Вы находитесь на странице: 1из 100

Better investing through factors, regimes and

sensitivity analysis


Cristian Homescu

This version: January 25, 2015

Recent periods of market turbulence and stress have created considerable interest in credible al-
ternatives to traditional asset allocation methodologies. It would be preferred if portfolios can be
decomposed into components that can be directly connected to independent risks and individually
rewarded by the market for their level of risk. This can be achieved through factor-based investing,
which relies on the observation that most return and risk characteristics for all asset classes can be
well explained by particular building blocks, or factors.
We describe main features of factors, factor investing and factor models, with emphasis placed on
ptactical topics such as selection of signicant factors associated to specic asset classes, dierentiating
between factors, anomalies or stylized facts, and preference for composite portfolios based on combining
factors. We have also analyzed implementation details and the factor risk parity strategy.
Then we consider improvements to factor-based investing through regime switching and sensitivity
analysis. We present theoretical and practical frameworks for Markov switching models and for sensi-
tivity analysis, and rely on representative examples to illustrate the benets of eciently incorporating
regimes and sensitivity analysis into portfolio management.
The nal section describes features of good testing procedures for portfolio behavior and performance,
in contrasts with possible testing pitfalls.


Email: cristian.homescu@gmail.com

Electronic copy available at: http://ssrn.com/abstract=2557236


Contents
Contents 2
1 Introduction 4
2 Factor investing 5
2.1 What are factors? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Factors and asset classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2.1 Literature overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.2 Risk factors in commodities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.3 Risk factors in currencies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2.4 Risk factors in bonds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2.5 Risk factors in global equities . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.2.6 Risk factors in private equity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2.7 Risk factors in real estate and REIT . . . . . . . . . . . . . . . . . . . . . . . . 18
2.3 Factor exposures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.4 Factor-based risk analysis, scenario simulation and stress testing . . . . . . . . . . . . . 19
2.5 Which factors to choose? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.6 Factors, anomalies or stylized facts? . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.7 How many factors are needed? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.8 Combining factors is a necessary ingredient in asset allocation . . . . . . . . . . . . . . 21
2.9 Risk factor parity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.10 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.10.1 Which asset universe to consider? . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.10.2 Dependency on weighting scheme . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.10.3 What is the optimal balancing frequency? . . . . . . . . . . . . . . . . . . . . . 28
2.10.4 Long only or long-short? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.10.5 Capacity constraints and transaction costs . . . . . . . . . . . . . . . . . . . . . 29
2.10.6 Investability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.11 Final thoughts on eective portfolio construction . . . . . . . . . . . . . . . . . . . . . 30

3 Factor models 30
3.1 Why do we need factor models? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.2 A primer on factor models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.3 Types of factor models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

4 Markov switching models for nancial markets 33


4.1 Reasons to consider Markov switching models . . . . . . . . . . . . . . . . . . . . . . . 34
4.2 A primer on Markov switching models . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.3 Estimating parameters of Markov switching models . . . . . . . . . . . . . . . . . . . . 35
4.4 Decoding the hidden states . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

5 Regime based investing 36


5.1 Literature review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
5.2 Finer points to be considered . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.3 Representative regime switching frameworks . . . . . . . . . . . . . . . . . . . . . . . . 39
5.3.1 First representative framework: two regimes applied to benchmark index . . . . 39
5.3.2 Second framework: more than two regimes applied to asset classes . . . . . . . 42
5.3.3 Third framework: regimes applied to factors . . . . . . . . . . . . . . . . . . . . 48

Electronic copy available at: http://ssrn.com/abstract=2557236


6 Sensitivity analysis in portfolio management 55
6.1 A primer on sensitivity analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
6.2 Sensitivity analysis and factor investing . . . . . . . . . . . . . . . . . . . . . . . . . . 56

7 Good testing procedures in portfolio management 60


7.1 Only backtesting is not enough . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
7.2 Backtest: a deeper discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
7.2.1 In-sample and out-of-sample . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
7.2.2 Overtting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
7.2.3 Minimum track record and backtest period . . . . . . . . . . . . . . . . . . . . 62
7.2.4 Assessing a backtest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
7.3 Multiple testing framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
7.3.1 Testing multiple hypotheses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
7.3.2 False discoveries and missed discoveries . . . . . . . . . . . . . . . . . . . . . . 64
7.3.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
7.3.4 When to stop testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
7.4 Evaluation based on Monte Carlo simulations . . . . . . . . . . . . . . . . . . . . . . . 67
7.5 Stress testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
7.5.1 Key questions and decisions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
7.5.2 Standard, historical or worst case? . . . . . . . . . . . . . . . . . . . . . . . . . 69
7.5.3 Entropy pooling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
7.5.4 Stressing covariance or correlation matrices . . . . . . . . . . . . . . . . . . . . 70
7.5.5 Reducing dimensionality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
7.5.6 Reverse stress testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

8 Conclusion 71
Bibliography 72
Appendix A: Minimum number of observations 83
Appendix B: Regression analysis to compute factor model parameters 85
Appendix C: Estimating covariance using a factor model 86
Appendix D: Parameter estimation for regime switching 88
Appendix E: Ecient sensitivity analysis 90
Appendix F: Time dependency of factor-based returns, exposures and correlations 92
Appendix G: Estimating carry factor 96
Appendix H: Protocol for factor identication 97
Appendix I: Monte Carlo simulation 98

Electronic copy available at: http://ssrn.com/abstract=2557236


1 Introduction
Traditional investment portfolios can typically be characterized by relatively static allocation and
by diversication across asset classes. In search of higher returns and better risk management and
diversication, institutional investors have augmented equity and xed income allocations with a wide
range of strategies (including full and partial long/ short, risk parity, and low-volatility) and alternative
investment classes.
However, market behavior in the past decade has uncovered shortcomings of the paradigm that
asset classes are as independent as possible, with little overlap and which, in aggregate, would cover
the investment universe with minimal gaps. Portfolio behavior during market turmoil has revealed that
traditionally constructed portfolios can be poorly diversied, with a pro-cyclical growth bias that may
lead to signicant drawdowns and losses. Moreover, traditional investing and rebalancing strategies
could not take into account dierent market regimes, and thus missing the portfolio customization
associated with regime-based investing.
With the lower risk tolerance of investors, an emphasis on risk-based investment style has emerged
since recent market turbulences. An eective risk management process allows the portfolio manager
to optimize the decision process during the investment cycle, and to put in place corrective measures
whenever necessary. To highlight the importance of risk management, Klement [2011] noted that,
in the long run, the investment will likely outperform the market even if it has much lower average
returns, provided that downside risks are greatly minimized through appropriate risk management and
prudent exit strategies.
Portfolio management from perspective of risk management was the main topic of the companion
paper Homescu [2014a]. Within the framework of risk, special attention must be given to protection
against extreme events, the so-called tail risk. This topic is rich and relevant enough to deserve its
own book Bhansali [2014], and was also the main subject of a companion paper Homescu [2014b].
Many traditional asset class and sub-asset-class correlations have trended upward over the past
decade, providing evidence against traditional usage of asset classes as primary portfolio building
blocks. These correlations rose to uncomfortable levels during the 200809 crisis, and seemingly dis-
parate asset classes (including alternative investment classes, such as hedge funds, commodities, and
infrastructure) moved in lockstep during the depths of the crisis, causing many investors to question
basic assumptions about traditional models and realize that such high correlations were due to the
signicant overlap in their underlying common risk factor exposures.
Ideally, investors could create portfolios using many components with independent risks that are
individually rewarded by the market for their level of risk. It is now widely accepted that most asset
classes could be broken down into building blocks (factors) that explain the majority of their return
and risk characteristics. While it is true that asset-based investing would provide an indirect way to
invest in factors, the main advantage of a direct factor-based investing approach is that factors can be
remixed into portfolios that are better diversied and more ecient than traditional methods would
allow.
The main goal of this report is to provide an in-depth analysis of the factor-based investment proces,
and also how to enhance it through regime investing and sensitivity analysis. The structure of the
article is as follows: the rst two chapters describe characteristics of factors, factor investing and factor
models. The next two chapters presents for Markov switching models and their usage in context of
regime investing. The principal topic of the next chapter is ecient deployment of sensitivity analysis
within portfolio management. The nal chapter addresses testing pitfalls and discusses features of
good testing procedures for portfolio behavior and performance.

4
2 Factor investing
Factor investing has become a widely deployed tool in the universe of investment strategies, as exem-
plied by many recent references written by practitioners and academics.
1

Bhansali [2011] argued that investors would be better o diversifying their exposures across risk
factors than asset classes. His perspective is based on the observation that macroeconomic forecasts
can be mapped easily to risk factors, whereas the mapping to asset classes (which are complex baskets
of risk factors) is obscured. Moreover, results indicate that four or ve underlying risk factors can
explain approximately 70% of the variation in most liquid assets.
Ang [2013b] considers that in very large portfolios it is very hard to nd excess returns that are not
related to factors, given that many mispricing opportunities are not scalable, manager decisions (both
internal and external) tend to be correlated as it is hard to nd truly independent portfolio strategies,
while security-specic bets are swamped at the portfolio level by macro-economic and factor risks. His
analogy was:

A farmer, for example, may certainly be able to select a farm with the best soil and conditions
for planting (farm security selection). But if there's a severe drought, having chosen the best
farm is not going to help (rain is the factor).

Thousands of correlated individual bets by managers eectively become large bets on factors. An
equity manager going long 1,000 value-oriented stocks and under-weighting 1,000 growth stocks does
not have 1,000 separate bets; he has one big bet on the value-growth factor. A xed income manager
who squeezes out the last bits of yield by nding 1,000 relatively illiquid bonds funded by short positions
in 1,000 liquid bonds (through the repo market) has made one bet on an illiquidity factor.
In a nutshell, factor investing aims to capture particular factor risk premia in a systematic way, for
example by building a factor index and replicating it, or by constructing a portfolio that gives exposure
to a range of risk factors. The objective is to combine factors to enhance the performance of portfolios.
Although constructing ex ante optimized portfolios using risk factor inputs is possible, there are
signicant challenges to overcome, including the need for active, frequent rebalancing; creation of
forward looking assumptions; and the use of derivatives and short positions. Exactly how risk factors
should be included in the portfolio construction process is still debated among practitioners, and
practical solutions are needed in order to allow investors to potentially incorporate risk factors in the
portfolio construction process while accommodating their constraints and existing investment processes.

2.1 What are factors?


As described in Podkaminer [2013]:

Factors are the basic building blocks of asset classes and a source of common risk exposures
across asset classes. Factors are the smallest systematic (or non-idiosyncratic) units that in-
uence investment return and risk characteristics. They include such elements as ination,
GDP growth, currency, and convexity of returns. In a chemistry analogy: If asset classes are
molecules, then factors are atoms.

1
While the reference list below is representative, it is by no means complete: Bender et al. [2013a], Amenc et al. [2014],
Amenc and Goltz [2014], Ang [2013a,b], Bambaci et al. [2013], Barber et al. [2014], Bhansali et al. [2012], Boudt
et al. [2014], Boudt and Peeters [2013], Bruneau et al. [2014], Carli et al. [2014], Cheung [2010, 2013], Cheung and
Mishra [2010, 2012], Cheung and Mittal [2009], Connor and Korajczyk [2010],Page [2010], Page and Taborsky [2011],
Perchet et al. [2014], Podkaminer [2013], Roncalli and Weisang [2012], Rosen and Saunders [2010], Werley [2011],
Werley and Speer [2011], Williams [2014], Witham [2012], Witte et al. [2013], Zivot [2011],Fama and French [2014,
1992], Fan et al., 2008, 2011], Gnedenko and Yelnik [2014], Hafner [2004], Huij et al. [2014], Jay et al. [2011], Jiang
and Martin [2013], Kang and Ung [2014], Koedijk et al. [2014], Y. et al. [2011], Martin [2013], Meucci [2010], Meucci
et al. [2014a,b], Minami [2013a,b], Ilmanen [2011].

5
Criteria to determine whether a systematic unit is a factor are listed in Ang [2013b], Ang et al. [2009]:

1. Have strong support in academic and practitioner research and economic justications

• Factors should have an intellectual foundation

• Research should demonstrate compelling rational logic, behavioral stories or both in ex-
plaining the factor eect

2. Have exhibited signicant premiums that are expected to persist in the future

• Factors are systematic by denition, arising from risk or behavioral tendencies which will
likely persist

• Factors' explanatory power should remain observable even if everyone knows about the
factors and many investors are pursuing the same factor strategies.

3. Have return history available for bad times

• Factor risk premiums exist because they reward the investors' willingness to suer losses
during bad times.

• Having some data points to measure worst-case scenarios is necessary for assessing risk-
return trade-os and risk management

• We also want a reasonable length of data to perform these exercises.

4. Be implementable in liquid, traded instruments

• Limiting ourselves to liquid securities does not make it impossible to construct a factor
illiquidity premium. One can collect an illiquidity premium by taking long positions in (rel-
atively) illiquid assets and short positions in (relatively) liquid assets, all in liquid markets.

• Illiquidity premiums within (liquid) asset classes are much larger than illiquidity premiums
across asset classes such as private equity or hedge funds

Moreover, a factor should only be considered for a benchmark if it is widely recognized. Many of
the factors that we recognize today (such as value-growth and momentum) were originally labeled
as anomalies, or alpha, as they could not be explained by existing (at the time) risk factors. As the
literature matured and there was growing acceptance of these strategies by institutional investors, they
became beta, or were treated as factor premiums in their own right.
We enumerate various types of factors employed in the literature:

• factors that reect forward-looking economic views: interest rates, slope of the yield curve,
corporate bond spreads, equity returns, changes in volatility, commodity returns, changes in
liquidity

• fundamental factors (momentum, value, size, etc) capture characteristics such as industry mem-
bership, country membership, valuation ratios, and technical indicators, and are usually incorpo-
rated in models, such as Fama-French or Carhart models (Fama and French [1992, 2014], Carhart
[1997])

• macroeconomic factors such as the ones described in Podkaminer [2013]: low/high growth,
low/high ination, ight to quality, surprises in ination, surprises in GNP

• statistical factors use statistical techniques such as principal components analysis (PCA) or
singular vakyue decomposition (SVD)

6
Following Bender et al. [2013a], we distinguish between generic factors and risk premia factors, with
the latter having earned a persistent signicant premium over long periods and reecting exposure to
sources of systematic risk. For instance all of the Fama-French factors count as risk premia factors
since the aim of those original studies was to isolate asset pricing drivers. However, not all of the Barra
Risk Model factors are candidate factors since the purpose of those models focuses on forecasting risk
and explaining fund performance. Examples of factors which are not considered good candidates for
long-term factor investing are Growth and Liquidity, which capture high growth stocks and highly
liquid stocks, because they have not outperformed the market over long periods.
The next 2 gures (from Podkaminer [2013]) are representative for the types of factors and factor
exposure across asset classes (in this case, US equity and US corporate bonds)

Figure 2.1: Major factor types

Figure 2.2: Factor exposure across asset classes

7
Bender et al. [2013a] identied the following six equity risk premia factors as most relevant, because
they have earned historically a premium over long periods, represent exposure to systematic sources
of risk, and have strong theoretical foundations.

Figure 2.3: Equity risk premia factors

MSCI indices were created accordingly, with following risk versus return characteristics:

Figure 2.4: Performance characteristics of MSCI Indiceses (from Bender et al. [2013b])

8
Asl and Etula [2012] have considered the following criteria:

1. each factor reects a distinct risk premium that is largely independent of others

2. each risk premium has a clear economic rationale

3. reward associated with each factor reects compensation for systematic risk in cross-section of
expected returns (assets that have higher exposure to factors are expected to earn higher returns)

4. each risk premium is investable, with factor returns achieved via long and short positions in
liquid, tradable assets or instruments

to identify six factors as sources of long-term return for a global investor, which are then deployed to
construct risk premium proles for various asset classes.

Figure 2.5: Six factors considered in Asl and Etula [2012]

Figure 2.6: Risk premium proles from Asl and Etula [2012]

Cazalet and Roncalli [2014] have found additional risk factors to be relevant, including:

• volatility
 dierent ways to look at volatility as a risk factor

 low volatility stocks outperform high volatility stocks

 stocks with a high sensitivity to innovation in aggregate volatility and with high idiosyncratic
volatility have low average returns

 betting-against-beta BAB risk factor (see Frazzini and Pedersen [2014])

9
• liquidity
 when liquidity level is measured by its bid-ask spread or trading volume, less liquid assets
outperform more liquid assets

 liquidity risk is a measure of the unexpected variation in total liquidity

• carry
 currency carry trade: investment that is long on currencies with high interest rates and
short on currencies with short interest rates will face a negative return on average

 for bonds: roll-down strategy and slope carry strategy

• quality
 success of the value strategy is explained by the strong performance of quality (strong value)
assets compared to distressed (weak value) assets

 Asness et al. [2013b] dened a quality security as  one that has characteristics that, all-else-
equal, an investor should be willing to pay a higher price for: stocks that are safe, protable,
growing, and well managed

• default risk
 spread between the return on a portfolio of long-term corporate bonds and the yield of
long-term government bonds

• co-skewness
 stocks with lower co skewness have higher expected returns

2.2 Factors and asset classes


While factors were originally described within the context of equity markets, many of them were found
to cut across asset class boundaries. For example, two of the most studied capital market phenomena,
value and momentum, were described in Asness et al. [2013a] in a generic way, applicable across many
asset classes:

• value eect describes the relation between an asset's return and the ratio of its long-run (or
book) value relative to its current market value

• momentum eect describes the relation between an asset's return and its recent relative per-
formance history

Moreover, the observation of both factors having to be used together, since data is best explained
by their combination while both strategies produce positive returns on average yet are negatively
correlated, remained valid for the additional asset classes considered.
Let us give a few examples.
Value-growth strategies buy cheap securities with high yields and sell expensive securities with low
yields. Foreign exchange's version of value-growth is carry: carry strategies go long currencies with
high interest rates and short currencies with low interest rates. In xed income, buying high yielding
bonds with long maturities funded by low yielding bonds with short maturities is called riding the
yield curve, and is related to the duration risk premium. In commodities, positive roll returns are
accomplished by buying cheap long-dated futures contracts that increase in price as their maturities
decrease.

10
Momentum is also pervasive and observed in equities, xed income, foreign exchange, real estate,
commodities (where it is often called CTA, for the type of fund which often pursues it), within and
across asset classes, and in asset classes all over the world. Similarly, selling volatility protection can
be done in options markets where the underlying securities are equities, xed income, commodities, or
foreign exchange.
Asness et al. [2013a] constructed zero-cost long-short portfolios that use the entire cross section of
securities within an asset class using weights based on value and momentum factors in a generic way.
For any security i = 1, ..., N at time t with signal Sit (value or momentum), securities are weighted
in proportion to their cross-sectional rank based on the signal minus the cross-sectional average rank
of that signal. Simply using ranks of the signals as portfolio weights helps mitigate the inuence
of outliers, but portfolios constructed using the raw signals are similar and generate slightly better
performance.
More specically, the weight on security i at time t was obtained as

PN !
S j rank [Sjt ]
wit = ci rank [Sit ] −
N

where the weights across all stocks sum to zero, representing a dollar-neutral long-short portfolio,
and ct is a scaling factor such that the overall portfolio is scaled to one dollar long and one dollar short.
Carry is another factor that could be described in a generic fashion. By dening carry as the return
to a security assuming its price/market conditions stays constant, Koijen et al. [2013] found that carry
predicts returns both in the cross section and time series for a variety of dierent asset classes including
global equities, global bonds, commodities, US Treasuries, credit, and options. Generic formulas for
computing the carry factor are given in Appendix G.

2.2.1 Literature overview


Relevant references include:

• international stocks and equity indices:Asness et al. [2013a], Koijen et al. [2013]

• individual commodities and commodity indices: Erb and Harvey [2006], Gorton et al. [2013],
Koijen et al. [2013], Blitz and de Groot [2014], Mire et al. [2012]

• currencies: Okunev and White [2003], Menkho et al. [2012], Asness et al. [2013a], Koijen et al.
[2013]

• corporate and government bonds: Asness et al. [2013a], De Jong and Driessen [2012], Lin et al.
[2011], Koijen et al. [2013], Houweling and van Zundert [2014]

• credit: Koijen et al. [2013]

• private equity: Kothari et al. [2012a,b]

• real estate and REIT: Mouchakkaa [2012]

Current research on main factors for various asset classes is summarized in following tables, following
Koedijk et al. [2014]:

11
Factors Found in Economic Rationale

Equity Compensation in excess of


Fixed income risk-free rate on a passive investment
Currency on a traditional broad source of risk
Ination Equity, Bonds, Currency, Commodities Measure for cost of living
Economic Growth Equity, Commodities

Table 1: Asset factors

Factor Found in Economic rationale

Carry Bonds, Currency Forward rates overestimate future spot rates


Trending Equity, Bonds, Currency, Commodities
Anomaly Mostly equity accruals anomaly, IPO anomaly
index-change anomaly

Table 2: Strategy factors

Factors Found in Economic Rationale


Equity, Bonds, Compensation for purchasing equity
Value Currency, Commodities below its intrinsic value, and the time
horizon needed before reverting to mean
Growth Equity, Bonds, purchase of stocks that have high
Currency, Commodities expected future growth rates
Momentum Equity, Bonds, Stocks that have done well in the
Currency, Commodities recent past will continue to do so
Low volatility Equity Low vol stocks earn high risk-adjusted
returns; investors overpay for risky stocks
Term Bonds positive dierence between long
term and short term interest rates
Credit Bonds Compensation for the risk of default
Volatility equity Protection sellers earn risk premiums embedded
in implied vol for providing insurance
Liquidity Equity, bonds
Emerging equity markets Equity
Convexity Bonds Bonds with a higher convexity
produce better results in volatile markets

Table 3: Style factors

2.2.2 Risk factors in commodities


The classic reasons for investors to allocate to commodities have been threefold:

1. to capture a potential commodity risk premium

2. to diversify a traditional equity/bond portfolio

3. to hedge ination risk

12
While traditional arguments for investing in commodities may not remained as valid in recent markets,
augmenting a portfolio with commodity factor premiums may signicantly increase performance. Such
factor premiums are documented to exist in the commodity market:

• momentum factor: Erb and Harvey [2006]

• carry factor is from Erb and Harvey [2006], Gorton and Rouwenhorst [2006], Koijen et al. [2013]

• low-volatility factor: Mire et al. [2012], Frazzini and Pedersen [2014].

• value: Asness et al. [2013a]

Returns for commodity futures are calculated following Asness et al. [2013a]. Each day one computes
the daily excess return of the most liquid futures contract, which is typically the nearest- or next
nearest-to-delivery contract, and then compounds the daily returns to a total return index from which
returns at a monthly horizon are calculated.
A methodology to construct the momentum, carry and low-risk factor premiums was described in
Blitz and de Groot [2014] using 24 individual commodities of the S&P GSCI index (which was also
used as a proxy for the commodity market premium during the calculations):

• six energy related commodities (crude oil, Brent crude oil, heating oil, gasoil, natural gas and
RBOB gasoline)

• seven metals (gold, silver, copper, aluminum, zinc, nickel and lead)

• eleven agricultural commodities (corn, soybeans, wheat, red wheat, sugar, coee, cocoa, cotton,
lean hogs, live cattle and feeder cattle).

They have ranked the commodities, at the end of every month, based on their past 12-month return
(for momentum), annualized ratio of nearby futures price to next-nearby futures price (for carry) and
past 3-year volatility using daily data (for low-risk)
2 . Next, the equally-weighted returns of the long-

only top 30 percent portfolio and the long/short top 30 percent minus bottom 30 percent portfolio
were evaluated over the subsequent month.
The value factor was computed in Asness et al. [2013a] as the log of the spot price 5 years ago
(actually, the average spot price from 4.5 to 5.5 years ago), divided by the most recent spot price,
which is essentially the negative of the spot return over the last 5 years.
Carry for a commodity futures contract was computed in Koijen et al. [2013] as

δt − rtf
Ct =
1 + rtf − δt

where rtf is the risk-free rate and δt is the convenience yield in excess of storage costs which can be
backed out from futures prices through formula

 
Ft = St 1 + rtf − δt

2
The long-short low-volatility factor is constructed in the spirit of BAB (Betting Against Beta) factor from Frazzini and
Pedersen [2014], with both low-risk top portfolio leveraged, and respectively high-risk bottom portfolio deleveraged,
to the average historical volatility of the long and short portfolio.

13
2.2.3 Risk factors in currencies
Currency returns can be related to and explained by various risk factors, with relevant references
including Okunev and White [2003], Menkho et al. [2012], Asness et al. [2013a], Koijen et al. [2013],
Prasetyo and Lo [2014].
The most included factors in the currency portfolio or index include:

• carry (interest-rate based):

 buy high interest rate currencies and sell low interest rate currencies

 positively correlated with credit and emerging market debt

• value (fundamental economics-based):

 long currencies that are undervalued relative to their fair value and short overvalued
currencies

 fair value dened as one of the following two alternatives:

∗ purchasing power parity (PPP) : prices of same goods and services in dierent coun-
tries

∗ balance of payments (BoP) consisting of current account (trade balance of goods and
services, income and current transfers), capital account (capital transfers, remittance
and acquisition/disposal of non-produced, non-nancial assets) and nancial account
(direct investment, portfolio investment and reserve account)

• momentum (price-based) or trend


 buy currencies with high returns and sell currencies with low returns in similar market
conditions

 performs best when markets show a persistent trend, and worst when markets are choppy
in tight and range-bound conditions

Additional factors are not yet widely considered, although there are studies arguing in their favor, such
as for beta based on volatility (currency option-based).
Characteristics of risk factor strategies were described as follows

Figure 2.7: Risk factor strategies for currencies (from Prasetyo and Lo [2014])

In the context of currencies as asset class, Asness et al. [2013a] dened the value measure as the
negative of the 5-year return on the exchange rate, measured as the log of the average spot exchange
rate from 4.5 to 5.5 years ago divided by the spot exchange rate today minus the log dierence in the

14
change in CPI in the foreign country relative to the U.S. over the same period. The currency value
measure is therefore the 5-year change in purchasing power parity.
For momentum they have computed the return over the past 12 months. We note that the most
recent month did not have to be excluded from calculations because currencies suer much less from
liquidity issues.
The carry of the currency dened in Koijen et al. [2013] is


rtf − rtf
Ct =
1 + rtf

with rtf the local risk-free rate and rtf the foreign risk-free rate.

2.2.4 Risk factors in bonds


Documented factors for bond markets include:

• Low-Risk (see Frazzini and Pedersen [2014], Houweling and van Zundert [2014], Ilmanen et al.
[2004])

• Momentum (seeAsness et al. [2013a], Houweling and van Zundert [2014], Jostova et al. [2013],
Pospisil and Zhang [2010] )

• Value (Asness et al. [2013a], Houweling and van Zundert [2014], Correia et al. [2012])

• Size (see Houweling and van Zundert [2014]

• Carry (see Koijen et al. [2013]

For factor-related calculations Houweling and van Zundert [2014] have considered two types of data
inclusion:

• only bond characteristics (rating, maturity, credit spread, etc.)

• a combination of bond characteristics, accounting data (leverage, protability, etc.), or equity


market information (equity returns, volatility, etc.).

The rst approach has the advantage of allowing inclusion of all bonds and not only of bonds issued
by companies with publicly listed equity, and thus facilitate the actual implementation of factors
in investment portfolios. The main advantage of second approach is potential improvement of the
performance of the resulting factor portfolio.
To dene the Size factor in the corporate bond market, one can use the total index weight of each
company, calculated as the sum of the market value weights of all its bonds in the index in that month.
Alternatively, one could dene Size as the market capitalization of the company's equity, the enterprise
value or the value of its assets, although this denition would require that the company has publicly
traded equity.
Low-risk factor portfolio can be constructed using both maturity and rating, as described in Ilmanen
[2011], Houweling and van Zundert [2014]: the rst step is selection of all bonds rated AAA to A- for
Investment Grade and all bonds rated BB+ to B- for High Yield, and then each month choose all
bonds shorter than M years such that the portfolio makes up a given percentage of the total number
of bonds.
The bond carry is computed in Koijen et al. [2013] using futures data, although in most cases this
procedure would have to rely on synthetic one-month futures due to characteristics of bond futures.
Liquid bond futures contracts are only traded in a few countries and, even when they exist, there are

15
often only very few contracts (typically just one). Further complicating matters is the fact that dierent
bonds have dierent coupon rates and the futures price is subject to cheapest-to-deliver options.
The bond carry is essentially the bond's yield spread to the risk-free rate (also called the slope of
the term structure) plus the roll down, which captures the price increase due to the fact that the
bond rolls down the yield curve.
We should note that bond carry could alternatively be computed under the assumption of a constant
bond price (leading carry to be the current yield if there is a coupon payment over the next time period,
otherwise zero) or the assumption of a constant yield to maturity (leading carry to be the yield to
maturity minus the risk free rate).
Regarding the value measure for bonds, Asness et al. [2013a] use the 5-year change in the yields of
10-year bonds, which is similar to the negative of the past 5-year return. Momentum is computed using
the past 12 months skipping the most recent month.
An alternative approach is considered in Houweling and van Zundert [2014]. The rst step is to
construct the Value factor portfolio each month, starting with a cross-sectional regression of credit
spreads on rating (AAA, AA+, AA, . . . , C) dummies and time-to-maturity. Then calculate the
percentage dierence between the actual credit spread and the tted credit spread for each bond, and
select a given percentage of bonds with the largest percentage deviations.
Regarding momentum eect in corporate bond market, Houweling and van Zundert [2014] have men-
tioned that results reported in the literature appear to be mixed. Momentum strategies have been shown
to generate prots in the High Yield market, but not for Investment Grade market, where correspond-
ing bond returns exhibit either reversal or no eect. The Momentum was dened in Houweling and
van Zundert [2014], Jostova et al. [2013] as the past 6-month return using a 1-month implementation
lag, with excess return versus duration-matched Treasuries used as return measure. The required
percentage bonds with the highest past returns are selected for the Momentum factor portfolio.

2.2.5 Risk factors in global equities


Asness et al. [2013a] have considered a universe of individual global stocks which exclude ADRs, REITs,
nancials, closed-end funds, foreign shares, and stocks with share prices less than $1 at the beginning of
each month, and limit the remaining universe of stocks in each market to a very liquid set of securities
that could be traded for reasonably low cost at reasonable trading volume size. Stocks were ranked
based on their beginning-of-month market capitalization in descending order.
For global equity indices they have considered country equity futures, to avoid including any returns
on collateral from transacting in futures contracts, and thus these returns can be seen as comparable
to returns in excess of the risk-free rate.
Koijen et al. [2013] has implemented a global equity carry strategy via futures. While one does
not always have an equity futures contract with exactly one month to expiration, one can interpolate
between the two nearest-to-maturity futures prices to compute a consistent series of synthetic one-
month equity futures prices. The carry of an equity futures contract was dened as the expected
dividend yield minus the local risk-free rate, multiplied by a scaling factor St/Ft (close to one).
Momentum was computed in Asness et al. [2013a] using the return over the past 12 months but
possibly excluding the most recent month, to avoid the 1-month reversal in stock returns due to
potential liquidity or microstructure issues.
The value measure was computed as follows:

• for individual stocks: use the signal of the ratio of the book value of equity to market value of
equity, or book-to-market ratio, with book values lagged 6 months to ensure data availability to
investors at the time, with the most recent market values used to compute the ratios.

• for indices: use the previous month's book-to-market ratio for the MSCI index of the country

16
2.2.6 Risk factors in private equity
Despite the proliferation of private equity (PE) as a mainstream asset class, there has been relatively
little research on the investment decision-making and asset allocation process employed by managers
of private equity portfolios (i.e., a portfolio that invests in a collection of individual private equity
partnerships). To quote from Kothari et al. [2012a]:

Is private equity truly a unique asset class to which widely established principles and asset-
allocation best-practices do not apply? If risk-type categories help improve diversication for
traditional asset classes, would they also be applicable to PE?

It turns out that, even for this asset class, a factor-based strategy systematically invested in PE fund
investments would generate a higher return for a given level of risk.
For example, Kothari et al. [2012a] have identied risk-type factors applicable to PE:

• Strategy/Product
 Dierent product types are typically representative of the dierent investment strategies
utilized by the General Partner when managing a specic fund.

 For example, a Buyout fund, which typically aims to purchase and improve underperforming
companies, has a much dierent risk and return prole than that of an early-stage Venture
Capital fund, which provides seed capital to startup companies.

• Fund Size
 The overall size (in AUM) of the individual private equity fund will inuence the size of the
underlying portfolios companies, and/or the number of investments in the fund.

 For example, Mega-Buyout funds have dierent risk/return characteristics than their smaller,
Mid-Market counterparts.

• Vintage
 The particular start date of a fund, the rst date in which that fund began to invest its
raised capital, can inuence the risk/reward prole.

 The economic environment during the period when the fund makes its investments can
inuence performance due to various factors, such as, for example, purchase multiple levels,
interest rates and the availability of credit.

• Industry focus
 Industries can be classied as cyclical or defensive, depending on how their prots are
aected by the business cycle.

 Funds that specialize or focus on specic industries can have drastically dierent risk/return
proles than other, more diversied fun

• Geographic Exposure
 Dierent geographical areas  U.S. versus non-U.S. for example  have dierent economic
growth rates and overall economic conditions may signicantly impact investment returns.

17
2.2.7 Risk factors in real estate and REIT
Mouchakkaa [2012] identied various risk factors associated to real estate investing:

• Real Estate Exposure: Investors' choice on overall allocation to real estate, within a multi-asset
class portfolio

• Vintage Year Exposure: Timing and pace of real estate exposure, including some important
considerations on measurement and monitoring of vintage year exposure

• Property Sector Exposure: Portfolio concentrations relative to managers, property types, ge-
ography, strategic risk classication, investment life cycle, and investment structure

• Leverage: Approach and exposure to leverage in investment ventures and portfolios, including
suggested metrics for measuring and monitoring leverage

Gold [2014] argued that individual REITs have exposures to:

• REIT Benchmark

• Property type

• Geographic region

• Market momentum

2.3 Factor exposures


According to Bender et al. [2013b], factor exposure captures the degree to which the index captures the
pure non-investable factor. While by denition the factors are the basic building blocks of investments,
gaining factor exposure may be challenging, since in many cases there is no natural way to invest
in many of them directly. An example of such a challenge is to obtain exposure to GDP growth. For
some other factors establishing exposures is simpler, although it may necessitate derivatives and/or
long/short positions.
Here are some examples of proxies for factor exposures:

• Inflation : Long nominal Treasuries index, short TIPS index

• Real interest rates : Long TIPS index

• Volatility : Long VIX Futures Index

• Value : Long developed country equity value index, short developed country growth index

• Size : Long developed country small-cap index, short developed country large-cap index

• Credit spread : Long U.S. high-quality credit index, short U.S. Treasury index

• Duration : Long Treasury 20+ year index, short Treasury 13 year index

• Emerging markets : Long global emerging market index, short global index

18
2.4 Factor-based risk analysis, scenario simulation and stress testing
Factor exposures are increasingly deployed (Asl and Etula [2012], Werley and Speer [2011], Ardia and
Meucci [2014], Meucci et al. [2014b, 2012], Meucci [2014, 2012]) to better capture real-world character-
istics such as fat tails and increased correlations in times of crisis, to create a custom benchmark for a
portfolio or fund, to perform risk analysis, or to construct stress testing and scenario simulations.
Traditional approaches tend to assume that asset returns are normally distributed and that con-
secutive returns are independent of one another. The reality is dierent, and is often associated with
sequences of bad returns, clusters of high volatility and increased correlations among risky assets dur-
ing periods of market turbulence. Such characteristics are much better captured by factor-based risk
analytics, which may also estimate portfolio's potential performance during historical stress periods
for which return data may not be fully available, e.g., 1973-1974 oil embargo or late 70s US high ina-
tion. More realistic factor-based distributions can also deliver Monte Carlo simulations with a better
chance to anticipate eects of future actions such as planned spending, inows, ination or taxes on
the distribution of future portfolio returns.

2.5 Which factors to choose?


One of the major questions for factor investing, discussed in Cazalet and Roncalli [2014], Jacobs and
Levy [2014], Pukthuanthong and Roll [2014], is how to concentrate the eorts on the factors that really
matter. For a long time, the standard model was the four factor model of Carhart [1997] which added
the momentum factor to the 3-factor model developed by Fama and French [1992].
More recently, other factors have emerged such as the volatility, low beta, quality or liquidity factors,
protability, to name just a few. For instance Subrahmanyam [2010] surveyed more than fty charac-
teristics that various papers contend are cross-sectionally related to mean returns. In fact, Cochrane
[2011] has recently referred to the situation as a zoo of new factors . Harvey et al. [2014] counted over
300 in academic papers, and this number has been increasing exponentially:

Figure 2.8: Factors and publications (from Harvey et al. [2014]

19
Are all these candidate factors relevant, or many
Such a data explosion begs the question:
of them were found simply because of extensive data mining?
Given the increasing number of potential factors, we may need quantitative protocols to decide on
the relevance and eect of factors. One such protocol was proposed in Pukthuanthong and Roll [2014]
for determining which factor candidates are related to both risk and returns, to neither, or to one but
not the other. For more details see Appendix H.
We conclude this section by noting that while factor allocations have the potential to change the
landscape of mandate structures by oering a new way to achieve exposure to systematic factors that
formerly could only be captured through active mandates, such allocations should be tailored to each
institution, to incorporate the institution's objectives and constraints (governance structure, horizon,
risk budget, and so on), and should include an assessment of the tradeo between investability and
factor exposure (which is tied to performance).

2.6 Factors, anomalies or stylized facts?


This question becomes increasingly important for practitioners and academics, e.g., Jacobs and Levy
[2014], Cazalet and Roncalli [2014]. In particular one needs to identify the risk factors which may
explain the cross-section of asset returns and is a compensation for taking an economic risk.
From one perspective we can think that the size factor is a risk factor, because small stocks are more
risky than large stocks. On the other hand, we can also think that the momentum factor is more an
anomaly. In this case, the abnormal return posted by this factor is not a risk premium, but can only
be explained by investment behaviors.
Moreover, a component may have both characteristics. Such an example is the value strategy,
which may be used to build aggressive portfolio by selecting distressed stocks, or to build defensive
portfolio by selecting quality stocks. Another example is given by January eect, which is a stylized
fact and can be seen as an anomaly or, alternatively, can become part of an investment strategy. It is
not a risk factor, however.
There are various reasons why some of the factors mentioned in the literature can be considered
anomalies, such as inability to be replicated or to predict returns out of sample (either in other time
periods or in other markets). It was also argued (Harvey et al. [2014]) that many factors have been
discovered because researchers frequently ignore the possibility that a certain number of factors are
bound to show statistically signicant results merely by chance. Extreme examples are given in Novy-
Marx [2014], with resulting showing that standard predictive regressions fail to reject the hypothesis
that returns depend on the party of the U.S. President, the weather in Manhattan, global warming,
El Nino, sunspots, or the conjunctions of the planets.
In more practical terms, a higher threshold can help weed out factors that appear valid but are
actually only the result of data mining or chance. Harvey et al. [2014] suggested using a t-statistic
of 3.0, rather than the traditional threshold of 2.0. However, even with this more stringent standard,
there is evidence for the market's much larger multidimensionality, as described in Jacobs and Levy
[2014], Green et al. [2013, 2014], Jacobs and Levy [1988], Lewellen [2015].
In a paper published 25 years ago, with factor returns puried via multivariate analysis Jacobs and
Levy [1988] have found that 9 of the 25 factors tested were signicant, with a t-statistic of 3.0 or
higher. The signicant factors included low price-to-earnings ratio, but not low share price; the sales-
to-price ratio, but not the book-to-price ratio; earnings surprises within the last month, but not those
in previous months; relative strength (price momentum); revisions in analysts' earnings estimates; and
return reversals. Small size was marginally signicant. Green et al. [2014] performed multivariate
testing on 100 factors and found 24 factors with t-statistics in excess of 3.0.
While the list of factors covered most of the factors which are now included under the smart beta
umbrella, we should observe that more factors have been identied as statistically signicant compared
to the number of factors generally pursued today by smart beta strategies. Interestingly, some popular

20
smart beta factors, such as size, book-to-price ratio, and price momentum, were not among the most
signicant factors, were not always identied as signicant, e.g., Green et al. [2014].

2.7 How many factors are needed?


Given the evidence for a larger number of signicant factors, it was argued in Jacobs and Levy [2014]
that smart beta strategies may need to include additional factors to the set of currently considered
factors (such as small size, value, price momentum, and low volatility), although the report also
recognized that this rrelatively small set of smart beta factors has performed well historically.
A multidimensional portfolio can achieve exposures to a large number of factors and is thus poised
to take advantage of more opportunities than a smart beta strategy that is based on only a few factors.
Furthermore, a multidimensional portfolio benets from diversication across numerous factors. It is
less susceptible than a smart beta portfolio to the poor performance of any one factor. As some factors
underperform, others may outperform, fostering greater consistency of performance.
The decision on which particular factors should be deployed in the portfolio construction is contingent
on comprehensive testing the portfolio performance and risk prole under comprehensive normal,
turbulent and stressed market conditions. Since a good test framework is such an important component
of successful portfolio management, we discuss it later in a chapter fully dedicated to this topic.

2.8 Combining factors is a necessary ingredient in asset allocation


There is strong intuition suggesting that multi-factor allocations will tend to result in improved risk-
adjusted performance. In fact, even if all factors to which the portfolio is exposed are positively
rewarded over the long term, there is extensive evidence that they may each encounter prolonged
periods of underperformance. More generally, the reward for exposure to these factors has been shown
to vary over time. If this time variation in returns is not completely in sync for dierent factors (see
Appendix F for evidence retrieved from literature), allocating across factors allows investors to diversify
the sources of their outperformance and smooth their performance across market conditions.
The example given in Bender et al. [2013a] is illustrative. During the period between 2001 and 2007,
the MSCI World Momentum, MSCI World Value Weighted, MSCI World Risk Weighted, and MSCI
World Equal Weighted Indexes all outperformed the MSCI World Index, while the MSCI World Quality
Index underperformed. From 2007-on, in contrast, the MSCI World Quality Index outperformed while
the MSCI World Momentum and MSCI World Value Weighted Indexes did not fare as well. Combining
Quality with Momentum and Value Weighted factor indexes for instance can yield a smoother ride
and diversify across multi-year cycles.
Because they reect systematic factors that respond to macroeconomic and macro market forces,
factor indexes can underperform the overall market for periods of time that may exceed an investment
committee's patience. However, many of these factors respond dierently to macroeconomic and macro
market forces, so they have historically low correlations which may yield strong diversication eects
for combining multiple factors in an allocation. Such a composite portfolio will have a factor-based
diversication which is largely independent on market regimes, while being much more eective during
nancial crises compared to traditional asset-based diversication Ilmanen and Kizer [2012], Bender
et al. [2010], Carli et al. [2014], Meucci et al. [2014a], Pappas et al. [2012], Bender et al. [2013a].
Some factors such as Value, Momentum, and Size have historically been pro-cyclical, performing well
when economy growth, ination and interest rates are rising. Other factors such as Quality and Low
Volatility have historically been defensive, performing well for a weak or falling macro environment.
In a nutshell, dierent factors work at dierent times, and this insight will also be utilized in the
context of regime-based investing described in Chapter 5. A well designed multi-factor portfolio would
provide the average level of returns of their components, while leading to strong risk reduction (in
relative terms), which eventually results in risk-adjusted performance that is well above average.

21
To illustrate, let's consider the correlations of the 7 MSCI factor indices mentioned above.

Figure 2.9: Correlations among MSCI indices (from Bender et al. [2013b])

Performance analysis of a portfolio obtained through a combination of 4 MSCI indices indicates


an Information Ratio substantially higher than the four individual indexes, reecting a much better
diversication, while maximum drawdown is signicantly lower.

Figure 2.10: Performance of a combination of MSCI indices (from Bender et al. [2013b])

Figure 2.11: Historical performance of a combination of MSCI indices (from Bender et al. [2013b])

22
Similar behavior can be observed in other asset classes when combining risk factors.
Results from Houweling and van Zundert [2014] indicate better performance for a composite portfolio
constructed by combining the Size, Low-Risk, Value and Momentum factors for U.S. Investment Grade
and U.S. High Yield corporate bonds.

Figure 2.12: Performance of market and composite portfolio for U.S. Investment Grade and U.S. High
Yield corporate bonds (from Houweling and van Zundert [2014])

Another example is from private equity (PDE) universe, with results obtained by Kothari et al.
[2012a,b] using a dataset comprised of ve private equity portfolios (representing nearly $200 billion in
private equity investments, or roughly 10% of the entire industry) belonging to top-tier institutional
investors: the California Public Employees' Retirement System (CalPERS), New York State Common
Fund, University of California Regents, Washington State Investment Board, and the University of
Texas Investment Management Company.
Relying on a number of risk-type factors for PE, namely Strategy/Product, Fund Size, Vintage,
Industry Focus, and Geographic Exposure, they have modied original portfolios to take advantage of
diversication benets within a broader asset class (in this case, private equity) and to improve per-
formance. On average, the Modied portfolios produce a 10% improvement across all of the individual
portfolio experiments.

Figure 2.13: Performances of original and modied PE portfolio (from Kothari et al. [2012a]

Figure 2.14: Performances of original and modied PE CalPERS portfolio (from Kothari et al. [2012a]

Allocating across more factors is also likely reduce trade turnover, given that some of the trades
necessary to pursue exposure to dierent factors may actually cancel each other out.

23
Consider the example of an investor who pursues an allocation across a value and a momentum tilt.
If some of the low valuation stocks with high weights in the value strategy start to rally, their weight
in the momentum-tilted portfolio will tend to increase at the same time their weight in the value-tilted
portfolio will tend to decrease. The eects will not be completely canceled out, but some reduction in
turnover can be expected through such natural crossing eects.

2.9 Risk factor parity


Practical limitations of traditional diversication approaches have recently lead to to academic and
practitioners alike to consider various heuristic methodologies, such as equally-weighted, minimum
variance, most diversied portfolio, equally weighted risk contributions, risk budgeting or diversied
risk parity strategies (see Homescu [2014a] for a review of the literature).
However such strategies can still produce portfolios that are highly concentrated in only one or two
true risk exposures (especially in equity risk), and therefore be under-diversied in other risk exposures.
Essentially having diversication in risk contribution from assets is generally not the same as having
diversication in the primitive sources of risk underlying asset returns, as argued inBhansali et al.
[2012], with results showing that two risk factors, driven by global growth and global ination, largely
dominate asset class risk and return.
Using principal component analysis, the orthogonalized representation of these two risk factors was
extracted from a sample universe of 9 conventional assets (U.S. equities, International equities, EM
equities, REITS, commodities, global bonds, U.S. long treasury, investment grade corporates and
high yield bonds). The rst two factors account for 68% of the variance in the co-movement of the
nine assets, while the other factors account for less than 10% each. It was also shown that the nine
asset classes have intuitive loading on the two dominant risk factors. The pro-cyclical assets such as
equities, commodities, REITS and high yield bonds exhibited signicant loading on the global growth
risk factor, while counter-cyclical assets such as global bonds, U.S. treasury and investment grade
corporates loaded heavily on the global ination risk factor.
A better approach would be to combine the features of factor-investing and risk budgeting, such that
diversication is obtained through decomposition of the portfolio's risk into risk factor contributions.
This composite approach, termed risk factor parity, was analyzed in Roncalli and Weisang [2012],
Williams [2014], Bhansali et al. [2012], Witham [2012], Romahi and Santiago [2012].
A diversied factor risk parity portfolio might look like this:

Figure 2.15: Example of a diversied factor risk parity portfolio (from Romahi and Santiago [2012])

24
Another connection between risk-based strategies and factor investing was identied in Leote de
Carvalho et al. [2012]. Various risk-based strategies, such as Minimum Variance, Maximum Diversi-
cation and Risk Parity, can be almost fully explained by their exposures to the low volatility factor,
the size factor and to a smaller extent the value factor. Extending the approach, it was shown how
portfolios of multiple Smart Beta indices could be replaced by more ecient robust portfolios with
targeted factor exposures.
One considerable advantage is that, since there is only a handful number of known factor anomalies
worth investing in, this approach makes life much easier compared to analyzing and selecting from
hundreds of Smart Beta indices available today. Another advantage is that investors can decide what
factors to include in their portfolios and what risk exposures to those factors they should aim at.
We conclude this section with a representative example of performance for a risk-based portfolio
augmented by 4 factors, constructed in Williams [2014]:

• Value Everywhere:
 this factor sorts across asset classes, and goes short or long in proportion with their relative
position in the sort.

 It is volatility balanced (not necessarily zero capital) trade going long value assets and short
growth assets.

 The assets include global equity indices, individual equities, government bonds, commodity
futures, and currencies.

• Momentum Everywhere:
 created using the same volatility balanced multi-asset sort as Value Everywhere

 securities are sorted along intermediate term price momentum.

• Betting Against Beta:


 goes long low beta assets and short high beta assets.

 The longs are levered so that the overall portfolio is market neutral.

 It includes global individual equities, corporate credit, global equity and corporate bond
indices, government bonds and commodity futures.

• Volatility Selling:
 The OBOE Put Index sells S&P 500 puts and collateralizes its exposures with T-bills.

 The number of puts sold each month is limited by the relative expensiveness of the puts
and also ensures that the T-bill positions can always nance any possible loss on the puts.

This strategy was found to have the best Sharpe ratio when compared to various practitioner portfolios.

Figure 2.16: Performance comparison of factor risk parity with other strategies (from Williams [2014])

25
2.10 Implementation
One needs to address many challenges for a successful implementation of factor investing. Until recently,
these factors were available by primarily investing in mutual funds, although investment vehicles (either
passively or actively managed) are now considered, and development of factor indexes (MSCI, Russell,
etc.) certainly will help.
The main question asked by practitioners (see Cazalet and Roncalli [2014], Bender et al. [2013a],
Kang and Ung [2014], Amenc and Goltz [2014]) is How to transform these academic risk factors into
investable portfolios? A successful answer relies on addressing various constraints due to the capacity,
turnover and transaction costs of these factor portfolios, while being done within a framework satisfying
regulatory, guideline or investor restrictions.
Factor allocations should be driven rst and foremost by the institution's investment objectives
and constraints (governance structure, horizon, risk budget, funding ratios, size of managed assets,
etc.). One institution may seek to enhance risk-adjusted returns, or limit downside risk, while another
institution might be trying to replicate the performance of certain investment style. Additionally,
candidate factors should be only those the institution believe will persist in the future. Only then the
institution's objectives and constraints together will drive the choice of factors among these candidates.
Ongoing performance monitoring is indispensable in ensuring that selected factor-based strategies
meet investor expectations and objectives. This would involve examining the portfolio's overall factor
exposure, sector biases and whether it has any secondary exposures to macroeconomic factors.
An illustration of key stages of decision making and implementation is shown below:

Figure 2.17: Key Stages of Decision Making and Implementation (from Kang and Ung [2014])

2.10.1 Which asset universe to consider?


Academic risk factors are generally obtained using a large asset universe. For example, Fama and
French compute their HML factor using following number of stocks:

Asia Pacic Europe Japan North America US

Big 326 615 591 888 846


Small 2646 4093 1840 2982 2385

Table 4: Number of stocks for HML factors

However, it is not realistic to think that an institutional investor will manage its small-cap exposure
according to the entire stock universe, because the liquidity risk is too high. Practitioners deploy
much fewer stocks, as seen in following table from Cazalet and Roncalli [2014], with column AFP
corresponding to the average number of stocks used by Asness et al. [2014] to build their QMJ factor,
while last column describes number of stocks in MSCI World index.

26
Figure 2.18: Number of stocks for factor estimation (from Cazalet and Roncalli [2014])

The choice of stock universe for computing factors may be quite relevant, as shown in next gure.
Since the methodology (Fama and French [2012]) to build the factors is the same in both cases, the
dierences are due entirely to the composition of stock universe utilized for each case.

Figure 2.19: Factors using Fama-French stock universe and S&P500 (from Cazalet and Roncalli [2014])

27
A similar conclusion is obtained for the MSCI Europe index.

Figure 2.20: Factors using Fama-French stock universe and MSCI Europe (from Cazalet and Roncalli
[2014])

2.10.2 Dependency on weighting scheme


The choice of weighting scheme when constructing a factor may be very signicant in the performance,
even for the case of having high correlation between same factor constructed with dierent schemes.
Some weighting schemes are backward-looking and may result in overweighting underperforming factors
and underweighting outperforming factors.
Such a situation is described in Cazalet and Roncalli [2014] when comparing value-weighted and
equally-weighted US factors obtained by Kenneth French. While SMB, WML and HML factors are
shown to be highly correlated when computed using the 2 approaches, the global performance diered
considerably, for SMB factor in AsiaPacic and Europe and HML factor in AsiaPacic and the US.

2.10.3 What is the optimal balancing frequency?


Answering this question in a successful manner would depend on many parameters, such as impact
on the portfolio turnover and also on trading costs, whether we are in normal or turbulent markets,
characteristics of risk factors (low-frequency, high frequency, etc.)

2.10.4 Long only or long-short?


Most products based on risk factors generally use a long-only portfolio, for following main reasons:

• Institutional investors use mainly long-only portfolios for their core investments, either by design
or by mandate.

• Feasibility of factor indexes, given that a short portfolio is more complicated and costly to manage
and may require the use of derivatives and OTC products.

• Some investors believe that a portfolio with a beta exposure and long/short strategies may be
equivalent to a portfolio with long-only exposures.

28
• it may become prohibitive or even impossible to short securities in times of crisis

However, the exposure to risk factors may be signicantly modied by imposing long-only constraints,
as shown in next gure:

Figure 2.21: Performance of factors in long-only and long/short strategies (from Cazalet and Roncalli
[2014]

2.10.5 Capacity constraints and transaction costs


Many studies on risk factor investing generally assume that the investor has no transaction costs or
capacity constraints. However, in practice these trading frictions may be highly inuential on the
portfolio performance, in particular when the turnover is high.
Frazzini et al. [2013] found that size, value, and momentum survive transactions costs at fairly
substantial sizes, but that short-term reversals do not. Break-even fund sizes for the Fama and French
long/short factors of size, value, and momentum were estimated to be103, 83, and 52 billion dollars
among U.S. securities, respectively, and 156, 190, and 89 billion dollars globally. Short-term reversal
strategies, on the other hand, would not survive transactions costs at sizes greater than $9 billion in
the U.S. or $13 billion globally. Moreover, a combination of value and momentum has even higher
capacity ($100 billion U.S., $177 billion globally) due to netting of trades across negatively correlated
positions.

2.10.6 Investability
Investability refers to how liquid and tradable a portfolio is. As described in Bender et al. [2013b],
there are multiple dimensions to investability, such as

29
• Tradability/Liquidity:

 quanties how liquid the stocks are in the index replicating portfolio and how tradable the
portfolio is.

 Metrics include days to trade individual stocks at rebalancings and during the initial con-
struction, and days to trade a certain portion of the portfolio

• Turnover/Cost of Replication: measures the turnover of the index at rebalancing which scales
with trading costs

• Capacity:

 quanties the percentage of a stock's free oat or full market capitalization the portfolio
would own

 Metrics include active share and maximum strategy weight

2.11 Final thoughts on eective portfolio construction


Eective portfolios must be designed not only to provide ecient management of risk and return but
also for genuine investability. Optimal control of rebalancing should be done to reduce turnover (and
hence of transaction costs), while weight and trading constraints may ensure higher capacity and higher
liquidity. Moreover, it was already discussed that portfolios based on larger number of factors have
additional reduction of turnover, since trades done during the rebalancing stage may partially oset
each other.
Model specic risks can be diversied through combining various weighting schemes. SciBeta indexes
(for details see Amenc and Goltz [2014], Gonzalez and Thabault [2013], Amenc et al. [2012]) can
be considered as good examples of how to achieve such combinations (Maximum Deconcentration,
Maximum Risk Diversication, Maximum Decorrelation, Ecient Minimum Volatility and Ecient
Maximum Sharpe Ratio) together with relative risk constraints.
To conclude, exposure to various factors whose premia behave dierently over time and across
market conditions provides for smoother outperformance. Moreover, natural crossing benets reduce
the turnover of multi-factor mandates relative to separate single factor mandates. Finally, the triple
diversication paradigm (based on factors, components and weighting schemes) leads to a neutral point
that greatly reduces the possibility of taking a sizable bet (which might be wrong) on the winning
components, the right weighting scheme, or the best factor tilt.

3 Factor models
The use of factor models in the nancial industry has gained increased popularity recently, due to the
continued market turmoil and asset price uncertainty and also because traditional approaches have
proven less reliable.
The factor risk models utilize a set of factors (or, in other words, explanatory variables) to explain
prices and returns. Such models were found to be rather successful in forecasting returns and estimating
covariances and co-movements.

3.1 Why do we need factor models?


One may ask why there is any need for factor or risk models when one can simply measure volatility,
covariance, correlation, etc. directly from the underlying asset returns data. It is a valid question, and
its answer relies on two observations:

30
• historical data is subject to a great deal of false relationships

• there is a data limitation issue (mathematicians refer to it as a degrees of freedom issue) in


the sense of not having enough historical data to determine statistically signicant quantities
(correlations, covariances, etc.), as illustrated in Appendix A.

As described in Glantz and Kissell [2014], factor models address two deciencies encountered when
using historical market data to compute covariance and correlation:

1. require much fewer observations for a reasonably accurate and stable estimate of covariance
matrix

2. use a set of common explanatory factors across all assets, allowing for apples-to-apples com-
parisons between factors.

3.2 A primer on factor models


A factor model has the form
M
X
rt = α0 + fjt bj + et
j=1

where

• rt is the asset return in period t

• α0 is a constant term

• fkt is the value of k th factor in period t

• bk is exposure to k th factor (also referred to as beta or sensitivity to factor)

• et is noise of asset in period t (in other words, return not explained by the model).

Parameters of the model are usually determined via ordinary least squares (OLS) regression analysis,
as detailed in Appendix B, sometimes augmented with a weighting scheme so more recent observations
have a higher weight in the regression analysis.
For number of time periods P and number of factors M, a factor model (on return of ith asset) can
be written in matrix notation as
ri = αi + F bi + ei
where
         
ri1 αi1 f11 f21 ··· fM 1 bi1 ei1
 ri2   αi2  f12 f22 ··· fM 2  b  ei =  ei2 
b =  i2
     
ri = 
 · · ·  αi =  · · ·
F = 
···  i  ···
 
  ··· ··· ···   ··· 
riP αiP f1P f2P ··· fM P biM eiP

A factor model across a universe of N assets can be written as

R = α + Fβ + e

31
where
     
r11 r21 ··· rN 1 α11 α21 ··· αN 1 f11 f21 ··· fM 1
 r12 r22 ··· rN 2 
  α12 α22 ··· αN 2    f12 f22 ··· fM 2 
R=
 ··· α= F = 
··· ··· ···   ··· ··· ··· ···   ··· ··· ··· ··· 
r1P r2P ··· rN P α1P α2P ··· αN P f1P f2P ··· fM P
   
b11 b21 ··· bN 1 e11 e21 ··· eN 1
 b12 b22 ··· bN 2   e12 e22 ··· eN 2 
β=
 ···
e =  
··· ··· ···   ··· ··· ··· ··· 
b1M b2M ··· bN M e1P e2P ··· eN P
The covariance matrix derived from the factor model has the expression (for details see Appendix
C):
C = β T cov(F )β + Λ
where Λ is the idiosyncratic variance matrix with each diagonal term given by variance of the
regression term for each asset, and cov (F ) is the factor covariance matrix, which is a diagonal matrix
in majority of practical situations.

3.3 Types of factor models


According to Glantz and Kissell [2014], factor models can be divided into four categories of models:

1. index models

2. macroeconomic models

3. cross-sectional or fundamental data models

4. statistical factor models.

For index models, macroeconomic factor models and fundamental data models one needs to specify
in advance either a set of explanatory factors or a set of company-specic factor loadings. Index
models and macroeconomic factor models are constructed from a set of specied factors and estimated
sensitivities to those factors, while fundamental data models rely on factor loadings (fundamental data)
and estimated explanatory factors.
In contrasts, the statistical factor models diers do not make any prior assumptions regarding the
appropriate set of explanatory factors and do not force any preconceived relationship into the model.

Index models
There are two forms of the index model commonly used in the industry:

1. single-index: based on a single major market index such as the S&P500.

2. multi-index: extends the general market index by incorporating additional information on the
asset's sector index and perhaps the asset's industry index.

Macroeconomic models
A macroeconomic multifactor model denes a relationship between stock returns and a set of macroe-
conomic variables such as GDP, ination, industrial production, bond yields, etc. The appeal of using
macroeconomic data as the explanatory factors in the returns model is that these variables are readily
measurable and have real economic meaning. However, such a model may not suciently capture the
most accurate correlation structure of price movement across assets, and usually do not do a good
job capturing the covariance of price movement across dierent regimes, e.g., normal and turbulent
markets.

32
Fundamental data models
This type of models estimate stock returns from a set of variables that are specic to each company,
based on fundamental and technical data, rather than through factors that are common across all
stocks. The fundamental data consists of company characteristics and balance sheet information. The
technical data (also called market-driven data) consists of trading activity metrics such as average
daily trading volume, price momentum, size, etc.
While it is intuitive to incorporate fundamental data into multi-factor models, this approach has
some limitations. First, data requirements are cumbersome, requiring analysts to develop models using
company-specic data. Second, it is often dicult to nd a consistent set of robust factors across stocks
that provide strong explanatory power.

Statistical factor models


These models do not have available in advance neither the explanatory factors nor sensitivities to
these factors. While such quantities are not readily observed in the market, they can be derived from
historical data for construction of this type of models.
The main advantage is that any preconceived bias is removed from the model. The main disadvantage
is that it does not provide portfolio managers with a set of factors to easily determine what is driving
returns, since the statistical factors do not have any real-world meaning.
There are three common techniques used in statistical factor models:

• Eigenvalue-eigenvector: based on a factoring scheme of the sample covariance matrix

• Singular value decomposition: based on a factoring scheme of the returns matrix of returns

• Factor analysis: based on a maximum likelihood estimate of the correlations across stocks.

4 Markov switching models for nancial markets


Markov Switching models M SM s are becoming increasingly popular in nance, due to their ability to
t the data, to lter unknown regimes and states on the basis of the data, to allow a powerful tool to
test hypotheses formulated in the light of nancial theories, and to their forecasting performance with
reference to both point and density predictions.
Regime switching models can match the tendency of nancial markets to often change their behavior
abruptly and the phenomenon that the new behavior of nancial variables often persists for several
periods after such a change. While the regimes identied by regime switching models are identied
by an econometric procedure, they often intuitively match dierent periods in regulation, policy, and
other secular changes. In empirical estimates, the regime switching means, volatilities, autocorrela-
tions, and cross-covariances of asset returns often dier across regimes, which allow regime switching
models to capture the stylized behavior of many nancial series including fat tails, heteroskedasticity,
skewness, and time-varying correlations. In equilibrium models, regimes in fundamental processes, like
consumption or dividend growth, strongly aect the dynamic properties of equilibrium asset prices and
can induce non-linear risk-return trade-os.
We can distinguish 3 main categories of Markov switching models:

• hidden Markov model (HMM)

• hidden semi-Markov model (HSMM)

• continuous-time hidden Markov model (CTHMM)

33
In an HMM, the distribution that generates an observation depends on the state of an unobserved
Markov chain. The transition probabilities of the Markov chain are assumed to be constant implying
that the sojourn times are geometrically distributed.
However, the memoryless property (time until the next transition out of the current state is indepen-
dent of the time spent in the state) of the geometric distribution is not always appropriate. Then one
may use the hidden semi-Markov model (HSMM) in which the sojourn time distribution is modeled
explicitly for each hidden state.
If the observations are not equidistantly sampled then a continuous-time hidden Markov model
(CTHMM) that factors in the sampling times of the observations needs to be applied. The advantages
of a continuous-time formulation include the exibility to increase the number of states or incorporate
inhomogeneity without a dramatic increase in the number of parameters.
For additional information the reader is referred to Guidolin [2012, 2011], Guidolin and Timmermann
[2006], Kim and Nelson [1999], Ang and Timmermann, Nystrup [2014].

4.1 Reasons to consider Markov switching models


As discussed in the literature survey Guidolin [2012], M SM s were considered either as a modeling
tool on the basis of a statistical reason (because the data ask for them) or, alternatively, for economic
motivations (because general knowledge of a phenomenon or a eld makes them plausible).
While the key challenge of a M SM lies in correctly inferring and predicting regime shifts (sometimes
called turning points), there is merit to the alternative consists of thinking of M SM s not only (or
mostly) as devices to anticipate regime switches but as a exible class of nonlinear tools, the usefulness
of which cannot be completely pinned to the fact that the Markov chain characterizing them can
be easily predicted. Of course, this is a very subtle distinction, because it remains odd to entertain
the possibility that a regime switching model may be useful in certain applications (e.g., portfolio
choice, risk management) even though it does not really help that much with forecasting the switching
dynamicsand yet this was found to be the case.
In practical terms this means that in order for M SM s to be very valuable tools the identication
of regime shifts has to be only good enough accurate.
Regarding the question of which asset classes theM SM s were found to be suitable for, the answer
seems to be ALL OF THEM: equity, xed income, currencies, commodities, government and corporate
bonds, real estate returns etc.

4.2 A primer on Markov switching models


We start with a generic K−regime hidden Markov model HMM with VAR(p) process for a N −dimensional
vector yt and St ∈ {1, 2, · · · , K} latent states can be described as (see:
p
X
yt = µSt + Aj,St yt−j + ΩSt t (4.1)
j=1

with t normally and identically distributed t ∼ N ID (0, IN ), µSt a N −dimensional vector and ΩSt
a N ×N matrix representing the state St factor in a regime dependent Cholesky factorization of the
covariance matrix
ΘSt = ΩSt ΩTSt
A non-diagonal ΩSt captures simultaneous co-movements, while dynamic (lagged) linkages across
dierent variables may be captured by the VAR(p) component.
The unobservable state process St is modeled as a Markov chain with the K×K probability transition
matrix Pt = {pij,t } dened as
pij,t = P r (St = j|St−1 = j, Ft )

34
where Ft = {yj }tj=1 and the transition matrix may be constant in time or have time dependent
probabilities.
The model can be described by a combination of 2 equations

St = f (St−1 , θt−1 ) + ηt
yt = h (St , θt ) + ΩSt t

If the assumption of geometrically distributed sojourn times is unsuitable, then hidden semi-Markov
models can be applied. HMMs and HSMMs dier only in the way that the state process is dened.
For HSMM the future states are only conditionally independent of the past states when the process
changes state.
HSMMs provide a considerable exibility compared to HMMs without complicating the estimation
signicantly, though the maximization of the term that depends on the parameters of the sojourn time
distributions typically requires a numerical solution. The selection of the most appropriate sojourn
time distribution can, however, be a complicated problem.
In continuous-time Markov chains CTHMM, transitions may occur at all times rather than at
discrete and equidistant time points. There is no smallest time step and the quantities of interest are
the transition probabilities.
The number of parameters increases linearly with the number of states. Thus, a CTHMM yields a
parameter reduction over its discrete-time analogue if the number of states exceeds three. The higher
the number of states, the larger the reduction. In addition, it is possible to incorporate inhomogeneity
without a dramatic increase in the number of parameters using splines, harmonic functions, or similar.
Another advantage of CTHMM is the exibility to use data with any sampling interval as the data
is not assumed to be equidistantly sampled. In a discrete-time model, the trading days are aggregated
and thus weekends and bank holidays are ignored so that Friday is followed by Monday in a normal
week. A CTHMM is able to recognize the longer time span between Friday and Monday, which could
lead to a dierent behavior.
For a successful deployment of Markov switching models there are essentially three main issues that
need to be addressed:

• Evaluate the likelihood of the observations given the parameters.

• Estimate the model parameters that maximize the likelihood of the observed data.

• Infer the most likely sequence of states.

4.3 Estimating parameters of Markov switching models


K N + pN 2 + N (N +1/2) + (K − 1) parameters. This number gets
 
Eq.(4.1) implies the estimation of
large when N is large. For instance, for K = 2, N = 8 and p = 1 we would need to estimate 218
parameters.
The rst step towards estimation and prediction of a M SM is to write it in state-space form as a
combination of 2 equations: measurement and transition (see Bloch [2014], Guidolin [2012])

St = f (St−1 , θt−1 ) + ηt transition


yt = h (St , θt ) + ΩSt t measurement

with independent error processes {ηt }and {t }.


The fundamental inference mechanism is Bayesian and consists in computing the posterior quantities
of interest sequentially in the following recursive calculation:

35
1. Letting ψt = {y1 , · · · , yt } be the information set up to time t, we get prior distribution corre-
sponding to the distribution of the parameters before any data is observed.
ˆ
p (St |ψt−1 ) = p (St |St−1 ) p (St−1 |ψt−1 ) dSt−1

2. Using the sampling distribution p (yt |St ) of the observed data conditional on its parameters. the
updating equation becomes

p (yt |S t−1 ) p (St |ψt−1 )


p (St |ψt ) =
p (yt |ψt−1 )

The updates provides an analytical solution if all densities in the state and observation equation are
Gaussian, and both the state and the observation equation are linear. If these conditions are met,
the Kalman lter provides the optimal Bayesian solution to the tracking problem. An illustration of
Kalman lter algorithm is provided in Appendix D.
Otherwise we require approximations such as the Extended Kalman lter, or Particle lter which
approximates non-Gaussian densities and non-linear equations. The particle lter uses Monte Carlo
methods, in particular Importance sampling, to construct the approximations. The approach that we
prefer is based on Kim lter and is described in Appendix D.

4.4 Decoding the hidden states


Decoding is the process of determining the states of the Markov chain that are most likely (under the
tted model) to have given rise to the observation sequence.
There are two dierent ways of inferring the most likely sequence of hidden states:

• locally by determining the most likely state at each time t based on smoothing probabilities

• globally by determining the most likely sequence of states using a specialized (usually Viterbi)
algorithm

Each approach has advantages and disadvantages.


Local decoding can lead to impossible state sequences since it does not take the transition prob-
abilities into account, but reduces the probability of misclassication compared to global approach,
especially at the beginning and at the end of the series or when number of states is larger. Global
decoding reduces the number of transitions compared to local decoding.
To give an example, analysis done in Bulla et al. [2011] for in-sample case showed that the estimated
conditional variances are almost identical for either approach. However, the selected approach was
based on Viterbi algorithm, since number of regime-switches was signicantly lower for the Viterbi
paths, which are eectively smoothed versions of the smoothing probabilities, and thus transaction
costs were reduced.

5 Regime based investing


Investors have long recognized that economic conditions frequently undergo regime shifts. The economy
often oscillates between a steady, low-volatility state characterized by economic growth and a panic-
driven, high-volatility state characterized by economic contraction, with evidence of such behavior
(sometimes having additional regimes) documented in market turbulence, ination, GDP or GNP, etc.
Financial markets periodically transition between normal-volatility and high volatility regimes, in
which risky assets appear to have unusually high risk that does not t long-run assumptions. Depending

36
on the persistence of volatility regimes and the ability to forecast them, investors may wish to adapt
their portfolios to account for observed changes in volatility and volatility forecasts. Economic logic
suggests that an investor would be willing to give up some expected portfolio return if the portfolio's
risk were lowered. Such investors may decide to de-risk the portfolio during the high-volatility storm
and re-risk once skies are clear.
Although asset class behavior can vary signicantly over shifting economic scenarios, traditional
strategic asset allocation (SAA) approaches either make no eort to adapt to such shifts or seek to
develop static all-weather portfolios that optimize eciency across a range of economic scenarios. On
the other hand, a dynamic approach based on regimes has the potential to take advantage of positive
economic regimes, while withstanding adverse economic regimes and reducing potential drawdowns.
Ecient ltering procedures for parameter estimation in context of regime switching are discussed
in Appendix D.

5.1 Literature review


A few examples illustrate the variety of regime-switching models used in literature.
Kritzman et al. [2012] considered a regime-switching model which incorporated indicators for market
turbulence, ination and economic growth. When applied to stocks, bonds, and cash, a strategy based
on the forecasted regimes was shown to reduce downside risk and improve the ratio of return to Value-
at-Risk (VaR) relative to a static strategy out of sample. Rather than making an assumption about
transaction costs, the authors reported the break-even transaction cost that would oset the advantage
of the dynamic strategy.
Escobar et al. [2013] analyzed portfolio strategies with investment style switching based on the value
of a crisis indicator:

1. turbulent time indicator, an out-of-sample indicator built using macroeconomic data and relying
on a Markov switching model to determine states of calm market, turbulent bullish market, and
turbulent bearish market (see Hauptmann et al. [2014])

2. recession indicator, out-of-sample indicator available from the FRED database

3. heuristic indicator, an in-sample indicator tailor-made from the history of nancial crises.

Dopfel and Ramkumar [2013] studied whether a managed volatility approach has advantages over
static strategic asset allocation through normal- and high-volatility regimes.
Wang et al. [2012] considered a regime switching framework with asset allocation in accordance
with one of two possible states of the world: a) normal risk (normal uncertainty: normal return
volatility and correlations); b) high risk (high uncertainty: high volatility and correlations).
Guidolin and Timmermann [2006] found that no less than four separate regimes (characterized as I)
crash; II) slow growth; III) bull; IV) recovery) are required to capture the joint distribution of stock
and bond returns. It was argued that a four-state regime switching model is not only supported by
statistical evidence indicating that having only two states is insucient, but also is a key determinant
of the portfolio weights and (out-of-sample) return performance.
Using monthly MSCI total returns from January 1970 to December 1997, Ang and Bekaert [2002]
concluded that the existence of a high-volatility bear market regime does not negate the benets of
international diversication and that the high-volatility, high-correlation regime tends to coincide with
a bear market. In a subsequent study, Ang and Bekaert [2004] included additional equity markets
from around the world and found that the regime-switching strategy dominates static strategies out-
of-sample for a global all-equities portfolio, and that the model proposes to switch primarily to cash
in a persistent high-volatile market. They also considered market timing based on a regime-switching
model in which the transition probabilities depended on a short-term interest rate.

37
Bauer et al. [2004] focused on monthly returns from January 1976 to December 2002 of a six-
asset portfolio consisting of equities, bonds, commodities and real estate. They observed changing
correlations and volatilities among assets, and demonstrated a signicant information gain by using
a regime-switching instead of a standard mean-variance optimization strategy. After accounting for
transaction costs, however, a substantial part of the positive excess return disappeared.
Ammann and Verhofen [2006] estimated the four-factor model of Carhart [1997], using monthly data
from January 1927 to December 2004. They found two clearly separable regimes with dierent mean
returns, volatilities and correlations. One of their key ndings was that value stocks provide high
returns in the high-variance state, whereas momentum stocks and the market portfolio perform better
in the low-variance state.
Hess [2006] examined the improvement of portfolio performance when imposing conditional Capital
Asset Pricing Model (CAPM) strategies based on regime forecasts from an autoregressive Markov
regime-switching model. On the basis of returns of the Swiss stock market and its 18 sectors from
January 1973 to June 2001, his results indicate that regime switches are a valuable timing signal for
portfolio rebalancing.
Bulla et al. [2011] examined the performance of a regime-based asset allocation strategy under
realistic assumptions, compared to a buy-and-hold strategy. Utilizing daily return series of major
equity indices in the United States, Japan and Germany over the past 40 years, the strategy proves
protable after taking transaction costs into account. Portfolio risk is lowered by an average of 41 per
cent for all markets under consideration, and annualized excess returns between 18.5 (S&P 500) and
201.6 (Nikkei) basis points can be realized by avoiding highly volatile periods.
A three-state multivariate Gaussian hidden Markov model was employed in Nystrup [2014] for a
strategic asset allocation decision, with asset classes considered being stocks, bonds, and commodities
(and without any risk-free asset). Three proved to be an optimal number of states to reproduce the
stylized facts of data and market behavior (including skewness, leptokurtosis, volatility persistence,
and time-varying correlations) with the most exceptional observations being allocated in a separate
state. A fourth state did not seem to improve the t signicantly compared to the added complexity
and two states were not enough to adequately capture the stylized behavior.
Darolles and Vaissie [2011] employed a regime switching dynamic correlation model to reduce the
downside exposure when markets collapse while capturing most of upside when the situation normalizes.
A Target Volatility strategy described in Hocquard et al. [2013] combined a payo distribution model
with a regime switching framework for tail risk management. A hidden Markov model (HMM) was
employed to identify the presence of three volatility regimes (high, medium, and low) and to estimate
parameters for the three regimes.
Fischer and Seidl [2013] showed the eects on on the asset allocation and on the absolute and
relative performance of portfolios due to employing a regime-switching framework (with two states
corresponding to bull and bear markets)
Liu and Chen [2014] used the regime switching technique to describe the time-varying uncertainty
set of the rst and second order moments.
Zakamulin [2014] tested two strategies based on unexpected volatility dened as the dierence be-
tween the forecasted volatility (one month ahead) using a GARCH(1,1) model and realized volatility.
The data included daily and monthly returns of the S&P 500 and the Dow Jones Industrial Average
index from 1950 through 2012. In the rst strategy, the weight of stocks relative to cash was changed
gradually on a monthly basis based on the level of unexpected volatility, whereas the second strategy
was either all in stocks or all in cash depending on whether the unexpected volatility was below or
above its historical average. Both strategies were found to outperform static strategies out of sample.
Additional references include Grobys [2012], Guidolin and Timmermann [2006], Guidolin [2011], Y.
et al. [2011], Roman et al. [2010], Vo and Maurer [2013], Papenbrock and Schwendner [2013], Ernst
et al. [2009], Bernhart et al. [2011], Tu [2010].

38
5.2 Finer points to be considered
There are various ner points that need to be considered when reviewing or implementing a regime-
based allocation study, such as:

• transaction costs

• optimal number of regimes

• out-of-sample testing

• inclusion of risk-free assets

Many studies do not include transaction costs, although frequent rebalancing has the potential to
greatly reduce the excess return of a dynamic strategy based on regime switching. Reporting the break-
even transaction, as done by Kritzman et al. [2012], may be one of the most meaningful procedures
as the transaction costs faced by private investors are likely to exceed those of professionals who can
implement dynamic strategies in a cost-ecient way using nancial derivatives like futures or swaps.
Testing a model on the same data that it was tted to does not reveal its actual potential. As
noted in Homescu [2014a], Harvey and Liu [2014a], Bailey et al. [2014a], the out-of-sample potential
is likely to be lower (than the in-sample performance), as investors do not have perfect foresight. It
is not unusual that non-linear techniques provide a good in-sample t, yet get outperformed by a
random walk when used for out-of-sample forecasting. A poor out-of-sample performance can also be
an indication that the data-generating process is non-stationary.
Various regime-based portfolio allocation studies were based on switching to a risk-free asset in
the most turbulent periods. Holding a risk-free asset obviously yields a risk reduction, but leverage
would be needed to signicantly increase the return. This nding was consistent with the observation
presented in Frazzini and Pedersen [2014], Asness et al. [2012] that safer assets oer higher risk-adjusted
returns than riskier assets because of leverage aversion.

5.3 Representative regime switching frameworks


There are 3 main categories of regime switching frameworks discussed in the literature, and represen-
tative examples are presented next.

5.3.1 First representative framework: two regimes applied to benchmark index


A two-state Markov-switching model with time-varying covariances and expected returns was consid-
ered in Saunders et al. [2013] towards managing a fund of hedge funds.
The return dynamics of N hedge fund indices and one global index depend on an unobservable
process {S(t)} with state space {0, 1} as follows:

Ri (t) = µi (S(t)) + σi (S(t))Wi (t), i = 1 · · · N


R0 (t) = µ0 (S(t)) + σ0 (S(t))W0 (t)

where {W0 , W1 , W2 , · · · , WN } is a multivariate normal variable with zero mean, unit standard devi-
ation and correlation matrix C(S(t)).
The state process S is modeled as a time-homogenous Markov chain with transition matrix
 

P = 1−p p 
q 1−q

where p , P (S(t) = 1|S(t − 1) = 0) and q , P (S(t) = 0|S(t − 1) = 1) .

39
The initial distribution of S is (η, 1 − η) ,with η , P (S(0) = 0), and assuming that the time period
is split into M steps then the model is completely parameterized by

θM SM G = (p, q, µ0 , µ1 , Σ0 , Σ1 , η) (5.1)

with p, q, η ∈ [0, 1], µ0 , µ1 ∈ RM +1 , Σ0 , Σ1 ∈ R(M +1)×(M +1) and the entries of the covariance matrices
are given by Σi,j (S(t)) = Cij (S(t))σi (S(t))σj (S(t))
Two M SM s are considered:

• general M SM , denoted M SM G, completely parameterized by (5.1)

• Gaussian mixture model, denoted GM M , for the particular case when S(t) is an iid, completely
parameterized by
θGM M = (p, µ0 , µ1 , Σ0 , Σ1 )

The state process S, indicating whether the market is in a stressed or normal regime, is not directly
observable based on market data. Consequently it must be estimated indirectly, using market observ-
ables as proxies to make inferences regarding its history and parameters. Estimation process relied on
a major stock index as market observable to identify turbulent market phases and had 3 steps:

1. Estimate the model parameters and the parameters for the state process S, when only historical
data of the global stock index are considered.

2. Compute the most likely sequence of states.

3. Separation into two market regimes is obtained by the state sequence detected in step 2.

The investors can put capital in four dierent indices which act as proxies for the following strategies:

1. Equity Hedge(EH) strategies maintain positions both long and short in primarily equity and
equity derivative securities by using both quantitative and fundamental techniques. Strategies
can range broadly in terms of diversication, levels of net exposure, leverage employed, holding
period, and concentrations of market capitalizations.

2. Equity M arket N eutral(EM N ) strategies are quantitative strategies that attempt to remain
neutral to moves in the overall market while proting based on skill in selecting which individ-
ual equities to buy and sell. Examples include factor-based strategies, and strategies based on
statistical arbitrage and high-frequency techniques.

3. M acro(M ) strategy managers trade based on hypotheses on the movements of macroeconomic


variables and the impact that these movements will have on nancial markets. They use both
quantitative and fundamental techniques, and long and short-term holding periods.

4. M erger Arbitrage(M A) strategies primarily focus on opportunities in equity and equity-related


instruments of companies which are currently engaged in a corporate transaction.

The MSCI World Index time series is used as the reference index for crisis identication within the
M SM G and the GM M frameworks.
Three types of investment strategies are considered:

1. strategy (denoted MSM) based on M SM G regime switching

2. strategy (denoted GMM) based on GM M regime switching

3. strategy (denoted BS) based on standard Black-Scholes model, with returns assumed to be i.i.d.
multivariate normal random variables

40
The setting of the case study is: The rst investment decision is made on March 15, 2004, and after
that the portfolio is reallocated every 10 trading days. For any (re-)allocation decision, the investor
can process all information available up to the investment date. There are 160 investment periods.

Figure 5.1: Portfolio performance (from Saunders et al. [2013])

There are two dierent investment patterns, one for normal times and the other for turbulent periods.
In normal times the MSM investor employs roughly the same portfolio split as the BS and GMM
investor, primarily weighted on Equity Hedge. The portfolio allocation changes signicantly during
periods of market turbulence as captured by the regime switching models:

• early investment periods 1, 2 and 5 (March to May 2004), reecting the aftermath of the burst
of the dot-com bubble

• between investment periods 56 and 60, which corresponds to a temporary market downturn of
the MSCI World Index starting in May 2006.

• after investment periods 85/86 (mid-July 2007), when MSCI World Index reaches the peak

• investment periods 111 to 120 (July 2008 to December 2008), around the 2008 nancial crisis

Both regime-switching models outperform the Black-Scholes model in the crisis periods, and this
outperformance is more pronounced for the MSM investor. The main dierence in performance
occurs in investment period 111, and arises from the early crisis recognition.

41
Figure 5.2: Analysis of investment periods 111 to 120 (from Saunders et al. [2013])

Moreover, the performance for regime switching models is better overall:

Figure 5.3: Overall performance and risk metrics (from Saunders et al. [2013])

5.3.2 Second framework: more than two regimes applied to asset classes
A three-state multivariate Gaussian hidden Markov model was employed in Nystrup [2014]. The
increased potential complexity of the model due to its continuous-time formulation was not observed
in practice, since the estimated models turned out to have a simple structure as only certain transitions
were seen to occur.
The estimation results identied a need for an adaptive approach, as the parameters of the esti-
mated models were far from remaining constant throughout the sample period. However, the adaptive
estimation method (based on exponential forgetting) more than compensated by avoiding a third state
or a conditional t-distribution in the high-variance state to capture the most exceptional observations.
Data utilized consisted of 3 indices: (a) J.P. Morgan Government Bond Index JPM GBI; (b) Mor-
gan Stanley Capital International All Country World Index MSCI ACWI and (c) Standard & Poor's
Goldman Sachs Commodity Index S&P GSCI.
Here are parameter estimates for various models tted to daily in-sample log-returns of MSCI ACWI:

Figure 5.4: Parameter estimates for m-state HMMs (from Nystrup [2014])

42
Figure 5.5: Parameter estimates for m-state HSMMs (from Nystrup [2014])

Figure 5.6: Parameter estimates for 4-state CTHMMs (from Nystrup [2014])

The inferred number of transitions during the 15-year in-sample period ending in 2008 using global
and local decoding is shown in next gure:

Figure 5.7: Number of transitions (from Nystrup [2014])

Based on the inferred number of transitions the four-state models and the multivariate models are
ruled out as possible candidates for a successful strategy. The decoded states show a good t to the
market behavior, as shown in next gure, with Cyan describing the bull state, yellow the bear state,
and red the recession state.

43
Figure 5.8: The MSCI ACWI and the decoded states using the Viterbi algorithm (from Nystrup [2014]).

Based on the results of the decoding, it was decided that it was not worth testing the HSMMs, the
four-state models, and the multivariate models.
In an in-sample setting with perfect foresight the Viterbi path is the obvious choice as it is the most
likely sequence of states with fewer transitions than the smoothed path. This is, however, not the case in
an out-of-sample setting based on one-step predictions. The arguments for using the Viterbi algorithm
rely on the knowledge of future returns. Out of sample, where only past returns are known, it makes
more sense to consider the smoothing probabilities in order to obtain the best possible prediction.
This facilitates a ltering procedure based on the condence in the predicted state rather than the
number of times the state has been predicted with no minimum delay in responding to regime changes.
A higher threshold value reduces the number of transitions at the same time as increasing the risk of
delaying the reaction to regime changes.
The predicted states based on adaptively estimated model with a probability lter with a 92%-
threshold are shown below, together with MSCI data. Cyan describes the bull state and yellow the
bear state.

44
Figure 5.9: The MSCI ACWI and the predicted states (from Nystrup [2014]).

To measure the performance of various portfolios 5000 or 10000 simulated scenarios were generated
from the models calibrated to historical data. Since a scenario is a realization of the future value of all
parameters that inuence the portfolio, a collection of scenarios should capture the range of variations
that is likely to occur in these parameters including the impact of the shocks that are likely to come.
The annualized in-sample performance metrics (realized return (RR), standard deviation (SD), max-
imum drawdown (MMD), and Sharpe ratio (SR)) of the thee indices was:

Figure 5.10: Annualized in-sample performance (from Nystrup [2014]).

The 3 models considered were:

• HM MN (2): 2-state Gaussian distribution HMM

• HM MN t (2):: 2-state t−distribution HMM

• HM MN (3): 3-state Gaussian distribution HMM

45
The rst set of optimal portfolio allocations was found by partitioning the in-sample returns based
on the Viterbi path, under the constraint that with no short positions allowed. For each state 5,000
one-year scenarios were bootstrapped from the returns belonging to that particular state.

Figure 5.11: Performance of strategies over the 15-year in-sample period after subtracting 0.1% trans-
action costs with no short positions allowed (from Nystrup [2014])

The other 2 sets of optimal portfolio allocations in a similar way taking into account dierent
constraints: short positions are allowed, and respectively deviations from the benchmark allocation are
limited to ±0.1

Figure 5.12: Performance of strategies over the 15-year in-sample period after subtracting 0.1% trans-
action costs with short positions allowed (from Nystrup [2014])

Figure 5.13: Performance of strategies over the 15-year in-sample period after subtracting 0.1% trans-
action costs with limited deviations from the benchmark allocation (from Nystrup [2014])

Annualized out-of-sample performance of the three indices for out-of-sample periods were:

Figure 5.14: Annualized out-of-sample performance

The predicted states based on adaptively estimated model with a probability lter with a 92%-
threshold are shown below, together with MSCI data. Cyan is the bull state, yellow is the bear state.

46
Figure 5.15: The MSCI ACWI and the predicted states for the out-of-sample period (from Nystrup
[2014]).

The out-of-sample performance of the three dierent strategies based on the decoded states of the
adaptively estimated models are shown next, for the period 2009-2014.

Figure 5.16: Performance of strategies over the out-of-sample period after subtracting 0.1% transaction
costs and with one-day delay in the portfolio adjustment (from Nystrup [2014])

The reported results are based on a one-day delay between the prediction of a state change and the
portfolio adjustment. That is, if the predicted state of day t+1 based on the closing price at day t is
dierent from the state that the allocation at day t is based on and the condence in the prediction is

47
above the 92%-threshold, then the allocation is changed at the closing of day t + 1. If the reallocation
could be implemented at the closing of day t, i.e. with no delay, then the RR would have been a lot
higher with the SD and MDD remaining largely unchanged.

Figure 5.17: Performance of strategies over the out-of-sample period after subtracting 0.1% transaction
costs and with no delay in the portfolio adjustment (from Nystrup [2014])

5.3.3 Third framework: regimes applied to factors


The rst two frameworks were considered for asset classes or benchmark indices. One could also
identify regimes directly on factors, and two representative examples are presented in this section.
Kritzman et al. [2012] have applied a Markov-switching model to partition history into meaningful
regimes based on measures of nancial turbulence (for both equities and currencies), ination, and
economic growth, and within this framework they have tilted portfolios dynamically in accordance
with the relative likelihood of a particular regime.
Financial market turbulence was dened as a condition in which asset prices behave in an uncharac-
teristic fashion given their historical pattern of behavior, including extreme price moves, decoupling of
correlated assets, and convergence of uncorrelated assets. Daily turbulence indices were computed for
U.S. equities using returns for the 10 S&P 500 sector indices and for G10 currencies using currency
returns versus the U.S. dollar.
Ination was measured using monthly percentage changes in the seasonally adjusted U.S. Consumer
Price Index for All Urban Consumers, while economic growth was quantied using quarter-over-quarter
percentage growth in the seasonally adjusted U.S. real gross national product, Both metrics were
considered over a period from 1947 through December 2009.
The model considered was a simple regime switching model with 2 regimes, with observations in each
regime assumed to be normally distributed with a given mean and standard deviation. In-sample
results reveal the presence of a normal regime and an event regime in each series, while standard
errors of each estimate for mu and sigma indicate that for each variable, the dierences in mean and
standard deviation across regimes are also statistically meaningful. Persistence was dened as the
estimated transition probability of staying in the current regime.

Figure 5.18: In sample estimation results of Markov-switching model (from Kritzman et al. [2012])

48
Each variable's historical probability of being in the event regime over time is largely consistent with
observed periods of market turbulence. We note that ination was classied as being in the event
regime because of uncertainty about the ination rate, and not because of observed rising ination.

Figure 5.19: In sample event probabilities (from Kritzman et al. [2012])

It was found that regimes in turbulence, ination, and economic growth have a signicant impact
on the relative performance of risk premiums, dened as dierence between performance during the
event regime compared and performance during the normal regime, normalized by standard deviation.

Figure 5.20: In sample relative risk premium performance (from Kritzman et al. [2012])

49
Out-of-sample analysis conrmed that a dynamic process (based on the regime switching model)
would outperform a static process.
Tactical asset allocation (TAA) was considered rst through following experiment, which consisted
of 3 steps done each month:

1. Calibrate the Markov-switching model using a growing window of data from inception up to that
point in time.

2. When the model indicated upcoming event regime(s), add defensive tilts for each predicted event
regime

3. Compare the performance of the dynamic portfolio of risk premiums and constant exposures and
roll the experiment forward.

The tilts considered were as follows:

Figure 5.21: Event regime tilts(fromKritzman et al. [2012])

Comparing performance of the 2 portfolios (static and dynamic based on regimes and tilts) shows
that dynamic process increased risk-adjusted performance by 41% while reducing signicantly downside
risk, as evidenced by its lower skewness, lower value at risk (VaR), and lower maximum drawdown.

Figure 5.22: Performance for static and dynamic TAA strategies(fromKritzman et al. [2012])

50
Given that many investors cannot access risk premiums directly and that strategic investors, such
as pension plans, must invest with a fairly long time horizon, additional tests were done for strategic
asset allocation (SAA), applying the same methodology as in the TAA example to allocation across
asset classes, with tilts considered as follows:

Figure 5.23: Event regime tilts(from Kritzman et al. [2012])

Comparison of the performance of the 2 portfolios (static, allocated 60% stocks and 40% bonds, and
dynamic based on regimes and tilts) shows that dynamic process increased risk-adjusted performance
by 41% while reducing signicantly magnitude and duration of large losses were also reduced.

Figure 5.24: Performance for static and dynamic SAA strategies (from Kritzman et al. [2012])

The second example (see Gkatzilakis and Sivasubramanian [2014], Angelidis and Tessaromatis [2014],
Amenc and Goltz [2014]) considered dynamic allocation of smart beta indices based on identifying
regimes to risk factors such as size, value, momentum and volatility. Evidence of such regimes for these
risk factors was presented in Gulen et al. [2011], Stivers and Sun [2010].
A simple regime switching model with 2 regimes, with observations in each regime assumed to be
normally distributed with a given mean and standard deviation, is used to estimate the regimes. Once
the regimes are classied, optimal portfolios are constructed for each regime.
The benchmark portfolios were SciBeta Long-Term United States Diversied Multi-Strategy Index
and SciBeta Long-Term United States Multi-Beta Multi-Strategy EW, with model specic risks di-
versied through combining various weighting schemes (Maximum Deconcentration, Maximum Risk
Diversication, Maximum Decorrelation, Ecient Minimum Volatility and Ecient Maximum Sharpe
Ratio) and imposing relative risk constraints (for details see Amenc and Goltz [2014], Gonzalez and
Thabault [2013], Amenc et al. [2012]).
Data used was based on long term smart beta SciBeta indices of the United States market from
1970 to 2012 with monthly frequency, with in-sample period from 1970 to 2002 (390 periods) and
out-of-sample from 2003 to 2012 (totaling 120 periods). The market risk premium, size premium,
value premium and momentum premium are obtained from Fama-French data library.

51
The 2 regimes (event and normal) considered for volatility, momentum, size and value correspond
to high volatility, low volatility, high momentum, low momentum, mid-cap, large-cap, value and growth.
The regimes could be classied according to monthly volatility of S&P 500 daily returns.
The number of event and normal regimes for in-sample period was:
Volatility Size Value Momentum

Number of event regimes 107 27 171 96


Number of normal regimes 283 363 219 294

In-sample results were:

Figure 5.25: Performance for in sample period (from Gkatzilakis and Sivasubramanian [2014])

In sample risk analysis conrms better performance of regime-based factor portfolio in terms of
maximum drawdown (Max DD) etc.

Figure 5.26: Risk analysis for in sample period (from Gkatzilakis and Sivasubramanian [2014])

Out-of-sample analysis also shows improved performance for the regime-based factor portfolio:

52
Figure 5.27: Performance for out-of-sample (from Gkatzilakis and Sivasubramanian [2014])

Annual performance is computed, following by averaging of the performance on a rolling window


basis. The regime-based portfolio outperforms the benchmark in 6 out of the 10 years.

Figure 5.28: Annual excess returns for out-of-sample (from Gkatzilakis and Sivasubramanian [2014])

The probabilities of outperformance for 1-year, 3-year and 5-year periods are calculated on a rolling
window basis. This is a measure that indicates the chance of beating the benchmark in a xed time-
frame regardless of which month of the year the investor invests in the portfolio.

53
Figure 5.29: Probability of Outperformance on a rolling window basis for out-of-sample period (from
Gkatzilakis and Sivasubramanian [2014])

Figure 5.30: Excess returns on a rolling window basis for out-of-sample period (from Gkatzilakis and
Sivasubramanian [2014]

Out-of-sample risk analysis is presented next:

Figure 5.31: Risk analysis for out-of-sample period (from Gkatzilakis and Sivasubramanian [2014]

54
6 Sensitivity analysis in portfolio management
We describe how portfolio management can be enhanced through sensitivity analysis SA. We provide
a primer on sensitivity analysis and especially how it can be eciently linked to the optimization
algorithm, and then we present the rational of using it in the framework of factor investing.

6.1 A primer on sensitivity analysis


SA is a very important tool (see Saltelli et al. [2004, 2008]) to analyze questions such as:

• How condent can I be about the results?

• Will the results change if I change the denition of the outcome?

• Will the results change if I change the method of analysis?

• Will the results change if we take missing data into account?

• How much inuence will minor data or parameter deviations have on the conclusions?

• What if the data were assumed to have a non-Normal distribution or there were outliers?

• How will the optimized solution be perturbed if I change values of parameters?

• How will the optimized value of cost functional be perturbed if I change values of parameters?

While all these goals may be important, the primary purpose of SA is to assess which variables play
dominant roles in the optimization, and to evaluate rst-order sensitivities (i.e., rst derivatives or
directional derivatives) of the solution vector with respect to perturbation parameters.
Let us consider the general optimization problem

min F (x, p)
x

subject to constraints
R(x, p) = 0,
with x ∈ RN the vector of optimization variables and p ∈RN the vector of parameters.
dF
We want to compute sensitivities of F with respect to parameters {pj }, namely
dpj :

N
dF X ∂F ∂xi ∂F
= +
dpj ∂xi ∂pj ∂pj
i=1

∂xi
The conventional, but rather inecient, approach to compute
∂pj is based on nite dierence ap-
proximation (exemplied below using a rst order approximation)


∂xi xi p∆pj − xi (p)

∂pj ∆pj

where ∆pj is a small variation on the j−th parameter and p∆pj = (p1 , p2 , · · · , pj + ∆pj , · · · , pM ).
This approach is computationally costly because it demands for one reference computation of

R(x, p) = 0 to compute x(p), and then one additional R(x, p∆pj ) = 0 to nd the necessary x p∆pj
for each component of the gradient. To compute the full gradient with respect to N optimization
variables, we would therefore have to solve the vector equation R=0 a total of N + 1 times.
More ecient approaches are described in Appendix E.

55
6.2 Sensitivity analysis and factor investing
There are many situations where sensitivities's values would provide very useful information for better
portfolio management from various perspectives: risk, transaction costs, parameters selection, factor
exposures, etc. Moreover, once factor exposures have been established, it then becomes a straightfor-
ward exercise to use them together with factor returns to create a custom benchmark portfolio.
It is our strong belief that using a framework with sensitivity analysis integrated within the op-
timization procedure, as described in previous section, would be highly benecial and cost-eective,
especially with the more ecient approaches described in Appendix E.
Unfortunately most portfolio allocation studies with sensitivity analysis either do not incorporate
sensitivity analysis or rely on a traditional approach of obtaining sensitivities by changing only one
parameter value while keeping the other constants, making the process much too cumbersome from
both implementation and computational point of view.
Having said that, we present a few examples from the literature with sensitivity analysis computed
with the traditional approach, to illustrate its usefulness even in such conditions.
Cazalet and Roncalli [2014] have considered a universe of three long/short risk factors plus the market
risk factor. It was also assumed that the short leg of the risk factors present the same characteristics
than the long leg of risk factors (same alpha, same beta and same idiosyncratic volatility). To compute
sensitivities to the parameters they have employed six sets of parameters, with only the changed value
displayed in the particular column.

Figure 6.1: Sets of parameters considered in Cazalet and Roncalli [2014]

The parameter sets changed respectively values of alpha of the second long-only risk factor α2+ ,
idiosyncratic volatility of the third risk factor e3+ ,
σ beta of the second risk factor β2 , the volatility σm
of the market risk factor, the market risk premium µm .
The results are shown below, with x∗j denoting the optimal weights of the j−th risk factor, SR(x∗ , r)
the Sharpe ratio of the optimal portfolio with respect to the risk-free
∗ ∗
asset, and µ (x |b), σ (x |b)
andIR (x
∗ |b) represent respectively the excess return, the tracking error volatility and the information

ratio of the optimal portfolio when the benchmark b is the market risk factor.

Figure 6.2: Long only solution (from Cazalet and Roncalli [2014])

56
Figure 6.3: Long-short solution (from Cazalet and Roncalli [2014])

The results indicate the dependence of the solution on estimation and stability of parameters (such
as alpha, beta or volatility of each risk factor) and on the universe of risk factors (inclusion or the
deletion of a specic risk factor would highly impact the solution).
Saunders et al. [2013] have considered fund-of-hedge-funds asset allocation using regime-switching
models. They have computed the sensitivity of the optimal portfolio weights with respect to the distress
regime probability, dened (depending on the model) either as the probability that any given period
turns out to be a crisis regime or the probability of transitioning to a crisis from a normal market
∂wj
state. This sensitivity
∂p , denoted the asset-weight distress sensitivity, measures the rate of change
of the optimal asset weight with respect to the distress regime probability.
Values of asset-weight distress sensitivities provides information towards identifying which of the
underlying assets should be under-weighted and which ones should be over-weighted if one has a
∂wj
forecast of the likelihood of occurrence of the crisis regime. The sign of
∂p will indicate to the
portfolio manager whether to under-weight () or over-weight ( + ) the specic asset.

Figure 6.4: Asset-weight distress sensitivities for MSM model (from Saunders et al. [2013])

57
Figure 6.5: Asset-weight distress sensitivities for GMM model (from Saunders et al. [2013])

Bender et al. [2013b] computed factor exposures using the Barra multi-factor models which estimate
factor portfolios using multivariate regressions. As shown in next gure, in some cases the sensitivities
indicate that the portfolio may have signicant exposure to factors other than the intended factor.

Figure 6.6: Factor exposures for MSCI factor indices (from Bender et al. [2013b])

58
Werley and Speer [2011] employed a generic factor risk framework to obtain greater transparency
when studying hedge funds, from a manager due diligence perspective as well as from a portfolio
construction perspective. Many managers are not willing to provide investors with full transparency
into their holdings, or do not provide intra-month estimates. This means that a returns-based analysis
is going to be built on monthly observations, and extra caution should be applied.
Through factor analytics the monthly portfolio or manager returns are separated into the systematic
or market risk (beta) a manager takes on, and the manager-specic risk or true value added of a
manager's investing (alpha).
The factor model is applied to a hedge fund, which did not provide its investors with full, portfolio-
level transparency. Te hedge fund's risk decomposition of the fund was rolled up into a universal
market factor set with 10 separate market factor exposures, with U.S. and international market equities
contributing 26% percent of the portfolio's total risk or volatility, commodities 7% and credit risk 17%.

Figure 6.7: Volatility decomposition of hedge fund (from Werley and Speer [2011] )

The manager component of the factor decomposition is essentially the residual risk that cannot
be explained by the other factors. For hedge funds, this is a close proxy for alpha risk, or the risk
associated with alpha generation.
Once factor exposures have been established, one can combine them together with factor returns to
create a custom benchmark for a portfolio or fund. The comparison between performance of considered
portfolio and respectively of its custom factor benchmark, shown below, illustrate that factor model
was successful in identifying the pattern of risk taking on a monthly basis.

Figure 6.8: Comparison performances for hedge fund and factor-based benchmark (from Werley and
Speer [2011] )

59
Sensitivity analysis would also prove useful to assess future risk, since stress testing and scenario
simulations using previously derived market factor exposures provide a unique perspective on portfolio
risk. The most important of portfolio's factor sensitivities could be applied to either a stress scenario,
such as September 2008, or a customized simulation reecting a more likely future scenario.

7 Good testing procedures in portfolio management


Designing good testing procedures is a very important component of any project relying on assumptions
and advanced computations, and portfolio management denitely requires such testing procedures.
That is even more relevant today, since quantitative investment teams routinely scan through petabytes
of nancial data, assisted by a host of technologies and advancing mathematical elds, looking for
patterns and for protable investment strategies.
We explain why simply relying on backtesting is not enough to fully quantify the strategies, and
sometimes may even be misleading. We then present alternatives, such as multiple-tests testing frame-
work, comprehensive stress testing and advanced Monte Carlo simulation.

7.1 Only backtesting is not enough


A backtest is a historical simulation of an algorithmic investment strategy, computing the series of
prots and losses that such strategy would have generated had that algorithm been run over that time
period. Popular performance statistics, such as the Sharpe ratio or the Information ratio, are used
to quantify the backtested strategy's return on risk and are studied by investors before deciding to
allocate capital to the best performing scheme.
Although backtesting is a powerful and necessary research tool, it can also be easily manipulated.
Harvey and Liu [2014a], Harvey [2014], Harvey and Liu [2014b], Bailey et al. [2014a,b], Bailey and
de Prado [2014]. For instance Bailey and de Prado [2014] argue that the most important piece of
information missing from virtually all backtests published in academic journals and investment oerings
is the number of trials attempted. Without this information, it is impossible to assess the relevance of
a backtest. To quote them:

Put bluntly, a backtest where the researcher has not controlled for the extent of the search
involved in his or her nding is worthless, regardless of how excellent the reported performance
might be.

7.2 Backtest: a deeper discussion


We investigate details, characteristics and pitfalls of backtesting strategies.

7.2.1 In-sample and out-of-sample


Regarding the measured performance of a backtested strategy, we have to distinguish between two very
dierent readings: in-sample (IS) and out-of-sample (OOS). The IS performance is the one simulated
over the sample used in the design of the strategy (also known as learning period or training set
in the machine-learning literature). The OOS performance is simulated over a sample not used in
the design of the strategy (a.k.a. testing set). A backtest is realistic when the IS performance is
consistent with the OOS performance.
It becomes apparent that is not enough to rely only on backtests to determine the performance and
robustness of the asset allocation strategy. While the quote from Masi [2014]  You never see a bad
back-test. Ever. In any strategy. may be too extreme, it conveys the observation that while any given
backtest will only show you what the results were for that particular market and rarely carry enough
weight to put the results to use in current market conditions.

60
As shown in Masi [2014], the overpromise that comes via back testing was nicely captured in a
chart displaying ETF performance pre and post creation. It was noted that ETF creators have been
back-tting strategies that have worked well in the past, but once a new ETF is launched, only 51%
outperform the broad US market. To be more specic, it was calculated that the average annualized
excess return for this series of ETFs is 10.3% in the 5 years before inception, but actually -1% in the
5 years afterward.
Perhaps the most common approach among practitioners is to require the researcher to hold-out a
part of the available sample (also called test set method). This hold-out is used to estimate the OOS
performance, which is then compared with the IS performance. If they are congruent, the investor has
grounds to reject the hypothesis that the backtest is overt. The main advantage of this procedure
is its simplicity. However, as described in Bailey et al. [2014a,b], this approach is unsatisfactory for
multiple reasons:

• if the data is publicly available, it is quite likely that the researcher has used the hold-out as
part of the IS.

• even if no hold-out data was used, any seasoned researcher knows well how nancial variables
performed over the OOS interval, and that information will be used in the strategy design,
consciously or not

• hold-out is clearly inadequate for small samples. The IS will be too short to t, and the OOS
too short to conclude anything with sucient condence.

• even if the researcher counts with a large sample, the OOS analysis will consume a large amount
of the sample to be conclusive, which is detrimental to the strategy's design. If the OOS is taken
from the end of a time series, we are losing the most recent observations, which often are the
most representative going forward. If the OOS is taken from the beginning of the time series,
the testing will be done on the least representative portion of the data.

• as long as the researcher tries more than one strategy conguration, overtting is always present.
The hold-out method does not take into account the number of trials attempted before selecting a
particular strategy conguration, and consequently hold-out cannot correctly assess a backtest's
representativeness. Hold-out leaves the investor guessing to what degree the backtest is overt.

7.2.2 Overtting
Given any nancial series, it is relatively simple to overt an investment strategy so that it performs well
IS. Overtting (see Bailey et al. [2014a,b]) is a concept borrowed from machine learning and denotes the
situation when a model targets particular observations rather than a general structure. For example, a
researcher could design a trading system based on some parameters that target the removal of specic
recommendations that she knows led to losses IS (a practice known as data snooping). After a few
iterations, the researcher will come up with optimal parameters, which prot from features that are
present in that particular sample but may well be rare in the population.
The meaningful answer to the question  is this backtest overt? is not a simple true or false, but
rather a non-null probability that depends on the number of trials involved (an input ignored by
hold-out). For example, Bailey et al. [2014b], Bailey and Lopez de Prado [2012] have evaluated the
probability that an estimated Sharpe ratio exceeds a given threshold in presence of non-Normal returns,
and have utilized this uncertainty-adjusted investment skill metric (termed Probabilistic Sharpe ratio,
or PSR) to:

1. allow establishing the track record length needed for rejecting the hypothesis that a measured
Sharpe ratio is below a certain threshold with a given condence level.

61
2. model the trade-o between track record length and undesirable statistical features (e.g., negative
skewness with positive excess kurtosis).

3. explain why track records with those undesirable traits would benet from reporting performance
with the highest sampling frequency such that the IID assumption is not violated.

4. enable the computation of the Sharpe ratio Ecient Frontier (SEF), which provides a way to
optimize a portfolio under non-Normal, leveraged returns while incorporating the uncertainty
derived from track record length.

A very important question is How easy us to overt backtest data? According to Bailey et al. [2014a,b],
Very Easy! If only 2 years of daily backtest data are available, then no more than 7 strategy variations
should be tried. If only 5 years of daily backtest data are available, then no more than 45 strategy
variations should be tried. In general, if a nancial analyst or researcher does not report the number
of trials N explored when developing an investment strategy, then it is impossible for a third party
to properly reconstruct the results, in the sense of correctly ascertaining the true level of reward
and risk. Indeed, it is shown in Bailey et al. [2014a,b] that for given any desired performance goal, a
nancial analyst just needs to keep trying alternative parameters for her strategy, and eventually she
will nd a variation that achieves the desired goal, yet the resulting strategy will have no statistically
valid predictive ecacy whatsoever
Another approach popular among practitioners consists in modeling the underlying nancial variable,
generate random scenarios and measure the performance of the investment strategy on those scenarios.
This presents the advantage of generating a distribution of outcomes, rather than relying on a single
OOS performance estimate, as the hold-out method does. The disadvantages are

• the model that generates random series of the underlying variable may also be overt

• it may not contain all relevant statistical features

• it has to be customized to every variable (large development costs).

Although there are many academic studies that claim to have identied protable investment strategies,
their reported results are almost always based on IS statistics. Only exceptionally do we nd an
academic study that applies the hold-out method or some other procedure to evaluate performance
OOS. It is argued in Harvey et al. [2014] that there are hundreds of papers supposedly identifying
hundreds of factors with explanatory power over future stock returns.
They conclude that backtest overtting is dicult to avoid, since any perseverant researcher will
always be able to nd a backtest with a desired Sharpe ratio regardless of the sample length requested.
As a result, and given that most published backtests do not report the number of trials attempted,
many of them may be overtted.
These conclusions echo the ones in Ioannidis [2005], although in a much more general setting:

for most study designs and settings, it is more likely for a research claim
to be false than true. Moreover, for many current scientific fields, claimed
research findings may often be simply accurate measures of the prevailing bias.

7.2.3 Minimum track record and backtest period


How long
The Minimum Track Record Length (MinTRL) was introduced to answer the question
should a track record be in order to have statistical confidence that its Sharpe ratio
is above a given threshold?.

62
The analogue to MinTRL in the context of overtting prevention when comparing multiple strategies,
termed Minimum Backtest Length (MinBTL), describes the length of time period needed to avoid
selecting a skill-less strategy among alternative specications.
Details on both quantities are given in Appendix A.

7.2.4 Assessing a backtest


Assessing the representativeness of a backtest is not an easy problem, as evidenced by the scarcity of
academic papers addressing a dilemma that most investors face.
A general framework is proposed in Bailey et al. [2014b] that adapts recent advances in experimental
mathematics, machine learning and decision theory to this very particular problem of assessing the
representativeness of a backtest. One advantage of the proposed solution is that it only requires time
series of backtested performance. It avoids the credibility problem (of preserving a truly OOS test-
set) by not requiring a xed hold-out, and swapping all IS and OOS sets. The approach is generic
in the sense of not requiring knowledge of the trading rule or forecasting equation. The output is
a bootstrapped distribution of OOS Sharpe ratios, as well as a measure of the representativeness of
the backtest that is called probability of backtest overtting (PBO). Although in their examples the
Sharpe ratio is always chosen to evaluate performance, the methodology can be applied to any other
performance measure.

7.3 Multiple testing framework


Most empirical studies which test portfolio performance are based on simple-testing procedures, e.g.,
DeMiguel et al. [2009a,b]. It is argued in Frahm et al. [2012], Harvey [2014] that is likely for such
studies to lead to wrong conclusions since the probability of rejecting a true null hypothesis can be
substantially larger than the nominal error rate of the simple test. Additionally, we would like to
identify all strategies which are outperformed by a specic strategy. The last question is a typical
application of multiple-testing procedures Romano et al. [2008].
In physical and medical sciences, multiple testing methods are routinely used in hypothesis testing,
as exemplied in Harvey [2014]:

• In medical research, a drug is given to many groups of people A p-value of 0.1% (t-ratio = 3.3)
is often used as the cuto rather than 5%

• In astronomy and physics, researchers test thousands or even millions of times for new planets
or particles  The high prole discovery of the Higgs Boson required a t-ratio of 5.0 (p-value <
0.0001%)

7.3.1 Testing multiple hypotheses


Hypothesis testing is the procedure of assessing whether variation between two sample distributions
can just be explained through random chance or not. If we have to conclude that two distributions
vary in a meaningful way, we must take enough precaution to see that the dierences are not just
through random chance. Relevant references include Harvey and Liu [2014a], Harvey et al. [2014],
Harvey [2014], Ferson and Chen [2014], Frahm et al. [2012], Harvey and Liu [2013]
There are 2 main types of errors in context of multiple hypotheses testing:

• Type I error, also known as a false positive: the error of rejecting a null hypothesis when it
is actually true. In other words, this is the error of accepting an alternative hypothesis (the
real hypothesis of interest) when the results can be attributed to chance. Plainly speaking, it
occurs when we are observing a dierence when in truth there is none (or more specically - no
statistically signicant dierence).

63
• Type II error, also known as a "false negative": the error of not rejecting a null hypothesis when
the alternative hypothesis is the true state of nature. In other words, this is the error of failing
to accept an alternative hypothesis when you don't have adequate power. Plainly speaking, it
occurs when we are failing to observe a dierence when in truth there is one.

In a nutshell, the Type I Error occurs when we choose a strategy that should have been discarded (a
false positive), and the Type II Error occurs when we discard a strategy that should have been chosen
(a false negative). Decision makers are often more concerned with false positives than with false
negatives because they would rather exclude a true strategy than risking the addition of a false one.
Such a mindset is denitely true for risk-averse investors, a lost opportunity is less worrisome than an
actual loss.
Suppose now that we are interested in analyzing multiple strategies on the same dataset, with the
aim of choosing the best, or at least a good one, for future application. A curious problem then
emerges: As we test more and more strategies, each at the same signicance level (probability of type I
error), the overall probability of choosing at least one poor strategy grows. This is called the multiple
testing problem, and it is so pervasive and notorious that the American Statistical Society explicitly
warns against it Bailey and de Prado [2014].
Researchers conducting multiple tests on the same data tend to publish only those that pass a
statistical signicance test, hiding the rest. Because negative outcomes are not reported, investors are
only exposed to a biased sample of outcomes. This problem is called selection bias, it is caused by
multiple testing combined with partial reporting, and it appears in many dierent forms according to
Bailey and de Prado [2014]:

• Analysts who do not report the full extent of the experiments conducted (le drawer eect)

• journals that only publish positive outcomes (publication bias)

• indices that only track the performance of hedge funds that didn't blow up (survivorship bias)

• managers who only publish the history of their (so far) protable strategies (self selection bias,
backlling)

What all these phenomena have in common is that critical information is hidden from the decision-
maker, with the eect of a much larger than anticipated Type I Error probability.

7.3.2 False discoveries and missed discoveries


The Type I error is the false discovery (investing in an unprotable trading strategy). The Type II error
is missing a truly protable trading strategy. Inevitably there is a tradeo between these two errors.
In addition, in a multiple testing setting it is not obvious how to jointly optimize these two types of
errors, such that we overlook only a minimal number of protable strategies when identifying trading
strategies that appear to be protable  but they are not. The good news is that perspective presented
in Harvey and Liu [2014b] may oer a solution: while using the single test criteria for multiple tests
induces a very large number of false discoveries (large amount of Type I error), the number of false
discoveries is greatly reduced at minimal cost to missing discoveries, as shown in next gure:

64
Figure 7.1: Optimal tradeo between false and missing discoveries (from Harvey and Liu [2014b])

Thus it appears that the primary challenge when testing multiple hypotheses is to guard against
false positive results, and two of the most common criteria in the statistics literature are:

1. family-wise error rate (FWER), describing the probability of making at least one false discovery

2. false discovery rate (FDR), which is the expected proportion of falsely rejected hypotheses.

The distinction between the two is very intuitive. In FWER, it is unacceptable to make a single
false discovery. This is a very severe rule but completely appropriate for certain situations. With the
FWER, one false discovery is unacceptable in 100 tests and equally as unacceptable in 1,000,000 tests.
In contrast, the FDR views unacceptable in terms of a proportion. For example, if one false discovery
was unacceptable for 100 tests, then 10 are unacceptable for 1,000 tests. The FDR is much less severe
than the FWER.
Two main categories of methods have been proposed to control the various error rates:

• simple adjustments: relies on controlling FWER or FDR under a general data structure by using
simple adjustments based only on summary t-statistics or p-values for each individual hypothesis
test. However, they are often too conservative (i.e., too few true discoveries).

• permutation sampling: resamples the entire dataset several times and constructs an empirical
distribution for the pool of test statistics. The empirical distribution then serves as the reference
distribution for determining the cuto values. This approach incorporates the data structure,
in particular the correlation structure among the individual test statistics. Therefore, it is less
conservative than the aforementioned methods. However, permutation tests are computationally
challenging. Moreover, permutation tests also require the knowledge of each individual dataset
based on which the t-statistic or p-value constructed. In cases when this information is not
available, permutation tests are not feasible.

Unfortunately, many interesting empirical inquiries in economics and science do not align with both
the simple adjustment and the permutation resampling approach. First, many economic variables are
inuenced by common shocks and are, hence, correlated. This means simple adjustment are overly
conservative and misleading as independence among test statistics is unrealistic. Second, when a
collection of previous studies are pooled together, we often do not have the luxury of having the
original dataset for each study. All we have is the single test statistic that summarizes the signicance
a research nding.

7.3.3 Examples
The procedures presented in Barras et al. [2010] and Ferson and Chen [2014] illustrate how the multiple
hypothesis approach can be deployed in context of asset management. Imagine that the population of

65
fund managers consists of three sub populations. A fraction of funds have zero alphas, a fraction are
good managers with positive alphas and a fraction are "bad" managers with negative alphas.Barras
et al. [2010] solved this problem by simulating the cross-section of funds' returns and using false
discovery rate, which is the expected fraction of the funds where the null hypothesis of zero alphas is
rejected, but for which the true alpha is zero. A number of studies subsequently applied their approach
to both mutual funds and hedge funds. Then the approach was rened in Ferson and Chen [2014] in
several directions:

• allow for less than perfect power in the tests and for the possibility that a test will confuse a bad
fund with a good fund.

• simultaneously estimate the true alphas and the fractions

The methodology is applied to both active US equity mutual funds during 1984-2011 and a sample
of hedge funds during 1994-2011. Both Barras et al. [2010] and Ferson and Chen [2014] nd that no
mutual funds have signicantly positive alphas after costs and expenses. In fact, using mutual funds
during 1984-2011, the best tting mixture of distributions model with three alpha values is one in which
there are two negative alphas and some zero alpha funds. The model with three alpha distributions
implies that about 35% of the mutual funds are good (zero alpha), about 30% are bad (alpha of
-0.06% per month) and the remaining 35% are ugly (alpha of -0.21% per month).
Harvey and Liu [2013] introduced a framework that allows for comparisons across many research
studies and explicitly controls for the correlation among the proposed variables. while using only sum-
mary statistics. The starting point is given by modeling the distribution of null hypotheses. Motivated
by standard Bayesian hypothesis testing, the framework uses a parametric mixture distribution to
succinctly capture how null and non-null hypotheses are drawn. Next, commonly used test statistics
are decomposed into the sum of score statistics, with usage the Pearson correlation among the con-
temporaneous score statistics to model the dependence among test statistics. Then, model parameters
are estimated by matching key moments of model implied and observed summary statistics. Finally,
threshold levels for hypothesis testing are found by equaling the exactly calculated Type I error rate
under the estimated parameter values to a prespecied signicance level.
The framework was applied in Harvey et al. [2014] for identifying the subset of relevant factors
within the large set (314 factors from 311 papers) that was collected from literature, and respectively
in Harvey and Liu [2014a] for evaluating backtests of portfolio strategies.

7.3.4 When to stop testing


While multiple testing is a very useful tool, it should not be abused. Multiple testing exercises should be
carefully planned in advance, so as to avoid running an unnecessarily large number of trials. Investment
theory, not computational power, should motivate what experiments are worth conducting. This begs
the question what is the optimal number of trials that should be attempted?
According to Bailey and de Prado [2014], an elegant answer to this critical question can be found
in the theory of optimal stopping, more concretely the so called secretary problem, or 1/e-law of
optimal choice. The key notion is that we wish to impose a cost to the number of trials conducted,
because every additional trial irremediably increases the probability of a false positive. In the context
of our discussion, it translates as follows: From the set of strategy congurations that are theoretically
justiable, sample a fraction 1/e of them (roughly 37%) at random and measure their performance.
After that, keep drawing and measuring the performance of additional congurations from that set,
one by one, until you nd one that beats all of the previous. That is the optimal number of trials, and
that best so far strategy the one that should be selected.
This result provides a useful rule of thumb, with applications that go beyond the number of strategy
congurations that should be backtested. It can be applied to situations where we test multiple

66
alternatives with the goal of choosing a near-best as soon as possible, so as to minimize the chances of
a false positive.

7.4 Evaluation based on Monte Carlo simulations


An illustration of the power and versatility of Monte Carlo simulation may convince the reader of
necessity to deploy such a tool within context of portfolio management.
The testing framework, described in Reust [2013], had to fulll a number of criteria:

• Include many dierent potential futures

• Include many dierent cash-ow patterns

• Respect path dependence (taxes one pays this year cannot be invested next year)

• Accurately analyze and assess the performance of the strategy in live conditions.

Testing was performed using many simulated potential future situations, with each strategy con-
fronted with both bootstrapped (sampled in chunks and reordered) historical data and novel simu-
lated data, and it is typical of how evaluation of portfolio strategies is usually done within a Monte
Carlo framework. The historical data allows inclusion of important aspects of return movements, like
auto-correlation, volatility clustering, correlation regimes, skew, and fat tails, while the simulated data
allows generation of novel potential outcomes, like market crashes bigger than previous ones, and
generally, future behaviors dierent than the past behavior.
We now describe methodologies which combine Monte Carlo framework with more specialized pro-
cedures. References include Jiang and Martin [2013], Pedersen [2014], Lewis [2011], Zivot [2011].
Lewis [2011] introduced an improved testing process for momentum (or relative strength) strategies,
which incorporated two elements:

• a continuous portfolio testing protocol managing portfolios similar to real world

• a Monte Carlo process overlaid on the continuous portfolio testing to insure robustness.

Zivot [2011] introduced Factor Model Monte Carlo(FMMC), which uses a tted factor model to simu-
late pseudo asset return data preserving empirical characteristics of risk factors and residuals. More-
over, the approach use full data for factors, rely on unequal history for assets to deal with missing
data, and estimate tail risk and related measures non-parametrically from simulated return data.
Then the algorithm is as follows:

1. Simulate values of the risk factors by re-sampling from full sample empirical distribution

2. Simulate values of the factor model residuals from tted non-normal distribution

3. Create factor model returns from factor models t over truncated samples, simulated factor
variables drawn from full sample and simulated residuals

A common problem in asset and portfolio risk and performance analysis is that the manager has such
a short history of asset returns that risk and performance measure estimates are quite unreliable. To
analyze how bad estimates based on short returns histories can be relative to those based on longer
histories, Jiang and Martin [2013] computed estimates for volatility, value-at-risk (VaR) and conditional
value-at-risk (CVaR) for an emerging markets (EM) hedge fund, using 10 years and 3 years data.

Volatility VaR for 95% CVaR for 95%

10 years data 47%(6%) 19%(3%) 35%(8%)


3 years data 20%(9%) 9.5%(9%) 10%(13%)

67
If 10 years of data are used, standard errors are small enough to render reasonable condence in
estimates. The shorter histories give rise to such larger standard errors rendering the risk estimates of
little use in the case of VaR and CVaR.
But with available long histories of many risk factors one can use a subset of them to construct a
high R2 risk-factor model for the asset returns, by tting a time series factor model for asset returns
with short histories whose risk and performance estimates are needed, using as predictor variables the
contemporaneous segment of a well-selected set of risk factors that have much longer histories.
Monte Carlo simulation may also be used to assess future performance of portfolio strategies, as
shown below.

Figure 7.2: Expected Risk-Return Prole of various strategies (from Teuber [2013])

For completion, Appendix I provides a short Monte Carlo primer.

7.5 Stress testing


The technique of stress testing is used to estimate the potential loss by submitting the model to extreme
variations in the parameters, corresponding to nancial disaster scenarios such as stock market crash,
collapse of the exchange rate, sudden increase in interest rates, etc. Both investors and regulators
require stress tests to assess whether portfolio managers quantify well enough the portfolio behavior
due to extreme variations of the risk factors to which the portfolios might be exposed to. Stress
tests can detect a portfolio's vulnerabilities and assess its expected reaction to market scenarios, and
consequently can add signicant value to an investment process. The stress scenarios must be adapted
to the nature of the portfolio's positions and its risks, and therefore any fundamental change in the
investment strategy should be accompanied by a recalibration of the crisis scenarios.
However, it can be challenging to determine and implement the most salient scenarios, which should
cover all the risk factors having a non-negligible inuence on the portfolio's value and handle correlation
changes between risk factors. Furthermore, the output of many stress tests is expressed in terms of
prot and loss (P&L), and this information is not directly actionable. The investor must translate
P&L into modied portfolio weights.
Good general overviews of stress testing are presented in Szylar [2014], Siddique and Hasan [2012],
while references for stress testing in context of portfolio management include Rebonato and Denev
[2012], Meucci [2012, 2014], Meucci et al. [2012], Siddique and Hasan [2012], Taleb et al. [2012], Cue
and Goldberg [2011], Ardia and Meucci [2014], Bruneau et al. [2014], Denev [2011], Pedersen [2014].

7.5.1 Key questions and decisions


According to Szylar [2014], there are a number of elements involved in the design of any stress scenario,
including:

68
• The choice of the type of risks to analyze (market, credit, interest rate, liquidity, etc.).

• Whether single or multiple risk factors are to be shocked.

• What parameter(s) to shock (prices, volatilities, correlations, etc.).

• By how much (based on historical or simulated scenarios).

• Over what time horizon.

One of the key decisions is how to calibrate the size of the shocks used for stress testing, since setting the
hurdle too low or too high might make the whole exercise meaningless. At a minimum, shocks should
be calibrated to the largest past movement in the relevant risk variables over a certain horizon (change
from peak to trough or deviation from trend). However, only relying on past data is not enough, and
stress scenarios should also include extreme events that might happen in the future. Alternatively,
with sucient data, we can attempt to estimate the joint empirical distribution of past deviations
from the trend of the relevant risk variables and use its quantiles as a starting point for simulating the
stress scenarios.

7.5.2 Standard, historical or worst case?


Traditionally, stress tests for market risk have been conducted by using severe but plausible scenarios
of possible states for market risk factors. The traditional approaches for scenarios utilize one of the
three methods:

1. standard: market risk factors are stressed by prespecied shocks, such as equity prices changing
by some standard deviations or increasing/decreasing oil prices to a certain level

2. historical: use market states related to historical time period with great market turbulence

3. worst-case: use extreme values of market variables

The traditional stress tests are easy to conduct but have important limitations. For example, the
historical scenarios oer reliable but likely less relevant information for risk management in the future;
the standard analysis may not be based on the changes in the market states that are close to a stress
event; and the worst-case analysis looks at the impacts of changes that are unlikely to occur.
Cue and Goldberg [2011] introduced a structured set of tools to envision and administer extreme
historical and hypothetical scenarios, and provided a way to incorporate the output of a portfolio
stress test directly into an investment decision. The framework was based on a scenario-constrained
mean-variance optimization that can incorporate extreme risk and non-Gaussian eects. It was shown
that a wide range of risk climates can be obtained from history even within a Gaussian framework, by
varying the analysis date and the responsiveness of covariance matrix estimation. Furthermore, any
historical covariance matrix can be modied to reect exogenous views on future correlations via the
introduction of a latent factor, which can represent a ight to quality, a change in liquidity, or another
transient eect that disrupts markets in a crisis.

7.5.3 Entropy pooling


The generalized Bayesian approach known as entropy pooling, introduced in Meucci [2008], is a
versatile framework to process market views and generalized stress-tests into an optimal "posterior"
market distribution. The inputs are an arbitrary market distribution ( prior distribution ) and gen-
eralized views on the market, such as views held by portfolio manager on spread, tail distribution,
correlation or stress tests. The output is a distribution called posterior distribution that is of minimum

69
deviation from the prior, but at the same time satises the generalized market views, often unlike the
prior.
The framework is extended in Ardia and Meucci [2014], Meucci et al. [2014b, 2012] to fully incor-
porate general views, such as inequality statements, nonlinear views and rankings based on market
observables. The limitation is that the posterior can be computed only with numerical methods,
making it challenging computationally, especially for large markets. To address this shortcoming, the
Factor Entropy Pooling is introduced as an enhancement which reduces the dimension of the
asset correlation structure using a factor model methodology and selects coordinates such that the
optimization target becomes unconstrained.
A procedure for multidimensional shock scenarios is described in Flood and Korenko [2013]. It
involves a grid search of sparse, well distributed, stress-test scenarios, and is viewed as a middle
ground between traditional stress testing and reverse stress testing. The methodology systematically
enforces internal consistency among the shock dimensions by sampling points of arbitrary severity from
a plausible joint probability distribution. The methodology may be tailored to situations confronted
in many regulatory stress tests, in which the loss function is not directly observable, in part because
the portfolio composition cannot be precisely observed without intensive eort.

7.5.4 Stressing covariance or correlation matrices


This is an important topic by itself, and requires careful consideration, because the correlation matrix
cannot be stressed in a straightforward way of changing only one element at a time, since the resulting
matrix may not be a valid correlation matrix anymore.
Galeeva et al. [2007, 2011] consider the following methods:

1. Bootstrapping: essentially a resampling of the historical timeseries

2. Element-wise perturbation: local perturbation of the correlation matrix elements

3. Angle perturbation: changes are applied to correlation parametrization based on angles

4. Eigenvalue perturbation: changes are applied to eigenvalues, while keeping eigenvectors the
same

The dierent methods lead to dierent results for the correlation VaR, which is not surprising since
there is clearly a lot of freedom in how to choose the perturbation and underlying densities. The
conclusion was that parametrized methods, based on either angles or eigenvalues, provide the most
appropriate tool to stress-test complicated portfolios of instruments and do it in a well-dened manner.

7.5.5 Reducing dimensionality


Instead of stressing all factors, one could reduce the computational burden by stressing only the most
important factors (chosen through PCA or some other procedure). Bruneau et al. [2014] considered
such an approach to test the impacts on the portfolio composition, expected return and CVaR.
Rebonato and Denev [2012], Denev [2011] introduced a framework for asset allocation in the presence
of stress events carried out in a coherent way instead of being done as an ad-hoc approach. They
proposed a procedure that blends standard statistical techniques such as correlations, copulae etc.
with a subjective Bayesian approach based on causality assumptions to model stress events. It is
argued approach is highly transparent, auditable and easy to be understood by non-specialists.
The framework is also constructed such that it can incorporate the views that exceptional events
facing the portfolio manager at any point in time are specic to the each individual crisis, and that past
regularities cannot be relied upon. It deals with exceptional returns by eliciting subjective probabilities,
and by employing the Bayesian net technology to ensure logical consistency. The portfolio allocation

70
is then obtained by utility maximization over the combined (normal plus exceptional) distribution of
returns.

7.5.6 Reverse stress testing


Reverse stress tests try to identify the risks that would lead an institution to fail. This is an appealing
idea in the sense that instead of starting from the existing standpoint and seeing how close one can go
toward the ridge of the cli without falling, reverse stress testing indicates what risks could be taken
to fall directly o the cli.
To quote from Szylar [2014]: That makes so much sense that you may wonder why we have
not carried out reverse stress tests for ages.
The main problem with reverse stress testing is  how to do it. There are so many reasons why an
institution would fail that it may take some time to determine meaningful stress tests. When we would
conduct other types of stress testing, we always would start from the known (the portfolio itself and
its VaR) and try to progress more or less in the dark to gauge the risks ahead. With a reverse stress
test, we would have to start from the unknown and try to gure out how we became lost on the way
home.

8 Conclusion
We have considered factor-based investing, starting with the main features of factors, factor investing
and factor models. We have discussed topics such as selection of factors, which factors are important
for specic asset classes, how to dierentiate between factors, anomalies or stylized facts, reasons to
prefer a combination of factors. We have looked into implementation details and have analyzed the
factor risk parity strategy, which combines factor-investing and risk parity.
It is our strong belief that factor-based investing can be greatly improved using regime switching and
sensitivity analysis. With this purpose in mind we have presented theoretical and practical frameworks
for Markov switching models and for sensitivity analysis, and have relied on representative examples
to illustrate the benets of eciently incorporating regimes and sensitivity analysis into portfolio
management.
Finally we describe features of good testing procedures for portfolio behavior and performance, while
also mentioning possible testing pitfalls.

71
Bibliography
N. Amenc and F. Goltz. Scientic beta multi-beta multi-strategy indices: Implementing multi-
factor equity portfolios with smart factor indices. http://docs.edhec-risk.com/
Available at
mrk/000000/Scientific_Beta_Library/External_use/White_papers/ERI_Scientific_Beta_
Publication_Scientific_Beta_Multi-Strategy_Factor_Indices.pdf, 2014.
N. Amenc, F. Goltz, A. Lodh, and L. Martellini. Diversifying the diversiers and tracking the track-
ing error: Outperforming cap-weighted indices with limited risk of underperformance. Journal of
Portfolio Management, 38(3):7288, 2012.

N. Amenc, R Deguest, F. Goltz, A. Lodh, and E. Shirbini. Risk allocation, factor investing and
smart beta: Reconciling innovations in equity portfolio construction. http://
Available at
www.edhec-risk.com/edhec_publications/all_publications/RISKReview.2014-07-09.0733/
attachments/EDHEC-Risk_Publication_Risk_Allocation_Factor_Investing_Smart_Beta.pdf,
2014.

M. Ammann and M. Verhofen. The eect of market regimes on style allocation. Financial Markets
and Portfolio Management, 20(3):309337, 2006.

A. Ang. Factor based approaches to risk parity. Available at SSRN: http://ssrn.com/abstract=


2277397, 2013a.

A. Ang. Factor investing. Available at SSRN: http://ssrn.com/abstract=2277397, 2013b.

A. Ang and G. Bekaert. International asset allocation with regime shifts. The Review of Financial
Studies, 15(4):11371187, 2002.

A. Ang and G. Bekaert. How do regimes aect asset allocation. Financial Analysts Journal, 60(2):
8699, 2004.

A. Ang and A.G. Timmermann. Regime changes and nancial markets. Available at SSRN: http:
//ssrn.com/abstract=1919497.
A. Ang, S.M. Schaefer, and N. Goetzmann. Evaluation of Active Management of the Norwegian Gov-
ernment Pension Fund Global. Available at http://www.regjeringen.no/upload/fin/statens%
20pensjonsfond/rapporter/ags%20report.pdf, 2009.
T. Angelidis and N. Tessaromatis. Global style portfolios based on country in-
dices. Available http://faculty-research.edhec.com/_medias/fichier/
at
edhec-working-paper-global-style-portfolios-based-on-country_1399629135936-pdf,
2014.

David Ardia and Attilio Meucci. Parametric stress-testing in non-normal markets via entropy pooling.
Available at SSRN: http://ssrn.com/abstract=2457459, 2014.

F.M. Asl and E. Etula. Advancing strategic asset allocation in a multi-factor world. The Journal of
Portfolio Management, 39(1):5966, 2012.

C. Asness, A. Frazzini, and L. H. Pedersen. Leverage aversion and risk parity. Financial Analysts
Journal, 68(1):4759, 2012.

Cliord S. Asness, Tobias J. Moskowitz, and Lasse Heje Pedersen. Value and momentum everywhere.
The Journal of Finance, LXVIIII(3):929985, 2013a.

72
C.S. Asness, A. Frazzini, and L.H. Pedersen. Quality minus junk. Available at http://ssrn.com/
abstract=2312432, 2013b.

C.S. Asness, A. Frazzini, R. Israel, and T.J. Moskowitz. Fact, ction and momentum investing. Journal
of Portfolio Management, 2014.

David H. Bailey and Marcos Lopez de Prado. The deated sharpe ratio: Correcting for selection bias,
backtest overtting, and non-normality. The Journal of Portfolio Management, 40(5):90107, 2014.

D.H. Bailey and M. Lopez de Prado. The Sharpe ratio ecient frontier. Journal of Risk, 15(2), 2012.

D.H. Bailey, J.M. Borwein, M. Lopez de Prado, and Q.J. Zhu. Pseudo-mathematics and nancial char-
latanism: The eects of backtest overtting on out-of-sample performance. Notices of the American
Mathematical Society, 61(5):458471, 2014a.

D.H. Bailey, J.M. Borwein, M. Lopez de Prado, and Q.J. Zhu. The probability of backtest overtting.
Available at SSRN: http://ssrn.com/abstract=2326253, 2014b.

J. Bambaci, J. Bender, R. Briand, A. Gupta, B. Hammond, and M. Subramanian. Harvesting Risk


Premia for Large Scale Portfolios. Analysis of Risk Premia Indices for the Ministry of Finance,
Norway. http://www.msci.com/products/indexes/strategy/factor/Harvesting_
Available at
Risk_Premia_For_Large_Scale_Portfolios_Public.pdf, 2013.
Brad M. Barber, Xing Huang, and Terrance Odean. Which Risk Factors Matter to Investors? Evidence
from Mutual Fund Flows. Available at SSRN: http://ssrn.com/abstract=2408231, 2014.

Laurent Barras, Olivier Scaillet, and Russ Wermers. False discoveries in mutual fund performance:
Measuring luck in estimated alphas. Journal of Finance, 65:179216, 2010.

R. Bauer, R. Haerden, and R. Molenaar. Asset allocation in stable and unstable times. Journal of
investing, 13(3):7280, 2004.

J. Bender, R. Briand, F. Nielsen, and D. Stefek. Portfolio of risk premia: A new approach to diversi-
cation. Journal Of Portfolio Management, 36(2):1725, 2010.

J. Bender, R. Briand, D. Melas, and R.A. Subramanian. Foundations of factor investing. Available at
http://www.msci.com/resources/pdfs/Foundations_of_Factor_Investing.pdf, 2013a.

J. Bender, R. Briand, D. Melas, R.A. Subramanian, and M. Subramanian. Deploying multi-factor


index allocations in institutional portfolios. Available at http://www.msci.com/resources/pdfs/
Deploying_Multi_Factor_Index_Allocations_in_Institutional_Portfolios.pdf, 2013b.
G. Bernhart, S. Hoecht, M. Neugebauer, M. Neumann, and R. Zagst. Asset correlations in turbulent
markets and their implications on asset management. Asia Pacic Journal of Operations Research,
28:123, 2011.

V. Bhansali. Tail risk hedging: Creating Robust Portfolios for Volatile Markets. McGraw-Hill, 2014.

Vineer Bhansali. Beyond risk parity. The Journal of Investing, 20(1), 2011.

Vineer Bhansali, Josh Davis, Graham Rennison, Jason C. Hsu, and Feifei Li. The risk in risk parity:
A factor based analysis of asset based risk parity. Journal of Investing, 21(3), 2012.

David Blitz and Wilma de Groot. Strategic allocation to commodity factor premiums. The Journal of
Alternative Investments, 17(2):103115, 2014.

73
D. Bloch. A practical guide to quantitative portfolio trading. Available at SSRN: http://ssrn.com/
abstract=2543802, 2014.

L. Bloechlinger. Power prices A regime switching spot/forward model with Kim lter estimation. PhD
thesis, Univ of St Gallen, 2008.

Kris Boudt and Benedict Peeters. Asset allocation with risk factors. Available at http://www.finvex.
com/images/pdf/finvex_wp_5%20final.pdf, 2013.

Kris Boudt, Wanbo Lu, and Benedict Peeters. Asset allocation with higher moments and factor models.
Available at http://www.finvex.com/images/pdf/finvex_wp%206%20final.pdf, 2014.

R.G. Brown. Dieharder: A Random Number Test Suite. Available at http://www.phy.duke.edu/


~rgb/General/dieharder.php, 2014.

Catherine Bruneau, Alexis Flageolle, and Zhun Peng. Risk factors, copula dependence and large size
http://www.iae-aix.com/affi2014/docs/phd/risk_factors_
portfolio management. Available at
copula_dependence_and_large_size_portfolio_management.pdf, 2014.
J. Bulla, S. Mergner, I. Bulla, A. Sesboue, and C. Chesneau. Markov-switching asset allocation: Do
prolestable strategies exist? J. Asset Management, 12(5):310321, 2011.

M Burger, B. Graeber, and G. Schindlmayr. Managing energy risk: an integrated view on power and
other energy markets. Wiley, 2008.

Mark M. Carhart. On persistence in mutual fund performance. Journal of Finance, 52:5782, 1997.

Tianie Carli, Romain Deguest, and Lionel Martellini. Improved risk reporting with factor-based
diversication measures. Technical report, EDHEC-Risk Institute, 2014.

Zelia Cazalet and Thierry Roncalli. Facts and fantasies about factor investing. Available at http:
//ssrn.com/abstract=2524547, 2014.

Wing Cheung. From Factor Ranking to the ABL Framework . Available at SSRN: http://ssrn.com/
abstract=1457025, 2010.

Wing Cheung. The Augmented Black-Litterman Model: A Ranking-Free Approach to Factor-Based


Portfolio Construction and Beyond. Quantitative Finance, 13(2), 2013.

Wing Cheung and Mayank Mishra. Crowded trades: A bayesian remedy for factor-based quants.
Available at SSRN: http://ssrn.com/abstract=1457026, 2010.

Wing Cheung and Mayank Mishra. Eective factor management with analytical precision. Available
at SSRN: http://ssrn.com/abstract=1677020, 2012.

Wing Cheung and Nikhil Mittal. Ecient bayesian factor mimicking: Methodology, tests and compar-
ison. Available at SSRN: http://ssrn.com/abstract=1457022, 2009.

John H. Cochrane. Discount rates. Available at http://www.nber.org/papers/w16972, 2011.

Gregory Connor and Robert A. Korajczyk. Factor models in portfolio and asset pricing theory. In
John Guerard, editor, Handbook of portfolio construction: contemporary applications of Markowitz
techniques, pages 401418. Springer, 2010.

M. Correia, S. Richardson, and I. Tuna. Value investing in credit markets. Review of Accounting
Studies, 17(3):572609, 2012.

74
Stacy Cue and Lisa R. Goldberg. Allocating assets in climates of extreme risk. Available at SSRN:
http://ssrn.com/abstract=1823630, 2011.

S. Darolles and M. Vaissie. The benets of dynamic risk management: Mitigating downside risk with-
out compromising long-term growth prospects. Available at SSRN: http://ssrn.com/abstract=
1947080, 2011.

F. De Jong and J. Driessen. Liquidity risk premia in corporate bond markets. Quarterly Journal of
Finance, 2(2):134, 2012.

V. DeMiguel, L. Garlappi, F. J. Nogales, and R. Uppal. A generalized approach to portfolio optimiza-


tion: Improving performance by constraining portfolio norms. Management Science, 55(5):798812,
2009a.

V. DeMiguel, L. Garlappi, and R. Uppal. Optimal versus Naive Diversication: how Inecient is the
1/N Portfolio Strategy. Review of Financial Studies, 22:19151953, 2009b.

A. Denev. Coherent asset allocation and diversication in the presence of stress events. Master's thesis,
University of Oxford, 2011.

J. A. Doornik. An improved ziggurat method to generate normal random samples. Available at


http://www.doornik.com/research/ziggurat.pdf, 2005.

Frederick E. Dopfel and Sunder R. Ramkumar. Managed Volatility Strategies: Applications to Invest-
ment Policy. The Journal of Portfolio Management, 40(1):2738, 2013.

D. Duy and J. Kienitz. Monte Carlo Frameworks: Building Customisable High-performance C++
Applications. Wiley, 2009.

C.B. Erb and C.R. Harvey. The strategic and tactical value of commodity futures. Financial Analysts
Journal, 62(2):6997, 2006.

C. Erlwein. Applications of hidden Markov models in nancial modeling. PhD thesis, Brunel University,
2008.

C. Ernst, M. Grossmann, M Hoecht, S. Minden, M. Scherer, and R. Zagst. Portfolio selection under
changing market conditions. International Journal of Financial Services Management, 4:4863, 2009.

Marco Escobar, Michael Mitterreiter, David Saunders, Luis Seco, and Rudi Zagst. Market crises and
the 1/N asset-allocation strategy. The Journal of Investment Strategies, 2(4):83107, 2013.

Eugene Fama and Kenneth R. French. Common risk factors in the returns on stocks and bonds.
Journal of Financial Economics, 1992.

Eugene Fama and Kenneth R. French. Size, value, and momentum in international stock returns.
Journal of Financial Economics, 105(3):457472, 2012.

Eugene F. Fama and Kenneth R. French. A ve-factor asset pricing model. Available at SSRN:
http://ssrn.com/abstract=2287202, 2014.

J. Fan, Y. Fan, and J. Lv. Large dimensional covariance matrix estimation via a factor model. Journal
of Econometrics, 147:186197, 2008.

Jianqing Fan, Yingying Fan, and Jinchi Lv. Large dimensional covariance matrix estimation via a
factor model. Available at SSRN: http://ssrn.com/abstract=957599.

75
Jianqing Fan, Yuan Liao, and Martina Mincheva. High-dimensional covariance matrix estimation in
approximate factor models. The Annals of Statistics, 29(6), 2011.

Wayne Ferson and Yong Chen. How many good and bad fund managers are there, really.
Availableat https://wpcarey.asu.edu/sites/default/files/uploads/department-finance/
wayne-fersonseminarapril-18-2014.pdf, 2014.
A.V. Fiacco and Y Ishizuka. Sensitivity and stability analysis for nonlinear programming. Ann. Oper.
Res., 1990.

Edwin Fischer and Immanuel Seidl. Portfolio Optimization: A Combined Regime-Switching and
Black-Litterman Model. Available at SSRN: http://ssrn.com/abstract=2205190, 2013.

Mark D. Flood and George Korenko. Systematic scenario selection: Stress testing and the nature of
uncertainty. Available at SSRN: http://ssrn.com/abstract=1262896, 2013.

Gabriel Frahm, Tobias Wickern, and Christof Wiechers. Multiple tests for the performance of dierent
investment strategies. AStA Advances in Statistical Analysis, 96(3):343383, 2012.

A. Frazzini and L.H. Pedersen. Betting against beta. Journal of Financial Economics, 111:125, 2014.

A. Frazzini, I. Ronen, and T.J. Moskowitz. Trading costs of asset pricing anomalies. Available at
SSRN: http://ssrn.com/abstract=2294498, 2013.

Roza Galeeva, Jiri Hoogland, and Alexander Eydeland. Measuring correlation risk. Available at
http://www.bbk.ac.uk/cfc/pdfs/conference%20papers/Wed/Correlation_final_2.pdf, 2007.

Roza Galeeva, Jiri Hoogland, and Alexander Eydeland. Measuring correlation risk for energy deriva-
tives. In Global Derivatives, 2011.

Georgios-Xanthakis Gkatzilakis and Sivagaminathan Sivasubramanian. Active allocation of smart beta


indices based on factor timing and regime switching. Master's thesis, EDHEC Risk Institute, 2014.

Morton Glantz and Robert Kissell. Multi-Asset Risk Modeling Techniques for a Global Economy in an
Electronic and Algorithmic Trading Era. Academic Press, 2014.

P. Glasserman. Monte Carlo methods in nancial engineering. Springer, 2003.

Boris Gnedenko and Igor Yelnik. Reconciling factor optimization with portfolio constraints. Available
at SSRN: http://ssrn.com/abstract=2446090, 2014.

Richard Gold. Modeling public and private real estate risk. Available at http://www.northinfo.com/
documents/588.pdf, 2014.

N. Gonzalez and A. Thabault. Overview of diversication strategy indices. Available at


http://docs.edhec-risk.com/ERI-Days-North-America-2013/documents/SciBeta_Overview_
of_Diversification_Strategy_Indices.pdf, 2013.
G.B. Gorton and K.G. Rouwenhorst. Facts and fantasies about commodity futures. Financial Analysts
Journal, 62:8693, 2006.

G.B. Gorton, F. Hayashi, and K.G. Rouwenhorst. The fundamentals of commodity futures returns.
Review of Finance, 17(1):35105, 2013.

J. Green, J. Hand, and F. Zhang. The supraview of return predictive signals. Review of Accounting
Studies, 18(3):692730, 2013.

76
J. Green, J. Hand, and F. Zhang. The Remarkable Multidimensionality in the Cross-Section of Ex-
pected U.S. Stock Returns. Available at SSRN: http://ssrn.com/abstract=2262374, 2014.

W.H. Greene. Econometric Analysis (7th Edition). Prentice, 2011.

K. Grobys. Active portfolio management in the presence of regime switching: what are the benets of
defensive asset allocation strategies if the investor faces bear markets? The Review of Finance and
Banking, 4(1):1531, 2012.

Massimo Guidolin. Markov Switching in Portfolio Choice and Asset Pricing Models: A Survey. In
Advances in Econometrics, pages 87178. Emerald Group Publishing, 2011.

Massimo Guidolin. Markov switching models in empirical nance. Available at https://ideas.repec.


org/p/igi/igierp/415.html, 2012.

Massimo Guidolin and Allan G Timmermann. Asset allocation under multivariate regime switching.
Available at SSRN: http://ssrn.com/abstract=940652, 2006.

H. Gulen, Y. Xing, and L. Zhang. Value versus growth: Time-varying expected stock returns. Financial
Management, 40:381407, 2011.

R. Hafner. Stochastic Implied Volatility: A Factor-Based Model. Springer, 2004.

C.R. Harvey. Backtesting. In Inquire Europe Spring Seminar, Vienna, 2014.

C.R. Harvey and Y. Liu. Multiple testing in economics. Available at SSRN: http://ssrn.com/
abstract=2358214, 2013.

C.R. Harvey and Y. Liu. Backtesting. Available at SSRN: http://ssrn.com/abstract=2345489,


2014a.

C.R. Harvey and Y. Liu. Evaluating trading strategies. The Journal of Portfolio Management, 40(5):
108118, 2014b.

C.R. Harvey, Y. Liu, and H. Zhu. ... and the Cross-Section of Expected Returns. Available at SSRN:
http://ssrn.com/abstract=2249314, 2014.

J. Hauptmann, A. Hoppenkamps, A. Min, and R. Zagst. Forecasting market turbulences using regime-
switching models. Financial Markets and Portfolio Management, 28(2):139164, 2014.

M.K. Hess. Timing and diversication: A state-dependent asset allocation approach. The European
Journal of Finance, 12(3):189204, 2006.

Alexandre Hocquard, Sunny Ng, and Nicolas Papageorgiou. A constant-volatility framework for man-
aging tail risk. The Journal of Portfolio Management, 39(2):2840, 2013.

C. Homescu. Many risks, one (optimal) portfolio. Available at SSRN: http://ssrn.com/abstract=


2473776, 2014a.

C. Homescu. Tail risk protection in asset management. Available at SSRN: http://ssrn.com/


abstract=2524483, 2014b.

Patrick Houweling and Jeroen van Zundert. Factor investing in the corporate bond market. Available
at SSRN: http://ssrn.com/abstract=2516322, 2014.

Joop Huij, Simon D. Lansdorp, David Blitz, and Pim van Vliet. Factor investing: Long-only versus
long-short. Available at SSRN: http://ssrn.com/abstract=2417221, 2014.

77
A. Ilmanen. Expected Returns: An investor guide to harvesting market rewards. Wiley Finance, 2011.

A. Ilmanen, R. Byrne, H. Gunasekera, and R. Minikin. Which risks have been best rewarded? The
Journal of Portfolio Management, 30(2):5357, 2004.

Antti Ilmanen and Jared Kizer. The death of diversication has been greatly exaggerated. The Journal
of Portfolio Management, 38:1527, 2012.

J. Ioannidis. Why most published research ndings are false. PLoS Medicine, 2005.

P. Jackel. Monte Carlo Methods in Finance. Wiley Finance, 2002.

Bruce I. Jacobs and Kenneth N. Levy. Disentangling equity return regularities: New insights and
investment opportunities. Financial Analysts Journal, 44(3):1843, 1988.

Bruce I. Jacobs and Kenneth N. Levy. Investing in a multidimensional market. Financial Analysts
Journal, 70(6), 2014.

E. Jay, P. Duvaut, S. Darolles, and A. Chretien. Multi-factor models and signal processing techniques:
Survey and examples. Available at SSRN: http://ssrn.com/abstract=1837315, 2011.

Yindeng Jiang and Doug Martin. Better Risk and Performance Estimates with Factor Model Monte
Carlo. Available at SSRN: http://ssrn.com/abstract=2295602, 2013.

G. Jostova, S. Nikolova, A. Philipov, and C.W. Stahel. Momentum in corporate bond returns. Review
of Financial Studies, 26(7):16491693, 2013.

Xiaowei Kang and Danie Ung. Practical considerations for factor-based asset allocation. Available at
SSRN: Available at SSRN: http://ssrn.com/abstract=2457228, 2014.

C Kim and C. Nelson. State space models with regime switching. MIT Press, 1999.

Joachim Klement. Investment management is risk management - nothing more and nothing less.
Available at SSRN: http://ssrn.com/abstract=2086214, 2011.

P.E. Kloeden and E. Platen. Numerical Solution of Stochastic Dierential Equations. Springer, 2010.

P.E. Kloeden, E. Platen, and H. Schurz. Numerical Solution of Stochastic Dierential Equations
Through Computer Experiments. Springer, 2003.

C.G. Koedijk, A.M.H. Slager, and P.A. Stork. Factor investing in practice: A trustees' guide to
implementation. Available at SSRN: http://ssrn.com/abstract=2391694, 2014.

R. Koijen, T. Moskowitz, L. Pedersen, and E. Vrugt. Carry. Available at http://www.nber.org/


papers/w19325, 2013.

R. Korn, E. Korn, and G. Kroisandt. Monte Carlo Methods and Models in Finance and Insurance.
CRC Press, 2010.

S.P. Kothari, Konstantin Danilov, and Gitanjali Meher Swamy. Generating superior performance in
private equity: A new investment methodology. Available at SSRN: http://ssrn.com/abstract=
2221203, 2012a.

S.P. Kothari, Konstantin Danilov, and Gitanjali Meher Swamy. Generating superior performance in
private equity: A new investment methodology. Journal of Investment Management, 11(3), 2012b.

78
Mark Kritzman, Sebastien Page, and David Turkington. Regime shifts: Implications for dynamic
strategies. Financial Analysts Journal, 68(3), 2012.

P. L'Ecuyer. TESTU01: empirical statistical testing of uniform random number generators. Available
at http://www.iro.umontreal.ca/~simardr/testu01/tu01.html, 2014.

P. L'Ecuyer and R. Simard. TestU01: A C Library for Empirical Testing of Random Number Gener-
ators. ACM Transactions on Mathematical Software, 33(22), 2007.

Pierre L'Ecuyer. Good parameter sets for combined multiple recursive random number generators.
Operations Research, 47(1):159164, 1999.

R. Leote de Carvalho, X. Lu, and P. Moulin. Demystifying equity risk-based strategies: A simple alpha
plus beta description. Journal of Portfolio Management, 38:5670, 2012.

Jonathan Lewellen. The cross section of expected stock returns. Critical Finance Review, 2015.

John Lewis. Relative strength and portfolio management. Available at SSRN: http://ssrn.com/
abstract=1998935, 2011.

H. Lin, J. Wang, and C. Wu. Liquidity risk and expected corporate bond returns. Journal of Financial
Economics, 99(3):628650, 2011.

J. Liu and Z. Chen. Regime-dependent robust risk measures with application in portfolio selection. In
2nd International Conference on Information Technology and Quantitative Management, 2014.

Marcos Lopez de Prado. The sharp razor: Deating the sharpe ratio by asking for a minimum track
record length. Available at SSRN: http://ssrn.com/abstract=2150879, 2014.

R. D. Martin. Robust covariances and distances: Common risk factor versus idiosyncratic outliers.
www.rinfinance.com/agenda/2013/talk/DougMartin.pdf, 2013.

Roger Masi. Helping Clients Hedge Market Risk: Four Important Considerations. Available at http:
//www.advisorperspectives.com/commentaries/macro_060414.php, 2014.

M. Matsumoto. Mersenne Twister Home Page. Available at http://www.math.sci.hiroshima-u.ac.


jp/~m-mat/MT/emt.html, 2014.

L. Menkho, L. Sarno, M. Schmeling, and A. Schrimpf. Currency momentum strategies. Journal of


Financial Economics, 106(3):660684, 2012.

A. Meucci, A. Santangelo, and R. Deguest. Measuring portfolio diversication based on optimized


uncorrelated factors. Available at SSRN: http://ssrn.com/abstract=2276632, 2014a.

Attilio Meucci. Fully exible views: Theory and practice. Risk, 21(10):97102, 2008.

Attilio Meucci. Enhancing the black-litterman and related approaches: Views and stress-test on risk
factors. Available at SSRN: http://ssrn.com/abstract=1213323, 2010.

Attilio Meucci. Stress-testing with fully exible causal inputs. Risk, 25(4):6165, 2012.

Attilio Meucci. Estimation and stress-testing via time- and market-conditional exible probabilities.
Available at SSRN: http://ssrn.com/abstract=2312126, 2014.

Attilio Meucci, David Ardia, and Simon Keel. Fully exible extreme views. Risk, 14(2):3949, 2012.

79
Attilio Meucci, David Ardia, and Marcello Colasante. Portfolio construction and systematic trading
with factor entropy pooling. Available at SSRN: http://ssrn.com/abstract=1742559, 2014b.

J. Mire, A.M Fuertes, and A. Fernandez-Perez. Commodity futures returns and idiosyn-
Available at SSRN: http://www.edhec-risk.com/edhec_publications/all_
cratic volatility.
publications/RISKReview.2013-06-03.4348/attachments/EDHEC_Working_Paper_Commodity_
Futures_Returns_F.pdf, 2012.
S. Minami. Optimal factor composition under sampling error of parameters - diversication to factor
portfolios. Available at SSRN: http://ssrn.com/abstract=2285878, 2013a.

S. Minami. Underestimation bias of risk on optimized portfolio by multifactor risk model - risk of long
short portfolio can be underestimated. Available at SSRN: http://ssrn.com/abstract=2254768,
2013b.

S. Mitra. Regime switching volatility calibration by the Baum-Welch method. Available at Arxiv:
http://arxiv.org/abs/0904.1500, 2009.

Paul Mouchakkaa. Real estate portfolio risk management and monitoring. Available at http://www.
morganstanley.com/realestate/Observatory.pdf, 2012.

Johnathan Mun. Modeling Risk: Applying Monte Carlo Risk Simulation, Strategic Real Options,
Stochastic Forecasting, and Portfolio Optimization. Wiley, 2010.

R. Novy-Marx. Predicting anomaly performance with politics, the weather, global warming, sunspots,
and the stars, journal of nancial economics. Journal of Financial Economics, 112(2):137146, 2014.

Peter Nystrup. Regime-based asset allocation: Do protable strategies exist? Master's thesis, Technical
University of Denmark, 2014.

J. Okunev and D. White. Do momentum-based strategies still work inforeign currency markets?
Journal of Financial and Quantitative Analysis, 38(2):425448, 2003.

Dessislava Pachamanova and Frank J. Fabozzi. Simulation and Optimization in Finance: Modeling
with MATLAB, @Risk, or VBA. Wiley, 2010.

Sebastien Page. The myth of diversication: risk factors versus asset classes. Available at http:
//media.pimco.com/Documents/Myth%20of%20Diversification.pdf, 2010.

Sebastien Page and Mark A. Taborsky. The myth of diversication: risk factors versus asset classes.
The Journal Of Portfolio Management, 37(4):12, 2011.

Jochen Papenbrock and Peter Schwendner. Handling risk on/risk o dynamics with correlation regimes
and correlation networks. Available at SSRN: http://ssrn.com/abstract=2254272, 2013.

Scott N. Pappas, Robert J. Bianchi, Michael E. Drew, and Rakesh Gupta. Risk-factor diversication
and portfolio selection. Available at SSRN: http://ssrn.com/abstract=2136827, 2012.

Magnus Erik Hvass Pedersen. Portfolio Optimization and Monte Carlo Simulation. Available at SSRN:
http://ssrn.com/abstract=2438121, 2014.

Romain Perchet, Raul Leote de Carvalho, and Pierre Moulin. Inter-temporal risk parity: a constant
volatility framework for factor investing. ssrn.com/abstract=2440654, 2014.

Hans Pirnay, Rodrigo Lopez-Negrete, and Lorenz T. Biegler. Optimal sensitivity based on IPOPT.
Math. Prog. Comp., 4:307331, 2012.

80
E. Platen. Numerical solution of stochastic dierential equations with jumps in nance. In MCQMC
2012, Sudney, 2012.

Eckhard Platen and Nicola Bruti-Liberati. Numerical Solution of Stochastic Dierential Equations
With Jumps in Finance. Springer, 2010.

Eugene L. Podkaminer. Risk factors as building blocks for portfolio diversication: The chemistry of
asset allocation. Available at http://www.cfainstitute.org/learning/products/publications/
irpn/Pages/irpn.v2013.n1.1.aspx, 2013.
L. Pospisil and J. Zhang. Momentum and reversal eects in corporate bond prices and credit cycles.
The Journal of Fixed Income, 20(2):101115, 2010.

http://www.russell.
Y Prasetyo and K Lo. Currency strategies and bond portfolios. Available at
com/documents/indexes/research/currency-strategies-and-bond-portfolios.pdf, 2014.
K. Pukthuanthong and R. Roll. A protocol for factor identication. Available at SSRN: http://ssrn.
com/abstract=2517944, 2014.

Riccardo Rebonato and Alexander Denev. Coherent asset allocation and diversication in the presence
of stress events. Journal of Investment Management, 2012.

Mike Reust. One massive monte carlo, one very ecient solu-
tion. Available https://www.betterment.com/blog/2013/11/25/
at
one-massive-monte-carlo-simulation-one-very-efficient-solution/, 2013.
Y. Romahi and K. Santiago. Diversication - Is it still the only Free Lunch? Alternative building
blocks for risk parity portfolios. Technical report, JP Morgan Asset Management Research Notes,
2012.

D. Roman, G. Mitra, and N. Spagnolo. Hidden Markov models for nancial optimization problems.
IMA J Manag Math, 21:111129, 2010.

J. P. Romano, A. M. Shaikh, and M. Wolf. Formalized data snooping based on generalized error rates.
Econometric Theory, 24:404447, 2008.

Thierry Roncalli and Guillaume Weisang. Risk parity portfolios with risk factors. Available at SSRN:
http://ssrn.com/abstract=2155159, 2012.

D. Rosen and D. Saunders. Risk factor contributions in portfolio credit risk models. Journal of Banking
& Finance, 34:336349, 2010.

A. Saltelli, S. Tarantola, F. Campolongo, and Ratto M. Sensitivity Analysis in Practice:. Wiley, 2004.

A Saltelli, M Ratto, T Andres, F Campolongo, J Cariboni, D Gatelli, M Saisana, and S Tarantola.


Global Sensitivity Analysis: The Primer. Wiley, 2008.

David Saunders, Luis Seco, Christofer Vogt, and Rudi Zagst. A fund of hedge funds under regime
switching. The Journal of Alternative Investments, 15(4):823, 2013.

Akhtar Siddique and Iftekhar Hasan, editors. Stress Testing: Approaches, Methods and Applications,
2012. Risk Books.

C. Stivers and L. Sun. Cross-sectional return dispersion and time variation in value and momentum
premiums. Journal of Financial and Quantitative Analysis, 45:9871014, 2010.

81
Avanidhar Subrahmanyam. The cross-section of expected stock returns: What have we learnt from
the past twenty-ve years of research? European Financial Management, 16:2742, 2010.

C. Szylar. Handbook of market risk. Wiley, 2014.

Nassim N. Taleb, Elie Canetti, Tidiane Kinda, Elena Loukoianova, and Christian Schmieder. A New
Heuristic Measure of Fragility and Tail Risks: Application to Stress Testing. Available at http:
//www.imf.org/external/pubs/ft/wp/2012/wp12216.pdf, 2012.

Timo Teuber. Dynamic risk parity - a smart way to manage risks. Technical report, Allianz Global
Investors, 2013.

J. Tu. Is regime switching in stock returns important in portfolio decisions? Management Science, 56:
11981215, 2010.

Huy Thanh Vo and Raimond Maurer. Dynamic asset allocation under regime switching, predictability
and parameter uncertainty. Available at SSRN: http://ssrn.com/abstract=2165029, 2013.

Peng Wang, Rodney N Sullivan, and Yizhi Ge. Risk-based dynamic asset allocation with extreme tails
and correlations. Journal of Portfolio Management, 38(4), 2012.

A. Werley and J. Speer. Factor Risk Management: A Generalized Methodology for Multi-asset Class
Portfolios. Technical report, J.P. Morgan Asset Management, 2011.

Anthony Werley. Factor risk management: A generalized methodology for multi-asset class portfolios.
Technical report, JP Morgan Endowments & Foundations Group, 2011.

R. Weron. Modeling and Forecasting Electricity Loads and Prices: A Statistical Approach. Wiley, 2006.

R. Weron. Heavy-tails and regime-switching in electricity prices. Mathematical Methods of Operations


Research, 69(3), 2009.

Peter Williams. Factor based approaches to risk parity. Available at SSRN: http://ssrn.com/
abstract=2398202, 2014.

Karen Witham. Risk factor investing, revealed: building balanced exposure to rewarded
https://www.blackrock.com/institutions/en-us/literature/publication/
risk. Available at
blk-risk-factor-investing-revealed.pdf, 2012.
J. H. Witte, D. Ples, and J. Corominas. The hidden risk factor. Journal of Mathematical Finance, 3:
2126, 2013.

Ma Y., MacLean L.and Xu K., and Zhao YG. Portfolio optimization model with regime-switching risk
factors for sector exchange traded funds. Pac J Opt, 7:455470, 2011.

Valeriy Zakamulin. The role of covariance matrix forecasting method in the performance of minimum-
variance portfolios. Available at SSRN: http://ssrn.com/abstract=2411493, 2014.

E. Zivot. Factor model based risk measurement and management. Available at www.rinfinance.com/
agenda/2011/EricZivot.pptx, 2011.

82
Appendix A: Minimum number of observations
We present minimum number of observations needed for various computations.

For meaningful estimation of covariance matrix


If a portfolio consists of N assets, the covariance matrix will be N × N, with N (N +1)/2 unique terms.
In order to estimate these total parameters, we need a large enough set of data observations to ensure
that the number of degrees of freedom is at least positive. Moreover, we are seeking to estimate the
value of each parameter, and a general rule of thumb is that there need to be at least 20 observations
for each parameter to have statistically meaningful results.
Assuming that we have NT observations for each of the N assets, after some algebra we obtain that
N T ≥ 10 × (N + 1), a number which is much larger than what is usually available for vast majority
of assets, as illustrated by next gure (from Glantz and Kissell [2014]).

Figure .1: Data Requirements for Constructing a Statistically Signicant Covariance Matrix

For meaningful Sharpe ratio


Bailey and Lopez de Prado [2012] derived an estimate of the Minimum Track Record Length (MinTRL)
How long should a track record be in order to have statistical
needed to answer the question
confidence that its Sharpe ratio is above a given threshold?,with MinTRL developed to
evaluate a strategy's track record (a single realized path).

For a given designated Sharpe ratio threshold SR


f and signicance level α the formula for MinTRL

83
is   2
 
f = 1 + 1 − γ3 SR γ 4 − 1 2 Z α
M inT RL α, SR c + SR
c (A.1)
4 c − SR
SR f

with SR
c the observed Sharpe ratio, Zα is the normal CDF corresponding to α,γ3 the skewness and
γ4 the kurtosis.
Lopez de Prado [2014] indicated that the longer track record must be:

• the smaller SR
c is

• the more negatively skewed returns are

• the greater the fat tails

• the greater our required level of condence.

Formula for MinTRL has some direct practical implications

1. if a track record is shorter than MinTRL, we do not have enough condence that the observed
Sharpe ratio is above the designated threshold

2. while a portfolio manager will be penalized because of her non-Normal returns, she can regain
the investor's condence over time (by extending the length of her track record).

It is important to note that MinTRL is expressed in terms of number of observations, not annual or
calendar terms. One should have in mind that the methodology relies upon equations which apply
to an asymptotic distribution, and Central Limit Theorem (CLT) is typically assumed to hold for
samples in excess of 30 observations. So even though a MinTRL may demand less than 2.5 years of
monthly data, or 0.5769 years of weekly data, or 0.119 years of daily data, etc. the moments inputted
in Eq. (A.1) must be computed on longer series for CLT to hold. This is consistent with practitioners'
standard practice of requiring similar lengths during the due diligence process.

For meaningful backtesting


The analogue to MinTRL in the context of overtting, termed Minimum Backtest Length (MinBTL),
was dened in Bailey et al. [2014a] as backtest length needed to avoid selecting a skill-less strategy
among M alternative strategies. In other words, it is the minimum period to prevent overtting when
comparing multiple strategies.
More specically, the Minimum Backtest Length (MinBTL, in years) needed to avoid selecting a
strategy with an InSample(IS) Sharpe ratio (denoted SRIS ) among M independent strategies with
an expected OutOf Sample(OOS) Sharpe ratio of zero is
 h h ii 
(1 − γ) Z −1 1 − M
1
+ γZ −1 1 − M ·exp(−1)
1
M inBT L ≈   < 2 ln M
SR IS
(SRIS )2

where Z is the standard normal CDF and γ ≈ 0.5772156649... is the Euler-Mascheroni constant.

84
Appendix B: Regression analysis to compute factor model parameters
The ordinary least squares (OLS) regression analysis to compute factor model parameters {bj }j=1···M
is:
M
X
rt = α0 + fjt bj + et
j=1

where rt is the asset return in period t, α0 is a constant term, fkt is the value of k th factor in period
t, bk is exposure to k th factor and et is noise of asset in period t (return not explained by the model).
Such an OLS regression analysis is required to have the following properties (see Greene [2011],
Glantz and Kissell [2014]) in order to be considered statistically correct

1. Error term has a mean of zero: E(et ) = 0

2. Variance of the error term for each asset is σe2

3. Variance of each factor is σf2k

4. Error term (residual) and each factor are independent E (efk ) = 0

5. Explanatory factors are independent E (fjt , fkt ) = 0

6. Error terms are independent for all lagged time periods E (et et−j ) = 0, i.e., no serial correlation
or correlation of any lags across the error terms.

7. Error terms across all stocks are independent, E (eit ejt ) = 0, i.e., the series of all error terms are
uncorrelated.

While properties 1-3 are true for a properly specied regression model, the other prop-
erties have to be tested. For property 5, if the factors are not independent, they can be
transformed into a new set of factors that are linearly independent.

85
Appendix C: Estimating covariance using a factor model
For number of time periods P and number of factors M, we outline procedure described in Glantz and
Kissell [2014] for a factor model (on return of ith asset) that is written in matrix notation as

ri = αi + F bi + ei

where
         
ri1 αi1 f11 f21 ··· fM 1 bi1 ei1
 ri2   αi2   f12 f22 ··· fM 2   bi2
 ei =  ei2 
  
ri = 
 · · ·  αi =  · · ·
F =  b = 
···  i  ···
 
  ··· ··· ···   ··· 
riP αiP f1P f2P ··· fM P biM eiP

A factor model across a universe of N assets can be written as

R = α + Fβ + e

where
     
r11 r21 ··· rN 1 α11 α21 ··· αN 1 f11 f21 ··· fM 1
 r12 r22 ··· rN 2   α12 α22 ··· αN 2   f12 f22 ··· fM 2 
R=
 ···
α =  F =  
··· ··· ···   ··· ··· ··· ···   ··· ··· ··· ··· 
r1P r2P ··· rN P α1P α2P ··· αN P f1P f2P ··· fM P
   
b11 b21 ··· bN 1 e11 e21 ··· eN 1
 b12 b22 ··· bN 2    e12 e22 ··· eN 2 
β=  ··· e= 
··· ··· ···   ··· ··· ··· ··· 
b1M b2M ··· bN M e1P e2P ··· eN P

The expected value of returns and, respectively, the excess returns, are given by

E[R] = α + F̄ β

R − E[R] = F − F̄ β + e

The covariance matrix C of returns is computed as


h i h T i
C = E (R − E[R])T (R − E[R]) = E
 
F − F̄ β + e F − F̄ β + e
h 2 i h 2 i
= E β T F − F̄ β + 2β T F − F̄ e + eT e] = β T E F − F̄ β + 2β T E F − F̄ e + E eT e
     

Using the properties in Appendix 8 we have that

βT E
  
F − F̄ e = 0
2
 
σe1 0 ··· 0
 0 σe22 ··· 0 
E eT e =
 

 ··· ···
 =Λ
··· ··· 
0 0 ··· 2
σeN
σf21 cov (f1 , f2 ) ··· cov (f1 , fM )
 
h 2 i  cov (f2 , f1 ) σf22 ··· cov (f2 , fM ) 
E F − F̄ =   = cov (F )
 ··· ··· ··· ··· 
cov (fM , f1 ) cov (fM , f2 ) ··· 2
σf M

86
with Λ the idiosyncratic variance matrix with each diagonal term given by variance of the regression
term for each asset, and cov (F ) the factor covariance matrix, which is a diagonal matrix in majority
of practical situations.
Thus covariance matrix derived from the factor model has the expression

C = β T cov(F )β + Λ

87
Appendix D: Parameter estimation for regime switching
Parameter estimation can be performed through ltering, which combines maximum likelihood method
with a lter or other algorithms, such as

• EM (expectation maximization) algorithm Erlwein [2008], Weron [2009, 2006]

• Kim Filter Weron [2006], Bloechlinger [2008]

• Baum Welch algorithm Mitra [2009]

• Hamilton lter Burger et al. [2008]

Our recommendation is to use the Kim lter, described below, which combines the Kalman lter,
the Hamilton lter and a collapsing procedure. It is an optimal estimator in the sense that no other
estimator based on a linear function of the information set yields a smaller mean square error. The
reader to chapter 5 of Erlwein [2008] and to Kim and Nelson [1999] for additional details.

Kalman lter
Since the Kalman lter is the main ingredient of the Kim lter, we start with its description.
The Kalman lter is an ecient estimator for a state vector of a linear dynamic system perturbed by
Gaussian white noise, using measurements that are linear functions of the system state but corrupted
by additive Gaussian white noise.The model has to be expressed in a state-space form characterized
by two equations:

• transition equation: describes dynamics of state variables which do not have empirical data

• measurement equation: represents the relationship linking the observable variables with the
non-observable variables

Thus, the Kalman lter makes it possible to evaluate the non-observable variables through an iterative
process, and it updates their values at each iteration using the new information. It is essentially a set
of equations that implements a predictor-corrector type estimator for the unobserved state variables.
During the prediction step, the Kalman lter forms an optimal predictor of the unobserved state
variable, given all the information available up to previous time point, to compute prior estimates.
During the updating step, new information becomes available and is used to update the prior estimates.
The resulting values are called posterior estimates.

Kim lter
In a similar manner we construct for Kim lter prior and posterior estimates for the unobserved
state vector and its associated mean square error, based on historical observations and conditional on
the regime at current and previous time. The Kim lter has the following 5 major steps:

1. The prior estimates are calculated using the regime-specic transition equations

2. Compute prediction errors

3. Compute the Kalman gains corresponding to each error term

4. Update the prior estimates

5. The posterior estimates are collapsed by applying a Hamilton lter

88
Curse of dimensionality is handled by collapsing the posterior estimates, which incorporate the infor-
mation at previous observation time point through a weighted average.
The whole algorithm incorporates maximum likelihood in the following way. We start with a fea-
sible choice of the parameters and initial state vector, with given covariance matrix. The prediction
equations are used to give the prior estimates. Given the historical market information on the observ-
able variables, the prediction error and its corresponding mean square error are then determined and
the prior estimates are updated. When the last observation is reached, the log-likelihood is evaluated
and a new parameter set is chosen such that the log-likelihood value is maximized (using a numerical
optimizer). The algorithm stops when the dierence between values of log-likelihood function at two
consecutive iterations is smaller than specied tolerance.
The next gure (from Bloechlinger [2008]) illustrates the whole algorithm, although applied to a
Kalman lter (which has only 4 steps). The main dierence is given by the additional collapsing step,
but otherwise the algorithm is similar.

Figure .2: Kalman lter algorithm

89
Appendix E: Ecient sensitivity analysis
We describe several ecient approaches to perform sensitivity analysis.
Let us consider the general optimization problem

min F (x, p)
x

subject to constraints
R(x, p) = 0
We have denoted the N −dimensional vector of optimization variables by x and the M −dimensional
vector of parameters by p.
dF
We want to compute sensitivities of F with respect to parameters {pj }, namely
dpj :

N
dF X ∂F ∂xi ∂F
= +
dpj ∂xi ∂pj ∂pj
i=1

Enhancing a NLP barrier-based optimization algorithm


The basic sensitivity strategy for Nonlinear Programming NLP solvers is derived through application
of the implicit function theorem to the KarushKuhnTucker KKT conditions of a parametric NLP.
As shown in Fiacco and Ishizuka [1990], sensitivities can be obtained from a solution with suitable
regularity conditions, merely by solving a linearization of the KKT conditions. Nevertheless, relaxation
of some of these regularity conditions induces singularity and inconsistency in this linearized system,
and may make the sensitivity calculation much more dicult.
With improvements in fast Newton-based barrier methods, such as IPOPT, sensitivity information
can be obtained very easily. As a result, fast on-line algorithms can be constructed where more
expensive optimization calculations can be solved in advance, and fast updates to the optimum can be
made on-line. Such an approach, presented in Pirnay et al. [2012], is based on a x-relax strategy that
modies the KKT system through a Schur complement strategy

Adjoint approach
The adjoint method allows to compute sensitivities much faster, and is described next.
The starting point is that the constraint R(x, p) = 0 must hold for any set of parameters p with
corresponding x, which implies that its Taylor expansion also has to vanish.

dR ∂R dx ∂R
= + =0
dpk ∂x dpk ∂pk
dF
We multiply on the left by any complex vector v and subtract the product from
dpk , which does not
change its value:
 
dF ∂F dx ∂F † ∂R dx ∂R
= + −v +
dpk ∂x dpk ∂pk ∂x dpk ∂pk
 
∂F ∂R ∂F ∂R dx
= − v† + − v†
∂pk ∂pk ∂x ∂x dpk

where v† denotes the complex conjugate for v.


This expression is the key to the adjoint method: if we are able to nd a v such that the multiplying
dx
factor of
dpk vanishes, we do not have to perform the expensive N +1 calculations.

90
Thus the adjoint method consists of 2 equations:

∂F ∂R
= v†
∂x ∂x
dF ∂F ∂R
= − v†
dpk ∂pk ∂pk

91
Appendix F: Time dependency of factor-based returns, exposures and
correlations
We present evidence presented in the literature of time variation for factor-based returns, exposures
and correlations.
The rst example (from Amenc and Goltz [2014]) describes the temporal behavior of size, momentum
and value factors for US market from 1973 to 2013. The Market factor is the daily return of the cap-
weighted index of all stocks that constitute the index portfolio in excess of the risk-free rate. Small
size factor is the daily return series of a cap-weighted portfolio that is long the smallest 30% of stocks
(by market cap) and short the largest 30% of stocks (by market cap) of the extended universe (i.e.
including small caps). Value factor is the daily return series of a cap- weighted portfolio that is long the
highest 30% and short the lowest 30% of stocks by B/M ratio in the investable universe. Momentum
factor is the daily return series of a cap-weighted portfolio that is long the 30% highest and short the
30% lowest 52 weeks (minus most recent 4 weeks) past return stocks in the investable universe.

Figure .3: Temporal behavior of size, momentum and value factors (from Amenc and Goltz [2014])

Temporal behavior of Fama-French factors and respectively Barra Equity factors are shown next

92
Figure .4: Temporal behavior of Fama-French factors (from Bender et al. [2013a])

Figure .5: Temporal behavior of Barra Equity factors (from Bender et al. [2013a])

While factor indexes have exhibited excess risk- adjusted returns over long time periods, over short
horizons factors exhibit signicant cyclicality, including periods of underperformance. Next gure
shows that each of the MSCI factor indices has experienced at a minimum a consecutive two-to-three
year period of underperformance. Some factors historically have undergone even longer periods; the
Small Cap or Low Size factor (captured by the MSCI World Equal Weighted Index in the gure) went
through a six-year period of underperformance in the 1990s.

93
Figure .6: Temporal behavior of MSCI Factor Indices (from Bender et al. [2013a])

Factor correlations can be volatile and unstable. Historical return correlations may not be as mean-
ingful in the future because relationships between the dierent factors change over time.
Illustrative examples include rolling three-year correlations of four factors

94
Figure .7: Three-Year Rolling Correlation of Four Dierent Factor Combinations (from Kang and Ung
[2014]

95
Appendix G: Estimating carry factor
Procedures to estimate the carry factor are described in Koijen et al. [2013], and are based on decom-
posing the return to any security into two components: carry and price appreciation. The carry return
can be thought of as the return to a security assuming its price/market conditions stays constant.

Estimating carry from future contracts


Consider a futures contract that expires in period t+1 with a current futures price Ft and spot price
of the underlying security St . Assume an investor allocates Xt dollars today of capital to nance each
futures contract (where Xt must be at least as large as the margin requirement). Next period, the
f f
value of the margin capital and the futures contract is equal to Xt (1 + rt ) + Ft+1 − Ft , where rt is
the risk-free interest rate today that is earned on the margin capital. Hence, the return in excess of
the risk-free rate is
Ft+1 − Ft
rt+1 =
Xt
The carry Ct of the futures contract is then computed as the futures excess return under the as-
sumption of a constant spot price from t to t + 1 and utilizing the observation that futures price expires
at the future spot price, i.e., Ft+1 = St+1 . Thus we obtain

St − Ft
Ct =
Xt
The carry for a full collateralized position (amount of capital Xt allocated to position is equal to
futures price Ft ) is computed as
St − Ft
Ct =
Ft

96
Appendix H: Protocol for factor identication
Pukthuanthong and Roll [2014] introduced a protocol for determining which factor candidates are
related to risks and which candidates are related to mean returns. Factor candidates could be related
to both risk and returns, to neither, or to one but not the other.
The protocol is composed of six steps to check necessary conditions and one nal step that exam-
ines a sucient condition. It was considered that any fundamental necessary condition for a factor
candidate must be related to the principal components of a (conditional) covariance matrix of returns.
However, this necessary condition does not distinguish between pervasive priced factors, those with
risk premiums, and non-pervasive, fully diversiable factors, which are related to covariances of some
individual assets but do not inuence the aggregate portfolio. This distinction is provided by the
sucient condition.
The steps of the protocol are as follows:

1. collect a set of N equities for the factor candidates to explain

• The equities in this set should be as close as possible to real assets.

• They should mostly be rms with a minimal amount of debt and with few positions in
derivatives.

2. extract the real return eigenvectors corresponding to the L largest eigenvalues from simultaneous
returns for the N real assets over some calendar interval

3. collect a set of K factor candidates using J ETFs as heterogeneous as possible and any K−J
factor candidates of interest.

4. using the L eigenvectors from step 2. and the K factor candidates from step 3., calculate the
time varying conditional real return covariance matrix

5. from the conditional covariance matrix computed in step 4. for each time period extract following
submatrices

a) conditional cross-covariance

b) conditional covariance of factor candidates

c) conditional covariance of real eigenvectors

6. Using the submatrices from step 5., compute canonical correlations between the factor candidates
and the eigenvectors for each date

7. run a pooled cross-section/time series panel with real returns as dependent variables and factors
that satisfy the necessary conditions as the explanatory variables, taking account of correlations
across assets and time

97
Appendix I: Monte Carlo simulation
Monte Carlo (MC) methods can be relied upon to perform risk and stress analysis, generate scenarios
or obtain benchmarks for testing and comparison. MC simulation is simple, exible, widely applicable,
and allows for straightforward parallelization (on computer grids, on graphical cards GPUs, or on cloud
systems).
Relevant references for Monte Carlo include:

• for quantitative nance: Korn et al. [2010], Glasserman [2003], Duy and Kienitz [2009]

• for portfolio management: Pachamanova and Fabozzi [2010], Mun [2010]

• for simulation of SDEs: Kloeden et al. [2003], Kloeden and Platen [2010], Platen and Bruti-
Liberati [2010], Platen [2012]

Random number generators


In order to simulate random variables (RVs) we have to rely on good random number generators to
obtain uniform RVs.
The following properties dene a good generator Korn et al. [2010]:

• a long period  how long before sequence repeats itself

• various statistical tests to measure randomness :

 TestU01 L'Ecuyer [2014], L'Ecuyer and Simard [2007]

 DieHarder Brown [2014](integrates Diehard and NIST)

• fast skip-ahead for ecient parallel implementation

Here are various examples of generators:

• good:

 Mersenne Twister Matsumoto [2014]: period of 219937 − 1, slower skip-ahead

 mrg32k3a L'Ecuyer [1999]: period of 2191 , fast skip-ahead

• not-so-good:

 RAND from older versions of Excel

 Ran0 from Numerical Recipes

Simulation of random variables from a specied distribution


We need ability to draw independent realizations of random variables having the desired distribution.
Usually generation of RVs is split into 3 stages (shown below with N (0, 1) as desired distribution)

1. generation of independent uniform U (0, 1) RVs

2. conversion into independent Normal N (0, 1) RVs

3. conversion into correlated Normal N (0, 1) RVs

98
If desired distribution is dierent than N (0, 1) the procedure can be employed along similar lines as
above.
The second step of the algorithm requires a conversion from uniform RVs to RVs of the desired
distribution. There are various algorithms for this conversion, as described in Korn et al. [2010]:

• Box-Muller method

• Marsaglia's Ziggurat method (with Doornik enhancement Doornik [2005])

• Inverse CDF transformation Acceptance rejection

Conversion to correlated RVs of desired distribution can be done using the following algorithm

• Simulate N independent Z1 , · · · , ZN from N (0, 1)


T
• Set X = ΣD Z , where correlation matrix Σ = ΣD ΣD

• Set Ui = Φ(Xi ), with Φ standard normal CDF

• Find Yi = Pi−1 (Ui )

One would need to check that the correlation/covariance matrix Σ is actually a valid correlation matrix.
If it is not, applying nearest correlation procedure would yield such a matrix.

Simulation for diusion SDEs


In some situations the underlying may be assumed to have dynamics described by a stochastic dier-
ential equation (SDE):
dX(t) = a(X, t)dt + b(X, t)dW (I.1)

with W one or multi-dimensional Brownian process


The usual method of choice for simulating X is the Euler-Maruyama scheme:

X̂j+1 = X̂j + a(X̂j , tj )∆tj + b(X̂j , tj )∆Wj

where ∆tj = tj+1 − tj , X̂j denotes approximation to X(tj ) and ∆Wj i.i.d. in N (0, ∆tj ).
It is conditionally stable and its order of weak convergence is 1.

Simulation of jump-diusion process


The dynamics of the underlying may be enhanced by adding jumps. Then we have a SDE which
incorporates a jump process.
The SDE with a given J(t) compound Poisson process is described by:

dX(t) = a (X(t)) dt + σ (X(t)) W (t) + c (X(t−)) dJ(t)

To simulate X we can use the Euler-Maruyama scheme specialized to such SDE.


With notation ∆tj = tj − tj−1 , the algorithm is:

1. Generate a random number Zj v N [0, 1]

2. Generate a random number Ξj v P n [λ · ∆tj ]

3. Simulate a random number Λj v L


p
4. Set X(ti ) = X(ti ) + a (X(tj )) ∆tj + σ ∆tj Zj + c (X(tj )) Λj Ξj

99
Variance reduction methods

For regular Monte Carlo methods the standard error only decreases as N, with N number of simu-
lations.
Method A has a more eective reduction of variance than method B if

A B
 
WA · V ar X̄N < WB · V ar X̄N

whereWA , WB denote the computational work PER SAMPLE for A or B


To improve the computational eciency one may employ various variance reduction methods Jackel
[2002], Glasserman [2003], Korn et al. [2010]:

• antithetic variates

• control variates

• stratied sampling

• Latin hypercube sampling

• Importance sampling

• moment matching

• quasi Monte Carlo

100

Вам также может понравиться