example 3 (An nth-to trigger basket over two asset classes). In this example, an
nth-to trigger basket consisting of six underlying TS is considered. The payment fre-
quency for premium payments is quarterly, the contract settles on 10 February 2009,
with a maturity of T = 5 years. The reimbursement rate is set to RIR = 50% for all TS
and for the CDS calibration. All TS contribute the same unit nominal to the portfolio.
All equity events are defined as τi = inf{t ≥ 0 : Si,t ≤ li }, where li = 0.2 · Si,0 for
the respective contract. Credit events are taken as τi = inf{t ≥ 0 : Vi,t ≤ di } for
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch08 FA
some suitable default threshold di , i = 1, 2, 3. Solving the model via a Monte Carlo
simulation requires the following steps.
(1) Initialize the simulation, i.e., choose the number of simulation runs SR and the
grid-size .
(2) For each simulation run n = 1, . . . , SR do:
(a) Simulate the vector-valued process (S1,t , S2,t , S3,t , V1,t , V2,t , V3,t ) on the grid
0 < < 2 < · · · < T .
(b) At each point of the grid, monitor for eventual trigger times below (or above,
for other contract specifications) the equity or default thresholds, respectively.
(3) Compute and store the discounted premium and trigger leg of the current simula-
tion run n.
(4) Estimate the expected discounted premium and trigger leg as the arithmetic mean
of the legs in each run.
(5) Set the (estimated) fair spread as the quotient of the estimated trigger and premium
leg.
Note that such a continuous model might underestimate the risk of extreme (joint)
movements. This fact is well known in the univariate case, see e.g., Chapter 1 of [20],
and is further amplified for joint movements, due to a lack of tail dependence of the
multivariate normal distribution. To overcome this, one might try to use jump processes
instead of geometric Brownian motions.
to model the stocks under Q. All equity events are again defined as τi = inf{t ≥
0 : Si,t ≤ li }. Credit events are defined as τi = inf{t ≥ 0 : exp (−λi t) ≤ Ui },
where U1 , U2 , U3 are dependent (univariate uniform) triggers and λ1 , λ2 , λ3 denote
the (constant) default intensities. Solving the model via Monte Carlo requires the
following steps.
(1) Initialize the simulation, i.e., choose the number of simulation runs SR.
(2) For each stock index, compute survival probabilities Q(τi ≥ t) from the parameters
of the respective process. For this, the required probability of a Brownian motion
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch08 FA
with drift not falling below some threshold level for a given amount of time is
given in [21], page 61.
(3) For each simulation run n = 1, . . . , SR do:
(a) Simulate the dependent trigger variables (U1S , . . . , U1C , . . . ) ∼ C.
(b) Compute the resulting trigger times.
(4) Compute and store the discounted premium and trigger leg of the current simula-
tion run n.
(5) Estimate the expected discounted premium and trigger leg as the arithmetic mean
of the legs in each run.
(6) Set the (estimated) fair spread as the quotient of the estimated trigger and premium
leg.
More details on the copula approach in the context of credit–risk modeling, including
asymptotic confidence intervals and the implementation of several nestedArchimedean
copulas, are given in [22].
where ψi (t) = exp(−t 1/ϑi ) and ϑi ∈ [1, ∞), i = 1, 2. A sufficient condition for (8.17)
being a copula, is that the nodes ψ2−1 ◦ ψ1 have completely monotone derivatives. For
the nested Gumbel copula, this condition is equivalent to ϑ2 ≤ ϑ1 . NestedArchimedean
copulas might be interpreted as follows: any two members of the groups stocks and
credit are coupled via an inner Archimedean copula, and any two members of different
groups are coupled via an outer Archimedean copula. Nested Archimedean copulas are
appealing for the modeling of cross asset portfolio derivatives, since members of the
same asset class can be coupled via inner copulas, which are then connected via one
outer copula. For simplicity, we choose a single parameter ϑ1 for both groups (stocks
and credit) and a second parameter ϑ2 ≤ ϑ1 for the outer copula. This implies that the
dependence within a group is at least as large as the dependence among members of
different groups. The resulting model prices as functions of the dependence parameters
are presented in Figs. 8.3 through 8.5.
An interpretation of the resulting model spreads is given in what follows:
1st to trigger spread (in bps) 2nd to trigger spread (in bps)
1 1
350 140
2
250
ρ
0.4 0.4 100
200
90
0.2 0.2
80
150
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
ρ1 ρ1
3rd to trigger spread (in bps) 4th to trigger spread (in bps)
1 1
70 50
0.8 0.8
60 40
0.6 50 0.6
30
ρ2
2
ρ
40
0.4 0.4
20
30
0.2 0.2 10
20
0 10 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
ρ1 ρ1
5th to trigger spread (in bps) 6th to trigger spread (in bps)
1 1
40
20
0.8 0.8
30
15
0.6 0.6
ρ2
2
ρ
20 10
0.4 0.4
10 5
0.2 0.2
0 0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
ρ1 ρ1
Figure 8.3 Model spreads from the structural approach (25,000 runs).
The {4,5,6}th-to trigger spreads are all increasing in both dependence parameters,
since more dependence (of either sort) makes multiple defaults more likely, with
according influence on spreads.
• Comparing the structural approach and the Gauss-copula approach to the Gum-
bel copula shows that kth-to trigger spreads for higher k are much larger for the
latter. This is explained by the positive upper-tail dependence of the Gumbel family.
Heuristically speaking, this property means that the trigger variables Ui have a pos-
itive probability for being simultaneously close to one. Translated into the pricing
framework, this implies multiple trigger events within one simulation run, hence,
positive spreads for higher kth-to trigger spreads. Extending the simple structural
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch08 FA
1st to trigger spread (in bps) 2nd to trigger spread (in bps)
1 400 1
140
0.8 350 0.8 130
300 120
0.6 0.6
110
ρ2
ρ2
250
0.4 0.4 100
200 90
0.2 0.2
80
150
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
ρ1 ρ1
3rd to trigger spread (in bps) 4th to trigger spread (in bps)
1 1
70 50
0.8 0.8
60
40
0.6 50 0.6
ρ2
ρ2
30
40
0.4 0.4
30 20
0.2 0.2
20 10
0 10 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
ρ1 ρ1
5th to trigger spread (in bps) 6th to trigger spread (in bps)
1 1
40
0.8 0.8 20
30
0.6 0.6 15
ρ2
ρ2
20 10
0.4 0.4
0.2 10 0.2 5
0 0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
ρ1 ρ1
Figure 8.4 Model spreads from the Gauss-copula model (250,000 runs).
model using jump processes might resemble this property. Another observation is
that spreads for the limiting case of independence agree within all models (up to a
small Monte Carlo noise).
• A final remark addresses the required computation time for both approaches. The
structural approach is based on the simulation of a path of a multivariate process
(on a fine grid) in each simulation run. Compared to the copula approach, which
only requires a sample from the respective copula in each run, it is therefore not
surprising that the computation time for the evaluation of one option price with
250,000 simulation runs differs massively. Our implementation in Matlab on a
standard PC requires about 3813 s for the structural and 11 and 18 s for the two
copula approaches, respectively.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch08 FA
1st to trigger spread (in bps) 2nd to trigger spread (in bps)
2 2
ϑ2
1.4 1.4 90
250
1.2 1.2 80
1 200 1
1 1.2 1.4 1.6 1.8 2 1 1.2 1.4 1.6 1.8 2
ϑ1 ϑ1
3rd to trigger spread (in bps) 4th to trigger spread (in bps)
2 60 2
40
1.8 50 1.8 35
30
1.6 1.6
40 25
ϑ2
ϑ2
1.4 1.4 20
30
15
1.2 1.2
20 10
5
1 1
1 1.2 1.4 1.6 1.8 2 1 1.2 1.4 1.6 1.8 2
ϑ1 ϑ1
5th to trigger spread (in bps) 6th to trigger spread (in bps)
2 2
30
1.8 25 1.8 15
1.6 20 1.6
10
ϑ2
ϑ2
15
1.4 1.4
10
5
1.2 1.2
5
1 1
1 1.2 1.4 1.6 1.8 2 1 1.2 1.4 1.6 1.8 2
ϑ1 ϑ 1
Figure 8.5 Model spreads from the Gumbel-copula model (250,000 runs).
8.5. OUTLOOK
interesting. Allowing for collateralized products referring to other asset classes than
debt offers new opportunities for investments and risk management.
Cross asset portfolio derivatives typically smooth the risk/return profile of volatile
asset classes. Whereas the classical Capital Asset Pricing Model (CAPM) is still the
appropriate model to optimize the underlyings regarding return, risk, and correlation,
this portfolio management style changes in the case of cross asset portfolio deriva-
tives. In the CAPM world, the risk/return optimization is already implemented in the
market portfolio, which is defined as being efficient, while correlation between the
portfolio constituents is defined as the co-movement of asset prices. The individual
risk perception is simply reflected by the investment share of the market portfolio,
and the risk-free asset. In a cross asset portfolio derivatives world, the underlying
portfolio is selectable without the usual µ/σ-constraints. Correlation is rather joint
trigger probability and, e.g., the specific tranche investment in case of a CTSO is done
in line with the individual risk perception. Correlation, defined as co-movement of
assets, is replaced by the correlation of tail events, which induces a different view on
the risk/return optimization. Systematic shocks, like the sub-prime crisis starting in
2007, are a crucial risk factor for a cross asset portfolio derivative, whereas market
fluctuations and cyclical moves are of minor importance for the performance of such
derivatives.
To sum up, cross asset portfolio derivatives might be an interesting investment
product or portfolio management tool. There should not be any concerns regarding
demand/supply patterns. Both, demand and supply for tail event risk is immense,
given the huge variety of players in this market and, e.g., the recent financial crisis or
the rise/decline of the oil price. Traditional tail-event risk-driven companies (insurance
sector) are probably skewed to selling this kind of risk to players like hedge-funds,
banks, and fund managers, who can implement the attractive trading positions, but
also optimize the multi-asset portfolios, using cross asset trigger derivatives.
8.6. CONCLUSION
for both insurance seller and buyer. A standardized market of these derivatives would
then allow to calibrate the models to the market’s perception.
Acknowledgments
We thank Dr. Jochen Felsenheimer of Assenagon and Dr. Philip Gisdakis of UniCredit
MIB for their initial thoughts on this product group. Moreover, we thank Marius Hofert
(Universität Ulm) for valuable remarks on earlier versions of the manuscript. Views
expressed in this paper are those of the authors and do not necessarily reflect the
positions of the respective employers.
References
[1] Schönbucher, PJ (2003). Credit Derivatives Pricing Models: Models, Pricing and Imple-
mentation, 1st edn. Wiley Finance Series, John Wiley and Sons, Inc.
[2] Hull, J (2004). Options, Futures, and Other Derivatives. London: Prentice-Hall Interna-
tional, Inc.
[3] Felsenheimer, J (2006). The wonderful world of CAOs (strategy update). Technical report,
UniCredit MIB Global Credit Research.
[4] Standard and Poor’s (2006). CDO spotlight: Global criteria for securitizations of funds of
hedge funds. Standard and Poor’s, Structured Finance.
[5] Brigo, D and F Mercurio (2001). Interest Rate Models — Theory and Practice. Springer,
Finance, Berlin: Springer.
[6] Benth, FE, JS Benth and S Koekebakker (2008). Stochastic Modeling of Electricity and
Related Markets. World Scientific.
[7] Zhou, C (2001). An analysis of default correlations and multiple defaults. Review of Finan-
cial Studies, 14, 555–576.
[8] Joe, H (1997). Multivariate Models and Dependence Concepts. New York: Chapman &
Hall.
[9] Nelsen, RB (1998). An Introduction to Copulas, 1st edn. Berlin: Springer.
[10] McNeil, AJ, R Frey and P Embrechts (2005). Quantitative Risk Management. Princeton
Series in Finance, Princeton University Press.
[11] Li, DX (2000). On default correlation: A copula function approach. The Journal of Fixed
Income, 9(4), 43–54.
[12] Schönbucher, PJ and D Schubert (2001). Copula-dependent defaults in intensity models.
URL http://papers.ssrn.com/sol3/papers.cfm?abstract_id= 301968. Working Paper.
[13] Marshall, AW and I Olkin (1988). Families of multivariate distributions. Jour-
nal of the American Statistical Association, 83(403), 834–841. URL http://www.
jstor.org/stable/pdfplus/2289314.pdf.
[14] McNeil, AJ and J Neslehova (2009). Multivariate Archimedean copulas, d-monotone func-
tions and l1-norm symmetric distributions. The Annals of Statistics, 37, 3059–3097.
[15] McNeil, AJ (2008). Sampling nested Archimedean copulas. Journal of Statistical Compu-
tation and Simulation, 78(6), 567–581.
[16] Hofert, M (2008). Sampling Archimedean copulas. Computational Statistics and Data
Analysis, 52(12), 5163–5174. doi:10.1016/j.csda.2008.05.019.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch08 FA
Part II
Alternative Strategies
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch09 FA
201
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch09 FA
202 Dersch
9.1. INTRODUCTION
9.2.1. Buy-and-Hold
Probably the most simple and most common risk management framework is the buy-
and-hold strategy. Not just among retail investors, either deliberately or not deliberately
buy-and-hold is a widespread approach. Many ambitious strategies will eventually drift
into passive sit-and-wait strategies as the investor sits out long periods of negative
performance or hesitates to take profits on time. Cheekily, buy-and-hold or strategic
trades very often stem from short-term tactical trades turned bad. The only reason why
this strategy remotely qualifies as some kind of protection strategy is that in the absence
of leverage the total loss is limited to the initial investment. However, buy-and-hold
has performed well over long periods and across many asset classes.
9.2.2. Stop-Loss
A stop-loss strategy is the first non-trivial step toward a risk management framework.
Here, we distinguish between an investment target IT on a present value base and a
target on a given time horizon T . In the first case, the position has to be switched into a
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch09 FA
204 Dersch
risk-free investment matching the investment horizon if the portfolio value falls below
the target value at any time
PVportfolio (t) > IT . (9.1)
In the second case, the portfolio value must not fall below
PVportfolio (t) > e−r(T −t) · IT (9.2)
or equivalently
PVportfolio (t) − e−r(T −t) · IT > 0. (9.3)
Here, T − t is the time until the investment horizon is reached and r is the risk-free
rate of this time span. The second case is slightly more complex. Here, we must
ensure that the investor reaches his/her investment target at maturity. Therefore, the
portfolio value must not fall below a certain floor value given by the right hand side
of Eq. (9.2). In case, the stop-loss level is reached, the portfolio must be liquidated
and invested in the risk-free asset. Investing the floor value with the risk-free rate will
ensure the given target value at maturity. The implementation of a stop-loss strategy
requires monitoring both the portfolio performance and the total return of the risk-
free investment. The risk-free investment is usually implemented with treasury bills or
bonds with a maturity matching the investment horizon. The re-investment of coupons
paid until maturity must also be considered. In case, the stop-loss level is reached,
a single portfolio re-allocation occurs. Strictly speaking, stop-loss is therefore not a
dynamic strategy.
In the above analysis, we assume that the investments and the target level are in
the same consolidation currency. Otherwise, the respective FX spot or forward rates
have to be additionally monitored. Please note that a stop-loss strategy with a target
value of zero is the same as the above buy-and-hold strategy.
maturity matching the investment horizon. The bond floor strategy is usually imple-
mented with a target value close to 100%. According to the above equation, a given
investment of EUR 100, a target value of EUR 101 in one year, and a risk-free rate of
2.5% would leave us with a risk budget of EUR 1.5 and a bond investment of EUR 98.5.
That means EUR 1.5 is invested with a buy-and-hold strategy in a risky investment.
A total loss of the risky investment would still guarantee the target amount of 101 EUR
in one year. For a target value close to 100%, the bond floor strategy — as compared
to the stop-loss strategy — swings the pendulum in the other direction of extreme risk
aversion with the consequences of little upside potential beyond the target value. There
are two things worth mentioning: The bond floor must not be larger than the amount
that may be earned with the risk-free investment and a bond floor of zero is the same
as the above buy-and-hold strategy.
206 Dersch
Table 9.1 Comparison of Simple Protection Strategies and Their Relation to CPPI.
Strategy Risk appetite Upside potential Short fall risk Relation to CPPI
In this section, we show sample simulations for the above protection strategies. Our
risky portfolio is the Dow Jones Euro Stoxx 50. The simulation covers a period of
close to 10 years. For comparative reasons, we use similar protection levels where
applicable.
• Table 9.2 summarizes the simulation parameter.
• Figure 9.1 shows the buy-and-hold strategy.
• Figure 9.2 shows the stop-loss strategy.
• Figure 9.3 shows the bond floor strategy on the DJ Euro Stoxx 50 index.
• Figure 9.4 shows the CPPI strategy with a target level of 100% and a multiplier of
6. The floor is calculated using a discount rate of 3.25%.
• Table 9.3 summarizes the simulation results.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch09 FA
180
160
140
120
100
80
60
40
20
0
1999 2000 2001 2002 2003 2004 2005 2006 2007 2008
Figure 9.1 Buy-and-hold strategy on the DJ Euro Stoxx 50 performance index for an initial
portfolio value of EUR 100 mn. The performance of the portfolio (in EUR mn) is the performance
of the index.
In this market environment, the bond floor strategy seems to perform best over the
complete period both in the absolute return and the size of the worst draw-down. The
three other strategies all reach the same maximal portfolio amount of EUR 156.3 mn
(up 56%), but fail to benefit at maturity. The stop-loss and CPPI strategy show a very
similar picture. They are both stopped out during the sharp market decline in mid-2002
and realize a slight loss as compared to the initial portfolio value. This translates into
a slightly negative annual return of −0.17% (CPPI) and — 0.30% (stop-loss). Buy-
and-hold ranks last with respect to final portfolio amount annual return, and suffers
the worst portfolio draw-down of 65% of the previous all-time high value.
Stop-loss and CPPI both fail the target by a small amount. This may be due to two
reasons. The first is fundamental and is caused by rapid market movements when the
position may only be liquidated below the theoretical stop level. We can reduce this
fundamental risk by monitoring the position intraday. The second reason is caused by
a simplification of our simulation framework. Monitoring the stop-loss or floor level
requires one to monitor the zero coupon bond with a maturity equal to the remaining
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch09 FA
208 Dersch
180
DJ EUR STOXX 50 risk-free
160
140
120
100
80
60
40
20
0
1999 2000 2001 2002 2003 2004 2005 2006 2007 2008
Figure 9.2 Stop-loss strategy on the DJ Euro Stoxx 50 performance index with a target level
of 100% at maturity. The risk-free investment is EUR overnight liquidity. The y-axis shows the
portfolio value in EUR mn.
160
DJ EUR STOXX 50 risk-free
140
120
100
80
60
40
20
0
1999 2000 2001 2002 2003 2004 2005 2006 2007 2008
Figure 9.3 Bond floor strategy on the DJ Euro Stoxx 50 index with a target level of 100% at
maturity. The risk-free investment is EUR overnight liquidity. The y-axis shows the portfolio
value in EUR mn.
180
DJ EUR STOXX 50 risk-free floor
160
140
120
100
80
60
40
20
0
1999 2000 2001 2002 2003 2004 2005 2006 2007 2008
Figure 9.4 CPPI strategy on the DJ Euro Stoxx 50 performance index with a target level of
100% at maturity. The risk-free investment is EUR overnight liquidity. The floor (black line) is
assumed to follow an annual rate of 3.25%. The y-axis shows the portfolio value in EUR mn.
Table 9.3 Historical Simulation of Four Different Protection Strategies Using the
Dow Jones Euro Stoxx 50. We Calculate the Annual Return Over the Simulation
Period and the Worst Draw-Down. A Draw-Down of, e.g., 65% Means That the
Portfolio Lost 65% of Its Previous All-Time-High.
The above simulations illustrate different protection strategies. However, they are not
suitable for practical use for a number of reasons:
210 Dersch
The volatility filter λ scales the size of the rebalance — either buy or sell — if a
rebalancing is triggered according to Eq. (9.8).
E = λ · |Eold (t) − Enew (t)|, (9.9)
with λ in the range of [0.7,1.0]. For λ smaller than one, we follow only a fraction of
the rebalancing amount. The idea behind the volatility filter is that the market trades
in a range rather than follows a trend.
The re-observation period t is also a transaction filter: The larger t, the more
time elapses between a test of Eq. (9.8). The re-observation period has to be in line
with . Small and a large t or very close to one and small t do not fit well
together. Typically, t is in the range of one day. This implies that we test Eq. (9.8) at
close, but ignore intra-day movements outside the range defined by . This reflects the
observation that the intra-day volatility is typically larger than the day-to-day volatility.
212 Dersch
The advantage of a lock-in is that gains that have been reached in the past are
protected. On the other hand, a lock-in may also cause a reduction in the exposure
due to a rising floor level and therefore limits participation in a future rise in the risky
portfolio.
214 Dersch
advantage of a market recovery if the risk budget is strongly depleted after a downturn
in the risky portfolio.
The situation is worse if this happens in the early stage of the strategy. One pos-
sibility to remedy this weakness is the minimum exposure CPPI. Here, the allocation
in the risky portfolio is held at a minimum guaranteed level. This ensures participa-
tion over the complete lifetime of the strategy. But the participation comes at a cost.
An additional option has to be purchased to guarantee the target level at maturity. In
this chapter, we focus on portfolio strategies without options and therefore follow a
different approach.
If we forgo the requirement of a target level at a fixed time horizon, we can
instead attempt to secure — at any time — a fraction, e.g., 80% of a past portfolio
value. In our advanced CPPI framework, this approach can be implemented by setting
the discount rate in Eq. (9.7) to zero. Now target and floor are the same. In order to
provide a risk budget, the initial target value has to be below 100% of the notional
amount invested. Adding the lock-in type all-time-high trailing results in a strategy
that has been described as Time-Invariant Portfolio Protection (TIPP) [4]. The main
characteristics are
The last point is a drawback of the approach. However, the investor can redeem the
investment at any time.
In the following, we carry out four simulations with and without transaction fil-
ters combined with and without transaction costs. The risky asset is the DJ Euro Stoxx
Selected Dividend 30 index. The simulation parameters are summarized in Table 9.4.
To illustrate the impact of transaction costs, we selected a period of a rising market. The
transaction costs are set to 60 bp of the transaction volume for the risky investment and
to 1 bp for the risk-free investment. The transaction costs are deducted from the port-
folio at the time of the transaction. In our simulation, we ignore the effect of slippage
and partial execution. Figure 9.6 shows one out of the four simulations using transac-
tion filters and considering transaction costs. Table 9.5 summarizes a comparison of
different simulations with and without transaction filters and transaction cost.
Without the transaction filter, there is a huge impact of transaction costs. The
difference is EUR 54.45 mn. This means that the naïve approach — Simulation 4 —
would suffer a drop of more than 20% of the final portfolio value in case transac-
tion costs have to be taken into account (Simulation 2). Using the transaction filter
results in much smaller dependency on transaction costs. The difference here is only
Table 9.5 Comparison of Different Historical CPPI Simulations With (Without) Trans-
action Filters, Simulations 1 and 3 (Simulations 2 and 4) and With (Without) Transaction
Costs, Simulations 1 and 2 (Simulations 3 and 4). For the Different Combinations, We
Show the Total Number of Rebalancing Steps, the Portfolio Value at Maturity and the
Total Transaction Costs.
Trans. Trans. # of PV final in Transaction cost
Simulation filter costs transactions EUR (mn) in EUR (mn)
216 Dersch
EUR 23.90 mn (9.4%). The above simulation clearly reveals the damaging effect of
transaction cost in the absence of the transaction filter. There is a primary effect caused
by pure transaction costs that are deducted from the portfolio. This accounts for EUR
−2.04 mn (Simulation 1, with transaction filter) and EUR −8.84 mn (Simulation 2,
without transaction filter). There is a secondary effect that shows the total effect of
transaction cost. A comparison of the final portfolio values shows that in case of
transaction filters the difference between portfolio values of EUR −23.90 mn contains
transaction costs in the amount of EUR −2.04 mn. In the absence of transaction filters,
the performance gap is EUR −54.45 mn, where EUR 8.84 mn is the pure transaction
cost. In this context, the transaction cost ratio
PV (tc) − PV (ntc)
tcr = (9.16)
TC
is a useful ratio to measure the secondary effect of transaction cost. Here, PV (tc) is the
performance with transaction cost and PV (ntc) without transaction cost and TC the
pure transaction cost incurred over the observation period. A tcr close to one means
that the performance difference with and without transaction cost is mainly caused by
pure transaction cost. A tcr of two means that the transaction cost causes a performance
reduction of twice the amount of the pure transaction cost. In our simulation, the tcr
is equal to 11.71 (with transaction filter) and 6.16 (without transaction filter). The tcr
depends on the compounding effect and the leverage of exposure. In a CPPI simulation
with a multiplier of 3, every EUR of risk capital (cushion) changes the exposure to
the risky portfolio by EUR 3. Transaction costs that reduce the exposure early in
the investment period may have a tremendous impact on the performance later on.
Ironically, in a falling market, transaction costs may even have a positive impact in the
presence of a high multiplier as they may force the early reduction of the exposure and
save the portfolio from otherwise higher losses. We also found that if the floor level
is reached early in a simulation, transaction costs become less significant as it makes
no difference whether a further falling market or transaction costs are the cause for a
decline in portfolio value. In general, we conclude that the effect of transaction costs
is not very easy to estimate. Here, simulations shed light on the effect.
Table 9.6 Parameter Settings for Three Different Advanced CPPI Simulations. The
Simulation Parameters Differ only in the Discount Rate for the Floor, the Lock-In Trigger,
and Lock-In Action (the Last Three Rows).
perform our simulation on the Dow Jones Euro Stoxx 50 index. Table 9.6 shows the
simulation parameters.
For comparative reasons, we first show the CPPI simulation without lock-in
(Fig. 9.5). As compared to the previous simulation shown in Fig. 9.4, we use a transac-
tion filter, consider transaction costs, and a target of only 80% as compared to 100%.
250
DJ EUR STOXX Select Dividend 30 risk-free floor
200
150
100
50
0
1999 2000 2001 2002 2003 2004 2005 2006
Figure 9.5 Historical CPPI simulation under realistic condition on the DJ Euro Stoxx Selected
Dividend 30 index using transaction filters and transaction costs. For a detailed specification,
please refer to the text above. The y-axis shows the portfolio value in EUR mn.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch09 FA
218 Dersch
180
DJ EUR STOXX 50 risk-free floor
160
140
120
100
80
60
40
20
0
1999 2000 2001 2002 2003 2004 2005 2006 2007 2008
Figure 9.6 Historical CPPI simulation on the DJ Euro Stoxx 50. The simulation parameters
are shown in Table 9.6. The areas show the allocation of the two assets. The black line is the
floor level. The y-axis shows the portfolio value in EUR mn.
Qualitatively, we find a similar result in Fig. 9.6 as before. The decline in the stock
market leads to a complete switch into the risk-free asset at the end of the investment
horizon. All previous gains are wiped out again. Due to the lower target of 80%, the
portfolio is still invested in the risky asset after the sharp decline in 2002 as compared
to a target level of 100%.
Figure 9.7 illustrates the impact of the all-time-high trailing. The floor is moved
upwards with a rising market. The initial target level of 80% trails the complete upward
market movement of 56%, resulting in a final portfolio amount of 133.8% of the initial
value. The simulation clearly demonstrates that trailing allows one to protect gains that
have been previously accumulated. The floor value is slightly missed for the reasons
already mentioned above.
As a consequence, the increased target level causes a complete exit from the risky
investment until the end of the investment horizon. The strategy protects gains, but is
unable to recover from the draw-down. This weakness is removed in the next simulation
shown in Fig. 9.8.
Compared to the previous simulation, we now discount the floor level with zero.
As a result, the floor is a horizontal line shifted upwards when new all-time-highs are
reached. The floor level marks the guarantee level. Discounting the floor with a rate of
zero reduces the risk budget because the floor is guaranteed instantaneously and not
at maturity.
The first half of the simulation period shown in Fig. 9.8 is similar to the simulation
in Fig. 9.7. The floor trails the rising index value and the following decline results in a
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch09 FA
180
DJ EUR STOXX 50 risk-free floor
160
140
120
100
80
60
40
20
0
1999 2000 2001 2002 2003 2004 2005 2006 2007 2008
Figure 9.7 Historical CPPI simulation on the DJ Euro Stoxx 50 using the all-time-high trailing
with monthly lock-in triggers. The further simulation parameters are shown in Table 9.6. The
areas show the allocation of the two assets. The y-axis shows the portfolio value in EUR mn.
The black line is the floor level. The floor value is slightly missed for reasons mentioned above.
200
DJ EUR STOXX 50 risk-free floor
180
160
140
120
100
80
60
40
20
0
1999 2000 2001 2002 2003 2004 2005 2006 2007 2008
Figure 9.8 Historical CPPI simulation on the DJ Euro Stoxx 50. The simulation parameters
are shown in Table 9.6. The areas show the allocation of the two assets. The black line is the
floor level. The y-axis shows the portfolio value in EUR mn.
reallocation to the risk-free investment. Here, this reallocation occurs earlier because
of the reduced risk budget and we therefore do not reach the same all-time-high in
the portfolio value. The second half of the simulation shows a significantly different
picture as compared to Fig. 9.7. The strategy participates in the rising market starting
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch09 FA
220 Dersch
in 2003. In 2005, the strategy is again allocated in the risky investment to a very high
proportion for a very short period of time. The guarantee level is moved upwards again.
As a result, we reach a final portfolio value of EUR 167.50 nm. For the given example,
the simulation shown in Fig. 9.8 is superior to the other strategies because of two
reasons:
The reason for the rising from the dead like behavior of this strategy is the fact that
even on the floor the risk-free asset continuously regenerates a cushion. Supported by
a large multiplier, the exposure is quickly scaled up again. By the end of 2005, a new
all-time-high levels are reached. They are trailed with a participation rate which equals
the target level (80%).
Table 9.7 summarizes the results for the three different simulations. The TIPP
strategy performs best as compared to the two other strategies.
Table 9.7 Comparison of Different Historical CPPI Simulations With and Without
Lock-In (Columns 2 and 3) and with Lock-In and Flat Floor (Column 4). For the Three
Combinations, We Show the Portfolio Return, the Final Value and the Final Floor Value.
The Floor Value for Simulation 2 is Slightly Missed for the Reasons Mentioned Above.
Strategy (1) CPPI (2) CPPI with lock-in (3) TIPP: lock-in and flat floor
Table 9.8 Parameters for an Historical Advanced CPPI Simulation With Leverage.
2.5
2.0
1.5
1.0
0.5
0.0
-0.5
Figure 9.9 Historical advanced CPPI simulation on the Credit Suisse Tremont Investable
Index. The y-axis shows the portfolio value in EUR mn. The leverage is limited to USD 0.6 mn.
The areas show the allocation of the two assets. Leverage shows up in a negative allocation of
the risk-free asset (light gray area). The gray-and-black-dashed lines are the performance of the
index and portfolio, respectively. The black line is the floor.
Figure 9.9 shows an historical simulation of the above strategy. The light gray
area represents the risk-free investment. Until 2008, we find a leveraged investment.
The allowed leverage of USD 600,000 is utilized to the maximum level at certain
times in 2000, 2005, 2006, 2007, and January 2008. Each increase of the target level
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch09 FA
222 Dersch
160
REX DAX risk-free DJ AIG Commodity Index floor
140
120
100
80
60
40
20
0
2000 2001 2002 2003 2004 2005 2006 2007 2008
Figure 9.10 Advanced historical CPPI simulation on a risky portfolio of three different low-
correlated asset classes. The four different shaded areas indicate the allocation over time. The
black line is the floor level.
Figure 9.10 illustrates how the allocations in different asset classes evolve over
time. From the end of 2000 until the beginning of 2005, the target level is constant.
There are two periods with a full investment in the risky asset (2000 and the end of 2004
until the end of 2005). The final portfolio value is EUR 131.32 mn. This corresponds to
a return of 3.0% p.a. mainly attributed to the years 2000 and 2005–2007. The present
value time weighted asset allocation is 45% Rex, 21% AIG Commodity Index, 17%
DAX, and 17% risk-free. This contrasts the current allocation of 22% Rex, 7% AIG
Commodity Index, 6% DAX, and 66% risk-free. This is intuitive as we are currently
very close to the floor level of EUR 123.95 mn.
224 Dersch
Table 9.10 Comparison of Various Characteristics of ETF’s Please See Also [5].
Bid-offer Management
ETF ISIN spread in bp fee in bp pa
adjusted from the Dow Jones Stoxx 600 universe based on their dividend yield and
dividend consistency. For an exact definition of the index, please refer to the cor-
responding description of Dow Jones. Due to the nature of the index creation, the
portfolio follows a more conservative anti-cyclical profit-taking strategy. Generally,
the underlying rational of the index, its dynamic and characteristics have to be taken
into account when setting the parameters of the dynamic protection strategy. Table 9.10
compares typical bid–offer spreads and management fees of various ETF. The shown
bid–offer spreads are snapshots and may vary from day to day. For further information,
please compare [5].
Bid–offer spreads reflect the characteristics of the underlying and are influenced
among other factors by liquidity and taxation issues. Larger bid–offer spreads in the
investment would favor the transaction filter with larger , t, and small λ.
We have demonstrated how the main drawback of CPPI, namely, its pro-cyclical behav-
ior, the lack of recovering potential once the floor has been hit, and the fixed investment
horizon can be overcome by introducing a number of advanced features. The above
strategies may be implemented with simple spread sheet models avoiding the usage
of options. However, the investor must be aware of the different risk aspects related to
this approach.
We illustrated how dynamic portfolio strategies empower an investor to replicate
fairly complex option profiles including path dependent look-back options. Under
certain conditions, the investor may save hedging costs, e.g., the option premium. But
the savings come at a price. The investor is left with the risk that the strategy will miss
the investment target. Buying an option allows one to lock in the implied volatility
at the time of purchase. If the realized volatility over the lifetime is higher than the
implied volatility of the option, the option is superior. Buying insurance — in the form
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch09 FA
Acknowledgment
The author is grateful to Thorsten Weinelt and UniCredit Research to support this
work, Peter Hieber for layout and formating, and to David Dakshaw for proof reading
the manuscript. Views expressed in this chapter are those of the author and do not
necessarily reflect positions of UniCredit Research.
References
[1] Bertrand, P and JL Prigent (2001). Portfolio insurance strategies: Obpi versus cppi. CERGY
Working Paper, 30.
[2] Black, F and R Jones (1987). Simplifying portfolio insurance. The Journal of Portfolio
Management, 14(1), 48–51.
[3] Black, F and AF Perold (1992). Theory of constant proportion portfolio insurance. Journal
of Economic Dynamics and Control, 16, 403–426.
[4] Estep, T and M Kritzman (1988). Tipp: Insurance without complexity. The Journal of
Portfolio Management, 14(4), 38–42.
[5] iShares (2008).
[6] Credit Suisse Tremont (2008).
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch09 FA
10
HOW GOOD ARE PORTFOLIO
INSURANCE STRATEGIES?
Portfolio insurance strategies are designed to achieve a minimum level of wealth while
at the same time participating in upward moving markets. The most prominent exam-
ples of dynamic versions are option-based strategies with synthetic put and constant
proportion portfolio insurance strategies. It is well known that, in a Black/Scholes
type model setup, these strategies can be achieved as optimal solution by forcing an
exogenously given guarantee into the expected utility maximization problem of an
investor with CRRA utility function. The CPPI approach is attained by the introduc-
tion of a subsistence level, the OBPI approach stems from an additional constraint
on the terminal portfolio value. We bring these results together in order to explain
when and why OBPI strategies are better than CPPI strategies and vice versa. We
determine the utility losses, which are caused by introducing a terminal guarantee
into the unconstrained maximization approach. In addition, we focus on utility losses,
which are due to market frictions such as discrete-time trading, transaction costs, and
borrowing constraints.
227
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch10 FA
10.1. INTRODUCTION
Portfolio strategies which are designed to limit downside risk and at the same time to
profit from rising markets are summarized in the class of portfolio insurance strate-
gies. Among others, [6] and [26] define a portfolio insurance trading strategy as a
strategy which guarantees a minimum level of wealth at a specified time horizon, but
also participates in the potential gains of a reference portfolio. Principally, one can
distinguish between two types of portfolio insurance strategies. A risky portfolio (or
benchmark index) is combined either with a risk-free asset or with a financial deriva-
tive. In particular, the first class includes dynamic versions of option-based portfolio
insurance (OBPI), stop-loss strategies, buy-and-hold strategies and constant propor-
tion portfolio insurance. The second class is mainly characterized by protective put
strategies, either in a static or rolling sense. Notice that, with the exception of the buy
and hold and the protective put, the above strategies are all dynamic in the sense that
they afford portfolio adjustments during the investment horizon. The concept of (syn-
thetic) option-based portfolio insurance is already introduced in [15] and [33]. The
constant proportion portfolio insurance (CPPI) is introduced in [12]. For the evolution
of portfolio insurance, we refer to [33].
The popularity of portfolio insurance strategies can be explained by various rea-
sons. On the side of institutional investors, there are regulatory requirements including
return guarantees as well as requisitions on the risk profile. For example, [1] considers
the problem of an institution optimally managing the market risk of a given exposure
by minimizing its Value-at-Risk using options. Amongst early papers on the optimality
of portfolio insurance are also [7] and [31]. More recently, [22] justifies the existence
of guarantees from the point of an investor through behavioral models. In particular,
they use cumulative prospect theory as an example, where guarantees can be explained
by a different treatment of gains and losses, i.e., losses are weighted more heavily than
gains, cf. [30] and [42].
Unfortunately, the justification of guarantees is less clear assuming that the
investor’s preferences can be described using the [43] framework of expected utility.
Dating back to [37], it is well known that in a Black/Scholes model setup and a constant
relative risk aversion (CRRA) utility function, the expected utility maximizing trading
rule is a constant mix strategy, i.e., a strategy where a constant fraction of wealth is
invested into the risky asset. In this case, an investment weight below one implies that
assets are bought when the asset price decreases. This is in sharp contrast to portfolio
insurance. In order to honor a terminal guarantee, the asset exposure is to be reduced if
the price of the risky asset decreases. Technically, it is straightforward to achieve CPPI
and OBPI strategies as the optimal solution of a modified utility maximization prob-
lem, which is based on an exogenously given guarantee. CPPI strategies are optimal for
an investor, who derives utility from the difference between the terminal strategy value
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch10 FA
and a given subsistence level. In contrast, the OBPI is optimal for a CRRA investor if
one exogenously adds the restriction that the terminal portfolio value is above the floor.
These results are well known in the literature. Without postulating completeness, we
refer to the works of [5, 6, 13, 16, 17, 21, 24, 26–28, 40, 41]. [20] considers the inverse
problem. They analyze if a specific dynamic strategy can be explained by solving the
maximization problem of an expected utility maximizing investor, i.e., they analyze if
a given investment strategy is consistent with expected utility maximization. In partic-
ular, they show that a strategy which implies a path-dependent payoff is not consistent
with utility maximization in a Black/Scholes-type model.
Another strand of the literature analyzes the robustness properties of stylized
strategies. Concerning the robustness of option hedges, we refer the reader to
[2, 8, 23, 25, 29, 34, 35]. The properties of continuous-time CPPI strategies are also
studied extensively in the literature, cf. [13] or [14]. A comparison of OBPI and CPPI
(in continuous time) is given in [9]. [44] also compares OBPI and CPPI strategies. In
particular, they derive parameter conditions implying second- and third-order stochas-
tic dominance of the CPPI strategy. The literature also deals with the effects of jump
processes, stochastic volatility models and extreme value approaches on the CPPI
method, cf. [10, 11]. An analysis of gap risk, i.e., the risk that the guarantee is vio-
lated, is provided in [3] and [19]. Gap risk is implied by introducing jumps into the
model or by market frictions such as discrete time trading. In practice, the gap risk
was already observable during the 1987 crash. In addition, the crash is sometimes even
explained or seen to be supported by the portfolio protection mechanisms. However,
there are also contradicting opinions, cf. [32]. Finally, there is also a a wide strand
of empirical papers, which measure the performance of portfolio insurance strategies.
For example, we refer to [18] who give an extensive simulation comparison of popular
dynamic strategies of asset allocation.
The following paper mitigates between expected utility maximization and the
comparison of stylized strategies. We start with an exposition of the three optimization
problems which imply constant mix, CPPI, and OBPI strategies as optimal. Instead of
giving a further justification for the existence of guarantees, we use the (well-known)
results of the optimization problems to explain the main differences between portfolio
insurance mechanisms. Comparing the terminal payoffs shows that both portfolio
insurance strategies, CPPI and OBPI, result in payoffs, which consist of a fraction of
the payoff of a constant mix strategy (which is optimal for the unconstrained CRRA
investor) and an additional term due to the guarantee. The additional term provides an
intuitive way to explain the main advantage of the OBPI approach as compared to the
CPPI approach. Intuitively, it is clear that the fraction of wealth which is put into the
optimal unconstrained strategy is linked to the price of the guarantee, i.e., the fraction
is less than one. In the case of the CPPI approach, the additional term is simply the
guarantee itself, i.e., the payoff of an adequate number of zero bonds. In contrast, the
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch10 FA
additional term implied by the OBPI is a put option where the (synthetic) underlying is
given by a fraction of the constant mix strategy and the strike is equal to the guarantee.
Obviously, the put is cheaper than the zero bonds. Therefore, an investor who follows
the OBPI approach puts a larger fraction of his wealth into the unconstrained optimal
portfolio than an investor who follows the CPPI approach. To asses the utility costs of
forcing a guarantee into the unconstrained problem, i.e., the utility costs from having
to use a suboptimal strategy, we compare the certainty equivalents of the different
strategies and calculate the loss rates.
In addition, we explain one major drawback of the OBPI method, which is due to
the kink in the payoff-profile caused by the option component. The terminal value of
the OBPI is equal to the guarantee if the put expires in the money. In contrast to the
CPPI method, this implies a positive point mass for the event that the terminal value
is equal to the guarantee. The probability is given by the real world probability that
the terminal asset price is below the strike of the put. Intuitively, it is clear that this
can cause a high exposure to gap risk, i.e., the risk that the guarantee is violated, if
market frictions are introduced. We illustrate this effect by taking trading restrictions
and transaction costs into account. It turns out that the guarantee implied by the CPPI
method is relatively robust. In contrast, the probability that the guarantee is not reached
under the corresponding synthetic discrete-time OBPI strategy is rather high.
The outline of the paper is as follows. In Sec. 10.2, we review the well-known opti-
mization problems yielding constant mix, CPPI, and OBPI strategies as optimal solu-
tions. We compare the optimal strategies and resulting payoffs, and we discuss some
advantages (disadvantages) of the different portfolio insurance methods. In Sec. 10.3,
we consider the utility losses caused by the introduction of strictly positive terminal
guarantees for a CRRA investor. In particular, we compare CPPI and OBPI strategies
according to their implied loss rate. We consider the effects of market frictions in
Sec. 10.4 where we focus on the loss rates, which are implied by discrete-time trading
and transaction costs. In addition, we compare the effects of these market frictions on
the protection mechanisms of CPPI and OBPI. In Sec. 10.5, we address the topic of
borrowing constraints and consider the capped version of CPPI strategies. Section 10.6
concludes the chapter.
All stochastic processes are defined on a stochastic basis (, F, F, P), which satisfies
the usual hypotheses. We consider two assets. The riskless zero bond B with maturity
T grows at a constant interest rate r, i.e., dBt = Bt rdt where BT = 1. The evolution
of the risky asset S, a stock or benchmark index, is given by a geometric Brownian
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch10 FA
motion
dSt = St (µdt + σdWt ), S0 = s, (10.1)
where W = (Wt )0≤t≤T denotes a standard Brownian motion with respect to the real
world measure P. µ and σ are constants and we assume that µ > r ≥ 0 and σ > 0. In
particular,
dSt = St (rdt + σdWt∗ ), S0 = s, (10.2)
where W ∗ = (Wt∗ )0≤t≤T denotes a standard Brownian motion with respect to the
equivalent martingale measure Q, i.e.,
dQ 1 µ−r 2
= e− 2 ( σ ) t− σ Wt .
µ−r
(10.3)
dP t
A continuous-time investment strategy or saving plan for the interval [0, T ] can be
represented by a predictable process (πt )0≤t≤T . πt denotes the proportion of the port-
folio value at time t, which is invested in the risky asset S. In the following, we also
refer to πt as the portfolio weight at time t. W.l.o.g., we consider strategies which
are self-financing, i.e., money is neither injected nor withdrawn during the investment
horizon [0, T ]. Thus, the fraction of wealth which is invested at time t in the zero bond
B is given by 1 − πt . Let V = (Vt )0≤t≤T denote the portfolio value process associated
with the strategy π, then the dynamics of V are given by
dSt dBt
dVt (π) = Vt πt + (1 − πt ) where V0 = x. (10.4)
St Bt
For the above model assumptions, it follows
dVt (π) = Vt [(πt (µ − r) + r)dt + πt σdWt ] where V0 = x. (10.5)
Alternatively, the strategies can be represented by the number of shares. Let ϕt =
(ϕt,S , ϕt,B )0≤t≤T where ϕt,S denotes the number of risky assets and ϕt,B the number of
zero bonds with maturity T , which are held at time t. In particular, we have
πt Vt (1 − πt )Vt
ϕt,S = and ϕt,B = , (10.6)
St Bt
where Bt denotes the t-price of the zero bond maturing at T . Traditionally, a strategy
specification via the portfolio weights is used in the context of portfolio optimization
while the convention of stating the number of shares is normally preferred in the context
of hedging.
In the case of a finite investment horizon T and no intermediate consumption
possibilities, the relevant optimization problem is given by
sup EP [u(VT (π))] subject to Eq. (10.5), (10.7)
π∈
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch10 FA
where denotes the set of all self-financing trading strategies. The utility function u
(u ∈ C2 ) is assumed to be strictly increasing and concave, i.e., u > 0 and u < 0. In the
following, we recall the well-known optimization problems, which justify three basic
strategy classes: constant mix (CM) strategies, constant proportion portfolio insurance
(CPPI) strategies, and option-based portfolio insurance (OBPI) strategies. In contrast
to a CM strategy, which is exclusively specified by a constant portfolio weight m, i.e.,
πtCM = m, portfolio insurance strategies incorporate a guarantee component which,
in the simplest case, is given by an amount GT , which is to be honored at the end of
the investment horizon T (T > 0). While a CPPI strategy is value-based in the sense
that the portfolio weights are exclusively specified by the current portfolio value (and
the present value of the guarantee), the (dynamic) OBPI approach is payoff and model
dependent. Here, the investment decisions are, in a complete model, given in terms of
the delta hedge of an option payoff. Formally, the three strategy classes are
CM = {π ∈ |πt = m, m ≥ 0} (10.8)
Vt − e−r(T −t) GT
CPPIG
= π ∈ πt = m ,m ≥ 0 (10.9)
Vt
t St
OBPIG = π ∈ πt = ,
Vt
∂
t = EQ [e−r(T −t) (h(ST ) − GT )+ |Ft ] . (10.10)
∂St
Notice that for GT = 0, we have CM = CPPI G . However, we refer to CPPI and OBPI
versions, where the guarantee is not a strategy parameter but GT > 0 is exogenously
given. In particular, we assume that GT < V0 erT .
For the OBPI approach, h is a functional which has to reflect the initial budget
constraint. In practice, h is typically a linear function of ST and the resulting payoff
equals a protective put strategy, where an initial investment in risky assets is protected
by corresponding puts. In general, h can be an arbitrary function, e.g., a power function.
In [39], a general discussion of how to choose h optimally with respect to different
utility functions is given.
Table 10.1 summarizes the optimization problems, which are suited to justify
the three strategy classes. Problem (A) is derived from the classic Merton problem,
cf. [37]. While Merton additionally considers optimal consumption, the optimization
problem (A) is limited to the terminal wealth. Nevertheless, due to the separability
of the investment and the consumption decision, the optimal investment strategy is
equal. Problem (B) introduces a subsistence level GT such that uB belongs to the
class of HARA utility functions. Problem (C) consists of the CRRA utility function
uA which is also used in problem (A) but poses an additional constraint on the ter-
minal value of the strategy, i.e., the constraint that the terminal strategy value must
be above or equal to the terminal guarantee GT . The solutions of the optimization
problems and their corresponding proofs are well known in the literature such that
we omit some technical parts of the proofs and refer to the literature given in the
introduction.
Intuitively, it is clear that the solutions of problems (B) and (C) are modifications
of the classic Merton problem (A), where a guarantee is exogenously forced into the
optimization problem, respectively the solution. Basically, the subsistence level in (B)
results in the optimization problem of (A) if the value process is reduced by the present
value of the terminal guarantee, i.e., the optimization problem (A) is given in terms
of the cushion process. Technically, the solution of problem (C) is more involved.
However, the solution of (C) is intuitive in the sense that the constraint on the terminal
value features an European option on the optimal payoff of (A), where the initial
investment must take into account the price of the option.
EP [uA (Vt2 (π∗ ))] = EP [EP [uA (Vt2 (π∗ ))|Vt1 (π∗ )]]
= EP [EP [uA (Vt1 (π∗ ) · V̂ 1 (t1 , t2 ))]|Vt1 (π∗ )]
= EP [uA (Vt1 (π∗ ))]EP [(V̂ 1 (t1 , t2 ))1−γ ],
i.e., the optimal portfolio weight in t < t1 must be equal for the two investment
horizons. This implies that the solution of problem (A) can be obtained by restricting
the strategy set to constant mix (CM) strategies such that it is enough to consider the
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch10 FA
maximization problem
Inserting σWT = ln ST
S0
− (µ − 21 σ 2 )T gives
m
ST
V0CM e(m(µ−r)+r− 2 m σ )T −m(µ− 2 σ )T
1 2 2 1 2
VTCM =
S0
= φ(V0CM , m)STm (10.13)
where
m
1 1
e(1−m)(r+ 2 mσ )T .
2
φ(x, m) := x (10.14)
S0
Notice that, as a function of the terminal asset price ST , the payoff VTCM is concave for
m < 1, linear for m = 1 and convex for m > 1. The expected utility is equal to
and
CT = φ(C0 , m)STm ,
as well as
Thus, the terminal value of the strategy can be interpreted in terms of an option on the
payoff h such that the optimal strategy π∗ is given in terms of the delta hedge, i.e.,
π ∗ ∈ OBPI . In the special case that GT = 0, the optimization problem reduces to the
∗
classic Merton case, i.e., it holds h(ST ; GT = 0) = φ(V0 , m∗ )STm . For GT > 0, the
optimal solution affords a reduction of the initial investment from V0 to Ṽ0 (Ṽ0 < V0 )
∗
such that h(ST ) = φ(Ṽ0 , m∗ )STm . The remaining money V0 − Ṽ0 is then used to
buy the put with payoff [GT − h(ST )]+ . A proof for the optimality of the payoff
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch10 FA
∗
∗
VT,C = GT + [φ(Ṽ0 , m∗ )STm − GT ]+ is given in [24]. The concavity of uA (x) =
(x1−γ )/(1 − γ) implies that for any payoff V̂T with V̂T ≥ GT
∗
uA (V̂T ) − uA (VT,C ) ≤ uA (VT,C
∗ ∗
)(V̂T − VT,C )
∗ −γ −γm
∗
∗ m∗ GT
φ(Ṽ0 , m ) ST (V̂T − VT,C ) for ST >
φ(Ṽ0 , m∗ )
= ,
−γ GT
GT (V̂T − GT )
∗
for STm ≤
φ(Ṽ0 , m∗ )
i.e.,
∗ −γm∗
uA (V̂T ) − uA (VT,C ) ≤ φ(Ṽ0 , m∗ )−γ ST ∗
(V̂T − VT,C )
−γm∗ −γ
− (φ(Ṽ0 , m∗ )−γ ST − GT )+ (V̂T − GT ). (10.21)
∗
Now, consider a portfolio VT (ε) = εVT,A + (1 − ε)ṼT , where ṼT is the terminal
wealth of any other strategy with the same initial wealth. The first-order condition of
optimization problem (A) implies
∂EP [u(VT (ε))]
= EP [u (VT (ε))(VT,A
∗
− V˜T )]|ε=1
∂ε ε=1
−γm∗
= EP [φ(V0 , m∗ )−γ ST ∗
(VT,A − ṼT )] = 0.
∗
Therefore, adding and subtracting the optimal unconstrained solution VT,A to the first
term on the right hand side of inequality (10.21) and taking expectations gives
−γm∗ −γm∗
φ(Ṽ0 , m∗ )−γ (EP [ST ∗
(V̂T − VT,A )] + EP [ST ∗
(VT,A ∗
− VT,C )]) = 0.
∗ ∗
Together with V̂T ≥ GT a.s. it follows EP [uA (V̂T ) − uA (VT,C )] ≤ 0 such that VT,C is
indeed the optimal solution w.r.t. problem (C).
∗
∗
The payoff VT,C = GT + [φ(Ṽ0 , m∗ )STm − GT ]+ can be replicated by a self-
financing strategy where the initial investment can be represented by the expected
discounted payoff under the uniquely defined equivalent martingale measure Q, i.e.,
∗
V0 = e−rT EQ [GT + (φ(Ṽ0 , m∗ )STm − GT )+ ]
+
∗ GT
= e GT + φ(Ṽ0 , m )EQ e−rT STm −
−rT ∗
.
φ(Ṽ0 , m∗ )
Ṽ0 has to be determined such that the initial cushion C0 = V0 −e−rT GT exactly finances
φ(Ṽ0 , m∗ ) power call options with power p = m∗ and strike K = (GT )/(φ(Ṽ0 , m∗ )),
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch10 FA
i.e.,
G
V0 − e−rT GT = φ(Ṽ0 , m∗ )PO 0, S0 ; m∗ , (10.22)
φ(Ṽ0 , m∗ )
where PO(t, St ; p, K) denotes the t-price of a power call with power p, strike K, and
maturity T . The pricing formula for power calls, cf. [45] or [36], is given by
2 p
− (1−p)σ (T −t)
−r(T −t) St e
2
= e
e−r(T −t)
St √
× N h1 t, √
p
− (1 − p)σ T − t
K
St
− KN h2 t, √ p
. (10.23)
K
ln z + (r + 12 σ 2 )(T − t) √
h1 (t, z) = √ ; h2 (t, z) = h1 (t, z) − σ T − t. (10.24)
σ T −t
∗
Recall that Ṽ0 ≤ V0 . A lower bound on Ṽ0 follows with [STm − GT /φ(Ṽ0 , m∗ )]+ ≤
∗
STm , i.e.,
G ∗
φ(Ṽ0 , m∗ )PO 0, S0 ; m∗ , ≤ φ(Ṽ0 , m∗ )EQ [e−rT STm ] = Ṽ0
φ(Ṽ0 , m∗ )
∗
Consider now the optimal portfolio weight πt,C of the dynamic strategy, i.e.,
φ(Ṽ0 , m∗ )PO t, St ; m∗ , φ(ṼG,m
T
∗)
St
∗
πt,C = 0
Vt
φ(Ṽ0 , m∗ )PO t, St ; m∗ , φ(ṼG,m
T
∗)
St
= 0
, (10.26)
e−r(T −t) GT + φ(Ṽ0 , m∗ )PO t, St ; m∗ , φ(ṼG,m∗ )
0
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch10 FA
σ2
p−1
St e 2 p(T −t) St √
PO
(t, St ; p, K) = p N h1 t, √ − (1 − p)σ T − t .
e−r(T −t) p
K
(10.27)
Finally, we compare the optimal investment weight implied by problem (C) with the
one of the unconstrained solution given by m∗ . Notice that
+
∗ GT ∗ GT
STm − ≥ STm −
φ(Ṽ0 , m∗ ) φ(Ṽ0 , m∗ )
implies that the denominator of the right hand side of Eq. (10.26) is larger than
∗ ∗ ∗
−1)(r+ 12 m∗ σ 2 )(T −t)
φ(Ṽ0 , m∗ )EQ [e−r(T −t) STm |Ft ] = φ(Ṽ0 , m∗ )Stm e(m
such that
∗ GT ∗ ∗ 1 ∗ 2
πt,C ≤ PO t, St ; m∗ , St(1−m ) e(1−m )(r+ 2 m σ )(T −t) .
φ(Ṽ0 , m∗ )
m∗ −1
∗ GT ∗ St ∗
(1−m∗ )σ 2 (T −t)
e− 2 m
1
PO
t, St ; m , ≤m .
φ(Ṽ0 , m∗ ) e−r(T −t)
Together, we have
∗
πt,C ≤ m∗ . (10.28)
Analogously to the problem (B), the terminal constraint in problem (C) also gives rise
to a reduction of the optimal unconstrained portfolio weight m∗ .
the optimization problems (A), (B), and (C) are summarized as follows
∗
∗
VT,A = φ(V0 , m∗ )STm (10.29)
∗
∗
VT,B = GT + φ(V0 − e−rT GT , m∗ )STm
V0 − e−rT GT ∗
= GT + VT,A (10.30)
V0
∗
∗ +
VT,C = GT + φ(Ṽ0 , m∗ )STm − GT (10.31)
∗ ∗
= φ(Ṽ0 , m∗ )STm + [GT − φ(Ṽ0 , m∗ )STm ]+
+
Ṽ0 ∗ Ṽ0 ∗
= V + GT − VT,A . (10.32)
V0 T,A V0
∗
The payoff VT,A corresponds to φ(V0 , m∗ ) power claims with power m∗ , where the
∗
number φ(V0 , m ) depends on the initial investment and the optimal investment weight
m∗ . The optimization problem (B) introduces a subsistence level which implies that
the number of power claims with power m∗ must be reduced to afford the risk-free
investment, which is necessary to honor the guarantee. In consequence, the portfolio
weight is lower than in the case of problem (A), cf. Inequality (10.19). The link between
the solutions of (A) and (B) is even more explicit if one considers the number of shares
in the asset which are held. Notice that the cushion dynamics, cf. Eq. (10.17) implies
that CtCPPI /C0CPPI = VtCM /V0 if the multiplier m of the CPPI is equal to the portfolio
weight of the CM strategy. In particular, this implies that the value of the cushion is
proportional to the value of the CM strategy, i.e., CtCPPI = (C0 /V0 )VtCM . This is also
true for the number of assets ϕt,S , i.e.,
C0 CM
CPPI
ϕt,S = ϕ (10.33)
V0 t,S
C0 CM
CPPI
ϕt,B = GT + ϕ , (10.34)
V0 t,B
where ϕt,B denotes the number of zero bonds with maturity T . In particular, it holds that
C0 CM
VtCPPI = e−r(T −t) GT + V . (10.35)
V0 t
The CPPI strategy can thus be interpreted as a buy-and-hold strategy of a constant
mix strategy with an additional investment into GT zero bonds, cf. also Eq. (10.31). A
similar reasoning applies to the solution of (C), which can be interpreted as a buy-and-
hold strategy of a constant mix strategy with an additional investment into a put with
strike GT , cf. Eq. (10.32). Obviously, the put is worth less than GT zero bonds such
that one can buy and hold more CM strategies in the case of the option based approach,
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch10 FA
i.e., Ṽ0 /V0 ≥ C0 /V0 . Intuitively, it is thus clear that the OBPI approach gives a better
result than the CPPI approach with respect to a utility function, which favors the CM
strategy with portfolio weight m∗ .
∗
In general, a modification of the payoff VT,A which honors the guarantee GT and
with t0 -price equal to V0 can be represented by
+
+
∗ G G
subject to EQ e−rT GT + α VT,A ∗
T T
GT + α VT,A −β −β = V0 .
α α
(10.36)
such that
1
sA,B = S0 e(r+ 2 (m−1)σ )T .
2
(10.37)
S0 = 1 V0 = 1
σ = 0.15 T = 10 GT = 1
r = 0.03 γ = 1.2
µ = 0.085 m = m∗ = 2.037
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch10 FA
3
Payoff
0
0.0 0.5 1.0 1.5 2.0 2.5 3.0
Terminal asset price
Figure 10.1 Optimal payoffs VT,A ∗ (solid line), V ∗ (dotted line), and V ∗ (dashed line),
T,B T,C
where the parameters are given as in Table 10.2.
We end this section by emphasizing one important consequence for the two pro-
tection mechanisms implied by the smooth and the kinked solutions, i.e., implied by
the assumptions that marginal utility jumps gradually or discontinuously to infinity.
∗
Notice that the smooth payoff VT,B implies that there is no probability mass on the
∗
event that the terminal value is equal to the guarantee GT , i.e., P(VT,B = GT ) = 0. In
∗
contrast, for the kinked payoff VT,C , it holds
∗ m∗ GT
P(VT,C = GT ) = P ST ≤
φ(Ṽ0 , m∗ )
m∗ GT
ST φ(Ṽ0 ,m∗ )
= P ln ≤ ln . (10.40)
S0 S0
where λ := (µ − r)/σ.
Intuitively, it is clear that a positive point mass on the event {VT = GT } might
indicate that the corresponding strategy is more sensitive to the introduction of gap
risk, which is caused by asset price jumps. This problem is considered in Sec. 10.4,
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch10 FA
where trading restrictions in the sense of discrete time trading and transaction costs
are introduced.
Consider for example the utility function uA (x) = x1−γ /(1 − γ) and a CM strategy
with optimal investment proportion m∗ , i.e.,
(1−γ)m∗ 1
CE T,A (πCM ) = φ(V0 , m∗ )(EP [ST ]) 1−γ . (10.43)
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch10 FA
p p
With EP [ST ] = S0 exp{(pµ − 12 p(1 − p)σ 2 )T } it follows
∗ ∗
µ− 12 γσ 2 m∗ (1−m∗ (1−γ)))T
CET,A (πCM ) = φ(V0 , m∗ )S0m e(m
∗
(µ−r)− 12 γ(m∗ )2 σ 2 )T
= V0 e(r+m
1 µ−r 2
= V0 erT e 2γ ( σ
) T
. (10.44)
Thus, the certainty equivalent is lower for an investor with subsistence level than an
investor without. We can also calculate the utility loss of an investor, who follows a
suboptimal strategy. It is described by the loss rate lT,i (π) of the strategy π and the
utility function i (i ∈ {A, B, C}) where
∗
CET,i
ln CET,i (π)
lT,i (π) := . (10.46)
T
CE ∗T,i denotes the certainty equivalent of the optimal strategy πi∗ = (πt,i ∗
)0≤t≤T while
CE T,i (π) the one of the suboptimal strategy π = (πt )0≤t≤T .
The loss rates with respect to the utility function uA (x) = x1−γ /(1 − γ) are
summarized in Table 10.3 for CM, CPPI, and OBPI strategies with strategy parameter
m. Notice that a loss rate, which is higher than the one of a risk-free investment, i.e.,
πtCM = 0, implies that the associated strategy is prohibitively bad. This critical loss
rate is equal to 12 γ(σm∗ )2 , cf. Table 10.3. For m = m∗ , it is obvious that lT,A (πOBPI ) <
lT,A (πCPPI ) because πCPPI is suboptimal w.r.t. the optimization problem (C) where
0.06 0.06
0.05 0.05
0.04 0.04
Lossrate
Lossrate
0.03 0.03
0.02 0.02
0.01 0.01
0.00 0.00
2.04 4 6 2.04 4 6
m m
Figure 10.2 Loss rates w.r.t. u = uA for CPPI (solid lines), OBPI (dashed) and CM (dotted)
strategies with varying parameter m. The parameters are given in Table 10.2. The investment
horizon is T = 10 years (left graph) and T = 20 years (right graph).
The utility loss implied by the guarantee is higher for the CPPI than for of the OBPI.
This is illustrated in Fig. 10.2. In addition, notice that the loss rates of the CM strategies
are symmetric in the sense that for > 0, a strategy parameter m = m∗ + implies
the same loss rate as the parameter m = m∗ − . In contrast, OBPI strategies yield
a lower loss rates in the case of m = m∗ + than for m = m∗ − . Intuitively, this
is clear since the protection feature implies that the portfolio weights of the portfolio
insurance strategies are too low compared to the optimal investment proportion m∗ ,
cf. Inequalities (10.19) and (10.28).
For the comparison of CPPI and OBPI, it is important to keep in mind that m∗
is the optimal OBPI parameter for u = uA . In contrast, m∗ is not the optimal CPPI
parameter for u = uA , but for u = uB which includes a subsistence level. In order to
compare the loss rates implied by CPPI and OBPI w.r.t. u = uA , it is thus necessary
to consider the maximization problem
(GT + φ(V0 − e−rT GT , m)STm )1−γ
max EP [uA (VT (π))] = max EP .
π∈CPPI m 1−γ
(10.47)
Figure 10.3 illustrates the loss rates for CPPI and OBPI strategies with varying param-
eter m. In addition, the minimal loss rates are summarized in Table 10.4, i.e., the loss
rates for OBPI strategies with m = m∗ and CPPI for m = m∗∗ , where m∗∗ is the opti-
mal solution for the maximization problem (10.47). Although the OBPI strategy with
parameter m = m∗ is, per construction, the uA -utility maximizing portfolio insurance
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch10 FA
0.08 0.08
0.06 0.06
Lossrate
Lossrate
0.04 0.04
0.02 0.02
0.00 0.00
2.04 4 6 8 10 12 14 16 18 1.63 4 6 8 10 12 14 16 18
m m
Figure 10.3 Loss rates implied by CPPI (solid lines) and OBPI (dashed lines) for varying m,
cf. Table 10.3. The risk aversion is γ = 1.2 (left figure) and γ = 1.5 (right figure). The other
parameters are given in Table 10.2.
Strategy γ\T 1 2 5 10 20
strategy, the additional loss of the CPPI strategy with parameter m = m∗∗ which is
measured by the difference of its loss rate and the one of the OBPI is rather low. Notice
that the differences remain approximatively equal for varying time horizons and risk
aversion parameters.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch10 FA
where K(m) := (m − 1)er n /m and ts := min{tk ∈ τ|Vtk < e−r(T −tk ) GT }. We set
T
where Vtτ,TAk+
(Ctτ,TA
k+
:= Vtτ,TA
k+
− e−r(T −tk ) GT ) denotes the portfolio (cushion) value
immediately after tk , i.e., the value net of transaction costs, which are proportional to
the asset price Stk . First, consider the portfolio value Vtτ,TA k
before transaction costs,
i.e., Vtτ,TA
0
:= V t0 and for k = 1, . . . , n
Vtτ,TA
k
:= ϕtτ,TA
k ,S
Stk + ϕtτ,TA
k ,B
Btk
!
mCtτ,TA
= max k−1+
, 0 Stk + (Vtτ,TA
k−1+
− ϕtτ,TA
k ,S
Stk−1 )er(tk −tk−1 )
Stk−1
Stk
= m max{Ctτ,TA
k−1+
, 0} − er(tk −tk−1 ) + Vtτ,TA
k−1+
er(tk −tk−1 ) . (10.50)
Stk−1
Consider now the adjustment to the proportional transaction costs, which are due
immediately after the trading dates. Assuming that the transaction costs are also due
at t0 and that the asset positions are transferred into a cash position immediately after
tn is consistent to the following definitions
Vtτ,TA
0+
:= Vt0 − ϕS,t
τ,TA
0+
θSt0 = Vt0 − mθCtτ,TA
0+
(10.51)
Vtτ,TA
k+
:= Vtτ,TA
k
− |ϕtτ,TA
k+ ,S
− ϕtτ,TA
k ,S
|θStk
St
= Vtτ,TA
k
− mθ max{Ctτ,TA
k+
, 0} − max{Ctτ,TA
k−1+
, 0} k
(10.52)
S tk−1
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch10 FA
Vtτ,TA
n+
:= Vtτ,TA
n
− ϕtτ,TA
n ,S
θStn
Stn
= Vtτ,TA
n
− mθ max{Ctτ,TA
n−1+
, 0} . (10.53)
Stn−1
With Eq. (10.51), it immediately follows Ct0 + = Ct0 /(1 + θm). Using Eqs. (10.52)
and (10.50) implies that Ctk + > 0 (k = 0, . . . , n − 1) and θ < 1/m, it holds
Stk+1
Ctk+1 + = Ctk+1
− mθ max{Ctk+1 + , 0} − Ctk + (10.54)
S tk
1 + θ Stk+1 m − 1 rT St
T
Ctk + m − e n for er n ≤ k+1
1 + θm S tk 1 + θm Stk
1 − θ Stk+1 m − 1 rT
= Ctk + m − e n otherwise
1 − θm Stk 1 − θm
S m − 1 rT St
Ctk + (1 − θ)m tk+1 − (m − 1)er n
T
for e n > k+1 .
Stk m(1 − θ) Stk
(10.55)
T
For Ctk + ≤ 0, it follows Ctk+1 + = Ctk+1 = er n Ctk + . It is worth mentioning that the
event {Vtτ,TA
n+
< GT } corresponds to the event that the adjusted cushion drops below
zero during the investment horizon. Notice that for Ctk + > 0
Stk+1 m − 1 rT
{Ctk+1 + < 0} ⇔ < e n =: Ak+1 .
Stk m(1 − θ)
Since the complementary of the event {∪n−1k=0 Ak+1 } is given by the event that all asset
T
price increments are above er n (m − 1)/(m(1 − θ)), it follows with the assumption that
the asset price increments are independent and identically distributed that
n
St1 m − 1 rT
P(Vtτ,TA < GT ) = 1 − P > e n
n+
St0 m(1 − θ)
where
ln (1−θ)m
+ (µ − r) Tn − 21 σ 2 Tn
d2TA (θ) := m−1
. (10.57)
σ Tn
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch10 FA
Setting V0 (ϕ; τ) := V0 (ϕ), the value process V(ϕ; τ) which is associated with ϕτ is
ξtdis
k+1
(ϕ; τ) := Vtk+1 (ϕ) − Vtk+1 (ϕ; τ) (10.60)
= (ϕtk+1 ,S − ϕtk ,S )Stk+1 + (ϕtk+1 ,B − ϕtk ,B )e−r(T −tk+1 ) .
Notice that negative costs refer to inflows while positive costs imply that further money
is needed to continue the strategy.
Taking proportional transaction costs into account also implies that there are trans-
action costs ξtTA
k+1
(ϕ; τ), which occur immediately after a trading date tk+1 , i.e.,
ξtTA
0
(ϕ; τ) := |ϕt0 ,S |θSt0
ξtTA
k+1
(ϕ; τ) = |ϕtk+1 ,S − ϕtk ,S |θStk+1 for k = 0, . . . , n − 2
and
ξtTA
n
(ϕ; τ) := |ϕtn−1 ,S |θStn .
Defining the payoff V̄T (ϕ; τ) according to the assumption that the inflows into the
strategy are lent according to the interest rate r and outflows are saved according to r
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch10 FA
gives
"
n−1
V̄T (ϕ; τ) := Vtn (ϕ) − ξtTA
0
er(T −t0 ) + (ξtdis
k+1
(ϕ; τ) + ξtTA
k+1
(ϕ; τ))er(tn −tk+1 ) .
k=0
(10.61)
0.010
Shortfall probability
0.008
Loss rate
0.006
0.004
0.002
0.000
3 4 5 6 7
m m
0.8
0.10
Shortfall probability
0.08 0.6
Loss rate
0.06
0.4
0.04
0.2
0.02
0.0
0.00
2.03704 4 6 2.04 4 6
m m
Figure 10.5 Loss rates w.r.t. u = uA (shortfall probabilities) in the Merton setup implied
by continuous-time OBPI (solid line), monthly OBPI without transaction costs (dashed lines)
and monthly OBPI with θ = 0.01 (dotted line) for varying m. The parameter setup is given in
Table 10.2.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch10 FA
gap risk, i.e., a strictly positive shortfall probability, gives a loss of minus infinity in the
case of u = uB and u = uC . Observe, that for both strategies, OBPI and CPPI, the loss
which is in the first instance caused by time-discretizing the strategies is rather low.
However, there is a huge impact caused by transaction costs. The effect is even more
pronounced for the OBPI than for the CPPI method. Intuitively, it is clear that in the
limit to continuous-time trading, the transaction costs eat up the cushion, which allows
a risky investment. The convergence of the value (cushion) process of the CPPI is for
example analyzed in [3]. Recall that loss rates above the one of a risk-free investment
are prohibitive. In the case of the basic parameter constellation, this critical value w.r.t.
u = uA is equal to 0.5γ(σm)2 = 0.056. In particular, if the strategy parameter m is
not chosen in a cautious way, portfolio insurance can get prohibitively bad because of
transaction costs.
Consider now the shortfall probability. Recall that, in contrast to the continuous-
time CPPI strategy, the continuous-time OBPI results in a strictly positive probability
that the payoff is not above the terminal guarantee. Thus, it is to be expected that
the shortfall probability of the OBPI is more sensitive to discrete-time trading and
transaction costs than the one of the CPPI, cf. the right graphs presented in Figs. 10.4
and 10.5. This effect is also illustrated in Fig. 10.6, where the distribution function of
the discrete-time versions of OBPI with m = m∗ and CPPI with m = m∗∗ is plotted.
However, to some extend the sensitivity of the OBPI to the gap risk measured by the
shortfall probability is also caused by the dicretization scheme. While the transaction
costs are financed via the cushion for the CPPI, this is not true in the case of the
discrete-time OBPI.
1.0 1.0
Distribution function
Distribution function
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0.0 0.0
1 2 4 6 8 1 2 4 6 8
Figure 10.6 Distribution functions of terminal values for OBPI (left) and CPPI (right) with
transaction costs (dotted line) and without transaction costs (solid line).
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch10 FA
where CtCCP := VtCCP − e−r(T −t) GT and w (w ≥ 1) denotes the restriction on the
investment proportion. In the following, we refer to borrowing constraints in the strict
sense, i.e., we assume that ω is set equal to one.
It is worth mentioning that the borrowing constraints introduce a path-dependence
and the payoff implied by the capped CPPI version cannot be stated as a function of the
terminal asset price as it is the case without borrowing constraints. Intuitively, it is to be
expected that the path dependence yields an additional utility loss in a Black/Scholes
type model setup. In order to calculate the loss rate, we consider the distribution of the
terminal value of the CCP strategy. Let f denote the density function of the terminal
value of the capped CPPI. Then it holds
# ∞
E[uA (πCCP )] = uA (v)f(v)dv
GT
# ∞ 1−γ
1
CE T,A (π CCP
)= v 1−γ
f(v)dv
GT
such that
( r+m+ (µ−r)− 21 γ(m∗ σ)2 )T
1 V0 e
lT,A (πCCP ) = ln 1−γ .
T $∞ 1
v1−γ f(v)dv
GT
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch10 FA
For details on the distribution of the capped CPPI, we refer to [4]. Basically, the
distribution can be obtained by considering the process (X)0≤t≤T , which is given by
the dynamics
where
1 (m − 1)V
µ−r 1 0
σ − 2 mσ x≤0
σ ln mG mC0 ≥ V0
0
(x) = and X0 = .
µ − r
1 1
(m − 1)C0
− σ x>0 ln mC0 < V0
σ 2 mσ G0
Along the lines of [4], one can show that the value process (VtCCP )0≤t≤T and cushion
process CtCCP 0≤t≤T are, for ω = 1, given by
m σXt
e Xt ≥ 0
VtCCP = Gt m − 1 (10.63)
1 + 1 eσmXt
Xt < 0
m−1
and
m σXt
m−1 e − 1 Xt ≥ 0
CtCCP = Gt .
1
emσXt Xt < 0
m−1
In particular, it holds
(m−1)v
1 ln mGt m
p dv v≥ Gt
σv σ m−1
P[VtCCP ∈ dv] = (m − 1)(v − Gt )
1 ln
Gt dv m
p v< Gt
σm(v − Gt )
σm m−1
0.08 0.08
0.06 0.06
Loss rate
Loss rate
0.04 0.04
0.02 0.02
0.00 0.00
2.04 4 6 8 10 12 2.04 4 6 8 10 12
m m
Figure 10.7 Loss rates w.r.t. u = uA implied by CPPI (solid lines) and capped CPPI (dashed
lines) for varying m. The risk aversion is γ = 1.2 (left figure) and γ = 1.5 (right figure), the
other parameters are as in Table 10.2. The constant line gives the loss rate of a stop-loss strategy.
10.6. CONCLUSION
also a benchmark for OBPI and CPPI strategies. The CPPI results from a subsistence
level, i.e., the utility is measured in terms of the difference of portfolio value and
guarantee instead of the portfolio value itself. In contrast, the OBPI results from the
additional constraint that the terminal payoff is above the guarantee or floor. Consider-
ing the associated payoffs, both approaches result in payoffs which consist of a fraction
of the payoff of the constant mix strategy and an additional term stemming from the
guarantee. Intuitively, it is clear that the fraction is linked to the price of the guarantee,
i.e., the fraction is less than one. The main difference between OBPI and CPPI can
easily be explained by the additional term. In the case of the CPPI approach, the addi-
tional term is simply the guarantee itself, i.e., the payoff of the adequate number of
zero bonds. In contrast, the additional term implied by the OBPI is based on a put on
the fraction of the constant mix payoff with strike equal to the guarantee. Obviously,
the put is cheaper than the zero bond itself. Thus, the OBPI fraction which is held of
the optimal unrestricted payoff is higher than in the case of the CPPI method. This is a
major advantage in terms of the associated utility costs, i.e., the loss in expected utility
which is caused by the introduction of a strictly positive guarantee. The utility costs are
measured and illustrated in terms of a loss rate linking the certainty equivalents of the
strict portfolio insurance strategies to the certainty equivalent of the optimal solution.
One major drawback of the OBPI method is due to its kinked payoff-profile. The
terminal value of the OBPI is equal to the guarantee if the put expires in the money. In
contrast to the CPPI method, this implies a positive point mass that the terminal value is
equal to the guarantee. This relevant probability is given by the real world probability
that the terminal asset prices are below the strike of the put. Intuitively, it is clear that
this can cause a high exposure to gap risk, i.e., the risk that the guarantee is violated, if
market frictions are introduced. We illustrate this effect by taking trading restrictions
and transaction costs into account. It turns out that the guarantee implied by the CPPI
method is relatively robust. However, the probability that the guarantee is not reached
under the corresponding synthetic discrete-time OBPI strategy is rather high.
Finally, we tackle the question of borrowing constraints. In a strict sense, borrow-
ing constraints imply that the proportion of asset must not be above one. In the case
of the CPPI method, the capped CPPI which simply states that the asset proportion is
adequately capped according to the borrowing constraint is of high practical relevance.
We give the distribution of the capped CPPI and illustrate the corresponding loss rate,
i.e., the loss which is due to borrowing constraints.
References
[1] Ahn, D-H, J Boudoukh, M Richardson and RF Whitelaw (1999). Optimal risk management
using options. Journal of Finance, 54(1), 359–375.
[2] Avellaneda, M, A Levy and A Parás (1995). Pricing and hedging derivative securities in
markets with uncertain volatilities. Applied Mathematical Finance, 2(2), 73–88.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch10 FA
[3] Balder, S, M Brandl andA Mahayni (2009). Effectiveness of CPPI strategies under discrete-
time trading. The Journal of Economic Dynamics and Control, 33, 204–220.
[4] Balder, S and A Mahayni (2008). Cash-lock comparison of portfolio insurance strategies.
Technical Report.
[5] Basak, S (1995). A general equilibrium model of portfolio insurance. Review of Financial
Studies, 8(4), 1059–1090.
[6] Basak, S (2002). A comparative study of portfolio insurance. The Journal of Economic
Dynamics and Control, 26(7–8), 1217–1241.
[7] Benninga, S and M Blume (1985). On the optimality of portfolio insurance. Journal of
Finance, 40(5), 1341–1352.
[8] Bergman, YZ, BD Grundy and Z Wiener (1996). General properties of option prices.
Journal of Finance, 51, 1573–1610.
[9] Bertrand, P and J-L Prigent (2002). Portfolio insurance strategies: OBPI versus CPPI.
Technical Report, GREQAM and Université Montpellier1, pp. 68–86.
[10] Bertrand, P and J-L Prigent (2002). Portfolio insurance: The extreme value approach to
the CPPI. Finance, 23(2).
[11] Bertrand, P and J-L Prigent (2003). Portfolio insurance strategies: A comparison of stan-
dard methods when the volatility of the stock is stochastic. International Journal of Busi-
ness, 8(4), 15–31.
[12] Black, F and R Jones (1987). Simplifying portfolio insurance. The Journal of Portfolio
Management, 14, 48–51.
[13] Black, F and AR Perold (1992). Theory of constant proportion portfolio insurance. The
Journal of Economic Dynamics and Control, 16(3–4), 403–426.
[14] Bookstaber, R and JA Langsam (2000). Portfolio insurance trading rules. The Journal of
Futures Markets, 8, 15–31.
[15] Brennan, MJ and ES Schwartz (1976). The pricing of equity-linked life insurance policies
with an asset value guarantee. Journal of Financial Economics, 3, 195–213.
[16] Brennan, MJ and ES Schwartz (1989). Portfolio insurance and financial market equilib-
rium. Journal of Business, 62, 455–472.
[17] Browne, S (1999). Beating a moving target: Optimal portfolio strategies for outperforming
a stochastic benchmark. Finance and Stochastics, 3, 275–294.
[18] Cesari, R and D Cremonini (2003). Benchmarking, portfolio insurance and technical anal-
ysis: A monte carlo comparison of dynamic strategies of asset allocation. Journal of Eco-
nomic Dynamics and Control, 27, 987–1011.
[19] Cont, R and P Tankov (2007). Constant proportion portfolio insurance in presence of
jumps in asset prices. Financial Engigeering No. 2007–10, Columbia University Center
for Financial Engigeering.
[20] Cox, CJ and HE Leland (2000). On dynamic investment strategies. The Journal of Eco-
nomic Dynamics and Control, 24, 1859–1880.
[21] Cox, JC and C-F Huang (1989). Optimal consumption and portfolio policies when the
asset price follows a diffusion process. Journal of Economic Theory, 49, 33–83.
[22] Døskeland, TM and HA Nordahl (2008). Optimal pension insurance design. Journal of
Banking and Finance, 32, 382–392.
[23] Dudenhausen, A, E Schlögl and L Schlögl (1998). Robustness of Gaussian hedges under
parameter and model misspecification. Technical Report, University of Bonn, Department
of Statistics.
[24] El Karoui, N, M Jeanblanc and V Lacoste (2005). Optimal portfolio management with
american capital guarantee. Journal of Economic Dynamics and Control, 29, 449–468.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch10 FA
[25] El Karoui, N, M Jeanblanc-Picqué and SE Shreve (1998). Robustness of the Black and
Scholes formula. Mathematical Finance, 8(2), 93–126.
[26] Grossman, SJ and J Villa (1989). Portfolio insurance in complete markets: A note. Journal
of Business, 62, 473–476.
[27] Grossman, SJ and J Zhou (1993). Optimal investment strategies for controlling drawdowns.
Mathematical Finance, 3, 241–276.
[28] Grossman, SJ and J Zhou (1996). Equilibrium analysis of portfolio insurance. Journal of
Finance, 51, 1379–1403.
[29] Hobson, D (1998). Volatility misspecification, option pricing and superreplication via
coupling. Annals of Applied Probability, 8(1), 193–205.
[30] Kahnemann, D and A Tversky (1979). Prospect theory: An analysis of decision under risk.
Econometrica, 47, 263–291.
[31] Leland, HE (1980). Who should buy portfolio insurance. Journal of Finance, 35(2),
581–594.
[32] Leland, HE (1988). Portfolio insurance and october 19th. California Management Review,
Summer, 80–89.
[33] Leland, HE and M Rubinstein (1976). The Evolution of Portfolio Insurance. In: Portfolio
Insurance: A guide to Dynamic Hedging, DL Luskin (ed.), pp. 3–10, Wiley.
[34] Lyons, TJ (1995). Uncertain volatility and the risk-free synthesis of derivatives. Applied
Mathematical Finance, 2, 117–133.
[35] Mahayni, A (2003). Effectiveness of hedging strategies under model misspecification
and trading restrictions. International Journal of Theoretical and Applied Finance, 6(5),
521–552.
[36] Mahayni, A and E Schlögl (2008). The risk management of minimum return guarantees.
Business Research, 1, 55–76.
[37] Merton, RC (1971). Optimal consumption and portfolio rules in a continuous time model.
Journal of Economic Theory, 3, 373–413.
[38] Munk, C (2006). Financial asset pricing theory. Lecture notes.
[39] Prigent, J-L and F Tahar (2006). Optimal portfolios with guarantee at maturity: Computa-
tion and comparison. International Journal of Business, 11(2), 171–185.
[40] Tepla, L (2000). Optimal portfolio policies with borrowing and shortsale constraints. The
Journal of Economic Dynamics and Control, 24, 1623–1639.
[41] Tepla, L (2001). Optimal investment with minimum performance constraints. The Journal
of Economic Dynamics and Control, 25(10), 1629–1645.
[42] Tversky, A and D Kahnemann (1992). Advances in prospect theory: Cumulative represen-
tation of uncertainty. Journal of Risk and Uncertainty, 5, 297–323.
[43] von Neumann, J and O Morgenstern (1944). Theory of Games and Economic Behavior,
1953 edn. Princeton, NJ: Princeton University Press.
[44] Zagst, R and J Kraus (2008). Stochastic dominance of portfolio insurance strategies —
OBPI versus CPPI. Technical Report.
[45] Zhang, PG (1998). Exotic Options — A Guide to Second Generation Options. World
Scientific.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch10 FA
11
PORTFOLIO INSURANCES, CPPI AND
CPDO, TRUTH OR ILLUSION?
Constant proportion portfolio insurance (CPPI) and constant proportion debt obligations
(CPDO) strategies have recently created derivative instruments, which try to protect a port-
folio against failure events and have only been adopted in the credit market for the last couple of
years. Since their introduction, CPPI strategies have been popular because they provide protec-
tion while at the same time they offer high yields. CPDOs were only introduced into the market
in 2006 and can be considered as a variation of the CPPI with as main difference the fact that
CPDOs do not provide principal protection. Both CPPI and CPDO strategies take investment
positions in a risk-free bond and a risky portfolio (often one or more credit default swaps). At
each step, the portfolio is rebalanced and the level of risk taken will depend on the distance
between the current value of the portfolio and the necessary amount needed to fulfill all the
future obligations.
We first analyze in detail the dynamics of both investment strategies and afterwards test the
safetyness of both products under a multivariate Lévy setting. More precise we first propose
a quick way to calibrate a multivariate Variance Gamma (VG) process on correlated spreads,
which can then be used to quantify the gap risk for CPPIs and CPDOs.
259
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch11 FA
11.1. INTRODUCTION
Financial institutions try to protect their portfolios against failure events and deriva-
tive instruments are a possible solution. Derivative instruments are a fast-growing
market in which alternative investment strategies such as Constant Proportion Portfo-
lio Insurance (CPPI) and Constant Proportion Debt Obligation (CPDO) are created.
Although these two recently developed instruments function in different ways when
deciding on their investment strategy, both investment funds attempt to provide a
portfolio insurance. More precisely, their strategy is to invest only a part of the cap-
ital in a risky asset and to invest the remainder in a safe way. The total value of
the portfolio at each time step will influence the position taken in the risky asset.
The decisions taken on the risk position at each time step aim to allow the investor
of a CPPI or CPDO to recover, at maturity, a given percentage of their initial cap-
ital, which allows them to benefit from a capital guarantee while participating in
the upside of an underlying asset. It could happen that the promised return is not
achieved. In this case, for the CPPI structure, the bank will have to cover the losses
at maturity while for the CPDO structure, the CPDO will unwind and the investor
will not receive the promised amount at maturity but only the remainder amount at
the time of unwinding. CPPI structures clearly safeguard a given percentage of the
invested capital for the investor, while for CPDO the investor appears to be taking
a risk. In the past, this risk has always been seen as very small and CPDOs have
been sold as very safe. Here, we will study those two products in depth in order to
answer the question whether CPDO and CPPI are really as safe and attractive as they
seem.
First, the concept of credit risk will be introduced. This is the risk that, after
agreeing on a certain contract, one of the involved parties will not fulfil its financial
obligations (such as paying a premium). Often the quality and price of financial prod-
ucts will heavily depend on this risk. Different ways to model this risk are presented and
can be used to price financial products. Next, a credit default swap (CDS) is introduced.
This financial instrument tries to provide a protection for credit risk by transferring it.
In exchange for a predefined cost, also called “spread”, a second party will cover the
losses one might suffer due to credit risk. The costs for such an insurance will clearly
depend on the size of the risk and the impact of the possible losses. Hence, the price
can be determined using the above-mentioned models for credit risk. There are no
obligations to trade credit default swaps so their changes in credit spread can be used
for speculations, for example: CDS spreads can be used as risky assets in investment
structures such as CPPI and CPDO.
The third section will be dedicated to a particular type of insurance portfolio, the
CPPIs. They are a derivative security with capital guarantee using a dynamic trading
strategy in order to incorporate the performance of a certain underlying product such
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch11 FA
as a simple stock or a CDS. They were introduced more than 10 years ago and are
frequently used.
The investment structure of CPPIs has been and continues to be a popular topic
for research. First, a multivariate jump-driven model, which can be applied for pricing
credit derivatives such as CPDOs, is discussed in detail in Sec. 11.4. Next, in Sec. 11.5,
a more general overview of the other recent developments regarding CPPIs is given.
All papers could be classified under three different fields of interest. A first group
of papers concentrates on proposing different models for pricing. The second group
concentrates on the estimation of the leverage factor in order to fix an upper bound
for the gap-risk, and the third group tries to extend the structure even more in order to
include an extra safety factor.
One of the recent new products based on the idea of a CPPI is the CPDO or
constant proportion debt obligation, which is discussed in Sec. 11.6. CPDOs are used
for credit portfolios comprising exposures to credit indices such as iTraxx and CDX.
The CPDO structure borrows many features from the CPPI structure, such as the
constant proportion. The main goal of a CPDO is to produce a high-yielding product
and this is achieved through a high degree of leverage. Contrary to a CPPI investment
strategy, leverage will be increased when the net asset value of the portfolio decreases
and descends below the target amount, but leverage will be decreased when the net
asset value of the portfolio increases and approaches the target. Once a CPDO reaches
this target amount, it will completely de-leverage. In this section, we will not only
focus on the dynamics of the structure but we will also give a short overview of the
research concentrating on this topic and discuss the question of the “safeness” of this
structure as has been highlighted recently in the news.
We will conclude by discussing in more depth the differences and similarities of
CPPI and CPDO. This should give an even better insight into the structure of both
financial instruments.
This section is intended to give a short introduction to the main financial concepts,
which will play a role in this work. First, the concept of credit risk is introduced and
different ways to model this risk are discussed. Next, a short introduction on credit
derivatives, i.e., more precise credit default swaps, is provided.
meet its obligations. A simple example is the case, where a single person signs a loan
with a bank. Here, it can happen that the person does not repay his debt according to
the agreement. In such a case, we say that default will have occurred. The risk that
such an event will happen is called credit risk and will always be spread over a certain
time length.
Taking a more global perspective, in finance we do not only deal with the situation
of a person and a loan but it will always be possible to characterize the credit risk in
terms of the following components: the obligor, the set of criteria defining the default,
and a time interval over which the risk is spread. When, for instance, we talk about
bonds, their default can be defined in several ways. It could, for example, be bankruptcy
but it might also be a rating downgrade of the company or failure to pay an obligation
(such as a coupon). But it could also concern the value of a firm — here a firm’s
value is linked to the value of its financial assets. In general one will look at the firm’s
asset value V = Vt , 0 ≤ t ≤ T and default will be defined as a boundary condition
on the asset value. For example a default event will occur if the value falls below a
certain fixed level L within the time horizon. Figure 11.1 presents two possible paths
250
200
150
t
V
100
50
0
0 1 2 3 4 5 6 7 8 9 10
τ T
Figure 11.1 Two possible paths for a firm’s asset value over time with T = 10. When the
black line occurs the obligor will default (when Vt < L) in the gray case the obligor will
survive.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch11 FA
for the firm value over time [0, 10] modeled through Black–Scholes, where µ = 0.05,
σ = 0.4, and S0 = 100 as explained below. In the case of the black line, the value of
the firm will fall below the lower bound, which is fixed at 20 just before t = 5 and
hence will default. In the case that the firm value follows the gray path, no default will
occur before T = 10.
As one will try to protect oneself from defaults, the size of credit risk will be
one of the crucial factors when determining the prices and hence the techniques for
estimating the probability of default of a reference entity within time T will be very
important.
Developing new models for the estimation of credit risk is an important topic in
the field of finance. The strong interest in this topic is also linked to the fact that in the
last couple of years the volume of instruments linked to credit risk traded on the market
has increased exponentially. Besides the increase in investments, there has also been
interest due to the Basel II Accord, which encourages financial institutions to develop
methods for assessing their risk exposure. Credit risk models are usually classified into
two categories: reduced-form models, including intensity-based models, and structural
models.
Intensity-based models, also known as hazard-rate models, focus directly on mod-
eling the default probability. The main idea of these models lies in the fact that at any
moment in time (as long as the contract is running) there is a probability that an obligor
might default. Default is defined at the first jump of a counting process with a certain
intensity. In practice, the models assume that the intensities of the default times follow
a certain process (stochastic or deterministic) and under those conditions the under-
lying default model can be constructed. This intensity of the process depends heavily
on the firm’s overall health and on the situation of the market.
The structural models, also known as firm-value models, link default events to the
value of the financial assets of the firm, such as in the example presented above. Credit
risk will hence depend on the model used for the value of the financial assets of the
firm and the criteria used for a default. This approach will almost always be used in
the remainder of this chapter when modeling the credit default.
A common way to model the time evolution of assets uses the following diffusion
process:
where Wt is a standard Brownian motion. Here, µ and σ > 0 are the so-called drift
and volatility factors.
The path of a Brownian motion is continuous and can easily be simulated when
one discretizes the time using very small steps t. The value of a Brownian motion
at time {nt, n = 1, 2, . . .} is obtained by sampling a series of independent standard
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch11 FA
The solution of the SDE given in Eq. (11.1) is called geometric Brownian motion and
is given by
The above way to model price changes of financial products is also referred to as
the Black–Scholes model. The main advantages of using the Black–Scholes model is
that it is easy to understand. Calculation of prices of derivatives under the model is
moreover typically not that time-consuming. A drawback is that it assumes normality
of the log returns of the financial assets, which is often not true in reality. Moreover,
the Black–Scholes model does not capture the possibility of sudden jumps, which do
occur in real live and often cause extra credit risk. Hence, a more flexible stochastic
process is often required to model reality in a better way.
It would be good to keep some properties of the Brownian motion such as inde-
pendence and stationarity of the increments but to drop the constraints of normality
and continuity of the paths. To create such a process, we must restrict ourselves to
the group of infinitely divisible distributions. For each infinitely divisible distribution
(with characteristic function φ(u)), a stochastic process can be defined which starts
at zero and has independent and stationary increments such that the distribution of
the increments over [s, s + t], s, t ≥ 0 has (φ(u))t as characteristic function. Such
processes are called Lévy processes, in honour of Paul Lévy, a pioneer of the theory.
(2) X0 = 0.
(3) The process has stationary increments.
(4) The process has independent increments.
Lévy process became very popular and are still more and more used in practice.
Examples of Lévy process are the (compound) Poisson process, the Gamma process,
the inverse Gaussian process, and the Variance Gamma process (VG). An in-depth
study of Lévy process in finance and how they can be applied can be found in [1].
Next, we will define an important group of credit derivatives, which are often used
in practice.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch11 FA
7000
6000
5000
4000
Euro
3000
2000
1000
1000
1 2 3 4 5 6 7 8 9 10
Year
Figure 11.2 Cash flows from the protection buyer for a 10-year CDS. Black: cash flows in
case no default occurs. Gray: cash flows in case there is a default at time t = 7.
and 50%. For the current example, we assume that the recovery rate will be equal to
R = 40%.
The annual amount paid by the protection buyer in this example is hence equal to
400 bp · 10,000 = 400 Euro and the payment of the protection seller in case of default
will be F(1 − R) = 6000 Euro. In Fig. 11.2, it is assumed that in the second scenario
the bond defaulted at the beginning of the seventh year.
Pricing models for CDSs based on Lévy processes can be found in [2] and [3].
In practice, CDS are not only used to reallocate the risk of a single asset (the
so-called single name CDS), but also a basket of assets might be considered. A credit
default swap index is a credit derivative used to hedge credit risk or to take a position on
a basket of credit entities. There are currently two main families of CDS indices: CDX
and iTraxx. CDX indices contain North American and Emerging Market companies,
whereas iTraxx contains companies from the rest of the world.
One of the most widely traded index is the iTraxx Europe index composed of the
most liquid 125 CDS referencing European investment grade credits. There are also
significant volumes, in nominal values, of trading in the HiVol index. HiVol is a subset
of the main index consisting of what are seen as the most risky 30 constituents at the
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch11 FA
time the index is constructed. Typically every six months new series of CDS indices
are issued.
Indices are like CDSs not only used as insurance against default risk but are also
traded on the market in a speculative way.
Portfolio Insurances are capital guarantee derivative securities that embed a dynamic
trading strategy in order to make a contribution to the performance of a certain under-
lying product (e.g., an asset, a CDS or a CDS index, …).
One particular type is considered here, the constant proportion portfolio insurance
e.g., [4], investing partially in a risk-free way and combining this with a risky asset.
The family of Constant Proportion Portfolio Insurance consists of investments
for which the amount necessary for guaranteeing a repayment of a fixed amount N at
maturity T is invested in a risk-free way, typically a bond, B, and only the exceeding
amount will be invested in one or more risky assets, S (i) . This way an investor can limit
its downside risks and maintain some upside potential. This type of portfolio insurance
has first been introduced by [5] and [6].
The product manager will take larger risks when his strategy performed well. But
if the market went against him, he will reduce the risk rapidly. The following factors
play a key role in the risk strategies an investor will take:
• Price: The current value of the CPPI. The value at time t ∈ [0, T ] will be denoted
as Vt .
• Floor: The reference level to which the CPPI is compared. This level will guarantee
the possibility of repaying the fixed amount N at maturity T , hence it could be seen
as the present value of N at maturity. Typically this is a zero-coupon bond and its
price at time t will be denoted as Bt .
• Cushion: The cushion is defined as the difference between the price and the floor:
e = Multiplier × Cushion%.
• “gap” risk: This is the probability that the CPPI value will fall under the floor.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch11 FA
The level of risk an investor will take at each time t is equal to the investment level as
long as the value of the CPPI exceeds the floor. For any time t, the future investment
decision will be made according to the following rule:
• If Vt ≤ Floor = Bt , we will invest the complete portfolio into the zero-coupon
bond.
• If Vt > Floor, we will invest a proportion equal to e in the risky asset portfolio.
It can easily be shown that under the assumption that the underlying asset price follows
a Black–Scholes model with continuous trading, there is no risk of going below the
floor and that the expected return at maturity of the CPPI is equal to [7]
E[VT ] = N + (V0 − Ne−rT ) exp(rT + m(µ − r)T ).
In practice, however, it is known that the probability of going below the floor is non-
zero. It might, for instance, happen that during a sudden downside move or due to
overnight changes, the fund manager might not be able to adjust the portfolio in time.
In the case of an event where the actual portfolio value falls under the floor, at maturity
the issuer will have to cover the difference between the actual portfolio value and the
guaranteed amount N. It is therefore of importance for the issuer of a CPPI note to be
able to quantify this risk, also called gap risk.
We will present an example of a possible cash flow for a CPPI with maturity T
equal to 10 years. For the sake of simplicity, we will consider only one underlying
asset with prices St and a risk-free asset, a zero-coupon bond Bt with a constant interest
rate r = 5%. We also assume that the initial price of the asset is equal to S0 = 100
and the prices over time are simulated from a Variance Gamma model, which will be
presented later (see Eq. (11.4)) with parameters σ = 0.5, ν = 0.25, and θ = 0.026.
For the CPPI process, the leverage or multiplier is fixed at 2.5 and the starting capital
is 100. We also consider that the CPPI at maturity repays the investor with at least the
initial capital. Figure 11.3 presents two examples of possible scenarios for the simple
CPPI. In the example on the left, the value of the CPPI will always stay above the floor,
while in the second example (on the right) at time τ a sudden drop of the risky asset
will result in a CPPI value below the floor, which is the gap risk. Note that different
scales are applied for the presentation of the results.
In this example, as the repayment at maturity of the initial value should be insured,
the floor will be 100 exp(−r(10 − t)) at each time t. For each step, the value of the
cushion is calculated and the portfolio is re-balanced according to the risk exposure.
The re-balancing is such that the bigger the difference between the CPPI value and the
floor, the higher the cushion value and the more risk one will take. The process will
stop at maturity or once a drop of the asset value occurs of such a level that the CPPI
value falls below or hits the floor. If such a drop happens the product manager will put
the risk exposure to zero and only invest in a risk-free way until maturity.
May 12, 2010
Portfolio Insurances, CPPI and CPDO, Truth or Illusion?
17:47
CPPI value stays above floor CPPI value drops bellow floor
300 300
WSPC/SPI-B913
Stockvalue
200 200
Value
Value
100 100
Stockvalue
0 0
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
120 200
CPPI
Bondfloor
100 150
Value
Value
80 100
CPPI
Bondfloor
60 50
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
60
Cushion 100 Cushion
40
Value
Value
50
b913-ch11
20
0
0
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
Time Time
Figure 11.3 Value of the risky asset (top), CPPI performance (central), and cushion (bottom) in case the CPPI value stays above the floor (left)
FA
and in case a sudden downwards jump occurs and the CPPI value drops bellow below the floor (right).
269
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch11 FA
This section is based on [8] and presents a possible way of modeling credit CPPI
structures. More precise, a dynamic Lévy model (VG) is set up for a series of correlated
spreads.
A recent innovation is the possibility to trade swaptions, an option granting its
owner the right but not the obligation to enter into an underlying swap (e.g., a CDS),
on indices. We will make use of these instruments to calibrate the underlying dynamic
spread models. The parameters of the model actually come from a two-step calibra-
tion procedure. First, a joint calibration of the correlated spreads on swaptions and
second by a correlation matching procedure. For the joint calibration, we make use of
equity-like pricing formulas for payers and receivers swaptions based on characteristic
functions and Fast Fourier Transform methods. To obtain the required correlation, we
set by a closed-form matching procedure the models correlation exactly equal to a
prescribed (e.g., historical) correlation. We then have a model in place that can gener-
ate very fast correlated spread dynamics under jump dynamics. The calibrated model
can be used to price a whole range of exotic structures. We illustrate this by pricing
credit CPPI structures. The CPPI structures considered in this section are such that the
invested capital is put in a risk-free bond and a position is taken on credit derivatives
indices (usually protection is sold). As discussed before, important in the handling of
CPPIs is assessing the risk that spreads of the underlying credit indices jump and lead
to so-called gap risk. Because of the built in jump dynamics, a better assessment of
this gap risk is made possible.
The distribution is known to be infinitely divisible and by standard Lévy theory one
thus can build out of such a distribution a process. The VG process is defined as
X(VG) = {Xt(VG) , t ≥ 0} with parameters σ, ν > 0, and θ. It is a process, which
starts at zero has stationary and independent increments, and for which the increments
(VG) √
Xs+t − Xs(VG) follow a VG(σ t, ν/t, tθ) distribution over the time [s, t + s] (for more
details see [1]). Theoretical background on Lévy process can be found in [10] and [11].
For some historical background on the VG model see [12].
Another way of constructing the VG process is by the technique of time chang-
ing. Here, we will start from a Gamma process. Recall that the density function of a
Gamma(a, b) distribution is given by
ba a−1
fGamma (x; a, b) = x exp(−xb), x > 0, and a, b > 0,
(a)
where (.) is the Gamma function. It is known that the distribution is infinitely divisible
and hence using the Gamma distribution one can build a process with independent and
stationary Gamma increments. Defining G = {Gt , t ≥ 0} as a Gamma process with
parameters a = b = 1/ν, the resulting Gt will follow a Gamma(at, b) distribution and
E[Gt ] = t. A VG process can be constructed by time-changing a Brownian Motion
with drift. More precisely, one can show (for more details see [1]) that the process
Xt(VG) = θGt + σWGt , t ≥ 0, (11.3)
where W = {Wt , t ≥ 0} is a standard Brownian motion independent from the Gamma
process, is indeed a VG process with parameters (σ, ν, θ).
The construction can be interpreted as if we now look at a Black–Scholes world
but now measured according to a new business clock (Gamma time). It has proven to
be very successful in the univariate setting, as the underlying VG distribution can take
into account, in contrast to the Normal distribution, skewness, and excess-kurtosis.
In [13], a VG model was used to accurately fit CDS curves.
With this alternative definition, we have an easy way to simulate a sample path
of the VG processes, which can be obtained by sampling a standard Brownian motion
and a Gamma process. The Gamma process can easily, like the Brownian motion, be
simulated at time points {nt, n = 1, 2, . . .} with t small. First generate independent
Gamma(at, b) random numbers {gn , n = 1, 2, . . .}. Then, the Gamma process can
be constructed by
G0 = 0 and Gnt = G(n−1)t + gn , n ≥ 1.
Similarly to Eq. (11.2) for the Brownian motion, the value over time of an asset using
the dynamics of a Variance Gamma process can now be modeled by
St = S0 exp(ωt + θGt + σWGt ), (11.4)
with ω = ν−1 log(1 − σ 2 ν/2 − θν).
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch11 FA
We work with multivariate extensions of the model along the technique described
in [14] and [15]. Additionally, we want to note that in [9], a symmetric version of a
multivariate VG is initiated.
To build the multivariate VG process a Gamma process G = {Gt , t ≥ 0} is needed,
where the parameters a and b are both fixed at 1/ν. Also an N-dimensional Brownian
motion is introduced W = {W t , t ≥ 0}, where W t = (Wt(1) , Wt(2) , . . . , Wt(N) ). It is
assumed that this process is independent of the Gamma process and that the Brownian
motions have a correlation matrix given by ρW = (ρijW , i, j, = 1, . . . , N)
(i) (j)
ρijW = E[W1 , W1 ].
Note that there is dependence between the Xt(i) ’s due to two causes: they are all con-
structed by a time change with a common Gamma time. This will mean that the
processes will all jump together, but jumps’ sizes can be different. Moreover, there
is dependency also built in via the Brownian motions. A straightforward calculation
shows that the correlation between two components is given by
(j) (j)
E[X1(i) X1 ] − E[X1(i) ]E[X1 ] θi θj ν + σi σj ρijW
ρij = = . (11.5)
(j)
Var[X1(i) ] Var[X1 ] σi2 + θi2 ν σj2 + θj2 ν
This clearly shows that, even if we assume the Brownian components to be independent
of each other, one still obtains a correlation between the different components because
of the common gamma time.
Suppose that we need to model, as later on will be the case, the evolution of N-
correlated spreads. Take for example the evolution of the iTraxx Europe main index
and its overseas opponent the Dow Jones CDX.NA.IG main index and/or the spreads
of their HiVol subsets.
We assume the following correlated dynamics for the evolution of N-dependent
spreads:
St(i) = S0(i) exp(ωi t + θi Gt + σi WG(i)t ) = S0(i) exp(ωi t + Xt(i) ), i = 1, . . . , N,
where
1
ωi = ν−1 log 1 − σi2 ν − θi ν .
2
These mean-correction terms, ωi , are in place because one can then easily show that
the spread processes are mean-reverting in the sense that we have for every t ≥ 0
E[St(i) ] = S0(i) .
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch11 FA
forward spread by the cost of this protection. More precisely, the adjusted forward
spread is given by
S T ∗ AT
(T, T ∗ ) = F0 (T, T ∗ ) +
(adj) (adj)
F0 = F0 ,
AT ∗ − A T
and the price of non-knockout payers and receivers are respectively given by
where
log(F0
(adj)
/K) + σ 2 T/2 √
d1 = √ and d2 = d1 − σ T .
σ T
Summarizing, we want to remark on the striking connection with vanilla option prices
in equity. Basically, pricing comes down to pricing under a Black–Scholes regime with
no interest rates and no dividends. However, the model has all the deficiencies of the
Black–Scholes framework: no-jumps, light-tails, symmetric underlying distribution,
etcetera.
Pricing of vanillas has already been worked out in full detail in equity settings by [17]
and by a slight adaption to the credit setting we can very fast calculate payer and
receiver swaptions using the Carr–Madan formula in combination with fast Fourier
transform methods. More precisely,
exp(−α log(K))
Payer(T, K) = A(T, T ∗ )
π
+∞
φ(v − (α + 1)i; T )
× exp(−iv log(K)) dv,
0 α2 + α − v2 + i(2α + 1)v
where the characteristic function of the log of the adjusted forward spread process at
maturity T is given by
(adj)
φ(u; T ) = E[exp(iu(log F0 + ωT + XT ))],
120
Spread
100
80
60
40
20
Jan-04 Time Mar-07
Figure 11.4 Evolution of Main and HiVol indices: iTraxx and CDX.IG.
Table 11.1 Correlation of Spreads: iTraxx (Main and HiVol) and CDX
(Main and HiVol).
• Buy protection on iTraxx Europe HiVol on the run (5 years) for 12 125
30
of the risky
exposure.
• Sell protection on DJ CDX.NA.IG Main On the run (5 years) for half of the risky
exposure.
• Buy protection on DJ CDX.NA.IG HiVol On the run (5 years) for 12 125
30
of the risky
exposure.
The risky exposure is actually taken in position in iTraxx Europe Main and DJ
CDX.NA.IG Main. By buying some protection on their HiVol components, the actual
risky exposure is reduced and one is only exposed to the non-HiVol names in both
indices. This is often done to eliminate as much as possible default risk. If a name
defaults, it will most likely be a HiVol name.
The initial 100M EUR are put on a risk-free bank account at a compound rate of
four percent.
Suppose the current quotes for the four components of our portfolio are these as
given in Table 11.3.
We rebalance after regular times, say daily and continue doing this until maturity
or until we have a negative cushion at a rebalancing date. In that case, all positions are
closed. We can however not pay back the principal amount since the portfolio value
is below the bond-floor. This is called gap risk (see Fig. 11.5). One of the aims of the
model is to calculate the gap risk or in other words the present value of these gaps.
We assume the following correlated VG dynamics for the spreads:
where Gt is a common Gamma Process such that Gt ∼ Gamma(t/ν, 1/ν) and Wt(i) are
correlated standard Brownian motions with a given correlation matrix ρW = (ρijW ).
t=0
CPPI
105
Deleverage Bond floor
Portfolio
100
95
Value
90
Gap risk
85 Leverage
80
75
0 1 2 3 4 5 6
Time
500 Deleverage
400
300
200
100
Leverage
0
0 1 2 3 4 5 6
Time
Figure 11.5 CPPI performance, leveraging, deleveraging, and gap risk in a multivariate VG
driven model.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch11 FA
In Fig. 11.7, one sees a typical picture of the correlated moves under the multivariate
VG jump dynamics.
80
Spreads (bp)
60
40
20
0
0 1 2 3 4 5 6
time
CPPIs received and still receive a lot of attention, not only from banks but also from
academia. Here, we try to give an overview of the results discussed in a list of papers.
Three main topics of research can be identified: the behavior of CPPI, limiting the risk,
and building insurances for this risk.
The first group of papers study in detail the behavior of CPPI strategies under spe-
cific conditions for the underlying portfolio. One way of pricing was already presented
in detail in the previous section.
The next group of papers try to measure the risk factors involved. In the first paper
discussed here, an upper bound for the multiplier m is searched for in such a way that
the investment in the risky portfolio is maximized under the condition that the gap risk
must stay under a certain limit. The second discussion presents an extended way to
calculate the VaR and GVaR of the CPPI portfolio.
The last group of papers concentrate on possible ways of extending the CPPI in
such a way that an insurance against the small but existing gap risk is build in. The
price and size of such an extra insurance will depend on the probability of hitting the
floor and hence ways to quantify this risk are also discussed.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch11 FA
For each of the groups, we discuss one example in more detail in the following
three subsections.
where T is the length of the time period covered and F [a,b] denotes the distribution
function of the variations Xk .
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch11 FA
If the distribution F [a,b] is not known, two different cases can be considered. Either
one can assume that no prior information is available or, alternatively, one might be
able to fix an upper bound due to anticipation. If no historical information is available,
extreme value statistics will be applied on the maximum value of any time period T
(0 < T < T ). EVT states that the distribution of the maximum of n observations
will converge to a limiting distribution with two parameters. After applying maximum
likelihood estimation the parameters of this limiting extreme value distribution, an
approximation of the upper bound of m is obtained. In the other case, where some
prior information is available, a portfolio manager might fix a maximal drop level b
based on his believes. Under this assumption of a maximal drop level, the calculation
of the upper bound for m becomes more complicated but the portfolio manager will
be able to use a higher multiplier m.
Finally, changes of the upper bound under the different assumptions are compared
when using S&P 500 index as a risky asset.
To finish the discussion, portfolios with CDOs are included. For the spread VaR
and default VaR, the previous results can be re-used. Besides those, also the VaR of
the profit and loss coming from daily moves of the correlation smile is of importance.
Here a model-free approach is proposed.
The above risk measures for each factor are finally combined using GVaR to give
a global estimation of risk. This is supposed to be stable and to offer great netting.
where (Wt )t is a standard Brownian motion, independent from the Poisson process
with measure of jumps (lt )t . It is also assumed that the expected value of jumps in
every finite interval [0, t] is finite and equal to λt (λ ≥ 0).
The paper concludes by extending the insurance in the sense that not only the first
time but every time the CPPI value reaches the adjusted floor, the amount K will be
invested in the CPPI portfolio.
Constant Proportion Debt Obligations (CPDOs) first appeared in August 2006 and are
a variation on the CPPI structure. They are used for credit portfolios comprising the
exposures to credit indices such as CDX and iTraxx. The CPDO’s risk exposure, just
as with the CPPI, is determined using a constant proportion approach and rebalances
its portfolio between the credit portfolio and a safe asset. The CPDO structure does
this with the aim of producing a high-yielding “AAA”-rated product.
A CPDO funds itself through the issuance of long-term debt paying timely coupon
and principal on the notes. The promised coupon is a spread above for example LIBOR.
The combination of a high coupon payment and the high rating have made CPDOs
very popular products.
First, we will try to explain the structure in detail. CPDOs are currently also a
popular topic in the news and a small overview is provided the next subsection. This
section is concluded presenting an example on rating of CPDOs.
• Riskfree: The amount of the total CPDO value invested in a risk-free way.
• PV(liabilities): The present value of the current liabilities will be the sum of all
discounted coupons still to be paid and the discounted value of the final principal
amount.
• Shortfall: The shortfall is the capital, which is missing to fulfill all the future obliga-
tions. It is defined as the difference between the present value of the liabilities and
the net asset value,
• CDS premium: The premium is the amount to be paid to the protection seller (such
as a CDS or CDS index) in exchange for the insurance. An increasing value of
the spread refers to an increased default probability of the underlying asset and
hence it will be negative as a default will lead to a decreasing NAV . Conversely, an
increasing spread results in a higher income.
• PV(CDS premium): The present value of all future premium payments up to matu-
rity.
• Leverage: The leverage refers to the degree of risk which will be taken at each time
step. A maximum level is often fixed at 15. The leverage is hence defined as
Shortfall
Leverage = min β , max(Leverage) , (11.6)
PV(CDS premium)
where β is a multiplier. Similar as in [21], page 18, in this document β will be fixed
at 1/Riskfree.
• Cash-in: In case the NAV is equal to or exceeds the target value (PV(liabilities)),
the necessary amount to cover all future liabilities is reached and hence the risky
exposure and leverage will be put to zero. From this point onwards the NAV will
be completely invested at the risk-free rate, with coupon and fees being paid until
maturity.
• Cash-out: if there are substantial losses and the NAV falls below a certain thresh-
old (often fixed at 10% of the initial investment), it will be said that cash-out has
occurred. In such a situation, the CPDO will unwind and the investor will receive
the remaining value.
The investment strategy follows the following steps. At every time-step, one should
check the discounted value of the future obligations (coupon and principal payment),
which is the amount which one tries to reach.
Next, as the CPDO exposes itself to risk by selling protection (CDSs or CDS
indices), the mark-to-market of this risky investment is checked and compared to the
price paid at the previous time step for this investment. An increase in the spread of the
risky CDS means that the underlying insurance has become more expensive and linked
to this it has become more likely that the seller of the protection will have to cover a
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch11 FA
future loss. Compared to the previous time-step, this can be considered as a loss for
the protection seller, as he appears to be underpaid for the protection it provides, and
a gain for the protection buyer.
Besides this gain or loss linked to the protection spread, the costs (or incomes)
related to the protection also need to be taken into account. Those costs are referred to
as the fee which is equal to the sum of all the CDS premium payments made by the
protection buyer for the insurance within the time step, taking into consideration the
possibility of a default event.
The total value of the CPDO (NAV ) at time t + 1 will hence be equal at the
accumulated value of the cash, invested risk free at time t, plus the value of the
risky asset at time t + 1 augmented with the gains (or losses) made by investing
and the fees collected (or paid) for the insurance in the time period [t, t + 1]. From this
value, we should subtract the coupon payments which need to be made to the CPDO
investor.
Based on the above value, and before going to the next step, the new leverage is
determined and according to this leverage new investment positions are taken in the
risky asset(s). If the NAV increases, the shortfall decreases and hence the leverage will
go down; whereas if the NAV decreases, the shortfall will increase and the leverage
will go up in order to try to fix the previous negative performance.
The above steps are repeated at each time step until cash-in or cash-out occurs,
or until maturity. As described in the definition, cash-in occurs when the total value of
the CPDO reaches or exceeds the current value of the future obligations. In this case,
the seller of a CPDO is sure to be able to fulfil all its future obligations and will from
then on only invest in a risk-free way. The probability for a CPDO that such an event
takes place will have a big impact on its rating. Cash-out occurs when the value of the
CPDO hits or shoots below a lower bound which is fixed at time t = 0. If this happens,
the CPDO will unwind and the investor will receive the remaining proceed. Such an
event could be called the gap risk and is comparable to the gap risk of a CPPI. The risk
of a cash-out event cannot be excluded, since in case of under-performance more risk
will be taken, which increases even more the possibility of arriving below the lower
bound when a downward jump occurs.
A similar situation as in Sec. 11.3 is used in order to present the CPDO functioning
in an example. For the risky asset, a CDS-index is considered with starting spread
S0 = 100, and the spreads follow the same Variance Gamma model as before. The
risk-free interest rate is fixed at r = 5%. For the CPDO structure, the coupon payment
is fixed at r + 2% and the maximum leverage is equal to 15. In this example, a cash-out
level of 10% of the initial investment is considered. Figure 11.8 presents two examples
of possible scenarios of a CPDO. In the left column, the target level is reached before
maturity at time t around 8 from that point onwards all cash will be invested in a
risk-free way. For the second example, in the right column, the CPDO value will drop
May 12, 2010
288
17:47
0.06 0.2
Spread Spread
0.04
WSPC/SPI-B913
0.1
0.02
0 0
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
150 150
CPDO
100 100 Target
CPDO Cashout
50 Target 50
Cashout
0 0
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
60 150
Shortfall Shortfall
40 100
20 50
0 0
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
6 6
Leverage Leverage
4 4
2 2
b913-ch11
Joossens and Schoutens
0 0
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
Figure 11.8 Value of the risky asset (CDS index) (first graph); the corresponding CPDO performance (second); shortfall at each time step (third)
and leverage taken at each time (final) in case of a cash-in situation (left) and a cash-out situation (right).
FA
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch11 FA
below the cash-out level just before time t = 5. This is similar to what, in the CPPI
case, we refer to as a gap. If such an event occurs, the CPDO will unwind and the
investor will receive all remaining cash.
In practice, risk positions will almost always be taken in CDS indices (iTraxx
and CDX). The benefits of this choice are summarized on Slide 10 of [22]. However,
questions about safety in the sense of the correctness of the high triple-A rating do
remain and this sensitive topic will be discussed in the next two subsections. Some
literature on CPDOs can be found in [21, 23–28].
100
iTraxx
CDX
50
0
0 1 2 3 4 5 6 7 8 9 10
150
CPDO
100 Target
Cashout
50
0
0 1 2 3 4 5 6 7 8 9 10
30
Shortfall
20
10
0
0 1 2 3 4 5 6 7 8 9 10
15
Leverage
10
0
0 1 2 3 4 5 6 7 8 9 10
Figure 11.9 First graph: value of the risky asset (CDS index); second: the corresponding
CPDO performance; third: shortfall at each time step; final: leverage taken at each time.
for their dynamics. The resulting model is then applied to simulate the evolution over
time of a CPDO.
Assume that the risk-free interest rate is fixed r = 0.04. For the CPDO structure
the coupon payment is fixed at r + 2% and the maximum leverage is equal to 15.
The initial quotes for our portfolio are iTraxx is quoted 24.625 bp and CDX is quoted
37.5 bp. The top graph of Fig. 11.9 presents a typical picture of the correlated spread
evolution under the multivariate VG jump dynamics, which is obtained after calibra-
tion. The corresponding path of the CPDO is presented in the second graph and the
third and fourth graphs present the corresponding shortfall and leverage positions at
each time.
In most of the current research documents, it is considered that the spreads follow
a Brownian motion which do not allow sudden jumps and hence it is less likely to hit
the barrier. If jumps are present in the spread dynamics, the value of the portfolio can
suddenly jump under the floor. Based on 10,000 Monte Carlo simulations using the
model as obtained from the calibration, we find that the likelihood of occurrence of a
cash-out event is around 1.94%, which is not as negligible as might be suggested in
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch11 FA
the literature. It looks that this percentage is closer to what is also observed in practice,
as discussed before.
When CPDOs were created, they were considered as a variation of the CPPI. They
borrow certain features such as a “constant proportion” approach to determine the
leverage and the re-balancing of the portfolio between the credit portfolio and the safe
asset.
On the other hand, they are also very different. Initially, the CPDO value will
be below the target value, while the CPPI manager tends to invest only the amount
exceeding the floor, which is needed to make the principal payment at maturity in a
risky way. Once decisions on the risky exposure need to be taken the idea is that a
CPDO investor will increase its risk as it is performing negatively while the CPPI will
decrease its risk as it is not performing well and approaches the floor level. In other
words, at each time step, the CPPI investor takes risk exposure positions based on the
amount of surplus the portfolio value has with respect to the floor value. The CPDO
investor, on the other hand, will at each time take risk exposure proportional to the
amount the CPDO portfolio is lacking in order to reach the target value.
Once the CPDO value reaches the target value, the manager will stop investing
in a risky way as there is sufficient capital to pay out all future liabilities and there is
no more need to create capital in a risky way. A CPPI manager will try to optimize its
profit but will stop taking risk at the moment that the CPPI value touches the floor, as
he is afraid to fall below by taking more risk.
Also the behavior at the final stage of the contract is different. A CPPI will,
at maturity, irrespective of the performance of the risky asset, receive the principal,
together with any positive return generated from the risky asset. In case of loss, when
the CPPI portfolio falls below the floor, the losses are covered by the seller of a CPPI
so that the investor will still receive the principal. An investor can hence always be sure
of receiving the principal and, in the case of good performance, even more. For the
CPDO, on the other hand, a target value is aimed for and in the case of a sufficiently
well performing risky asset a cash-in event will occur and the investor will receive all
promised coupon and principal payments. But when the risky asset does not perform
well and a cash-out event occurs, the CPDO will unwind before maturity. In such a
situation only the remaining amount will be paid out to the investor. Investors will
want to know the size of this risk at the time of investing in a CPDO and they will
use its rating as an indication. This way, the rating of a CPDO becomes important and
will have an effect on the price of a CPDO. In practice, the risk of a cash-out event has
apparently been under-estimated.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch11 FA
11.8. CONCLUSIONS
As the market of structured credit products grew over the last years, also the request
of protection mechanisms in structured credit transactions stays high and hence, also
here, a continuous evolution can be observed. CPPI and CPDO are recently developed
alternative investment strategies which aim to provide a protection.
CPPIs first came into use around 10 years ago and promise a pre-defined principal
payment at maturity. A constant proportion rule is applied to decide the investment
strategy. At every time step, the investment in the underlying risky asset and the safe
asset is re-balanced in order to optimize the profit.
CPDOs were introduced in 2006 and are intended to be safe, high-yielding instru-
ments. A similar constant proportion rule is used for their investment strategies and
they will invest in a risky asset by selling protection (such as CDX and iTraxx). Simi-
larly to the CPPI, re-balancing will be done at every time step until the targeted value
has been reached.
The aim of this chapter was to create an in depth view of the dynamics and risks
linked to both alternative investment strategies. Hence, first we have tried to explain
step by step how they both function and how investment decisions are made.
Understanding the dynamics well helps to identify remaining “safety gaps” and
allows an investor to get an idea about the size and possibility of experiencing such a
gap. For the CPPI, the possibility that the value of the total portfolio will fall below the
floor exists and will create a loss, while for the CPDO a loss occurs when a cash-out
event occurs.
In both cases, there is a strong interest in quantifying this risk. In a first group of
papers, the researchers concentrate on quantifying the risk using specific conditions.
Next, some propose ways to limit the multiplier factor in order to limit the risk, while
other papers suggest the possibility of investing on an insurance to cover this risk.
But also for those new developments a good and robust way of quantifying the risk is
necessary. It could be concluded that the gap risk for CPPIs should not be neglected
but safety nets can be used to avoid suffering from it.
As CPDOs are still very new, the field of research is still limited, and as they have
only been used on the market for a couple of years, their performance in the real world
has only been observed over a short time. As recent experience has shown, in real life
CPDOs do not seem to be as safe as they were expected to be. Clearly, there is still a
strong need to quantify the risk of cash-out events in a more realistic manner, and a
great deal of research remains to be done in this field.
References
[1] Schoutens, W (2003). Lévy Processes in Finance: Pricing Financial Derivatives. John
Wiley & Sons.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch11 FA
[2] Cariboni, J (2007). Credit Derivatives Pricing Under Lévy Processes. PhD thesis,
Katholieke Universiteit Leuven.
[3] Madan, D and W Schoutens (2007). Break on through to the single side. Technical Report,
Section of Statistics Technical Report 07-07, (2007). URL http://perswww.kuleuven.be/
∼u0009713/LevyCDS.pdf.
[4] Overhaus, M, A Bermudez, H Buehler, A Ferraris, C Jordinson and A Lamnouar (2007).
Equity Hybrid Derivatives. John Wiley & Sons.
[5] Black, F and R Jones (1987). Simplifying portfolio insurance. Journal of Portfolio Man-
agement, 48–51.
[6] Perold,A (1986). Constant portfolio insurance. Unpublished manuscript, Harvard Business
School.
[7] Cont, R and P Tankov (2007). Constant proportion portfolio insurance in presence of jumps
in asset prices, SSRN eLibrary. URL http://ssrn.com/paper=1021084.
[8] Garcia, J, S Goossens and W Schoutens (2008). Let’s Jump Together: Pricing of Credit
Derivatives. Risk Magazine, September.
[9] Madan, D and E Seneta (1990). The variance Gamma model for share market returns.
Journal of Business, 63, 511–524.
[10] Bertoin, J (1996). Lévy Processes. Vol. 121, Cambridge Tracts in Mathematics, Cambridge:
Cambridge University Press.
[11] Sato, K (2000). Lévy Processes and Infinitely Divisible Distributions. Cambridge Univer-
sity Press.
[12] Seneta, E (2007). The early years of the variance-gamma process. Advances in Mathemat-
ical Finance. Boston: Birkhäuser.
[13] Cariboni, J and W Schoutens (2004). Pricing credit default swaps under Lévy models.
UCS Report 2005-02, K. U. Leuven.
[14] Luciano, E and W Schoutens (2006). A multivariate jump-driven financial asset model.
Quantitative Finance, 6(5), 385–402.
[15] Leoni, P and W Schoutens (2008). Multivariate smiling, Wilmott Magazine, March Issue.
[16] Pederson, C (2004). Introduction to Credit Default Swaptions, Lehman Brothers.
[17] Carr, P and D Madan (1999). Option pricing and the fast fourier transform. Journal of
Computational Finance, 2(4), 61–73.
[18] Bertrand, P and J Prigent (2002). Portfolio insurance: The extreme value approach to the
CPPI method. Finance, 23(01A13), 68–86.
[19] Brun, J and L Prigneaux (2007). VaR approach for credit CPPI and counterparty risk.
Quant Congress, Quant Congress.
[20] Prigent, JL and F Tahar (2005) CPPI with cushion insurance. SSRN eLibrary. URL
http://ssrn.com/paper=675824.
[21] Standard&Poor’s (2007). Quantitative modeling apporach to rating index CPDO struc-
tures. Technical Report, Standard&Poor’s.
[22] ABN-AMRO (2007). Surf CPDO: A breakthrough in synthetic credit investments. Struc-
tured Products Forum-Tokyo.
[23] Varloot, E, G Charpin and E Charalampidou (2006). CPSO an asset class on its own or a
glorified bearish rated equity, UBS Investment Research — European Structured Credit.
[24] Lucas, D and R Manning (2007). A CPDO primer. UBS Investment Research — CDO
Insight.
[25] Varloot, E, G Charpin and E Charalampidou (2007). CPDO insights on rating actions
and manager impacts using a new formula. UBS Investment Research — European Struc-
tured Credit.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch11 FA
[26] Cont, R and C Jessen (2008). Constant propoertion debt obligations. Financial Risks
Inernational Forum.
[27] Saltuk, Y and J Goulden (2007). CPDOs and the upcoming roll. Technical Report, JP
Morgan.
[28] Linden, A, M Neugebauer, S Bund, J Schiavetta, J Zelter and R Hardee (2007). First
generation CPDO: Case Study on Performance and Rating. Structured credit global special
report, Derivative Fitch.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch12 FA
12
ON THE BENEFITS OF ROBUST
ASSET ALLOCATION FOR CPPI
STRATEGIES
In recent years, new ideas for the robustification of the traditional Markowitz frontier
have appeared in the literature. Based on one of these ideas — the so-called robust
counterparts — we introduce the concept of the robustified efficient frontier. As mean–
variance efficient portfolios are frequently used as risky assets for CPPI strategies,
we investigate the behavior of such strategies under estimation risk. Based on a toy
example, we explain the main idea how the concept of a robustified frontier can be
used to improve the performance of CPPI strategies. For this purpose, we compare
the theoretical performance of CPPI strategies based on the original and the robust
efficient frontier.
295
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch12 FA
12.1. MOTIVATION
During the last years, dynamic investment strategies have gained more and more impor-
tance in quantitative asset management. Until now, traditional asset management was
mainly based on the choice of a strategic asset allocation (SAA), which has been fixed
for one year in most cases. Nowadays, there is a move toward dynamic asset alloca-
tion (DAA), mainly driven by portfolio insurance concepts and the influence of option
pricing theory on investment strategies, like the best-of-two concept of [9]. Further, it
has been observed that the utilization of a dynamically rebalanced allocation allows
for a better-tailored risk/return profile for the investor, as described in [1].
In the following study, we focus on one of the most popular portfolio insurance
concepts — a constant proportion portfolio insurance (CPPI) strategy. This concept
was probably first introduced by Perold in 1986, see [24] (see also [25] as well as [5]),
followed by a large variety of theoretical and empirical studies, comparison to alter-
native strategies, etc. For a recent and extensive overview of related work, we refer to
the book [27], Chapter 9 and the references therein. Meanwhile, the properties as well
as the pros and cons of CPPI like strategies are well known and discussed in detail,
see also [27], Chapter 9. However, in our opinion, one rather relevant practical issue
has not yet found much attention by researchers — the influence of estimation risk
on the performance of CPPI strategies. In contrast to this observation, there is a good
amount of research on the influence of estimation risk on static one-period mean–
variance optimization, see [8] and the references contained therein. Hence, we focus
on the investigation of the influence of this estimation risk on the performance of CPPI
strategies as well as a mitigation of this risk by the robustification of mean–variance
efficient portfolios.
The remainder of this paper is organized as follows: first, we introduce the tradi-
tional CPPI strategy. A few numerical examples will show that the choice of the risky
asset has a significant impact on the investment performance of the CPPI strategy
and that especially the allocation according to mean–variance optimal portfolios adds
value to the strategy. Thereafter, it is shown that this behavior may deteriorate in the
presence of estimation risk. Therefore, robustified mean–variance efficient portfolios
are introduced and their impact on the investment performance is investigated.
In contrast to most academic research, our exposition is based on a discrete time frame-
work instead of a continuous time setting. This setup was chosen for several reasons.
First, in continuous time, CPPI strategies only exhibit a shortfall risk in the presence
of discontinuous price paths (i.e., in jump models). Second, estimation risk is much
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch12 FA
more important in discrete time. Third and most important, practical implementations
are always based on discrete time setups.
Definition 12.1. Let us assume that we are given a discrete time arbitrage-free financial
market (, F, F, P) with time steps t and final horizon T . Further, assume that on
this market, there are I tradeable assets, e.g., stocks, bonds, or equity indices, with
prices {St }t=0,...,T , St ∈ R+
I
following some adapted (discrete time) stochastic process.
This process will be specified in more detail in Assumption 12.3.
Remark 12.2. Please note that the probability measure P represents the true real world
measure and should not be confused with some martingale measure. Furthermore, we
do not assume that the market is complete.
Assumption 12.3. For the remainder of this paper, we will assume that the discrete
one-period asset returns St are i.i.d. in time and follow a multivariate elliptically
contoured distribution, i.e.,
Si,t − Si,t−t
St ∼ E(µ, , ψ) with Si,t := (12.1)
Si,t−t
independent of t, with parameters µ ∈ R+
I
, ∈ SI+ , and characteristic generator ψ
with 2ψ (0) = −1. Further, we assume that St is bounded, i.e.,
St ∈ [µ − 1 , µ + 1 ], especially St > −11.
Remark 12.4. In Assumption 12.3, the one-period mean µ and covariances are
used to describe the distribution of St . However, if not explicitly stated otherwise,
all tables and figures are scaled to annualized numbers through division by t.
Remark 12.5. We need to point out that the usual assumption of multivariate nor-
mal (or student t) returns contradicts the assumption of St ∈ R+ I
. We have cured
this by restricting the distribution of St to [µ − 1 , µ + 1 ]. For practical purposes,
P[Si,t ≤ −1] is negligibly small anyway for sufficiently small t, even if we were
in a multivariate normal or student t framework.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch12 FA
As we want to focus on estimation risk, we assume in the following that the char-
acteristic generator ψ is known (i.e., no model risk is involved) but that the parameters
µ and are not given a priori. Hence, one has to rely on estimators for these quantities.
Therefore, we will only consider parameter uncertainty and no model uncertainty for
the ease of this exploration.
For the formulation of the standard CPPI strategy, we need to make some addi-
tional assumptions about the financial market. First, we introduce two riskless assets
in the financial market; a cash account and a zero bond. Second, we define what in
practice is known as a risky asset. Whereas the cash account and the zero bond enhance
the financial market by two additional assets, all risky assets will be completely repro-
ducible by the existing tradeable assets (see Formula (12.2) below) and do therefore
not enlarge the market.
We will see in the following that for the standard CPPI strategy, only the zero bond
Z is actually necessary, the cash account C will only become important when dealing
with the problem of inter-temporal risk budgeting or in case of general liabilities, see
Remark 12.20.
Remark 12.7. In practice, there are a few different ways how the asset allocation for
the risky asset can be determined:
• The most simple way uses the naïvely diversified portfolio, i.e.,
1 1 1 1
w= ,..., = 1, i.e., wi = .
I I I I
This portfolio is often used as a benchmark against which other methodologies are
tested and compared. The advantage of this choice is that it does not rely on any
estimate of future returns or volatilities (i.e., covariance matrix of asset returns).
• More sophisticated allocations are based on the traditional Markowitz optimization.
At the inception of the CPPI strategy (i.e., at t = 0), an asset allocation is chosen
from the efficient frontier based on mean–variance optimization. For this approach,
estimates for the expected returns µ and the corresponding covariances need to
be available. Given these estimates, superior investment results should be obtained
due to the usage of efficient portfolios. The most popular choices are
— The global minimum variance portfolio
— The maximum Sharpe ratio portfolio
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch12 FA
More details on the setup of the mean–variance optimization will be given in the next
subsection.
Besides these initially fixed asset allocations, dynamically updated allocations are
also possible within the framework of CPPI strategies. As these are also subject to the
same estimation issues as the above myopic portfolio choice, we later focus on this
simple myopic setup.
0.11
0.105
0.1
Excess return (p.a.)
0.095
0.09
Efficient frontier
0.085
Min. variance portfolio
Naive portfolio
Max. sharpe ratio portfolio
0.08
0.13 0.14 0.15 0.16 0.17 0.18 0.19 0.2 0.21 0.22
Volatility (p.a.)
Remark 12.10. Two portfolios have been of particular interest thus far:
(1) The minimum variance portfolio w∗cl (0), i.e., the portfolio at the left end of the
frontier where λ = 0
(2) The maximum Sharpe ratio portfolio — in this context — given by w∗cl (λS ) where
w∗cl (λ) µ
λS = arg max .
λ∈[0,1] w∗cl (λ)w∗cl (λ)
We have illustrated both portfolios together with the naïve portfolio in Fig. 12.1.
Remark 12.11. Under the above assumptions, problem (Pλ ) is equivalent to the clas-
sical Markowitz formulation
max w µ (MVσ )
w∈W
s.t. ww ≤ σ 2 ,
where σ and λ are in a monotonous one-to-one relationship. For an exact definition of
the equivalence, we refer to [32].
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch12 FA
for the first I tradeable assets S, the cash acount C, the zero bond Z, and the risky asset
X, resp.
Remark 12.12. In the following, the CPPI strategy will only invest in the zero bond
and in the risky asset, i.e., we have ϕtS = 0 as well as ϕtC = 0. Nevertheless, the
tradeable assets S are required to construct the risky asset X.
• Let us denote the difference in the holdings before and after trading, the traded
lots, as
ϕt := ϕt − ϕt−t .
Based on the riskless and the risky asset, we can define the main features of a CPPI
strategy. Its core idea lies in the dynamic allocation of capital between the riskless and
the risky asset. The proportion of the riskless asset in the portfolio is chosen in such
a way that a prespecified minimum level of wealth is guaranteed at time T . This final
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch12 FA
Let us start with a feasible initial value of the floor L0 , cf. (12.4). Then it holds V0 >
E0 = V0 − L0 > 0 and therefore
E0 V0 − E0 L0
ϕ0X = , ϕ0Z = = = Vmin .
X0 Z0 Z0
A closer investigation of the simple CPPI strategy yields that — independent of the
returns of the risky asset or the zero bond — the lot sizes ϕtZ remain constant over time:
ϕtZ = ϕ0Z = Vmin . Therefore, since the strategy is self-financing, it must also hold that
ϕtX = ϕ0X . This means that in its simple version the CPPI strategy does not need to
rebalance the portfolio — here we mean rebalancing between riskless and risky asset,
the risky asset itself may of course need some rebalancing.
Remark 12.15. In general, the above simple CPPI strategy is not very well suited to
achieve the returns exceeding the risk free rate. Table 12.1 illustrates that although
the underlying risky asset has a rather broad annual return distribution ranging at least
from −0.20 to 0.50, the annual return of the CPPI remains within tight bounds from
0.01 to 0.06. This is immediately clear from the property of the simple CPPI strategy
that it is just a buy-and-hold strategy between the zero bond and the risky asset, i.e., at
each time t the fraction fZ0 is invested in the zero bond, whereas only 1 − fZ0 is
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch12 FA
• The simulations of the discrete one-period asset returns are based on a normal
distribution N (µt, t) with µ = µexc + r. The normal distribution has been cut
off to guarantee Assumption 12.3 in the sense of Remark 12.5 simply by neglecting
any violating scenario.
Algorithm 12.17. Given all the necessary information at time t, the standard CPPI
strategy can be formulated as
(1) Compute the present value of the floor Lt = Zt · Vmin = f · Zt · V0 .
(2) Compute the cushion or risk budget Bt := (Vt − Lt )+ .
(3) Compute the leveraged exposure as Et := min(mBt , Vt ).
(4) Invest Vt − Et in the zero bond and Et in the risky asset, i.e.,
Et V t − Et
ϕtX = , ϕtZ = .
Xt Zt
Remark 12.18. In the above version, the exposure is bounded by the current portfolio
wealth, however, it is possible to introduce an upper boundary level 0 < b < 1 and to
restrict Et by Et = min(mBt , bVt ).
Remark 12.19. The cap of Et by Vt in the standard CPPI strategy restricts an invest-
ment in the risky asset beyond the wealth of the strategy, i.e., short positions in the
zero bond are prohibited. Note that this cap was not necessary in the simple version
of the CPPI strategy.
As we see from Fig. 12.2 and Table 12.2, the multiplier m has a significant influence
on the performance of the investment strategy. We can observe that the average return
E[VT /V0 − 1] increases with increasing m up to a certain level. From there onwards,
no significant increase in average performance is possible. We can also observe that
the standard deviation behaves accordingly.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch12 FA
0.18
0.16
0.14
0.12
0.1
0.08
0.06
0.04
Figure 12.2 Average return and standard deviation of the standard CPPI strategy on the naïve
risky asset with f = 0.95 for different multipliers m.
In contrast to the simple CPPI strategy, for m > 1, it might happen that VT ≤ LT ,
i.e., the floor is no longer guaranteed. This probability of a shortfall is investigated in
more detail in the next subsection. Please note that this cannot happen in continuous
time as long as the price paths are continuous, see [27].
In addition to the multiplier m, also the floor f has a significant impact on the
performance and the risk of the strategy. In Tables 12.3 and 12.4, an overview of the
dependence is given.
It can be observed that with increasing multiplier the expected return of the CPPI
strategy increases, since more capital can be invested in the risky asset which accounts
for higher returns. The same line of argumentation holds for the floor: the lower the
floor, the more budget for riskier investments and thus the higher the expected return.
Table 12.3 Expected Return of the Standard CPPI Strategy on the Naïve Risky
Asset for Different Multipliers m and Different Floors f .
Table 12.4 Standard Deviation of the Standard CPPI Strategy on the Naïve Risky
Asset for Different Multipliers m and Different Floors f .
Analogously, with larger investments in the risky asset (i.e., higher multiplier and/or
lower floor), the volatility also increases.
Tables 12.5 and 12.6 illustrate the interrelation of the shortfall probability and the
expected shortfall with the multiplier m and the floor f for a fixed risky asset (the
naïve one).
As expected, it can be observed that the shortfall probability increases with increas-
ing multiplier m. The floor influences the shortfall probability in a similar way: the
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch12 FA
Table 12.5 Shortfall Probability of the Standard CPPI Strategy on the Naïve Risky
Asset for Different Multipliers m and Different Floors f .
m=1 0 0 0 0 0 0
m=9 0 0 0 0 0 0
m = 11 0 0.02% 0.02% 0.02% 0.02% 0.02%
m = 13 0.08% 0.16% 0.26% 0.28% 0.32% 0.38%
m = 15 0.48% 0.70% 1.14% 1.78% 2.08% 2.32%
m = 17 1.82% 2.62% 3.88% 5.62% 6.60% 7.32%
m = 19 3.80% 5.78% 8.30% 12.00% 14.22% 16.00%
m = 21 6.68% 10.10% 14.22% 20.12% 24.08% 27.20%
m = 23 9.16% 13.74% 19.62% 28.22% 34.06% 38.52%
m = 25 11.64% 17.56% 24.94% 35.44% 43.26% 48.54%
m = 30 15.86% 24.04% 33.42% 46.94% 56.54% 63.02%
Table 12.6 Expected Shortfall of the Standard CPPI Strategy on the Naïve Risky
Asset for Different Multipliers m and Different Floors f .
m=1 — — — — — —
m=9 — — — — — —
m = 11 — −0.0015 −0.0011 −0.0008 −0.0006 −0.0004
m = 13 −0.0008 −0.0030 −0.0022 −0.0016 −0.0012 −0.0022
m = 15 −0.0030 −0.0026 −0.0029 −0.0022 −0.0018 −0.0018
m = 17 −0.0025 −0.0022 −0.0024 −0.0022 −0.0018 −0.0015
m = 19 −0.0026 −0.0023 −0.0023 −0.0021 −0.0019 −0.0017
m = 21 −0.0026 −0.0023 −0.0023 −0.0021 −0.0020 −0.0019
m = 23 −0.0027 −0.0025 −0.0025 −0.0022 −0.0022 −0.0021
m = 25 −0.0028 −0.0027 −0.0027 −0.0024 −0.0025 −0.0023
m = 30 −0.0034 −0.0033 −0.0034 −0.0031 −0.0031 −0.0031
closer the floor to the initial wealth, the higher the shortfall probability. Further, the
expected shortfall remains approximately constant.
Remark 12.23. In practice, the multiplier m is deduced from the consideration of the
shortfall risk by a Monte Carlo simulation. The multiplier is chosen as large as pos-
sible to obtain a sufficient average investment return, while keeping a pre-determined
level of shortfall risk, i.e., low shortfall probability and low expected shortfall. For this
purpose, practitioners usually run simulation studies like the above to obtain a suit-
able multiplier. It should be noted that there are analytical approximations available
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch12 FA
(see [27]), which link the multiplier to the volatility or percentiles of the one-period
return distribution. However, these approximations are only valid if Et = mBt , i.e., for
sufficiently small multipliers, and do not longer hold if Et is capped at Vt . In the latter
case, the approximation largely overstates the potential shortfall risk. Therefore, one
has to rely on Monte Carlo simulations.
0.11
0.105
0.1
Excess return (p.a.)
0.095
0.09
Efficient frontier
Min. variance portfolio
0.085
Naive portfolio
Max. sharpe ratio portfolio
Naive portfolio efficient
0.08
0.13 0.14 0.15 0.16 0.17 0.18 0.19 0.2 0.21 0.22
Volatility (p.a.)
0.35
Exp. return naive
Exp. return efficient
0.3 Vol naive
Vol efficient
Shortfall prob. naive
0.25 Shortfall prob. efficient
0.2
0.15
0.1
0.05
0
5 10 15 20 25 30
Multiplier (m)
Figure 12.4 Improvement of the standard CPPI (f = 0.95) by replacing the naïve portfolio
with an efficient one.
Table 12.8 Expected Return of Selected CPPI Strategies for Different Floors
f and multiplier m, Such that the Shortfall Probability is Below 1%.
already introduced canonical candidates from the previous section — the minimum
variance portfolio, the maximum Sharpe ratio portfolio, the naïve portfolio, and its
mean–variance efficient replacement. It can be observed that the three mean–variance
optimal portfolios clearly dominate the inefficient naïve portfolio, both in terms of
expected return and Sharpe ratio. Comparing the efficient portfolios with each other is
not as easy. Taking the point of view of the asset manager, the risk of the asset manager
is clearly the cost of shortfall. These costs are roughly given by the product of shortfall
probability and expected shortfall. With shortfall probability below 1% and expected
shortfall roughly equal to 30 basis points, these costs are below one basis point. On the
other hand, as the asset manager wants to offer a strategy with a high expected return,
the maximum Sharpe ratio portfolio is clearly the best choice. This portfolio offers
a similar return as the efficient version of the naïve portfolio, but possesses smaller
standard deviation, and should hence be preferred.
Remark 12.24. Although the above analysis looks very promising, a few details
should not be overlooked.
• The CPPI strategy needs to be implemented in the simulation framework as close as
possible to the real-world setup, as distinct features of such strategies may influence
the results of simulation studies. Among these are for example trading filters (mainly
directly linked to the traded lots ϕt ), fees (usually depending on Vt ), slippage
effects (depending on the traded volume vt ), ratchets, allowance for short positions,
additional risks from overlay management, planned cash injections, etc.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch12 FA
• The joint distribution of the zero bond and the risky asset has to be modeled rather
exactly. Especially the tail behavior plays a major role in the quantification of the
shortfall risk, whereas the estimation of the expected return is important for the
communication of the investment target return.
• The choice of the composition of the risky asset and thus the multiplier m are
subject to the same estimation risk. If the parameters of the return distribution are
uncertain, it is well known that the mean–variance efficient portfolios have a rather
poor performance (see next section). In this case, the originally valid idea of using
optimized portfolios may revert into the contrary.
Unfortunately, while the first point can be covered with sufficient accuracy, the esti-
mation of both the distribution and the distribution parameters are always subject to
uncertainty. As shown in [4], [7] or [16] mean–variance efficient portfolios strongly
depend on the input data. Due to the inherent uncertainty in the estimates, the resulting
portfolios as well as their estimated returns and volatilities are thus not reliable, see
for example [23].
Thus, a natural question in this context is: how much can the results differ in the
presence of estimation risk? As already mentioned in the introduction, we focus on
the estimation risk for the parameters and assume for simplicity that the distribution
family itself is known.
Remark 12.25. Bertrand and Prigent have considered a possible estimation of the
multiplier by extreme value theory in [3]. Although this research is probably the closest
one to this study, its line of thinking is solely concerned with shortfall risk. They
concentrate on the choice of the multiplier by looking at the corresponding percentiles
of the one-period return distributions by extreme value theory. Our study goes into a
different direction, as we simultaneously cover shortfall risk and expected return, but
do so via parameter uncertainty.
0.14
0.12
0.1
0.08
0.06
0.1 0.12 0.14 0.16 0.18 0.2 0.22
Volatility (p.a.)
frontier based on the random parameters µ̂ and . ˆ This situation will lead to the
numbers tagged perceived or random. Finally, the dotted line is the efficient fron-
tier the investor will actually obtain, i.e., optimizing the portfolios according to the
expected parameters µ̂ and ,ˆ but realizing the risk–return profile with the true market
parameters µ and . Table 12.10 summarizes the return and volatility characteris-
tics of the minimum variance and the maximum Sharpe ratio portfolios on the three
frontiers.
Therefore, in Fig. 12.5, two kinds of estimation risk can be noticed. First, the
asset allocation is calculated based on µ̂ and ˆ and thus return µ̂X and volatil-
ity σ̂X are obtained based on the optimized allocation. As shown in [17], these are
biased estimators compared to the optimal values µX and σX and on average it holds
that µ̂X > µX and σ̂X < σX . This also gets obvious from Fig. 12.5. Further, as
the wrong asset allocation is implemented, an inferior portfolio performance can be
achieved in the market, see also [17] for more details. This can again be observed
in Fig. 12.5, where the true original frontier dominates the portfolios calculated by
the asset manager. Although we have picked only one representative toy example,
in [17], it is shown that this represents the average situation. Taking the estimates µ̂
and ˆ together with the corresponding risky asset compositions, the simulation of
the standard CPPI can be performed as usual Table 12.11. From this, optimal mul-
tipliers are obtained for both portfolios. The multipliers have been chosen in such
a way that the corresponding shortfall probabilities are roughly 1%. This choice of
multipliers yields the following perceived performance and standard deviation, see
Tables 12.12 and 12.13. Based on these multipliers and the wrong allocation, we run
the simulation again, this time with the true market parameters. Then, instead of the
above figures, we obtain the corresponding actual figures, see Tables 12.14–12.16.
The most noteworthy differences are both in the expected performance and in the
shortfall probability. Although the shortfall probability was believed to be below 1%,
it significantly increased to about 5%. This is not only worrying from the point
of view of the asset manager, who actually faces higher costs to cover shortfall,
but as well from the perspective of the client, who has a much higher probability
of ending up in the absorbing state than intended. Even more troublesome is the
expected return of the formerly superior maximum Sharpe ratio portfolio. Due to
the error maximization, the expected return of this portfolio turned out to be unre-
liable and the gap to the perceived performance is significant. In contrast to this,
the minimum variance portfolio shows a rather stable behavior in terms of expected
return.
Based on this toy example, it is already obvious that a more prudent choice of
multiplier and a consideration of the estimation risk in the composition of the risky asset
is absolutely necessary. Therefore, in the next section, we will introduce a framework
which not only provides more robust asset allocations, but at the same time gives
prudent estimates of risk and return which can be used in the composition of the CPPI
strategy.
In the last years, several attempts have been made to improve the estimation of
mean–variance efficient portfolios. A meanwhile pretty well-known way to robus-
tify portfolios is the resampling technique introduced and made popular by Michaud,
see [23], based on concepts as, e.g., given in [16]. Another rather popular approach is
the usage of robust estimators, as, e.g., investigated in [18], [19] or [26]. The third line of
thinking is based on the robust counterpart idea introduced by Ben-Tal and Nemirovski,
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch12 FA
see [2]. Several authors have considered instances varying from theoretical to practical
settings, see e.g. [6], [13], [20], [31] and [33].
with U being the (joint) uncertainty set for the unknown parameters (µ, ). In analogy
to the classical setting, the optimal solution of (RPλ ) will be denoted by w∗rob (λ).
Remark 12.27. It can be shown that the resulting robust portfolios are unique for
0 ≤ λ < 1, see [29], Proposition 5.2, for general convex and compact uncertainty
sets U.
Most applications of the robust counterpart approach are based on statistical confi-
dence sets around the corresponding point estimates, see [2], [6], [13], [20], [22],
or [32]. In general, the most intuitive confidence sets coming from statistics are confi-
dence ellipsoids, as for example nicely motivated in [22], Sec. 2.4.3. Typically, the
center of the ellipsoid is determined as the estimated value of the point estimate
and its shape is described using an according estimate for the estimator’s covariance
matrix. In addition, the size of the uncertainty set is calculated from an appropriate
percentile to achieve the desired level of confidence. Alternatively, several different
expert opinions (or likewise different estimators) can be used to form an uncertainty
set, see [20]. In most practical instances, the robustification is focused on the uncer-
tainty in the mean, while the covariance matrix is assumed to be known with certainty,
which is well supported by a number of theoretical results on the impact of both
uncertainties.
likelihood estimator, although very popular, is not the only unbiased consistent esti-
mator, which may be used for elliptical distributions. In fact, there exist several other
common estimators for the mean of an elliptical distribution used by practitioners.
For example one could consider the following four consistent estimators for the return
besides the maximum likelihood estimator:
This means that on top of the investment risk, i.e., risk from the stochastic asset returns,
an additional penalty for estimation risk is considered in the optimization. This penalty
grows linearly with the uncertainty in the estimators, i.e., with δ, as well as linearly
in the importance λ of the return term. Although this is a very appealing approach
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch12 FA
in practical settings, for the aim of the study, it is easier to work with an alternative
specification of uncertainty sets.
U1 := {r | (r − µ̂) ˆ
ˆ −1 (r − µ̂) ≤ δ2 } × {}, (12.6)
where δ is chosen appropriately to the target confidence level of the uncertainty set.
Again, no uncertainty of the covariance matrix is considered.
A more advanced version of the confidence set also considers the uncertainty of
the covariance matrix. In the latter case, applying similar ideas as in [22], it is possible
to show (see [32]) that the natural confidence ellipsoid around the joint maximum
ˆ approximately has the following form
likelihood estimator for both (µ̂, )
−1
U2 = (r, C) | (r − µ̂) ˆ −1 (r − µ̂) + 2 ·
ˆ 2 (C − ) ˆ − 12 2 ≤ δ2 .
ˆ (12.7)
tr
Using any one of the above uncertainty sets a surprising result can be proved by a close
investigation of the necessary and sufficient first-order optimality conditions (i.e., the
KKT conditions), see [28] or [32]: the robust efficient frontier is exactly the classical
efficient frontier up to a certain risk (i.e., volatility) level depending on the choice
of the parameter δ. In other words, the original classical efficient frontier is robust
in itself, as long as we do not move too far to the right. This can be interpreted as
if robustification is able to identify the unreliable upper part of the efficient frontier
which is cut off. The following theorems, taken from [32] formally state the described
result. Let
δ
H(θ) = 1 − θ + θ √ ,
S
2 κ
K(θ) = max (1 − θ) 1 + δ (1 − κ) + θδ ,
κ∈[0,1] S−1 S
where in this context S denotes the number of observations used for the point estima-
tion. Although K is given in terms of an optimization problem, it is computed rather
easily by any one-dimensional optimization routine.
Theorem 12.28. Let U = U1 as in (12.6) and let θ ∈ [0, 1]. Then the optimal solution
w∗rob (θ) of (RPθ ) equals the optimal solution w∗cl (λ) of the original problem (Pλ ) for
λ =
H (θ) = θ/(θ + H(θ)).
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch12 FA
Theorem 12.29. Let U = U2 as in (12.7) and let θ ∈ [0, 1]. Then the optimal solution
w∗rob (θ) of (RPθ ) equals the optimal solution w∗cl (λ) of the original problem (Pλ ) for
λ =
K (θ) = θ/(θ + K(θ)).
• The cut off risk level for the robust frontier can be determined by solving the classical
formulation with
1
λmax =
H (1) =
K (1) = √ < 1.
1 + δ/ S
The optimal portfolio w∗cl (λmax ) then determines the cut off risk level as
σmax = w∗cl (λmax )w ˆ ∗cl (λmax ).
• Considering the special point θ = 0, i.e., the (robust) minimum variance portfolio,
the corresponding λ is also zero, i.e., the minimum variance portfolio is already
robust in itself. For U1 the robust counterpart reduces to
min ww,ˆ
w∈W
which is obvious as the minimum variance portfolio does not depend on the uncertain
parameter µ. If uncertainty of the covariance matrix is considered, then the robust
counterpart reduces to
2
ˆ
min K(0) w w = min 1 + δ
ˆ
w w.
w∈W w∈W S−1
for the uncertainty set U2 . This means that the consideration of the uncertainty
of the covariance matrix is reflected in the factor K(0). Since this factor is larger
than 1, a higher volatility (i.e., a larger objective value) is expected for the minimum
variance portfolio in the robust setting. This increase in the risk can be interpreted
as a measure for estimation risk and can be incorporated to obtain an improved
estimation of the multiplier m.
• The results of both theorems hold true due to the fact that the matrices for measuring
the portfolio’s risk and for describing the shape of the uncertainty set U have the
same structure and hence the expressions for the risk and the penalty for estimation
risk can be combined. Otherwise we would end up as in the previous subsection,
where different penalty terms for estimation risk appear in the objective function.
The effect of such a robustification for an investor is that her position on the efficient
frontier moves to the left, toward the minimum variance portfolio. Hence, robustifi-
cation leads to more conservative portfolio choices that are nevertheless efficient —
simply with respect to a different trade-off parameter.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch12 FA
(1) Calculate the robust efficient frontier according to the previous section. Choose
an arbitrary portfolio w from the robust frontier, preferably the robust minimum
variance or the robust maximum Sharpe ratio portfolio.
(2) Determine the robustified parameters
1
µ̃ = µ̂ − δ √ ˆ
w, ˆ
˜ = K(0)2 · .
ˆ
S ww
(3) Run the CPPI strategy with the robustified parameters and fix the multiplier m as
usual, i.e., at an acceptable level of shortfall risk.
The choice K(0) is deduced from the fact that this factor appears in the robustified
version of the minimum variance problem and represents the penalty for the uncertainty
in . For all other portfolios, K(λ) is smaller and could be calculated, but for simplicity,
we use K(0) instead. The correction of the expected returns is motivated by the fact
that the worst-case return in the confidence set is given by µ̃, although this is a rather
pessimistic approach which may need some refinement.
If we apply Algorithm 12.31 to the previous setting for realistic choices of robusti-
fication parameters, δ = 1 and S = 52, we obtain µ̃ and K(0) as in Table 12.17. It can
be observed that the robustified return µ̃ is overly pessimistic and cannot be used to
estimate the expected performance of the portfolios or the CPPI strategy. Nevertheless,
it is still well suited for the determination of the multiplier m, as for this purpose, an
adverse market setting leads to a conservative choice. Further, the volatility plays the
major role in the setting of the multiplier anyway.
Using the robustified allocations for the risky assets and running the CPPI simu-
lation with the robustified parameters, we get the following multipliers, illustrated in
Table 12.18. The multiplier was again chosen in such a way that the corresponding
shortfall probability is below 1%. Comparing these multipliers to those in Table 12.11,
we see that the above multipliers are significantly smaller than those derived from the
estimated, but non-robustified, parameters.
Remark 12.32. From the illustration of the robust frontier, see Fig. 12.6, it can be
noticed that the robust frontier is usually rather short. This means that large deviations
from the minimum variance portfolios are usually not possible within the robust frame-
work. Therefore, we do not expect large differences between the (robust) minimum
variance and the robust maximum Sharpe ratio portfolio.
0.11
0.105
0.1
Excess return (p.a.)
0.095
0.09
0.085
Efficient frontier
Robust efficient frontier
0.08
0.13 0.14 0.15 0.16 0.17 0.18 0.19 0.2 0.21 0.22
Volatility (p.a.)
Figure 12.6 Illustration of the classical and the robust efficient frontier.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch12 FA
shortfall probability for the robust minimum variance portfolio compared to the non-
robust one. In case of the CPPI strategies based on maximum Sharpe ratio portfolios,
the expected returns of the robust versions are significantly higher compared to the
non-robustified risky asset allocations and much closer to the optimal ones. However,
even with robustification, CPPI on maximum Sharpe ratio portfolios actually can-
not beat those on the minimum variance portfolio, which in turn still beats the naïve
strategy. This means that portfolio optimization still adds value to the investment
process, but that estimation risks have to be taken very seriously. For completeness,
Table 12.23 contains the actual standard deviations of selected non-robust and robust
CPPI strategies. As expected, all robust versions have a lower standard deviation than
their corresponding non-robust strategy.
12.5. CONCLUSION
In this study, we have shown that under full information CPPI strategies should rely on
mean–variance efficient portfolios as risky assets. Based on Monte Carlo simulations,
suitable multipliers can be found. We have demonstrated that estimation risk may
have a significant impact on the performance of the CPPI strategy. A potential remedy
by robust mean–variance optimal portfolios seems to be beneficial for the overall
performance of the strategy, especially for the shortfall risk of the asset manager and
the average return. Nevertheless, our results also show that optimization beyond the
minimum variance portfolio is heavily influenced by estimation risk. Based on these
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch12 FA
preliminary results, it can be concluded that a more detailed and thorough analysis
needs to be carried out to investigate estimation risk and its cures in a broader context.
Acknowledgments
The authors want to express their gratitude to the editors for the opportunity to make this
contribution. Further, the second author appreciates the hospitality at the HVB Insti-
tute for Mathematical Finance, TU München, especially the inspiring annual research
seminars which fostered part of the research on robust allocations.
Disclaimer
The views expressed in this article are the authors’ personal opinions and should not be
construed as being endorsed by MEAG MUNICH ERGO AssetManagement GmbH
or Hypo Real Estate Holding AG.
References
[13] Goldfarb, D and G Iyengar (2003). Robust portfolio selection problems. Mathematics of
Operations Research, 28(1), 1–38.
[14] Huber, P-J (1981). Robust Statistics. New York: John Wiley & Sons.
[15] Ingersoll, J-E (1987). Theory of Financial Decision Making. Maryland: Rowman &
Littlefield.
[16] Jorion, P (1992). Portfolio optimization in practice. Financial Analysts Journal, 48(1),
68–74.
[17] Kan, R and DR Smith (2008). The distribution of the sample minimum–variance frontier.
Mangement Science, 54(7), 1364–1380.
[18] Kempf, A and C Memmel (2003). On the estimation of the global minimum variance
portfolio. SSRN Working Paper.
[19] Lauprête, GJ,AM Samarov and RE Welsch (2002). Robust portfolio optimization. Metrika,
55(2), 139–149.
[20] Lutgens, F (2004). Robust Portfolio Optimization. PhD Thesis, Maastricht University.
[21] Markowitz, H (1952). Portfolio selection. Journal of Finance, 7(1), 77–91.
[22] Meucci, A (2005). Risk and Asset Allocation. Berlin: Springer.
[23] Michaud, RO (1998). Efficient Asset Management. Boston: Harvard Business School Press.
[24] Perold, A (1986). Constant portfolio insurance. Unpublished manuscript.
[25] Perold, A and W Sharpe (1988). Dynamic strategies for asset allocation. Financial Analysts
Journal, 44, 16–27.
[26] Perret-Gentil, C and M-P Victoria-Feser (2004). Robust mean–variance portfolio selection.
FAME Research Paper no. 140.
[27] Prigent, J-L (2007). Portfolio Optimization and Performance Analysis. Chapman & Hall.
[28] Scherer, B (2007). Can robust portfolio optimisation help to build better portfolios? Journal
of Asset Management, 7(6), 374–387.
[29] Schöttle, K (2007). Robust Optimization with Application in Asset Management. Disser-
tation Technische Universität München.
[30] Schöttle, K and R Werner (2005). Costs and benefits of robust optimization. Working Paper.
[31] Schöttle, K and R Werner (2006). Towards reliable efficient frontiers. Journal of Asset
Management, 7(2), 128–141.
[32] Schöttle, K and R Werner (2009). Robustness properties of mean–variance portfolios (to
appear in: Optimization).
[33] Tütüncü, RH and M Koenig (2004). Robust asset allocation. Annals of Operations
Research, 132(1–4), 157–187.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch13 FA
13
ROBUST ASSET ALLOCATION
UNDER MODEL RISK
In this chapter, we propose a robust asset allocation methodology, when there is some
ambiguity concerning the distribution of asset returns. The investor considers several
prior models for the assets distribution and displays an ambiguity aversion against
them. We have developed a two-step ambiguity robust methodology that offers the
advantage to be more tractable and easier to implement than the various approaches
proposed in the literature. This methodology decomposes the ambiguity aversion into
a model-specific ambiguity aversion as well as relative ambiguity aversion for each
model across the set of different priors. The optimal solutions inferred by each prior are
transformed through a generic absolute ambiguity function ψ. Then, the transformed
solutions are mixed together through a measure π that reflects the relative ambiguity
aversion of the investor for the different priors considered. This methodology is then
illustrated through the study of an empirical example on European data.
327
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch13 FA
13.1. BACKGROUND
Note that in this framework, even if the agent has several priors as reference, he/she
does not show any aversion toward the co-existence of different models to take her
decision. He or she is neutral toward model uncertainty as there exists no ambiguity
about the set of models considered and their likelihood to occur.
However, as shown by Ellsberg in [3], the decision makers show a more averse
behavior when betting on events for which outcomes are ambiguous (i.e., when there
is also some uncertainty regarding the underlying model) than when betting on events
for which the outcomes are only risky (i.e., the underlying model is well known).
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch13 FA
A financial illustration of the Ellsberg paradox is the risk premium paradox: investors
tend to invest more in their local market even though the expected return is lower than
for foreign markets, the reason being that the investors add an ambiguity premium
to risky assets (investors prefer investing in assets located in their geographical zone,
because they can better apprehend their return distribution).
For this reason, the SEU framework fails to take into account this additional
source of aversion of investors. Various approaches have been developed in the liter-
ature to take into account this aversion toward model uncertainty in the investment
decision process. Among them, Gilboa and Schmeidler in [4] proposed a min–max
approach leading to a very conservative decision rule based upon the worst case
model. More recently, Klibanoff et al. in [5] introduced a generalized model. They
consider an increasing, concave transformation function characterizing the investor
ambiguity aversion through a parameter γ. The optimal weights vector is then deter-
mined as
φ∗ ≡ argmax {EQ [u(Xφ , λ)], γ}π(Q). (13.3)
φ
Q∈Q
The main feature of this model is that it unifies all the previous approaches accounting
for model ambiguity. However, this theoretical approach can be very challenging to
implement in practice for different reasons, including the calibration of the various
parameters. Indeed, no distinction is made between specific ambiguity aversion for a
given model (“How good is this specific model to represent the reality?”) and general
ambiguity aversion for the whole class of models (“How much can all the models
explain the reality?”). Moreover, solving explicitly Program (13.3) can be extremely
difficult, even numerically, especially in the multi-dimensional case or when adding
some constraints on the portfolio allocation. Note that Klibanoff et al. only give a
simple numerical example for a portfolio with three assets in their paper [5], whereas
practitioners often consider portfolios with hundreds of assets. We have compared their
example to our methodology in [6] and we also provide a more complex theoretical
example that can be solved in close form if using our methodology. Finally, such an
approach lacks some flexibility in the sense that if the investor considers a new model,
he or she has to re-compute the program entirely. To overcome those limitations, we
propose a robust, general framework for decision making under uncertainty. We are
not aiming at finding the optimal solution for a given criterion but more at finding a
robust solution.
We propose a new approach to model ambiguity that is altogether more flexible, easier
to compute, more robust and tractable than the methods proposed in the literature. This
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch13 FA
• Absolute ambiguity: this refers to the ambiguity the investor has for a given model.
We first solve the optimization problem assuming that the model considered is the
true model. We thus compute a distorted expected value of the deduced optimal
weights, transformed by an Absolute Ambiguity Robust Adjustment (AARA) func-
tion denoted ψ. Note that the absolute adjustment is made on the solution and not
on the choice criterion. This allows us for some additional flexibility in the use of
each model.
• Relative ambiguity: this expresses the relative ambiguity the investor has among his
or her different models: in a second step, we aggregate the adjusted optimal weights
computed for each model through a RelativeAmbiguity RobustAdjustment (RARA)
function, denoted π.
Let us denote φQ ≡ argmaxφ EQ [u(Xφ , λ)]. Then the ARA portfolio allocation
φARA ≡ (φiARA )i∈{1,...,N} is obtained as
φiARA ≡ ψ{φQi , γ}π(Q), i ∈ {1, . . . , N} (13.4)
Q∈Q
N
and φ0ARA ≡ 1 − i=1 φiARA .
Due to its specific nature, the risk-free asset has no model risk associated with
it (its future value is known with certainty). Therefore, it plays a specific role in the
ambiguity-adjusted optimal asset allocation. It can be assimilated to a refuge value in
the following sense: the more the investor is averse to ambiguity, the more he or she
will invest in the risk-free asset. In this sense, as the “disinvested” part of the wealth
from the risky assets is transferred to the risk-free asset, the adjusted weight of the
risk-free asset corresponds to the amount of money the investor is reluctant to invest
in risky assets because of her aversion toward model risk.
We describe below the characteristics of the functions ψ and π.
The function ψ must satisfy some key properties to be consistent with the ratio-
nality of the investor:
• Definition of the function ψ: The investor treats the absolute ambiguity aversion
with the same type of transformation across all the different models (ψ is the same for
all the models). What distinguishes the absolute ambiguity aversion transformation
across the models is the specific ambiguity aversion parameter γ Q the investor
attributes to each model. As the optimal weights obtained for each model φQ are
bounded by 1, ψ(1, γ Q ) represents the maximum weight the investor will assign
to any asset after the AARA transformation. We note aQ ≡ ψ(1, γ Q ) and a ≡
maxQ∈Q aQ . Therefore, ψ is defined on the set of optimum model dependent weights
[−1; 1] × Q onto a set [−a; a] of transformed weights. Note that the investor can
set the value of aQ and deduce the value of γ Q depending on the explicit form he or
she chooses for ψ.
ψ : [−1; 1] × Q → [−a; a]
(13.5)
(φ, Q): → ψ(φQ , γ Q )
The following properties apply for the risky assets only, i.e., for i ∈ {1, . . . , N}.
The function ψ is then an odd function symmetric around zero. We can assume that a
long-short investor has the same aversion to positive or negative weights of the same
absolute value: ∀φi ∈ [−1; 1], ψ(−φi ) = −ψ(φi ). The AARA function penalizes
the scale of the optimal weights of a given model without discriminating between
negative and positive weights.
• Invariant point: There is no ambiguity aversion for a zero weight: ψ(0) = 0. If
the model Q assigns no weight on a given asset, the transformation ψ should not
modify the “neutrality” of the model Q toward this asset.
• Limit behavior: When the investor is infinitely averse to ambiguity, it will prevent
her from trading as he or she trusts none of his or her models, and therefore all
the portfolio weights should be defaulted to zero. On the contrary, if the investor
is neutral to ambiguity, the function ψ should leave the model-dependent weights
invariant.
We used a very similar function than the one applied by Klibanoff et al. to account
for ambiguity. The function ψ can be any classical S-Shape function, which has the
nice property of being concave (convex for negative values), symmetric, and mono-
tonic (similar attributes as for classical utility functions). What really characterizes the
function ψ is its ambiguity parameter γ that accounts for the concavity of the function
ψ and therefore the ambiguity aversion of the investor (Fig. 13.1).
After the transformation through the AARA function ψ and the RARA function π, the
ARA weight of any risky assets is defined as
∀i ∈ [1; N], φiARA = ψ(φiQ , γ Q )π(Q)
Q∈Q
Here we see that the measure π is relative, as it takes into account all the different models
performances, whereas γ is absolute, as it only considers the performance of a given
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch13 FA
model. Note that the absolute ambiguity parameter is −γ rather than γ, according to the
shape of the function ψ given in Sec. 13.2.1. Both the relative ambiguity parameter π
and the absolute ambiguity parameter −γ are proportional to the performance measure
considered: the higher the performance measure, the higher the parameters. Note that
it is for the sake of simplicity that we chose to parametrize similarly π and γ, what
matters is that the absolute and relative ambiguity aversion be positively correlated
with the performance measure considered.
To compare different asset allocations for different models, and also the impact of
the ambiguity aversion to the different weights assigned to each asset, we present in
the following section some properties and give some definitions. We also propose a
measure of distance between two different asset allocations that we will use to compare
different asset allocations in our empirical example developed in the last section. The
Ambiguity Robust Adjustment refers to the combine adjustment by the AARA to the
optimal weights independently computed for each prior and by the RARA performed
to combine those adjusted weights. These measures can be used to describe how to
measure the ambiguity aversion of an investor toward a model in general, and more
specifically toward a single asset.
For the following definitions, let φQ be the asset allocation conditional on model
Q ∈ Q and φARA be the ARA asset allocation.
Definition 13.2 (Portfolio Distance). Let us consider two models Q1 and Q2 in the
set of priors Q. We define the distance measure δ between the two models as
N
δ(φQ1 , φQ2 ) = |φiQ1 − φiQ2 |
i=0
δ(φQ1 , φQ2 ) represents the turnover value to rebalance the investor’s portfolio from the
asset allocation φQ1 to the asset allocation φQ2 .
assets Si and Sj ,j, i ∈ [1, N], if WAAi < WAAj , we say that the investor shows more
ambiguity aversion toward the asset j than toward the asset i.
AAA(Q) represents the theoretical turnover to rebalance the investor’s asset allocation
on the risky assets i = 1, . . . , N from the optimal weights obtained if the prior Q is
assumed to be equal to the unknown real model P to the Absolute Ambiguity Adjusted
portfolio, taking into account the investor absolute aversion against his or her prior Q.
We deduce that the total value of the investor’s Weighted Ambiguity Adjustment toward
all her priors is
N
AAA(Q) ≡ |φiQ − ψ(φiQ , γ Q )|π(Q).
Q∈Q i=1
Definition 13.5 (Relative Ambiguity Adjustment). We define the value of the Rel-
ative Ambiguity Adjustment (RAA) of an investor as
N
ARA Q
RAA(Q) ≡
φi − φi π(Q) ,
i=1 Q∈Q
which represents effectively the turnover between the Robust Ambiguity Portfolio and
the Subjective Expected Utility Portfolio (the SEU portfolio allocation is defined as:
φiSEU ≡ Q∈Q φiQ π(Q)).
In this section, we detail the results of some tests we run to evaluate the performance
of ambiguity robust portfolios, compared to other classical optimized portfolios (we
consider optimized portfolios as well as other type of factor model portfolios, and
therefore we mainly focus on the ambiguity transformation and not on the initial
optimization under each prior). The tests were run on European assets. We collected
daily close prices from January 2000 to April 2008 (equivalent to 2167 business days).
Our set of securities is the set of Eurostoxx 600 constituents that were trading across the
whole period. For this empirical study, we consider a null risk-free rate (the cash is not
remunerated), as we consider a portfolio total return, as would be done for strategies
without benchmarks such as hedge funds or proprietary groups strategies. We assume
that our transactions costs, fees, and slippage correspond to three basis point of the daily
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch13 FA
turnover of the strategy (i.e., the total costs and fees are assumed to be in percentage
three basis point of the daily turnover, for instance if the turnover of a strategy at a
given date is 50% and its return is 10 basis points, the costs and fees amount to 1.5
basis points return and therefore the startegy return for this day after cost is 8.5 basis
points). We run a back test on historical data when the investor re-balances daily her
portfolio by re-estimating the different models and re-setting her optimal investment
weights over an estimation window of 120 days (we choose a daily rebalancing, as
we consider daily returns for the asset time series). The performances of the different
strategies were then evaluated. We find that our Ambiguity Robust approach allows
the investor to achieve superior performance compared to that achieved by classically
optimized portfolios.
• The equally weighted portfolio (EW): gives an equal weight to all the risky assets.
We define the EW portfolio asset allocation as
1
∀i ∈ {1, . . . N} : φtEW,i = .
N
• The minimum variance portfolio (MN): is the fully invested Markowitz effi-
cient portfolio with minimum variance, obtained when the investor minimizes the
expected variance of the portfolio (i.e., minφ φ φ, where represents the covari-
ance matrix of the stock returns). The MN portfolio allocation is defined as
−1
t−w,t−1 1
φtMN =
1 −1
t−w,t−1 1
where t−w,t−1 is the empirical covariance matrix estimated over the window [t −
w, t − 1] and 1 is the N-vector of ones.
• The mean variance portfolio (MV): is the fully invested, maximum Sharpe, mean–
variance Markowitz efficient portfolio, obtained when the investor maximizes the
empirical quadratic expected utility (i.e., maxφ µφ −φ φ, where µ is the empirical
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch13 FA
mean and the empirical covariance matrix of the stock returns, note that we
consider a risk aversion equal to 1). The MV allocation is defined as
−1
t−w,t−1 µt−w,t−1
φtMV = ,
1 −1
t−w,t−1 µt−w,t−1
where t−w,t−1 is the empirical covariance matrix estimated over the window [t −
w, t − 1] and µt−w,t−1 is the empirical vector of mean returns estimated over the
same window.
• The CAPM portfolio weights (CA): we base our CAPM portfolio on the Jensen
alphas. We estimate the CAPM betas over the estimation window. Consider-
M
ing rt−w:t−1 as the vector of the Eurostoxx 600 market returns over the period
[t − w, t − 1], the beta of the risky asset i is therefore estimated at time t as
i M
cov(rt−w:t−1 , rt−w:t−1 )
βti ≡ M
.
var(rt−w:t−1 )
We then compute the Jensen alpha as the difference between the observed return at
time t of the asset i and the beta-adjusted market return:
αit ≡ rti − βti rtM .
We then define the CAPM weights as the weighted average alphas across all the
risky assets considered:
αi
φtCA,i ≡ N t j
.
j=1 αt
• The CAPM uncertain portfolio (UC): we define the UCAPM weights as the CAPM
weights adjusted by the variance of the CAPM residuals:
αit
var(αi. )
φtUC,i ≡ j .
N αt
j=1 var(αj. )
More formally, we compute the relative financial ratio FR of the asset i among its
peers assets j in the sector S:
αFR,i,S
t ≡ µ(FR.,S
t ) − FRt ,
i
where µ(FR.,S t ) is the average of the given financial ratio among the assets in the
sector S. And therefore we get the following asset allocation, by scaling the alphas
per sector, so that the investor invests the same amount of money in each sector:
αFR,i
t
φtFR,i ≡ N FR,j,S
,
S
j=1 |αt |
where FR stands either for PE, CF, or PB and NS is the number of assets belonging
to sector S.
• The subjective expected utility portfolio (SEU): we define by π the vector of
Relative Ambiguity Aversion Weights given to the models considered. In the present
empirical example, we assume that Q∈Q π(Q) = 1, therefore the SEU portfolio
is defined as the different models weights weighted by π.
∀i ∈ {1, . . . , N} : φtSEU,i ≡ φQ,i π(Q).
Q∈Q
• The ambiguity robust portfolio (RA): finally, the optimal ambiguous portfolio
is defined as the different models weights adjusted by the Absolute Ambiguity
Adjustment ψ and weighted by the vector of Relative Ambiguity Aversion
Weights π:
∀i ∈ {1, . . . , N} : φtARA,i ≡ ψ(φQ,i , γ Q )π(Q).
Q∈Q
• Sharpe ratio (SHR): that represents at time t the ratio of the empirical mean return
of a portfolio over its empirical standard deviation over the period [t − w, t − 1].
We denote:
φ
φ µ(rt−w,t−1 )
SHRt ≡ φ
.
σ(rt−w,t−1 )
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch13 FA
• Sortino Price ratio (SOR): that represents at time t the ratio of the empirical mean
return of a portfolio over the empirical standard deviation of its negative returns
over the period [t − w, t − 1]:
φ
φ µ(rt−w,t−1 )
SORt ≡ φ φ
.
σ(rt−w,t−1 {rt−w,t−1 < 0})
• Gain Loss ratio (GLR): which is the ratio of total positive returns over total negative
returns:
φ
φ {rt−w,t−1 > 0}
GLRt ≡ φ .
{rt−w,t−1 < 0}
• Winner Loser ratio (WLR): similar to the Gain Loss ratio, it is the ratio of the
number of total positive returns over the number of total negative returns:
φ φ
φ rt−w,t−1 {rt−w,t−1 > 0}
WLRt ≡ φ φ
.
rt−w,t−1 {rt−w,t−1 < 0}
The following two measures are used to compare different portfolio performances.
φ
N
T/Ot ≡ |φti − φt−1
i
|.
i=0
13.4.3. Results
First we describe the individual performances of the eight models considered. Then, we
show how the Subjective Expected Utility portfolios of Savage (as defined in Eq. (13.2)
and where the single strategies are linearly weighted by the weighted average of their
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch13 FA
different performance measures) outperform the individual strategies. We use the SEU
portfolios as benchmarks for our ARA methodology, as the Klibanoff et al. model is
almost impossible to compute for a large universe of assets. Finally, we display the
performances of the ambiguity robust portfolios, parametrized by the four different
performance measures considered. We conclude that the ambiguity robust portfolios
outperform by far the non-ambiguous SEU portfolios.
UC PE CF PB
EW MN MV CA
performance measure of the different strategies over the historical returns between date
t − 120 and t − 1. Even though the strategies UC and CA returns underperform the EW
strategy return over the whole period considered (the equally weighted portfolio has the
biggest total return of almost 70%), there exist some periods where the reverse is true
(in 2002 for instance, the CA and UC portfolios perform better). This confirms that the
relative weight of each models should be dynamic, as proposed in our ambiguity robust
methodology. In order to have numbers of similar magnitude, we scale the inverse
Sharpe and Sortino measures by a reference ratio of three annualized: in practice, it
means that an investor considers a portfolio for which the return is three times as big
as its risk as benchmark. We have also capped the performance measures by three
to prevent a strongly performing model dominating the others. Formally, we have:
SHR = min(max(0, SHR 3
), 3) and SOR = min(max(0, SOR 3
), 3).
performance measures: Sharpe, Sortino, Gain Loss and Win Lose ratios). In Table 13.3,
we display the statistics of the Ambiguity Robust portfolios estimated with the γ and
π measures. The RA portfolios outperformed the best CA strategy portfolio in terms
of all four different performance measures considered.
The Absolute Ambiguity Aversion (AAA) and the Relative Ambiguity Aversion
(RAA) measures allow us to quantify our ambiguity adjustment. The biggest ambiguity
adjustments are made for the less performing strategies (based on Win Lose and Gain
Loss ratios). As we can see, the Ambiguity Robust Portfolios outperform all the SEU
portfolios, meaning that the ψ adjustments enhance the performance of ambiguity
averse investors portfolios.
13.5. CONCLUSION
Our aim was to provide the reader with a simple and easy to implement methodology
for practitioners who want to allocate in a robust way their portfolio of assets when
they have several models for the asset returns distributions but they are ambiguous
about each of them (i.e., they do not fully trust them). We propose a simple, robust,
and systematic method of combining the different weights computed conditionally per
model. The concrete methodology we develop enables to account for ambiguity via
practical empirical measures, such as the performance measures of each conditional
portfolio.
Our method is very different from the approach proposed in the literature and in
particular that of Klibanoff et al. as we do not proceed to a unique complex optimization
that is extremely challenging to solve in practice. Our approach is more practical and
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch13 FA
very industry-oriented: practitioners can easily define the optimal weights for each of
the models they consider, then they can mix those prior conditional weights, taking into
account the absolute and relative ambiguity they have against each model, as reflected
by the concavity of ψ and the weight π(Q) attributed to each model Q.
The parametrization we propose for the Absolute Ambiguity Aversion and the
Relative Ambiguity Aversion in our empirical example is by no means optimal. How-
ever, we have shown that it can greatly enhance the performance of some classical
portfolio strategies. The proposed robust ambiguity portfolio is more stable, with a
lower turnover, and less risky, with a higher certainty equivalent ratio, than unadjusted
portfolios as well as SEU portfolios. The ARA methodology smoothens and reduces
the risk of classical strategies. One of the main features of the ARA methodology is to
provide a dynamic adjustment for model uncertainty. As the parametrization depends
on the past performance of the strategy, it would be interesting to study how the ARA
model would behave in major downturns, and adapt to different market regimes.
References
14
SEMI-STATIC HEDGING
STRATEGIES FOR EXOTIC
OPTIONS∗
In this chapter, we give a survey of results for semi-static hedging strategies for exotic
options under different model assumptions and also in a model-independent frame-
work. Semi-static hedging strategies consist of rebalancing the underlying portfolio
only at certain pre-specified timepoints during the lifetime of the hedged derivative,
as opposed to classical dynamic hedging, where adjustments have to be made con-
tinuously in time. In many market situations (and in particular in times of limited
liquidity), this alternative approach to the hedging problem is quite useful and has
become an increasingly popular research topic over the last years. We summarize the
results on barrier options as well as strongly path-dependent options such as Asian or
345
May 19, 2010 16:20 WSPC/SPI-B913 b913-ch14 FA
lookback options. Finally, it is shown how perfect semi-static hedging strategies for
discretely observed options can be developed in quite general Markov-type models.
14.1. INTRODUCTION
In the famous work [10] of 1973, it was shown how a standard European option can
be replicated by dynamically trading the underlying asset and investing in the riskless
bond. The so-called delta hedging strategy is (with some modifications) still among
the most widely used methods to manage and reduce the risk inherent in writing
an option. However, already in [60] static “hedging” methods were applied to relate
different option prices, and also the Put–Call–Parity, which at least goes back to [76],
can be interpreted as a consequence of a particular static hedging strategy. Recently,
in [39], some pitfalls of the dynamic replication strategies were outlined and it was
argued in favor of the more robust static replication methods, which in the meantime
attained quite some attention in the practical and academic literature, starting in the
mid-1990’s with [14, 37].
The semi-static hedging approach structurally differs from the dynamic counter-
part in terms of hedging instruments and trading times. While dynamic hedges base
on the assumption of being able to trade continuously in time, semi-static portfolios
do not need to be adjusted dynamically but only at pre-specified (stopping) times.
Furthermore, in contrast to classic dynamic hedging, they typically use plain vanilla
(or other liquidly traded) options as hedging instruments. Studies comparing the per-
formance of semi-static and dynamic hedging strategies (as for example undertaken
in [43, 64, 79]) indicate that the semi-static ones are in many situations more robust
and able to outperform their dynamic counterparts.
Here, we aim to give an overview of the most prominent semi-static hedging
strategies for options written on some asset (e.g., a stock or an exchange rate), for
which some standard European options are liquid. In this chapter, we focus on fully
discrete strategies, i.e., we exclude the possibility of dynamic hedging (for combined
dynamic-static hedging strategies see [52, 53]). In principle, the outlined methods
directly translate to the case where the options are written on baskets of assets, but one
has to be aware that plain vanilla options with the basket as the underlying are needed
to apply the strategies in practice.
Strategies, for which the composition of the portfolio is not changed after the
initial composition except investing gains in the riskless asset are called static, and
those with a finite number of transactions are termed semi-static (semi-static strategies
can be further subclassified as done in [55]).
We also distinguish between model-dependent and model-independent strategies.
The former depend on assumptions on the asset price process (like continuity) or need
May 19, 2010 16:20 WSPC/SPI-B913 b913-ch14 FA
a fully specified model, while the latter are correct in any market, which is frictionless
in the following sense:
Definition 14.1 (Frictionless market). A market is called frictionless, if
• The no-arbitrage assumption holds.
• All investors are price-takers.
• All parties have the same access to relevant information.
• There are no transaction costs, taxes, or commissions.
• All assets are perfectly divisible and can be traded at any time.
• There are no restrictions on short-selling.
• Interest rates for borrowing and lending are identical.
Throughout the text, we assume frictionless markets. Although the methods work for
time-dependent interest rates as well, we (mainly for notational convenience) further
assume that the riskless interest rate is a constant r.
Definition 14.2 (Weak and strong path-dependence). An option is said to be
weakly path-dependent if the dynamics of the option price depend only on asset price
and time, i.e., the stochastic differential equation (SDE) for the option price differs
from the SDE of a vanilla option only in the boundary conditions. Otherwise the option
is said to be strongly path-dependent.
While for weakly path-dependent options, quite general methods are available, the
theory is somewhat more involved for strongly path-dependent options. Nevertheless,
we present hedging strategies for lookback options, as well as for Asian options, and
discretely sampled options in general.
The rest of the chapter is structured as follows. Section 14.2 introduces a method
to synthesize arbitrary (sufficiently regular) payoffs that depend solely on the asset
price at some later time T . In Sec. 14.3 techniques for weakly path-dependent options
are outlined. The focus lies in particular on barrier options, which are the most liquid
and best understood exotic options of this kind. Section 14.4 deals with two examples
of strongly path-dependent options (the lookback and the Asian option) and shows
some robust hedging strategies. In Sec. 14.5, discretely sampled options, which may
be weakly path-dependent (e.g., discretely monitored barrier options) or strongly path-
dependent (e.g., Asian options) are considered. Finally, Sec. 14.6 concludes and points
out some potential future research topics.
Now consider options with a path-independent payoff. To fix ideas let us denote the
price process of some asset S up to time T by ST = {Ft }0≤t≤T and the payoff function
May 19, 2010 16:20 WSPC/SPI-B913 b913-ch14 FA
Note that (14.1) holds for any K ∗ and even extends to cases where p is not twice
differentiable (e.g., for a convex payoff function, p should then be understood as left
derivative and p as a positive measure).
For hedging purposes, we can rewrite (14.1) as
where F denotes the forward price of the asset. In this way, the payoff of the contingent
claim p is decomposed into four parts, two of which are hedgeable by static positions
in the riskless bond and a forward on the asset. The other two are synthesized by
“infinitesimal” positions in standard European options. In total we get the following
static hedging strategy (cf. [22]):
This means that if European calls and puts with maturity T are liquid for all strikes,
any (sufficiently regular) contingent claim on the time-T -price of S can be replicated
May 19, 2010 16:20 WSPC/SPI-B913 b913-ch14 FA
EQ [e−r(T −t) p(ST )|Ft ] = e−r(T −t) p(K∗ ) + e−r(T −t) (Ft − K∗ )p (K∗ )
∞ K∗
+ p (x)C(Ft , t; x, T )dx + p (x)P(Ft , t; x, T )dx, (14.2)
K∗ 0
which is obtained by setting p(S ) = (S − K)+ and choosing K∗ > K (Ft denotes the
time t-forward price of the asset). Note that (14.1) furthermore implies the following
static hedging portfolio for the call payoff:
• A long position in a put with the same strike K and maturity T as the call
• A long position in a forward contract
• Investing exp(−rT )(Ft − K) in the bond
It is easy to see that the payoff of this portfolio at time T is the same as provided by
the vanilla put, and the hedging strategy is therefore correct.
Now, given a set of traded option prices, one can derive no-arbitrage bounds
for the price of contingent claims by solving a semi-definite optimization problem
(see [9,51], where also more general options are considered). The dual of this problem
is then to find the most expensive sub- and the cheapest super-replicating strategy using
static positions in liquid options. The latter question is obviously highly relevant in
practical applications, as it shows an (in some sense) optimal way to cover the payoff in
any case.
Alternatively the right hand side of (14.1) may be approximated by a sum, where
the sampling points correspond to the liquid strikes. However, then one cannot ensure
that the payoff of the hedging portfolio is sufficient to cover the payoff of the contingent
claim, unless the weights of the liquid options are chosen with great care.
The prototype of a weakly path-dependent option is the barrier option, which, starting
from [14], has attracted by far the most attention in the literature of semi-static hedging.
The payoff of a barrier option generally depends on whether the asset price has reached
some pre-specified region up to maturity or not.
For example, the Down-and-Out call (DOC) with parameters strike K, barrier
B < S0 , and maturity T has the payoff
In a similar manner the payoff of Up-and-Out and Up-and-In calls (UOC and UIC)
are defined by
and
Since the plain vanilla option payoff obviously dominates the barrier option payoff,
a first simple super-replicating portfolio for the considered barrier option is a long
position in a standard European option with the same strike and the same maturity.
Let us denote the time-t-price of a Down-and-Out and a Down-and-In call with
strike K, barrier B, and maturity T by DOC(Ft , t; K, T, B) and DIC(Ft , t; K, T, B),
respectively. Since the sum of the payoff of a DOC and a DIC with the same parameters
equals the payoff of a plain vanilla call, we have by the law of one price
Analogous equations hold for UOC and UIC prices and naturally for the corresponding
put options as well.
In the remainder of this section, the DOC serves as the illustrating example.
Extensions to other option types are often analogous.
In practice, early-exercise or cancellation features additionally play a role. The
interested reader is referred to [65] for an optimization based treatment of super-
and sub-replicating strategies for American options and to [33, 34] or [8] for hedging
strategies of installment options.
Let us assume that the barrier option with maturity T is written on the time T
forward price FT of an asset and let us use the forward price as the underlying. For
convenience, the prices of all options are w.l.o.g. expressed with respect to this forward
price in the rest of the section.
Any barrier option can be adapted to this notation, simply by replacing the original
constant barrier by a time-dependent one. The first hitting time τ is defined as the first
time for which the forward asset price process is at or below the (now possibly time-
dependent) barrier B(t),
τ = inf{t ≤ T : Ft ≤ B(t)},
• Buy a standard European call struck at K and maturing at T and sell a European put
with the same parameters.
• The first time the forward price touches the barrier close the positions.
• If the barrier is never hit, hold the portfolio until expiry.
May 19, 2010 16:20 WSPC/SPI-B913 b913-ch14 FA
This perfectly replicates the payoff of the DOC, as follows by considering the two
different sacenarios:
• The barrier is not reached until maturity: then the payoff of the portfolio is the same
as the one of the DOC, i.e., (ST − K), as necessarily ST = FT ≥ B = K;
• The barrier is hit at τ < T : then
C(Fτ , τ; K, T ) = P(Fτ , τ; K, T ),
because of the Put–Call–Parity (14.3) (recall that Fτ = K). Thus the cost of closing
the positions is 0 and the payoff of the DOC is replicated also in this case.
The price of the DOC is hence given by
DOC(F0 , 0; K, T, K) = C(F0 , 0; K, T ) − P(F0 , 0; K, T ). (14.6)
This hedging strategy is rather simple (and — due to the involved continuity assump-
tion — model-dependent). But it is not straight-forward to extend this kind of strategy
to other types of barrier options (an exception is the barrier exchange option, see [57]).
Thus there is the need for a more general approach.
One of the most famous semi-static hedging strategies for barrier options was
originally developed by Carr and Bowie [14] (see also [15–17]) and is based on a
relation between standard European calls and puts with different strikes:
Assumption 1. Denote the current time by t. For any T ≥ t the implied volatility
smile (for maturity T ), plotted as a function of the log-moneyness, is symmetric. More
precisely: let K and H√ be such that the geometric mean of K and H is the current
forward price Ft , i.e., KH = Ft , then
σimp (K) = σimp (H ).
We shortly sketch the proof of the above theorem: recall that the price of a call in the
Black-Scholes model (with volatility σ) is given by
C(Ft , t; K, T ) = e−r(T −t) Ft (d+ (Ft /K)) − K(d− (Ft /K)) ,
where Ft is the forward price of the asset,
log(x) ± 12 σ 2 (T − t)
d± (x) = √ ,
σ T −t
May 19, 2010 16:20 WSPC/SPI-B913 b913-ch14 FA
and
√ (x) is the cumulative distribution function of the normal distribution. Due to
KH = Ft , the above is easily seen to be equivalent to
K −r(T −t)
C(Ft , t; K, T ) = e H(−d− (Ft /H )) − Ft (−d+ (Ft /H ))
H
K
= P(Ft , t; H, T ).
H
Thus as long as the implied volatility for vanilla options with strike K and H is identical
(which is Assumption 1), the theorem remains valid.
Necessary and sufficient conditions on an asset price model to fulfill Assumption 1
are given in [20]. A first example is of course the Black-Scholes model. Another is the
following local volatility model for the forward price:
Condition (14.7) implies that the local volatility at a fixed time s plotted as a function
of the log-returns is symmetric around 0, which explains also the symmetry of the
implied volatility surface.
Starting from Theorem 14.1, in [14] it is investigated how the Put–Call–Symmetry
can be used to hedge barrier options (a similar relation between different call and put
prices, termed Put–Call–Reversal, was used in [6] to hedge long-term call options).
Consider a DOC with strike K, maturity T , and barrier B < K. By definition, the
payoff of this option is (ST − K)+ , if the barrier was never hit and 0 otherwise. In order
to hedge the final payoff, a vanilla call with the same strike and maturity as the DOC
should be bought. Clearly this is a super-hedge and another vanilla option can be sold
such that one obtains a portfolio with value 0 at the barrier. Assuming the conditions
for the Put–Call–Symmetry at the first hitting time τ, we know that a put with the
same maturity and strike B2 /K fulfills the geometric mean condition of the theorem.
Therefore
K
C(B, τ; K, T ) = P(B, τ; B2 /K, T ).
B
Hence a portfolio consisting of a long position in the corresponding vanilla call and a
short position in K/B puts with strike B2 /K has value 0, if the forward price is exactly
at the barrier. Consequently, a replicating portfolio for a DOC is to take a long position
in a standard European call with strike K and maturity T and a short position in K/B
puts with strike B2 /K and the same maturity. The corresponding trading strategy is to
hold the portfolio until expiry, if the barrier is never hit, or to close the positions (at 0
May 19, 2010 16:20 WSPC/SPI-B913 b913-ch14 FA
cost) at τ ≤ T . Note that for the correctness of the replication strategy it is necessary
that the forward price is exactly B at the first hitting time (which explains the continuity
assumption).
Summing up, under the no-arbitrage principle this leads to the equation:
K
DOC(F0 , 0; K, T, B) = C(F0 , 0; K, T ) − P(F, 0; B2 /K, T ). (14.8)
B
This hedging procedure can be extended to other types of barrier options but then
typically becomes more complicated. For example, for an UOC one would also need
to take a position in a binary call (i.e., an option, which pays 1 if the asset price at
maturity exceeds the strike, and 0 otherwise) and the latter one can be hedged by the
strategies outlined in the previous section. If the assumption of a constant barrier on
the forward price is relaxed, super- and sub-replication strategies can be derived. For
the details we refer to [17] (see also the nice summary [69]). In [56], the case when
only a finite number of options with prespecified strikes are available is considered and
conditions under which a strategy based on the Put–Call–Symmetry is a semi-static
superhedge are identified.
Another hedging strategy for barrier options has its roots in papers of Derman
et al. [36,37]. In order to point out the main idea of the Derman–Ergener–Kani (D–E–K)
algorithm, a binomial model is considered first, i.e., the asset price process is assumed
to follow a binomial tree. In this setting the D–E–K-algorithm for a DOC with maturity
T , strike K, and barrier B is as follows:
• Identify the boundary nodes, i.e., the maturity and the barrier nodes and denote the
resulting time grid by t1 < · · · < tn ;
• Buy a standard European option replicating the payoff of the barrier option if the
barrier boundary is not hit during the lifetime of the barrier option. In the case of a
DOC with B < K this would be a standard European call with the same strike and
maturity as the barrier option;
• Choose n options with different maturities Ti and strikes Ki having payoff 0 if the
barrier was not hit. In the example of a DOC, one could for instance choose puts
with strikes Ki ≤ B and maturities Ti ≤ T . Then solve the linear equation system
Vi · w + C(B, ti ; K, T ) = 0, i = 1, . . . , n (14.9)
for w, where
Vi := (P(B, ti ; K1 , T1 ), P(B, ti ; K2 , T2 ), . . . , P(B, ti ; Kn , Tn ))
denotes the vector of the time-ti -prices of the n additional options and · is the scalar
product.
Note that due to (14.9) the portfolio consisting of the option replicating the final payoff
and wi options with strike Ki and maturity Ti can be sold at zero cost at every barrier
node and thus the semi-static trading strategy of holding the portfolio until expiry in
May 19, 2010 16:20 WSPC/SPI-B913 b913-ch14 FA
case the barrier has not been hit, or to close the positions at the first hitting time, is a
perfect replication of the barrier option.
In the binomial framework the D–E–K-algorithm thus gives a perfect hedge for
all kinds of barrier options — in particular also for double-barrier options and options
with a time-dependent barrier — as long as enough different standard European options
are liquid (i.e., at least as many as there are barrier nodes). However, since the prices
in (14.9) are model-dependent (as they have to be calculated using a model — in the
above case the binomial model), the entire portfolio is model-dependent.
For continuous-time models, the approach outlined in [37] is to discretize (in time)
the asset price process and hedge the values at the discrete monitoring times. In [78],
the effectiveness of such a discretization for stochastic volatility is examined and it is
found that it works well for a small volatility of volatility, but not satisfactory in the
opposite case combined with discontinuous final payoff of the barrier option (as for
an UOC). We come back to better alternatives in the next section.
A possible limit strategy for continuous-time price models can be achieved by
assuming standard options with all strikes and maturities to be liquid. This strategy was
developed by Andersen et al. [5] (see also [61]), who observed that the key assumption
on the price model is that the price of European options, besides their parameters, only
depends on the current time t and asset price Ft (this assumption for instance rules out
stochastic volatility models). The approach is, as the D–E–K-algorithm, very powerful
in terms of the barrier structure and allows for arbitrary (sufficiently regular) functions
of time.
For the ease of exposition we assume that the forward price again follows a local
volatility model in the risk-neutral world (the method can be extended to more general
Markov-type models), i.e.,
dFt
= σ(Ft , t)dWt , (14.10)
Ft
where σ is now just assumed to be positive and sufficiently regular to admit a unique
solution of Eq. (14.10).
Consider again the standard example of a DOC on the forward price of an asset
with dynamics as specified in (14.10). Then the following PDE specifies the price
function G(Ft , t) of a DOC with strike K, barrier B, and maturity T started at time t
(and has in particular not reached the barrier up to the time t):
1
Gt (F, t) + σ 2 (F, t)F 2 GFF (F, t) = 0, t < T, F > B
2
G(F, t) = 0, t < T, F ≤ B (14.11)
+
G(F, T ) = (F − K) ∀F,
Then
in case the asset price has not crossed the barrier before t. If the function σ in (14.11) is
sufficiently regular, G(F, t) is twice differentiable with respect to F , except at F = B,
where the (formal) second derivative is given by δ(F − B)G+ F (B, t) (δ denoting the
+
Dirac Delta measure and GF the right derivative). Using (14.11) and a generalized
Meyer–Itô-formula (see [68]), one finds:
T
(FT − K)+ = G(F0 , 0) + 11{Ft >B} GF (Ft , t)Ft σ(Ft , t)dWt
0
T
1
+ G+ B
F (B, t)dLt ,
2 0
+
T
1 + B
= E[(FT − K) |Fτ ] − E G (B, t)dLt Fτ ,
0 2 F
May 19, 2010 16:20 WSPC/SPI-B913 b913-ch14 FA
where the right-hand-side is exactly the time-τ-price of the portfolio. Note that at this
point it is crucial that the prices of European options only depend on the asset price at
time τ in order to interpret the expectations as prices. Since
τ
1 τ +
G(F0 , 0) + 11{Ft >B(t)} GF (Ft , t)dFt + G (B, t)dLBt = G(B(τ), τ)
0 2 0 F
and G(B(τ), τ) = 0 by definition, the value of the portfolio at time τ is 0. Thus we can
close the positions at zero cost and replicate the payoff of the DOC.
This replication strategy can be generalized to all kinds of terminal payoffs, barrier
regions, rebates, and also to jump-diffusion processes. We refer to [5] for details.
An interesting difference to hedging strategies using the Put–Call–Symmetry is
that here not only the correctness, but also the strategy itself depends on the model.
More precisely: the number G+ F (B(t), t) has to be calculated using a model, while the
strategy outlined before is independent of the exact specification of the local volatility,
as long as Assumption 1 is fulfilled.
to an inequality, such that the trading strategy becomes a super-replication and then
minimize the cost of this hedge.
γ̃ − K K−B
≤ C(B − x, 0; K, T ) − C(B − x, 0; B, T ) − C(B − x, 0; γ̃, T )
γ̃ − B γ̃ − B
≤ 0,
May 19, 2010 16:20 WSPC/SPI-B913 b913-ch14 FA
where the first inequality follows from the Put–Call–Parity (14.3) and the second
one from the convexity (in K) of the call price (or alternatively by a direct look at
the final payoff of the last portfolio).
The portfolio is thus a subreplicating one for any γ̃ and hence γ̃ can be chosen such
that the initial value V of the portfolio is maximized. Since
γ̃ − K K−B
V = C(F0 , 0; K, T ) − P(F0 , 0; B, T ) − C(F0 , 0; γ̃, T )
γ̃ − B γ̃ − B
K−B
= C(F0 , 0; K, T ) − P(F0 , 0; B, T ) + (P(F0 , 0; K, T ) − C(F0 , 0; γ̃, T )),
γ̃ − B
the optimal γ̃ is given by
P(F0 , 0; B, T ) − C(F0 , 0; β, T )
γ̃ = argmaxβ>K .
β−B
If K < B, then a sub-replicating portfolio for the DOC is given by:
The trading strategy is the same as before and the subreplication property is again
shown by distinguishing two cases:
• If the barrier is not hit, we have for the payoff of the strategy
B−K
(ST − F0 ) + (F0 − K) − (γ − ST )+ ≤ (ST − K),
γ −B
where the right-hand side corresponds to the payoff of the DOC, since B > K;
• If the barrier is hit at time τ, then Ft = B − x. Thus the time-τ-value of the long
position in the forward is exp(−r(T − τ)) (B − x − F0 ) and hence for Vτ we find
B−K
− P(B − x, τ; γ, T )
γ −B
B−K
≤ e−r(T −τ) (B − K) − P(B, τ; γ, T )
γ −B
B − K −r(T −τ)
≤ e−r(T −τ) (B − K) − e (γ − B) = 0,
γ −B
where the first inequality follows from the monotonicity of the put price and the
second one from the Put–Call–Parity (14.3).
May 19, 2010 16:20 WSPC/SPI-B913 b913-ch14 FA
Similarly as in the case K > B it can be shown, that the optimal choice for γ, i.e., the
one maximizing the initial value of the portfolio, is given by
P(F0 , β, 0, T )
γ = argminβ>B .
β−B
Super-replicating portfolios are in both cases found by omitting the barrier feature in
the final payoff, i.e., they are given as the trivial super-hedges consisting of
• A long position in a call with maturity T and strike K, if K ≥ B
• A long position in a call with maturity T and strike B and in B − K binary calls (a
path-independent option with payoff 11{ST ≥B} ) with strike B, if K < B
Of course the sub- and super-hedging strategies also imply price bounds on the DOC:
If K ≥ B we have
γ̃ − K K−B
C(F0 , 0; K, T ) − P(F0 , 0; B, T ) − C(F0 , 0; γ̃, T )
γ̃ − B γ̃ − B
≤ DOC(F0 , 0; K, T, B) ≤ C(F0 , 0; K, T ),
while in the case K < B we have
B−K
e−rT (F0 − K) − P(F0 , 0; γ, T ) ≤ DOC(F0 , 0; K, T, B)
γ −B
≤ C(F0 , 0; B, T ) + (B − K)BC(F0 , 0; B, T ),
where BC(Ft , t; B, T ) denotes the time-t-price of a binary call with strike B, and γ̃, γ
are as before.
The particular role of these sub- and super-replication strategies is identified in
the following result:
Theorem 14.3.1 ( [12]). Assume that European options for all strikes with maturity
T are liquid and that there are no options available with maturity less than T. Then the
bounds on the price of the DOC are sharp. This means that there are forward price
processes that are martingales and consistent with the prices of the European options
for which the price of the DOC is given by the lower (resp. upper) bound. In those
market models, the hedging strategies are perfect.
The proof of this theorem is closely related to the Skorokhod problem (see also [67]).
the corresponding hedging strategies) can be used to hedge them. For Asian options
robust hedging strategies are shown, which are related to hedging strategies for basket
options. A perfect (model-dependent) strategy for discretely sampled options (DSO),
which include Asian or cliquet options, is outlined in Sec. 14.5.
n−1
+ DOC(F0 , 0; Ki , T, Hi+1 ) − DOC(F0 , 0; Ki , T, Hi ) . (14.15)
i=1
This portfolio indeed matches the payoff of the ERDC exactly, since if Hi+1 < mt ≤ Hi
(recall mt = inf 0≤s≤t Ss ), the sum in (14.15) starts with i + 1 and the leading term
May 19, 2010 16:20 WSPC/SPI-B913 b913-ch14 FA
n−1
+ DOC(F0 , 0; Ki , T, Ki+1 ) − DOC(F0 , 0; Ki , T, Ki ) . (14.16)
i=1
Consider now the floating strike lookback call, defined at the beginning of the section.
Recall that the payoff at maturity T is given by ST −mT , where mT is the minimum price
of the asset up to T . This type of lookback calls can be seen as RC with a continuum
of roll-down barriers and strikes. Due to this continuum of roll-down barriers we are
not able to find an exact hedge, but we give super- and sub-hedges for lookback calls.
It is clear that an RC with K0 = F0 and Kn = 0 undervalues the LO, because the
strike of the lookback option at time t can only be below or equal the strike of the RC.
Therefore a model-independent sub-hedge for a LO is given by:
= DOC(F0 , 0; F0 , T, K1 ) + DIC(F0 , 0; Kn , T, Kn )
n−1
+ DOC(F0 , 0; Ki , T, Ki+1 ) − DOC(F0 , 0; Ki , T, Ki ) .
i=1
(14.17)
Again the barrier options might be hedged using the strategies of Sec. 14.3. Adding
more roll-down barriers obviously increases the quality of the sub-hedge, i.e., a tighter
bound is obtained.
May 19, 2010 16:20 WSPC/SPI-B913 b913-ch14 FA
In order to find a super-hedge for the LO, an ERDC with Ki = Hi+1 for 1 ≤ i ≤ n
and Hn+1 = 0 can be used. Such an ERDC is a super-hedge, because its strike is below
or equal to the strike of the LO throughout the lifetime of the option. Hence we have
= DOC(F0 , 0; H1 , T, H1 )
n+1
+ DOC(F0 , 0; Hi , T, Hi ) − DOC(F0 , 0; Hi , T, Hi−1 ) . (14.18)
i=2
Another portfolio having the same payoff as the one stated above is given by:
n−1
+ (Hi − Hi+1 ) DIB(F0 , 0; Hi , T )
i=1
+ Hn DIB(F0 , 0; Hn , T ), (14.19)
where DIB is a down-and-in digital option that pays 1 at expiry if the barrier was
hit before maturity. As for the sub-hedge, the bounds become tighter the more Hi are
added.
Note that (14.17) as well as (14.18) and (14.19) are model-independent super-
and sub-hedging portfolios, respectively.
For fixed-strike lookback options one can use similar arguments and digital options
to replicate the payoff (see [13]). Then again, the hedging strategies for weakly path-
dependent options can be used.
It is worth noting that in [48] the best model-independent bounds for lookback
options is calculated. However, the ideas are similar to the ones used for the derivation
of the robust strategies for hedging barrier options and we omit the details here.
Taking the operator (·)+ inside the sum in (14.20) clearly gives an upper bound and
hence a static super-replication strategy for the AC in terms of European calls, i.e.,
+
n
+
n
Sti − nK = (St1 − K1 ) + · · · + (Stn − Kn ) ≤ (Sti − Ki )+ , (14.21)
i=1 i=1
n
whenever i=1 Ki = nK.
A simple first choice for the strikes of the calls is Ki = K ∀ 1 ≤ i ≤ n, which
is, however, not optimal. The cheapest (an hence best) choice was originally found
in [74] for complete markets and later generalized in [1, 3] (see also [25, 80]) and it
turns out that the concept of comonotonicity is a helpful tool to this end:
FX1 ,X2 ,...,Xn (x1 , x2 , . . . , xn ) = min{FX1 (x1 ), FX2 (x2 ), . . . , FXn (xn )}.
Note that the right-hand side of the above equation is a copula (often called lower
Frechet-copula). For a general introduction to this field and proofs of the properties
used below, see [40, 41].
To simplify notation, let us assume that the marginal distribution functions are
strictly increasing. Let (X1 , . . . , Xn ) be a non-negative random vector with marginal
distribution functions FXi and suppose (Y1 , . . . , Yn ) to be the comonotone vector with
the same marginal distributions. Setting S C = ni=1 Yi the following holds:
n
FS−1
C (x) = FX−1i (x), 0 ≤ x ≤ 1. (14.22)
i=1
n
FSC (m) = FXi FX−1i (FS C (m)) , m ≥ 0. (14.23)
i=1
Let (Y1 , . . . , Yn ) be the comonotone vector with the marginal distribution functions
FSt1 , . . . , FStn and S C the sum of the Yi ’s. The following inequality holds for all
May 19, 2010 16:20 WSPC/SPI-B913 b913-ch14 FA
n
i=1 Ki = nK:
n +
n
n
FSC (nK) = E Yi − nK ≤ E (Yi − Ki )+ = Sti (Ki ).
i=1 i=1 i=1
Using Eq. (14.23) and the relation Sti (Ki ) = exp(rti )C(F0 , 0; Ki , T ) we finally find:
n
n
exp(rti )C(F0 , 0; FS−1
ti
(FS C (nK)), T ) ≤ exp(rti )C(F0 , 0; Ki , T ).
i=1 i=1
As a consequence, the optimal choice for the call strikes in (14.21) is:
Ki = FS−1
t
(FS C (nK)).
i
Note that
n
FS−1
t
(FS C (nK)) = FS−1
C (FS C (nK)) = nK,
i
i=1
for all strictly increasing marginal cumulative distribution functions. Since the inverse
function of FS C is given by (14.22) and the right-hand side is strictly increasing in x, it
is computationally straight-forward to calculate FS C (nK). Thus this hedging strategy
is also simple to evaluate.
Numerical studies carried out in [1] for popular Lévy models and in [3] for stochas-
tic volatility models suggest that the obtained bounds become tighter the deeper the
options are in the money.
Of course the concept of comonotonicity can also be applied to other types of
options, for which the payoff depends on a sum of possibly dependent random vari-
ables, e.g., basket options. For more details we refer to [25, 49, 50]. In those papers, in
particular, the (realistic) case is considered, when there are only finitely many strikes
liquid in the market and one has to find the optimal combination of those.
This last question and especially the problem of finding a sub-replicating portfolio
(which is associated to a lower price bound for theAsian option) is also addressed in [2].
Lower price bounds on the AC price were first considered by Curran [31] and Rogers
and Shi [70], who pioneered a quite accurate method to determine lower price bounds
in the Black-Scholes model based on the following idea: Jensen’s inequality gives
n +
n +
1 1
St − K ≥ E Sti |Z − K , (14.24)
n i=1 i n i=1
or variants thereof are very popular (because the distribution of the geometric
mean is explicitly available in the Black-Scholes model) and leads to tight lower
price bounds, since the arithmetic and geometric average are strongly correlated
(see [31, 66, 70, 77, 80]).
However, in contrast to upper price bounds based on the concept of comono-
tonicity, this method in general does not imply a subreplication strategy. Never-
theless, in [2] it is shown that using Z = St1 together with (14.22) yields a
robust sub-replicating portfolio consisting of ni=1 e−r(T −ti ) /n calls with maturity t1
n
and strike nK/( j=1 er(tj −t1 ) ). The corresponding trading strategy is to do nothing
when S1 ≤ nK/ ni=1 er(ti −t1 ) or to buy ni=1 e−r(T −ti ) /n assets in the case that
n r(ti −t1 )
S1 > nK/ i=1 e . The cost for this trade is exactly the payoff of the options
in the portfolio plus Ke−r(T −t1 ) , which one should borrow. Then at each monitoring
time ti one should sell e−r(T −ti ) /n assets and invest the gain in the riskless bank account.
At maturity T of the Asian call, the payoff of the trading strategy is
n
1
St − K 11{ni=1 er(ti −t1 ) St1 >nK} ,
n i=1 i
In Sec. 14.2 a hedging portfolio for path-independent options was presented. As the
payoff of discretely monitored options (DSO) depends only on a finite number of
monitoring times, such options can be understood as path-independent between the
monitoring times and it is possible to construct a semi-static hedging strategy with
adaptations of the portfolio only at the monitoring times. This kind of strategies was
developed by Carr and Wu [24] and Joshi [55] (see also [54]). Similar to Sec. 14.2, we
assume that standard European options are liquid for all strikes and monitoring times
of the DSO. However, to apply the techniques for the path-independent options we
need the extra assumption that the asset price process is Markovian. In particular, the
prices of standard European options may only depend on the current option price.
We consider a recursion algorithm that can be used to price such options and to
obtain some Greeks (namely the Delta and the Gamma) of the exotic option through the
corresponding ones of plain vanilla options. This algorithm is based on the assumption
that the price of the exotic option depends, in addition to the current price of the asset,
only on some summary statistic of the historic asset prices at the monitoring times
which is measurable with respect to the filtration generated by the asset price and
updated only at the monitoring times of the option. If more than one statistic is needed
May 19, 2010 16:20 WSPC/SPI-B913 b913-ch14 FA
to describe the price of the exotic option, the method is still feasible, but for notational
convenience we focus on the case with a single one.
As an illustrating example, an AC with strike K, monitoring times t1 , . . . , tn , and
maturity T = tn is considered, but any other similar DSOs (like e.g., cliquets, or
discretely monitored barrier and lookback options) could serve as well.
Let us denote the summary statistic of a generic DSO at time ti by Xi . Due
to the assumptions on X, Xi is Fti -measureable and Xt = Xi for ti ≤ t ≤ ti+1 ,
i = 0, 1, . . . , n − 1. Furthermore we assume that
Xi+1 = f(Xi , Sti+1 ), (14.25)
which is fulfilled for all DSOs we are aware of.
For the AC, Xi is the running average Ai of the asset prices at time ti , i.e.,
1
i
Xi := Ai = St , ∀ 1 ≤ i ≤ n.
i j=1 j
where gS and gSS denote the first and second derivative of g with respect to its second
argument. Note that, as in Sec. 14.2, the above describes a static hedge for the DSO at
the time tn−1 and therefore also its time-tn−1 -value Vtn−1 is settled.
Vtn−1 (Xtn−1 , Stn−1 ) = g(Xn−1 , K ∗ ) + gS (Xn−1 , K ∗ )(Ftn−1 − K∗ )
∞
+ gSS (Xn−1 , x)C(Stn−1 , tn−1 ; x, T )dx
K∗
K∗
+ gSS (Xn−1 , x)P(Stn−1 , tn−1 ; x, T )dx, (14.27)
0
and
1
Vtn−1 (Xtn−1 , Stn−1 ) = C(Stn−1 , tn−1 ; nK − (n − 1)An−1 , T ),
n
where C(Ft , t; K, T ) := e−r(T −t) (Ft − K) for K < 0 with Ft denoting again the
forward price.
The Markov property together with (14.27) then implies that the value Vtn−1 of the
DSO at time tn−1 only depends on Xtn−1 and Stn−1 and using again (14.25) we actually
have (with a slight abuse of notation)
Thus we can use the right-hand side of (14.27) as the new payoff function of a (mod-
ified) DSO and iterate the procedure until we reach the current time t0 and obtain a
replicating portfolio. The associated hedging strategy is to hold this portfolio until
expiry t1 of the standard European options involved and invest the payoff of these
options to form the new replicating portfolio for the European-type option Vt2 (X1 , St2 ).
This strategy is of course self-financing.
Note that the replicating portfolio does not change until t1 , and thus one can
calculate the price of the DSO, as well as the derivatives with respect to the asset price,
with this portfolio. Unfortunately the derivatives with respect to other parameters of
the DSO price in general cannot be calculated in the same manner, since changing
those parameters would also have an effect on the portfolio.
It is worth noting that the weights of the standard options in the hedging portfolio
at the current time t0 are determined by the Gamma of the DSO at time t1 .
The derivation of this hedging strategy of course relies on the Markov assumption
on the model and the liquidity of arbitrary standard European options. However, even
if those assumptions may fail in practice, the approach gives an idea of how to obtain
approximative hedging strategies. More precisely: if neither the Markov assumption
nor the liquidity assumption is fulfilled, one might not be able to hedge the DSO
perfectly. However, if alternatively a hedging strategy is chosen in order to minimize
a certain risk measure for the hedging error in (14.26), the semi-static strategy might
still outperform classical dynamic hedging strategies. Recently, this kind of approach
was applied e.g., in [7].
May 19, 2010 16:20 WSPC/SPI-B913 b913-ch14 FA
References
[1] Albrecher, H, J Dhaene, M Goovaerts and W Schoutens (2005). Static hedging of Asian
options under Lévy models. Journal of Derivatives, 12, 63–72.
[2] Albrecher, H, P Mayer and W Schoutens (2008). General lower bounds for arithmetic
Asian option prices. Applied Mathematical Finance, 15, 123–149.
[3] Albrecher, H and W Schoutens (2005). Static hedging of Asian options under stochastic
volatility models using fast Fourier transform. In Exotic Options and Advanced Lévy
Models, A Kyprianou et al. (eds.), pp. 129–148. Chichester: Wiley.
[4] Allen, S and O Padovani (2002). Risk management using quasi-static hedging. Economic
Notes, 31, 277–336.
[5] Andersen, LBG, J Andreasen and D Eliezer (2002). Static replication of barrier options:
Some general results. Journal of Computational Finance, 5, 1–25.
[6] Andreasen, J and P Carr (2002). Put call reversal. Working Paper, New York University.
[7] Becherer, D and I Ward (2008). Optimal static hedging of equity and credit risk using
derivatives. Working Paper, Humboldt University, Berlin.
[8] Ben-Hameur, H, M Breton and P François (2006). A dynamic programming approach to
price installment options. European Journal of Opertational Research, 169, 667–676.
[9] Bertsimas, D and I Popescu (2002). On the relation between option and stock prices:
A convex optimization approach. Operations Research, 50(2), 358–374.
[10] Black, F and M Scholes (1973). The pricing of options and corporate liabilities. Journal
of Political Economy, 81, 637–654.
[11] Breeden, DT and RH Litzenberger (1978). Prices of state-contingent claims implicit in
option prices. Journal of Business, 51(4), 621–651.
May 19, 2010 16:20 WSPC/SPI-B913 b913-ch14 FA
[12] Brown, H, D Hobson and LCG Rogers (2001). Robust hedging of barrier options. Math-
ematical Finance, 11, 285–314.
[13] Buchen, PW and O Konstandatos (2004). A new method of pricing lookback options.
Mathematical Finance, 15, 245–259.
[14] Carr, P and J Bowie (1994). Static simplicity. Risk, 7(8), 45–50.
[15] Carr, P and A Chou (1997). Breaking barriers. Risk, 10, 139–145.
[16] Carr, P and A Chou (2002). Hedging complex barrier options. Working Paper, New York
University.
[17] Carr, P, K Ellis and V Gupta (1998). Static hedging of exotic options. The Journal of
Finance, 53, 1165–1190.
[18] Carr, P and R Lee (2007). Realized volatility and variance: Options via swaps. Risk, May,
76–83.
[19] Carr, P and R Lee (2010). Hedging variance options on continuous semimartingales.
Finance and Stochastics (forthcoming).
[20] Carr, P and R Lee (2008). Put-call symmetry: Extensions and applications. Mathematical
Fincane 19(4), 523–560.
[21] Carr, P and D Madan (2005). A note on sufficient conditions for no arbitrage. Finance
Research Letters, 2, 125–130.
[22] Carr, P and D Madan (2006). Towards a theory of volatility trading. In Risk Book on
volatility, R Jarrow (ed.). Berlin: Risk.
[23] Carr, P and W Schoutens (2008). Hedging under the Heston model with jump-to-default.
International Journal of Theoretical and Applied Finance, 11, 403–414.
[24] Carr, P and L Wu (2002). Static hedging of standard options. Working Paper, Courant
Institute, New York University.
[25] Chen, X, G Deelstra, J Dhaene and M Vanmaele (2008). Static super-replicating strategies
for a class of exotic options. Insurance: Mathematics and Economics, 42(3), 1067–1085.
[26] Coleman, TF, Y Kim, Y Li and M Patron (2007). Robustly hedging variable annuities with
guarantees under jump and volatility risk. Journal of Risk and Insurance, 7, 347–376.
[27] Coleman, TF, Y Li and M Patron (2006). Hedging guarantees in variable annuities (under
both market and interest rate risk). Insurance: Mathematics and Economics, 38, 215–228.
[28] Cousot, L (1969). Conditions on option prices for absence of arbitrage and exact calibra-
tion. Journal of Banking & Finance, 31, 3377–3397.
[29] Cox, AMG and J Oblój (2008). Robust hedging of double touch barrier options. Working
Paper, available at http://arxiv.org/abs/0808.4012.
[30] Cox, AMG and J Oblój (2009). Robust pricing and hedging of double no-touch options.
Working Paper, available at http://arxiv.org/abs/0901.0674.
[31] Curran, M (1994). Valuing Asian and portfolio options by conditioning on the geometric
mean price. Management Science, 40(12), 1705–1711.
[32] Davis, M and D Hobson (2007). The range of traded option prices. Mathematical Finance,
17, 1–14.
[33] Davis, M, W Schachermayer and R Tompkins (2001). Pricing, no-arbitrage bounds and
robust hedging of installment options. Quantitative Finance, 1, 597–610.
[34] Davis, M, W Schachermayer and R Tompkins (2002). Installment options and static
hedging. Journal of Risk Finance, 3(2), 46–52.
[35] Delbaen, F and W Schachermayer (2006). The Mathematics of Arbitrage. Berlin Heidel-
berg: Springer.
[36] Derman, E, D Ergener and I Kani (1994). Forever hedged. Risk, 7, 139–145.
May 19, 2010 16:20 WSPC/SPI-B913 b913-ch14 FA
[37] Derman, E, D Ergener and I Kani (1995). Static options replication. The Journal of Deriva-
tives, 2, 78–95.
[38] Derman, E, I Kani and M Kamal (1997). Trading and hedging local volatility. Journal of
Fincancial Engineering, 6, 233–268.
[39] Derman, E and NN Taleb (2005). The illusion of dynamic replication. Quantitative
Finance, 5, 323–326.
[40] Dhaene, J, M Denuit, MJ Goovaerts, R Kaas and D Vyncke (2002). The concept of
comonotonicity in actuarial science and finance: Applications. Insurance: Mathematics
& Economics, 31(2), 133–161.
[41] Dhaene, J, M Denuit, MJ Goovaerts, R Kaas and D Vyncke (2002). The concept of comono-
tonicity in actuarial science and finance: Theory. Insurance: Mathematics & Economics,
31(1), 3–33.
[42] Dupire, B (1997). A unified theory of volatility. Working Paper, BNP Paribas.
[43] Engelmann, B, M Fengler and R Poulsen (2007). Static versus dynamic hedges: An empir-
ical comparison for barrier options. Review of Derivatives Research, 9, 239–264.
[44] Fink, J (2003). An examination of the effectiveness of static hedging in the presence of
stochastic volatility. The Journal of Future Markets, 23(9), 859–890.
[45] Gastineau, G (1994). Roll-up puts, roll-down calls, and contingent premium options. The
Journal of Derivatives, 1, 40–43.
[46] Giese, AM and J Maruhn (2007). Cost-optimal static super-replication of barrier options —
an optimization approach. Journal of Computational Finance, 10, 71–97.
[47] Green, RC and RA Jarrow (1987). Spanning and completeness in markets with contingent
claims. Journal of Economic Theory, 41, 202–210.
[48] Hobson, D (1998). Robust hedging of the lookback options. Finance and Stochastics, 2,
329–347.
[49] Hobson, D, P Laurence and TH Wang (2005). Static-arbitrage optimal subreplicating
strategies for basket options. Insurance: Mathematics & Economics, 37(3), 553–572.
[50] Hobson, D, P Laurence and TH Wang (2005). Static-arbitrage upper bounds for the prices
of basket options. Quantative Finance, 5(4), 329–342.
[51] Hodges, SD and A Neuberger (2001). Rational bounds on exotic options. Working Paper,
Warwick University.
[52] Ilhan, A, M Jonsson and R Sircar (2008). Optimal static-dynamic hedges for exotic
options under convex risk measures. Working Paper, available at http://ssrn.com/abstract=
1121233.
[53] Ilhan, A and R Sircar (2005). Optimal static-dynamic hedges for barrier options.
Mathematical Finance, 16, 359–385.
[54] Joshi, MS (2002). The Concepts and Practice of Mathematical Finance. Cambridge:
Cambridge University Press.
[55] Joshi, MS (2002). Pricing discretely sampled path-dependend options using replication
methods. Working Paper, available at www.quarchome.org.
[56] Kraft, H (2007). Pitfalls in static superhedging of barrier options. Finance Research Letters,
4, 2–9.
[57] Lindset, S and S-A Perrson (2006). A note on a barrier exchange option: The world’s
simplest option formula? Finance Research Letters, 3, 207–211.
[58] Maruhn, J and EW Sachs (2006). Robust static super-replication of barrier options in
the Black-Scholes model. In Robust Optimization-Directed Design (RODD), AJ Kurdila,
PM Pardalos and M Zabarankin (eds.), pp. 127–143. Springer.
May 19, 2010 16:20 WSPC/SPI-B913 b913-ch14 FA
[59] Maruhn, J and EW Sachs (2009). Robust static hedging of barrier options in stochastic
volatility models. Mathematical Methods of Operations Research, 70(3), 405–433.
[60] Merton, RC (1973). Theory of rational option pricing. Bell Journal of Economics and
Management Science, 4, 141–183.
[61] Mijatović, A (2008). Local time and the pricing of time-dependent barrier options. Working
Paper, available at: http://arxiv.org/pdf/0809.1747.
[62] Nachmann, D (1998). Spanning and completeness with options. Review of Financial
Studies, 31, 311–328.
[63] Nalholm, M and R Poulsen (2006). Static hedging and model risk for barrier options.
Journal of Futures Markets, 26, 449–463.
[64] Nalholm, M and R Poulsen (2006). Static hedging of barrier options under general asset
dynamics: Unification and application. Review of Derivative Research, 13, 46–60.
[65] Neuberger, A (2007). Bounds on the American option. Working Paper, Warwick Business
School.
[66] Nielsen, JA and K Sandmann (2003). Pricing bounds onAsian options. Journal of Financial
and Quantitative Analysis, 38(2), 449–473.
[67] Oblój, J (2004). The Skorokhod embedding problem and its offspring. Probability Surveys,
1, 321–392.
[68] Peskir, G (2005). A change-of-variable formula with local time on curves. Journal of
Theoretical Probability, 18, 499–535.
[69] Poulsen, R (2006). Barrier options and their static hedges: Simple derivations and exten-
sions. Quantitative Finance, 6, 327–335.
[70] Rogers, LCG and Z Shi (1995). The value of an Asian option. Journal of Applied Proba-
bility, 32, 1077–1088.
[71] Ross, S (1976). Options and efficiency. Quarterly of Economics, 90, 75–89.
[72] Sbuelz, A (2005). Hedging double barriers with singles. International Journal of Theoret-
ical and Applied Finance, 8, 393–407.
[73] Schweizer, M and J Wissel (2008). Term structures of implied volatilities: Absence of
arbitrage and existence results. Mathematical Finance, 18, 77–114.
[74] Simon, S, MJ Goovaerts and D Dhaene (2000). An easy computable upper bound for the
price of an arithmetic Asian option. Insurance: Mathematics & Economics, 26, 175–184.
[75] Siven, J and R Poulsen (2009). Auto-static for the people: Risk-minimizing hedges of
berrier options. Working Paper, University of Gothenburg.
[76] Stoll, HR (1969). The relationship between put and call prices. Journal of Finance, 24,
801–824.
[77] Thompson, GWP (1999). Fast narrow bounds on the price of Asian options. Working Paper,
University of Cambridge.
[78] Toft, KB and C Xuan (1998). How well can barrier options be hedged by a static portfolio
of standard options? Journal of Financial Engineering, 7, 147–175.
[79] Tompkins, R (1997). Static versus dynamic hedging: An evaluation of hedge performance
via simulation. Netexposure, 1(2), 1–28.
[80] Vanmaele, M, G Deelstra, J Liinev, J Dhaene and MJ Goovaerts (2006). Bounds for the price
of discrete arithmetic Asian options. Journal of Computational and Applied Mathematics,
185, 51–90.
May 19, 2010 16:20 WSPC/SPI-B913 b913-ch14 FA
375
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch15 FA
15.1. INTRODUCTION
A classical question in Mathematical Finance is how the issuer of an option can hedge
her risk by trading in the underlying. To tackle this problem in incomplete markets,
we consider variance-optimal hedging, cf. [5, 18, 22] and the references therein for a
survey of the extensive literature. Variance-optimal hedging of a contingent claim H
means that one minimizes the expected squared hedging error
E[(v0 + ϕ • ST − H)2 ]
over all initial endowments v0 and trading strategies ϕ, where ϕ • ST represents the
cumulated gains resp. losses from trading ϕ up to the expiry date T of the claim.
In this chapter, we consider the above problem in affine stochastic volatility models.
These generalize Lévy processes by allowing for volatility clustering and are capable
of recapturing most of the stylized facts observed in stock price time series.
For Lévy processes, variance-optimal hedging has been dealt with using PDE
methods by [6] and by employing Laplace transform techniques in [4,10]. The approach
of [10] has subsequently been extended to affine models by [12,13,17] if the discounted
asset price is a martingale and by [14] in the general case. However, whereas [4, 10]
incorporate both continuous and discrete rebalancing, the results for affine processes
have focused on continuous trading thus far.
The present study complements these results by showing how to deal with discrete-
time variance-optimal hedging in affine models. Since only finitely many trades are
feasible in reality, this analysis is important in order to answer the following questions:
(1) How should discrete rebalancing affect the investment decisions of the investor,
i.e., to what extent should she adjust her hedging strategy?
(2) How can one quantify the additional risk resulting from discrete trading, i.e., by
how much does the hedging error increase?
The general structure of variance-optimal hedging in discrete time has been thoroughly
investigated by [21]. However, examples of (semi-) explicit solutions seem to be limited
to the results of [4, 10] for Lévy processes and [2] for some specific diffusion models
with stochastic volatility. Here, we show how to extend the Laplace transform approach
of [10] to general affine stochastic volatility models. Similarly as in [12, 13], we
focus on the case where the discounted asset price process is a martingale. Numerical
experiments using the results of [10] and [14] indicate that the effect of a moderate
drift rate on hedging problems is rather small.
This article is organized as follows. In Sec. 15.2, we summarize for the convenience
of the reader the general structural results of [21] on variance-optimal hedging in dis-
crete time, reduced to the case where the underlying asset is a martingale. Subsequently,
we explain how the Laplace transform approach can be used in general discrete-time
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch15 FA
models in order to obtain integral representations of the objects of interest. Section 15.4
turns to the computation of the integrands from Sec. 15.3 in affine stochastic volatil-
ity models. We show how to compute all integrands in closed form for the optimal
initial capital and hedging strategy. This parallels results for continuous-time hedging
in [4, 10] and [12, 13] and for discrete-time hedging in [4, 10]. Somewhat surprisingly,
the expressions for the corresponding hedging error turn out to be considerably more
involved than in the continuous-time case and cannot be computed in closed form. We
propose two approaches to circumvent this problem: First, we determine a closed form
approximation, whose error becomes negligible as the number of trades tends to infin-
ity. As an alternative, we put forward a simple Monte–Carlo scheme to approximate
the hedging error via simulation. Section 15.5 contains some numerical examples for
the time-changed Lévy models introduced by [3].
S = S0 exp(Xt ), S0 ∈ R+ , t ∈ T0 ,
E[ST2 ] < ∞,
as well as
to rule out degenerate cases. Our goal is to compute the variance-optimal hedge for a
given contingent claim H in the following sense.
As noted in the introduction, we restrict ourselves to the case where the stock price
is a martingale.
Proof. This follows from [21, Sec. 4.1] by making use of the martingale properties of
S and V .
Notice that if the initial capital is fixed at v0 ∈ R rather than being part of the
optimization problem, the same strategy ϕ is still optimal if S is a martingale (cf. [21,
Sec. 4.1]). However, the corresponding hedging error increases by (v0 − V0 )2 in this
case.
In order to derive formulas that can be computed in concrete models, we use the
Laplace transform approach, which has been introduced to variance-optimal hedging
by [10]. The key assumption on the contingent claim is the existence of an integral
representation in the following sense.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch15 FA
Assumption 15.4. Suppose that the payoff function of the claim is of the form H =
f(ST ) for some function f : (0, ∞) → R, such that
R+i∞
f(s) = sz l(z)dz,
R−i∞
Example 15.5. Most European options admit an integral representation of this kind.
For example, for the European call with payoff function f(s) = (s − K)+ , we have,
R+i∞
1 K1−z
f(s) = sz dz,
2πi R−i∞ z(z − 1)
for any R > 1 by [10, Lemma 4.1]. More generally, the Bromwich inversion formula
as in [10, Theorem A.1] ascertains that l is typically given by the bilateral Laplace
transform of x → f(exp(x)), cf. [10] for more details and examples.
Theorem 15.6. We have H ∈ L2 (P) and the corresponding option price process is
given by
R+i∞
Vtn = V(z)tn l(z)dz, tn ∈ T 0 ,
R−i∞
Proof. The first assertion follows from Assumption 15.4 and Fubini’s Theorem along
the lines of [13, Lemma 3.3 and Proposition 3.4]. The second can be derived analo-
gously using the Cauchy–Schwarz Inequality and E[St2n ] < ∞, tn ∈ T .
For the hedging error, Proposition 15.3 and Theorem 15.6 yield the following
similar integral representation.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch15 FA
If, for tn ∈ T ,
∞ ∞
E[|J2 (tn , R + ix1 , R + ix2 )|]|l(R + ix1 )||l(R + ix2 )|dx1 dx2 < ∞, (15.2)
−∞ −∞
R+i∞ R+i∞
− E[J2 (tn , z1 , z2 )]l(z1 )l(z2 )dz1 dz2 .
tn ∈T R−i∞ R−i∞
Theorem 15.6 and Corollary 15.7 show that in order to compute semi-explicit formulas
of the discrete variance-optimal capital, hedging strategy and hedging error, one must
be able to compute conditional exponential moments of the process X. This suggests to
consider models whose moment generating function E[exp(uXtn )] is known in closed
form. Here we use affine processes in the sense of [7].
Assumption 15.8. Suppose that (yt , Xt )t∈T0 is the restriction to discrete time of a
semimartingale which is regularly affine w.r.t. y in the sense of [7, Definitions 2.1
and 2.5]. This means that the characteristic function of (y, X) has exponentially affine
dependence on y, i.e., there exist mappings j : T × iR2 → C, j = 0, 1 such that,
for t ≥ s and (u1 , u2 ) ∈ iR2 ,
In [11], it is shown that a large number of stochastic volatility models from the
empirical literature fit into this framework. Examples include the models of Heston [9]
and Barndorff-Nielsen and Shephard [1] as well as their extensions to time-changed
Lévy models by [3]. A particular specification of this general class of models is given
by the following OU-time-change model:
for a mean reversion speed λ > 0, a Lévy process L with Lévy exponent ψL , and an
increasing Lévy process Z with Lévy exponent ψZ , i.e.,
In this case, (y, X) is affine by [11, Sec. 4.4] and in view of [11, Corollary 3.5], we have
1 − e−λt L
1 (t, u1 , u2 ) = e−λt u1 + ψ (u2 ),
λ
t
0 (t, u1 , u2 ) = ψ Z (1 (s, u1 , u2 ))ds.
0
0 (t, u1 , u2 )
aλ b − 1 (t, u1 , u2 )
b log + tψ L
(u ) if bλ = ψL (u2 ),
bλ − ψ L (u2 ) b − u1
2
=
b
−aλ (e λt
− 1) + t if bλ = ψL (u2 ),
λu1 − ψ L (u2 )
where log denotes the distinguished logarithm in the sense of [19, Lemma 7.6], i.e.,
the branch is chosen such that the resulting function is continuous in t.
S := {z ∈ C2 : Re(z) ∈ (−∞, (M ∨ 0) +
) × ((2R ∧ 0) −
, (2R ∨ 2) +
)},
for some
> 0 and M := sup{21 (T − tn , 0, r) : r ∈ [R ∧ 0, R ∨ 0], tn ∈ T0 }.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch15 FA
Example 15.11. By the proof of [12, Theorems 3.3, 3.4], Assumption 15.10 holds in
the OU-time-change models from Example 15.9, if the Lévy exponents ψ L and ψZ
admit analytic extensions to {z ∈ C : Re(z) ∈ ((2R ∧ 0) −
, (2R ∨ 2) +
)} resp.
{z ∈ C : Re(z) ∈ (−∞, M +
)} for some
> 0. For example, if L is chosen to be an
NIG process with Lévy exponent
ψL (u) = uµ + δ( α2 − β2 − α2 − (β + u)2 ),
for µ ∈ R, δ, α > 0, β ∈ (−α, α) in the Gamma-OU-time-change model from
Example 15.9, one easily shows that the Lévy exponents ψ Z and ψ L admit analytic
extensions to {z ∈ C : Re(z) ∈ (−∞, b)} resp. {z ∈ C : Re(z) ∈ (−α − β, α − β)}.
Consequently, checking the validity of Assumption 15.10 amounts to verifying
M < b, 2R > −α − β, 2R ∨ 2 < α − β,
for M = ((1 − e−λT )/λ)2 max{ψL (R ∧ 0), ψ L (R ∨ 0)} ≥ 0 in this case.
By [7, Theorem 2.16(ii)], Assumption 15.10 implies that the exponential moment
formula (15.3) holds for all z ∈ S. In particular, S is square-integrable. We proceed
by providing sufficient and essentially necessary conditions that ensure the validity of
the martingale Assumption 15.2 and the non-degeneracy condition (15.1).
Example 15.13. For OU-time-change models, the martingale conditions (15.4) read
as ψL (1) = 0, i.e., exp(L) has to be a martingale. For example, in the NIG-OU models
from Example 15.11, this means
µ = δ( α2 − (β + 1)2 − α2 − β2 ).
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch15 FA
From now on, Assumptions 15.10 and 15.12 are supposed to be in force. Combined
with Theorem 15.6, Assumption 15.10 allows us to compute the variance-optimal
initial capital v0 and the variance-optimal hedging strategy ϕt at time t by performing
single numerical integrations.
Moreover, for tn ∈ T ,
R+i∞
V(z)tn−1 exp(κ0 (tn , z) + κ1 (tn , z)ytn−1 ) − 1
ϕtn = l(z)dz,
R−i∞ Stn−1 exp(δ0 + δ1 ytn−1 ) − 1
where, for j = 0, 1,
T
δ j = j , 0, 2 ,
N
T T
κj (t, z) := j , 1 (T − t, 0, z), z + 1 − j , 1 (T − t, 0, z), z .
N N
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch15 FA
Proof. The formula for V(z) follows immediately from Assumption 15.10 and [7,
Theorem 2.16(ii)]. Analogously, we obtain
and
with
T
κ0 (t, z) = 0 , 1 (T − t, 0, z), z + 1
N
T
−0 T − t + , 0, z + 0 (T − t, 0, z),
N
T T
κ1 (t, z) = 1 , 1 (T − t, 0, z), z + 1 − 1 T − t + , 0, z .
N N
We now consider the expression for the minimal expected squared hedging error
in Corollary 15.7. The first term J1 represents the variance of an unhedged exposure
to the option. In view of Assumption 15.10, it can be computed by evaluating a double
integral with the following integrand.
We now turn to the second term in the formula for the hedging error in Corol-
lary 15.7. Suppose for the moment that (15.2) holds. By Eqs. (15.6) and (15.7), we
have
(eκ0 (tn ,z1 )+κ1 (tn ,z1 )ytn−1 − 1)(eκ0 (tn ,z2 )+κ1 (tn ,z2 )ytn−1 − 1)
× .
eδ0 +δ1 ytn−1 − 1
E[J2 (tn , z1 , z2 )]
= (I(κ1 (tn , z1 ) + κ1 (tn , z2 ), tn , z1 , z2 )eκ0 (tn ,z1 )+κ0 (tn ,z2 ) + I(0, tn , z1 , z2 )
− I(κ1 (tn , z2 ), tn , z1 , z2 )eκ0 (tn ,z2 ) − I(κ1 (tn , z1 ), tn , z1 , z2 )eκ0 (tn ,z1 ) )
where
I(u, tn , z1 , z2 )
e(u+1 (T −tn−1 ,0,z1 )+1 (T −tn−1 ,0,z2 ))ytn−1 +(z1 +z2 )Xtn−1
:= E .
exp(δ0 + δ1 ytn−1 ) − 1
(eν0 (z1 )+ν1 (z1 )y0 − 1)(eν0 (z2 )+ν1 (z2 )y0 − 1)
E[J2 (T, z1 , z2 )] = V(z1 )0 V(z2 )0 ,
e0 (T,0,2)+1 (T,0,2)y0 − 1
for νj (z) = j (T, 0, z + 1) − j (T, 0, z), j = 0, 1, which allows to compute the static
hedging error by evaluating a double integral.
For Lévy processes, we have δ1 = 0. Hence the denominator in the expression
for I reduces to a constant in this case and the expectations can be computed using
(15.3). This leads to the formula obtained in [10].
For affine models, one can verify that as the number of trading times N tends
to infinity, the argument of the expectation in the expression for I converges to an
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch15 FA
exp(δ0 + δ1 yt ) − 1 ≈ δ0 + δ1 yt ,
which seems reasonable as the number N of trading dates tends to infinity, because
both δ0 = 0 (T/N, 0, 2) and δ1 = 1 (T/N, 0, 2) converge to zero in this case. We
obtain the following first-order approximation.
S := {z ∈ C2 : Re(z) ∈ (−∞, (M ∨ 0) +
) × ((2R ∧ 0) −
, (2R ∨ 2) +
)},
for some
> 0 and M := M ∨ 21 (T/N, M/2, R + 1).
(2)
exp(M yt + 2RXt )
E < ∞. (15.8)
δ0 + δ1 yt
Then (15.2) is satisfied and for n = 1, . . . , N, z1 , z2 ∈ R + iR,
u ∈ 1 (T − tn−1 , 0, z1 ) + 1 (T − tn−1 , 0, z2 )
+ {κ1 (tn , z1 ) + κ1 (tn , z2 ), κ1 (tn , z1 ), κ1 (tn , z2 ), 0},
we have
exp(uytn−1 + (z1 + z2 )Xtn−1 )
E
δ0 + δ1 ytn−1
−uδ0 /δ1 1
e δ1 δ0
+ us exp us + χ0 (s) + χ1 (s)y0 ds if δ0 , δ1 = 0,
δ1 δ0 δ1
0
1
= 1 1 + us
exp(ϑ0 (s) + ϑ1 (s)y0 )ds if δ0 = 0,
δ1 0 s
1
exp(0 (tn−1 , u, z1 + z2 ) + 1 (tn−1 , u, z1 + z2 )y0 ) if δ1 = 0,
δ0
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch15 FA
where, for j = 0, 1,
T δ1
δ j = j , 0, 2 , χj (s) := j tn−1 , log(s) + us, z1 + z2 ,
N δ0
E[V(z2 )tn Stn |Ftn−1 ] − V(z2 )tn−1 Stn−1
×
. (15.9)
S
tn−1
≤ S0R e0 (T −tn−1 ,0,R) e 2 M ytn−1 +RXtn−1 ,
1
where we have used Assumption 15.10 and [7, Theorem 2.16(ii)] for the equality.
Together with (15.9), this implies
exp(M ytn−1 + 2RXtn−1 )
|J2 (tn , z1 , z2 )| ≤ C ,
δ0 + δ1 ytn−1
for some constant C > 0 which does not depend on ω and z1 , z2 . Consequently, (15.8)
and Assumption 15.4 yield (15.2). The second part of the assertion now follows along
the lines of the proof of [13, Theorem 4.2] under the stated assumptions.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch15 FA
1 1 1
= − ,
δ0 + δ1 yt + 1
(δ
2 0
+ δ 1 yt )2 δ0 + δ1 yt δ 0 + 2 + δ1 y t
1
exp(δ0 + δ1 yt ) − 1 ≈ δ0 + δ1 yt + (δ0 + δ1 yt )2
2
follows directly from Theorem 15.16.
Instead of using the closed-form approximation proposed above, one can
eschew semi-explicit computations and instead calculate the hedging error using a
Monte–Carlo simulation as in [6]:
1 K
(4) Use the empirical mean K k=1 J0 (ωk ) as an estimator for J0 .
In addition to its simplicity, this approach has the advantage of approximating the
entire distribution of the hedging error, rather than just its mean. On the other hand,
computation time is increased.
In order to illustrate the applicability of our formulas and examine the effect of discrete
trading, we now investigate a numerical example. More specifically, we consider the
NIG-Gamma-OU model from Examples 15.9 and 15.11. As for parameters, we use the
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch15 FA
values estimated in [16] using the generalized method of moments, adjusting the drift
rate µ of L in order to ensure the martingale property of S:
β = −16.0, α = 90.1, δ = 85.9, µ = 15.0,
λ = 2.54, a = 0.847, b = 17.5.
By Example 15.11, Assumption 15.10 and the prerequisites of Theorem 15.16 are
satisfied for European call options and R = 1.1. Henceforth, we consider a European
call with discounted strike K = 100 and maturity T = 0.25 years. The results for
the variance-optimal initial hedge ratio ϕ0 for N = 1 (static hedging) and N = 12
(weekly rebalancing) are shown in Figs 15.1 and 15.2.
For static hedging, the impact of discretization seems to be quite pronounced,
in particular for out-of-the-money options. Also notice that this effect turns out
to be substantially bigger for the NIG-Gamma-OU than for the Black–Scholes
model. For weekly rebalancing, the effect of discretization on the initial hedge ratio
already becomes marginal. More specifically, the difference between the discrete- and
continuous-time variance-optimal hedging strategies is barely visible in Fig. 15.2.
Figure 15.3 shows a simulated path of the discrete variance-optimal hedges for
N = 1, 3, 12, 60.
0.9
0.8
0.7
0.6
Hedge ratio
0.5
0.4
0.3
0.2
N=1
0.1 N=3
N=12
N=60
0
0 0.05 0.1 0.15 0.2 0.25
Time
Figure 15.3 A simulated path of optimal hedge ratios for N = 1, 3, 12, 60.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch15 FA
We now turn to the minimal expected squared hedging error, which is depicted
for N = 1 (static hedging) to N = 60 (daily rebalancing) in Fig. 15.4. As the number
N of trading dates tends to infinity, the discrete hedging errors approach the respective
continuous-time limits both in the Black–Scholes model and in the NIG-Gamma-OU
model. Naturally, this limit vanishes in the complete Black–Scholes model. As noticed
above, the static hedging error for N = 1 can be computed without using any approx-
imations. For N ≥ 2, the discrete-time hedging error in the given NIG-Gamma-OU
model is approximated surprisingly well by the sum of the respective continuous-
time hedging error and the corresponding discrete-time hedging error in the Black–
Scholes model. In fact, the maximal absolute difference is smaller than 0.045. If such
an approximation can be used for the specific model at hand, computation time can
often be drastically reduced by evaluating the formulas from [10, 12] instead of The-
orem 15.16.
Note that the discrete hedging errors in the NIG-Gamma-OU model have been
approximated using Theorem 15.16. Since the corresponding results for a simulation
study using one million Monte–Carlo runs differ by less than 2.5% for N = 1, . . . , 60,
we do not show them here. However, in Fig. 15.5, we use the results of the Monte–
Carlo study to depict an approximation of the distribution of the hedging error for
Figure 15.5 Approximated distribution of the hedging error for N = 1, 12, 60.
N = 1, N = 12, and N = 60. Apparently, not only the variance of the hedging error,
but also its law depend crucially on the rebalancing frequency.
Acknowledgments
The first and fourth authors gratefully acknowledge financial support through Sach-
beihilfe KA 1682/2-1 of the Deutsche Forschungsgemeinschaft.
References
[5] Černý, A and J Kallsen (2007). On the structure of general mean–variance hedging strate-
gies. The Annals of Probability, 35, 1479–1531.
[6] Cont, R, P Tankov and E Voltchkova (2007). Hedging with options in presence of jumps.
In Stochastic Analysis and Applications: The Abel Symposium 2005 in honor of Kiyosi Ito,
F Benth, G Di Nunno, T Lindstrom, B Øksendal and T Zhang, (eds.), pp. 197–218. Berlin:
Springer.
[7] Duffie, D, D Filipovic and W Schachermayer (2003). Affine processes and applications in
finance. The Annals of Applied Probability, 13, 984–1053.
[8] Filipović, D and E Mayerhofer (2009).Affine diffusion processes: Theory and applications.
Radon Series on Computational and Applied Mathematics, 8, 1–40.
[9] Heston, S (1993). A closed-form solution for options with stochastic volatilities with
applications to bond and currency options. The Review of Financial Studies, 6, 327–343.
[10] Hubalek, F, L Krawczyk and J Kallsen (2006). Variance-optimal hedging for processes
with stationary independent increments. The Annals of Applied Probability, 16, 853–885.
[11] Kallsen, J (2006). A didactic note on affine stochastic volatility models. In From Stochastic
Calculus to Mathematical Finance, Y. Kabanov, R Liptser, and J Stoyanov (eds.), pp. 343–
368. Berlin: Springer.
[12] Kallsen, J and A Pauwels (2009). Variance-optimal hedging for time-changed Lévy pro-
cesses. To appear in Applied Mathematical Finance.
[13] Kallsen, J and A Pauwels (2009). Variance-optimal hedging in general affine stochastic
volatility models. To appear in Advances in Applied Probability.
[14] Kallsen, J and RVierthauer (2009). Quadratic hedging in affine stochastic volatility models.
Review of Derivatives Research, 12, 3–27.
[15] Lamberton, D and B Lapeyre (1996). Stochastic Calculus Applied to Finance. London:
Chapman & Hall.
[16] Muhle-Karbe, J (2009). On Utility-Based Investment, Pricing and Hedging in Incomplete
Markets. Dissertation, Technische Universität, München.
[17] Pauwels, A (2006). Varianz-optimales Hedging in affinen Volatilitätsmodellen. Disserta-
tion, Technische Universität, München.
[18] Pham, H (2000). On quadratic hedging in continuous time. Mathematical Methods of
Operations Research, 51, 315–339.
[19] Sato, K (1999). Lévy Processes and Infinitely Divisible Distributions. Cambridge:
Cambridge University Press.
[20] Schoutens, W (2003). Lévy Processes in Finance. New York: Wiley.
[21] Schweizer, M (1995). Variance-optimal hedging in discrete time. Mathematics of Opera-
tions Research, 20, 1–32.
[22] Schweizer, M (2001). A guided tour through quadratic hedging approaches. In Option
Pricing, Interest Rates and Risk Management, E Jouini, J Cvitanic, and M Musiela (eds.),
pp. 538–574. Cambridge: Cambridge University Press.
May 12, 2010 17:47 WSPC/SPI-B913 b913-ch15 FA
INDEX
395
May 12, 2010 18:8 WSPC/SPI-B913 b913-Index FA
396 Index
CPDO/constant porportion debt obligation, freight futures, 71, 78, 79, 81, 86
259–261, 285–292 freight rates, 74, 75, 77, 80–82, 90
CPPI, 201–203, 205–207, 209–211,
213–224 Gamma distribution, 271
CPPI strategies, 295, 296, 300, 304, 305, Gamma-OU process, 381
311, 321, 323, 324 gap risk, 229, 230, 241, 246, 247, 251, 259,
CPPI/constant proportion portfolio 268, 270, 277–279, 281–284, 287, 292
insurance, 259–261, 267–270, 275, 278, Global irradiance, 53, 55–58, 60–62
281, 282, 284, 285, 289, 291, 292 granger causality, 82, 83
Credit default swap index/CDS index, 154, guarantees, 228–230, 242
155, 157, 161, 169
Credit default swap/CDS, 148, 149, Heath–Jarrow–Morton, 94
153–155, 157, 158, 160–163, 165, 169, Hull-White model, 56
170
CreditGrades, 125 index certificates, 124, 126, 135
cross asset portfolio derivatives, 175–177, inflation, 51–53, 55–57, 65
182, 191, 195 inflation protection, 52, 53, 55
cushion, 267–269, 275, 277, 282, 284, 305, infrastructure, 51–53
307 issuer risk, 123–125, 130, 132, 139
iTraxx, 261, 266, 272, 275–277, 285, 289,
default risk, 123 290, 292
defaultable knock-out put options, 126
discount certificates, 126 Jensen’s Inequality, 107–110
discretely sample options, 347, 362, 367
dynamic asset allocation, 296 Lévy process, 264
dynamic portfolio insurance, 201 Laplace transform approach, 376, 378
leverage, 261, 267, 268, 275, 285–288, 290,
Economic Scenario Generator, 56 291
efficient portfolios, 295, 296, 299, 300, 312, LIBOR model, 98
313, 316, 324, 325 limited partnership fund, 51, 53
estimation risk, 295, 296, 298, 313, 314, log-normal (shifted), 126
316, 320, 321, 324, 325 lookback options, 346, 347, 364, 368
ETF, 201, 203, 210, 223, 224
European options, 348, 349, 355, 357, 358, Macroeconomic factors, 56, 57
361, 363, 367, 369, 370 Margrabe Formula, 117
(extended) roll-down call, 362 Markov-Switching model, 3, 4, 11, 12, 19
expected shortfall, 308, 309, 312 Markov transition matrix, 4, 6, 7
expected-squared hedging error, 375 maximum Sharpe ratio portfolio, 299, 301,
explanatory variables, 72, 79–82, 312, 314, 316, 321, 322
84–86, 90 mean-variance optimization, 299, 300, 310,
exposure, 303, 305, 307 311
extreme value theory (EVT), 282 method of moments, 11
minimum variance portfolio, 299, 301, 312,
floor, 303, 305, 307–309 316, 320, 322–324
forward freight agreement, 77, 78 model-independent strategies, 359
freight derivatives, 71, 77, 78 Monte Carlo simulation, 13
May 12, 2010 18:8 WSPC/SPI-B913 b913-Index FA
Index 397