Академический Документы
Профессиональный Документы
Культура Документы
Contents
Abstract: ........................................................................................................................................................ 3
Chapter 1: Selection of quality factors and building the quality score ......................................................... 3
1.1 Introduction .................................................................................................................................. 3
1.2 Explanation of the variables................................................................................................................ 4
1.3 Methodology of selection of factors ................................................................................................... 5
1.4 RESULTS............................................................................................................................................... 5
1.4.1 STATIC APPROACH ....................................................................................................................... 5
1.4.2 DYNAMIC APPROACH................................................................................................................... 7
1.4.3 Remarks & Decisions ........................................................................................................................ 7
1.5 Valuation measures for the quality stocks.......................................................................................... 7
1.5.1 Methodology ................................................................................................................................ 7
1.5.2 Results .......................................................................................................................................... 8
1.6 Valuation measures for momentum stocks ........................................................................................ 9
1.6.1 Methodology ................................................................................................................................ 9
1.6.2 Results ........................................................................................................................................ 10
1.6.3 Remarks: .................................................................................................................................... 11
Chapter 2: Correlation measures across factors......................................................................................... 11
2.1. Introduction and definitions ....................................................................................................... 11
2.2 Percentages of common names........................................................................................................ 12
2.2.1 Methodology .............................................................................................................................. 12
2.2.2 Results ........................................................................................................................................ 12
2.2 Indexes performance properties: ..................................................................................................... 14
Chapter 3: Dynamic asset allocation strategies .......................................................................................... 17
3.1 Introduction ...................................................................................................................................... 17
3.2 Methodology..................................................................................................................................... 18
3.2.1 Allocation styles ......................................................................................................................... 18
3.2.2 Forecasting returns .................................................................................................................... 20
3.3. MYOPIC RETURNS ....................................................................................................................... 22
3.3.1 𝜎 − problem .............................................................................................................................. 22
1
3.3.2 MV allocation ............................................................................................................................. 27
3.3.3 ERC allocation ............................................................................................................................ 29
3.4 RISK Adjusted Returns....................................................................................................................... 31
3.4.1 𝜎-problem .................................................................................................................................. 31
3.4.2 MV allocation ............................................................................................................................. 32
3.4.3 ERC allocation ............................................................................................................................ 34
3.5 Vector Auto-regression Returns ....................................................................................................... 36
3.5.1 𝜎-problem .................................................................................................................................. 36
3.5.2 MV allocation ............................................................................................................................. 37
3.5.3 ERC allocation ............................................................................................................................ 39
3.5.4 Remarks regarding VAR stability and back-testing ............................................................. 40
3.6 Risk-weighted returns (GARCH(1,1))................................................................................................. 43
3.7: Information ratio.............................................................................................................................. 44
3.7.1: Methodology............................................................................................................................. 44
3.7.2 Myopic returns .................................................................................................................... 44
3.7.3: VAR(1) forecasted returns results............................................................................................. 47
3.7.4 REMARKS for VAR allocation results: ......................................................................................... 49
3.7.5 GARCH(1,1) forecasted returns results ...................................................................................... 50
3.7.6 Remarks...................................................................................................................................... 51
3.8 Final conclusion on allocations ................................................................................................... 51
Chapter 4: APPENDIX: ................................................................................................................................. 51
Matlab code for the functions ERC, MV, RPB, 𝜎 −problem for allocation strategies ............................ 51
RPB and ERC portfolio: ........................................................................................................................ 51
MV portfolio ........................................................................................................................................ 53
SOURCES: ................................................................................................................................................ 54
2
Abstract:
The following work will be composed of three parts:
1. The selection of the factors that explain best the implied risk premium (Implied Cost of Equity)
across stocks and industrial sectors. Definition of “Quality” and “quality score”
2. Consistency check between the correlations of performances of representative stocks for each
factor and the correlations of performances of their corresponding MSCI indexes.
If the consistency is valid, we can reduce the problem of allocation across hundreds of stocks, to
allocation across indexes.
3. Analysis of various allocation strategies across indexes depending on forecast models and
allocation problem (minimum-variance, 𝜎 −targeted maximum return problems, or risk
budgeting problems (in particular ERC portfolios)).
Data availability for a factor = the average over all stocks of the CAPEX Universe, of the number of
non-NaN values = 1 – average percentage of NaN’s.
The CAPEX Universe = the stocks which have complete historical information about EPS (Earnings
per Share) and CAPEX (Capital Expenditures) throughout the last 10 years.
The summary of data availability for relevant variables is as follows:
3
MSCI also provides three factors which define the quality score.
The factors are: variability of earnings per share, leverage and profitability.
It also defines the quality score as the average of the z-scores of each of the above three factors.
For details of the methodology of computation of the quality z-score, see page 6/17 (Section 2.2) of this
link: https://www.msci.com/eqb/methodology/meth_docs/MSCI_Quality_Indices_Methodology.pdf
DEFINITION :
We define Cost of Equity of a firm as the compensation demanded by the market for owning
the asset and bearing the risk of ownership.
We can put Cost of Equity in relation with other variables to be studied further:
Cost of Equity = 𝐷𝑃𝑆 (𝑛𝑒𝑥𝑡 𝑦𝑒𝑎𝑟)/𝑃𝑟𝑖𝑐𝑒 (now) + growth rate of dividend (DPS = dividend per
share)
Variability of earnings per share = annualized standard deviation of EPS weekly of % changes.
𝑵𝒆𝒕 𝑫𝒆𝒃𝒕 𝑵𝒆𝒕 𝑰𝒏𝒄𝒐𝒎𝒆
𝑳𝒆𝒗𝒆𝒓𝒂𝒈𝒆 = 𝑻𝒐𝒕𝒂𝒍 𝑬𝒒𝒖𝒊𝒕𝒚
, 𝑷𝒓𝒐𝒇𝒊𝒕𝒂𝒃𝒊𝒍𝒊𝒕𝒚 = 𝑻𝒐𝒕𝒂𝒍 𝑨𝒔𝒔𝒆𝒕𝒔.
We define Implied Cost of Equity of a firm as the internal rate of return that equates the asset’s
market value to the present value of its expected future cash flows.
(1), (2) – the 2 momentum scores are chosen according to pages 4, 5 of the article published here:
https://www.msci.com/eqb/methodology/meth_docs/MSCI_Momentum_Indexes_Methodology_Sep20
14.pdf
4
1.3 Methodology of selection of factors
We will proceed in two steps:
1. STATIC APPROACH: We will choose the best 3 factors explaining implied cost of equity
from the factors mentioned above, by studying one-factor regression models as
follows:
a.
M1. 𝑪𝒐𝑬 = 𝒂𝟏 + 𝒃𝟏 ⋅ 𝑳𝒆𝒗𝒆𝒓𝒂𝒈𝒆, M2. 𝑪𝒐𝑬 = 𝒂𝟐 + 𝒃𝟐 ⋅ 𝑷𝒓𝒐𝒇𝒊𝒕𝒂𝒃𝒊𝒍𝒊𝒕𝒚
M3. 𝑪𝒐𝑬 = 𝒂𝟑 + 𝒃𝟑 ⋅ 𝒔𝒕𝒅𝒅𝒆𝒗(𝑬𝑷𝑺 % 𝐜𝐡𝐚𝐧𝐠𝐞𝐬),
M4. 𝑪𝒐𝑬 = 𝒂𝟒 + 𝒃𝟒 ⋅ 𝜷, M5. 𝑪𝒐𝑬 = 𝒂𝟓 + 𝒃𝟓 ⋅ 𝑬𝑷𝑺 𝑴𝒐𝒎𝒆𝒏𝒕𝒖𝒎 𝒔𝒄𝒐𝒓𝒆,
M6. 𝑪𝒐𝑬 = 𝒂𝟔 + 𝒃𝟔 ⋅ 𝑷𝒓𝒊𝒄𝒆 𝑴𝒐𝒎𝒆𝒏𝒕𝒖𝒎 𝒔𝒄𝒐𝒓𝒆,
M7. 𝑪𝒐𝑬 = 𝒂𝟕 + 𝒃𝟕 ⋅ 𝒔𝒕𝒅𝒅𝒆𝒗(𝑷𝒓𝒊𝒄𝒆 % 𝒄𝒉𝒂𝒏𝒈𝒆𝒔)
1.4 RESULTS
1.4.1 STATIC APPROACH
Table 1.1 (Average of 𝑅 2 across industry sectors, over 1997 – 2017 period)
Sector M1 M2 M3 M4 M5 M6 M7
Auto 32.8% 32.6% 24.5% 8.8% 3.8% 7.1% 25.2%
Basic Resources 17.1% 23.3% 12.1% 4.3% 6.0% 8.3% 13.4%
Chemicals 14.3% 24.0% 22.3% 8.8% 6.8% 10.4% 23.9%
Construction 16.6% 14.5% 19.7% 10.8% 10.2% 12.7% 10.7%
Finance 17.4% 16.4% 9.1% 9.2% 16.4% 20.0% 19.6%
Food 26.9% 27.3% 12.9% 6.4% 6.3% 7.5% 19.0%
Health Care 17.2% 23.3% 13.5% 5.6% 8.8% 10.8% 14.4%
Industrial Goods 22.3% 24.6% 16.3% 7.1% 6.5% 9.5% 19.2%
5
Media 16.4% 23.8% 14.1% 7.6% 7.0% 9.3% 13.1%
Oil & Gas 19.2% 14.9% 11.5% 7.5% 7.0% 12.6% 11.7%
Personal Goods 22.8% 22.5% 11.6% 6.3% 6.2% 10.0% 12.8%
Real Estate 9.7% 12.7% 9.2% 9.0% 4.1% 10.0% 11.6%
Retail 15.6% 20.5% 10.4% 12.0% 10.6% 13.1% 9.3%
Technology 18.5% 29.3% 19.2% 5.7% 10.3% 7.6% 12.4%
Telecommunications 24.5% 27.3% 15.8% 9.3% 1.9% 5.2% 13.2%
Travel & Leisure 17.2% 24.8% 10.8% 9.6% 7.9% 9.0% 11.5%
Utilities 19.6% 21.6% 9.7% 6.2% 6.8% 11.9% 8.4%
Average 19.3% 22.5% 14.3% 7.9% 7.4% 10.3% 14.7%
Table 1.2 (Average of 𝑅 2 across industry sectors, over 2007 – 2017 period)
Sector M1 M2 M3 M4 M5 M6 M7
Auto 32.9% 34.8% 31.1% 10.9% 5.1% 11.4% 50.2%
Basic Resources 19.0% 23.7% 19.7% 8.9% 5.6% 11.7% 18.1%
Chemicals 14.5% 24.1% 26.6% 9.4% 10.4% 10.5% 37.9%
Construction 16.6% 14.8% 21.7% 14.5% 14.8% 18.7% 16.4%
Finance 17.7% 16.4% 12.8% 6.6% 16.5% 25.5% 22.3%
Food 27.0% 27.5% 19.6% 11.6% 8.2% 8.0% 36.0%
Health Care 17.7% 23.7% 18.7% 8.3% 9.1% 15.0% 27.1%
Industrial Goods 22.4% 24.7% 20.2% 11.5% 10.8% 12.4% 25.6%
Media 16.4% 23.8% 17.8% 12.1% 7.1% 12.0% 25.6%
Oil & Gas 19.2% 15.1% 19.9% 11.1% 9.4% 13.9% 17.5%
Personal Goods 23.0% 22.8% 19.4% 12.8% 10.1% 10.4% 27.4%
Real Estate 9.7% 12.7% 19.5% 7.7% 9.5% 17.5% 19.2%
Retail 15.8% 20.8% 11.7% 11.7% 10.5% 12.5% 16.7%
Technology 18.6% 29.5% 26.7% 12.3% 9.3% 8.2% 33.9%
Telecommunications 24.7% 28.0% 18.8% 8.3% 7.2% 13.2% 17.3%
Travel & Leisure 17.4% 24.7% 13.0% 8.8% 7.6% 11.5% 18.3%
Utilities 19.7% 21.8% 16.7% 9.7% 12.9% 16.2% 15.4%
Average 19.5% 22.9% 19.6% 10.4% 9.7% 13.4% 25.0%
6
1.4.2 DYNAMIC APPROACH
1.5.1 Methodology
1. We compute the quality scores at each point in time and for each stock as the average between the z-
scores of EPS variability, Profitability and Leverage.
2. We take the average of quality scores across the history of each stock
4. We compute the average of P/E ratio and the average of growth at each point in time for the quality
selection and for the entire universe.
7
1.5.2 Results
8
REMARK:
The quality stocks are more expensive overall than regular stocks, but they display a better growth and
implied cost of equity.
2. We compute the risk-adjusted price momentums by dividing with the 3Y annualized standard
deviations of each stock’s returns.
9
1.6.2 Results
10
1.6.3 Remarks:
On the momentum stocks, we remark an inverse behavior as we have seen in the quality stocks.
The Cost of equity is significantly lower on momentum stocks than overall on the market, display a lower
growth rate and have a slightly higher P/E ratio.
DEFINITION:
For a given threshold 𝛼 ∈ (0,1), two factors 𝐹1 , 𝐹2 , and at time 𝑡, we define 𝑃𝐶𝑁(𝐹1 , 𝐹2 , 𝛼, 𝑡):=
𝑐𝑎𝑟𝑑(𝐹1 ∩𝐹2 )
[𝛼⋅𝑛]
where 𝑛 = the number of stocks present in our universe, and 𝐹1 , 𝐹2 are the stocks lying in the
𝛼 −tail of the universe of stocks, from each factor point of view.
Example
Suppose 𝐹1 := small cap, 𝐹2 : =low vol, n=500, then 𝑃𝐶𝑁(𝐹1 , 𝐹2 , 10%, 𝑡) =the number of common stocks
from these with the lowest market capitalization and these with the lowest annualized standard
𝑐𝑎𝑟𝑑(𝐹1 ,𝐹2 )
deviation of price returns = 50
.
11
Objective
The objective is to find a correspondence between the correlation of the index (relative) performances
and the PCN pairwise functions. The point is to show that the indexes reflect somehow our choices of
stocks.
We hope that the hierarchies of the performance correlations and the PCN functions to be the same (or
very similar) between all the pairs of factors, so that we can ease our study in our asset allocation
strategies by reducing the universe to a few indexes.
Instead of comparing the values of each factor, we transform all quantities (except quality score) into z-
scores.
Then we compare the 𝛼 −tails of each factor, at each point in time and we take the stocks presenting
the highest (lowest) z-scores as follows:
Price Momentum, Growth, Value, Quality, High dividend (highest) and Small Cap (Size), Low vol (the
lowest).
1
1. Price Momentum score = ⋅ (𝑃𝑟𝑖𝑐𝑒 𝑀𝑜𝑚 6𝑀𝑠𝑐𝑜𝑟𝑒 + 𝑃𝑟𝑖𝑐𝑒 𝑀𝑜𝑚 12𝑀𝑠𝑐𝑜𝑟𝑒 ) where
2
𝑃𝑡−1 −𝑃𝑡−7 𝑃𝑡−1 −𝑃𝑡−13
𝑃𝑡−7 𝑃𝑡−13
𝑃𝑟𝑖𝑐𝑒 𝑀𝑜𝑚6𝑀 = , Price Momentum (12M) = where 𝜎 =standard deviation
𝜎 𝜎
of the weekly returns of the price over the last year (or last 3 years (optionally)).
2. Value = P/E (Price / Earnings share);
3. Quality score = average of the z-scores of: standard deviation of Earnings per Share returns,
Leverage z-score and Profitability z-score.
4. High dividend yield = trailing 12 months dividends per share.
5. Size = Market Capitalization
6. Low vol = Earnings per share weekly returns variability over the last year.
We will compute the average of PCN for the thresholds 𝛼 = 10%, 20%, 30% across time.
2.2.2 Results
Table 1.1 (Average Percentage of common names for 10% threshold
Avg CPN 10% Price MMT Growth Value Quality Low Vol High dvd Small cap
Price MMT 100.0% 9.8% 15.3% 15.4% 8.8% 10.5% 10.1%
Growth 9.8% 100.0% 35.2% 22.0% 0.2% 6.2% 20.6%
Value 15.3% 35.2% 100.0% 18.2% 7.0% 6.7% 12.5%
Quality 15.4% 22.0% 18.2% 100.0% 9.2% 12.3% 18.6%
Low vol 8.8% 0.2% 7.0% 9.2% 100.0% 7.8% 5.6%
High dvd 10.5% 6.2% 6.7% 12.3% 7.8% 100.0% 2.1%
Small cap 10.1% 20.6% 12.5% 18.6% 5.6% 2.1% 100.0%
12
Period: January 2001 – October 2017
Avg CPN 20% Price MMT Growth Value Quality Low Vol High dvd Small cap
Price MMT 100.0% 20.0% 28.6% 25.1% 20.2% 21.0% 19.0%
Growth 20.0% 100.0% 41.2% 27.0% 2.3% 15.4% 30.3%
Value 28.6% 41.2% 100.0% 28.3% 17.7% 19.6% 19.5%
Quality 25.1% 27.0% 28.3% 100.0% 18.9% 18.7% 26.9%
Low vol 20.2% 2.3% 17.7% 18.9% 100.0% 20.1% 13.6%
High dvd 21.0% 15.4% 19.6% 18.7% 20.1% 100.0% 4.5%
Small cap 19.0% 30.3% 19.5% 26.9% 13.6% 4.5% 100.0%
Period: January 2001 – October 2017
Avg CPN 30% Price MMT Growth Value Quality Low Vol High dvd Small cap
Price MMT 100.0% 31.1% 40.7% 34.9% 32.6% 31.4% 29.4%
Growth 31.1% 100.0% 47.5% 33.3% 10.9% 27.0% 37.7%
Value 40.7% 47.5% 100.0% 37.3% 29.2% 28.8% 27.7%
Quality 34.9% 33.3% 37.3% 100.0% 25.9% 28.6% 34.9%
Low vol 32.6% 10.9% 29.2% 25.9% 100.0% 32.6% 22.8%
High dvd 31.4% 27.0% 28.8% 28.6% 32.6% 100.0% 11.2%
Small cap 29.4% 37.7% 27.7% 34.9% 22.8% 11.2% 100.0%
One can remark that the Value and Growth are well correlated from the point of view of common stocks
presence for each threshold (see the figure below).
Figure 1: Percentage of common names for the best 10% - 30% Growth and Value stocks
13
We will compare the hierarchies of correlations of the relative performances of the indexes with the
hierarchy of the CPN functions.
In the table below are mentioned the MSCI trackers for the indexes used throughout this presentation,
order preserved when building the matrices of correlations, (relative) returns and other measures.
14
We will present here the relative evolution of the indexes, with respect to their initial value.
In the below table are presented the average relative correlations between the factor performances
along the period January 2001- October 2017.
Table 2.1
AVG rel corr Price MMT Growth Value Quality Low Vol High dvd Small cap
Price MMT 100.0% 52.0% -52.2% 40.4% 34.2% -16.2% 15.1%
Growth 52.0% 100.0% -99.9% 68.2% 38.9% -33.7% -3.2%
Value -52.2% -99.9% 100.0% -68.2% -39.2% 33.8% 2.9%
Quality 40.4% 68.2% -68.2% 100.0% 57.1% -1.1% -13.5%
Low vol 34.2% 38.9% -39.2% 57.1% 100.0% 27.0% -1.1%
High dvd -16.2% -33.7% 33.8% -1.1% 27.0% 100.0% -13.9%
Small cap 15.1% -3.2% 2.9% -13.5% -1.1% -13.9% 100.0%
15
Hierarchy of relative correlations (absolute value, decreasing order)
I will add the first 5 pairs for PCN (10%), PCN (20%), PCN (30%) and relative performance correlations.
Table 2.2
The constant of the hierarchies is the (Growth, Value) pair which marks the highest correlation from
both measurements (Percentage of Common Names and relative performance correlation). Also the
composition of the classifications is similar.
In the figure below I represent the PCN functions on weekly data, as well as correlations of the weekly
relative performances.
16
Period: January 2001 – October 2011
CONCLUSION:
The index performances reflects the stock factor selection, so we can study allocation strategies based
only on the factor indexes.
Now we go further to the most interesting part of my work: dynamic asset allocation along factors.
Let 𝜇 = 𝐸[𝑅] and Σ = 𝐸[(𝑅 − 𝜇)(𝑅 − 𝜇)𝑇 ] be the vector of expected returns and the covariance matrix
of asset returns. The expected return of the portfolio is therefore 𝜇(𝑥) = 𝐸[𝑅(𝑥)] = 𝐸[𝑥 𝑇 𝑅] =
𝑥 𝑇 𝐸[𝑅] = 𝑥 𝑇 𝜇 whereas its variance is equal to:
𝑇
𝜎 2 (𝑥) = 𝐸 [(𝑅(𝑥) − 𝜇(𝑥))(𝑅(𝑥) − 𝜇(𝑥)) ] = 𝐸[(𝑥 𝑇 𝑅 − 𝑥 𝑇 𝜇)(𝑥 𝑇 𝑅 − 𝑥 𝑇 𝜇)𝑇 ] =
17
𝜎12 𝜎12 … 𝜎1𝑛
Σ = 𝜎22 𝜎22 … 𝜎2𝑛 is the covariance matrix of (relative) returns of indices/stocks.
… … … …
𝜎
( 𝑛1 𝜎𝑛2 … 𝜎𝑛2 )
𝜇 = (𝜇1 , 𝜇2 , … , 𝜇𝑛 ) is the vector of the expected returns.
As in any portfolio allocation strategy, there are always two inputs: the expected returns, and the
covariance matrix of returns.
1. Forecast the returns (find 𝜇) (using either gross estimations / econometric models).
2. Using these results and the covariance matrix of the returns, using either closed-formulas or
numerical methods, we find the strategy 𝑥 = (𝑥1 , 𝑥2 , … , 𝑥𝑛 ).
1. 𝜎 − targeted 𝜇 −problem.
2. ERC portfolio
3. MV portfolio
For the ERC portfolio we will use standard deviation of returns as proxy for risk and risk-
decomposition.
3.2 Methodology
3.2.1 Allocation styles
Min-variance portfolio:
The problem considered is:
𝑻
(𝑷𝟏 ): 𝒂𝒓𝒈𝒎𝒊𝒏𝒙∈𝑹𝒏 𝝈(𝒙) such that {𝟏 𝒙 = 𝟏
𝒙≥𝟎
The structure depends only on the covariance matrix as given in the relation below.
Σ−1 1
𝑥𝑀𝑉 = 1𝑇 Σ−11
Here 𝜎(𝑥) = 𝑥 𝑇 Σ𝑥, where Σ = covariance matrix of the relative returns (could be
weekly/monthly, quarterly, semi-annual or yearly).
18
𝝈 − targeted, 𝝁 − 𝒑𝒃 portfolio:
The problem considered is:
𝟏𝑻 𝒙 = 𝟏
𝑻
(𝑷𝟐 ): 𝒂𝒓𝒈𝒎𝒂𝒙𝒙∈𝑹𝒏 (𝒙 𝝁) subject to the constraints: {𝝈(𝒙) ≤ 𝝈∗
𝒙≥𝟎
𝜎 = targeted standard deviation of relative returns.
If we consider as risk measure, the standard deviation of the portfolio, 𝜎(𝑥) = √𝑥 𝑇 Σ𝑥, 𝑥 ∈ 𝑅 𝑛
𝜕𝜎(𝑥)
we define as marginal risk of the asset 𝑖, .
𝜕𝑥𝑖
Also we define the risk-contribution of asset 𝑖, as the product between the corresponding
weight and the marginal risk of asset i.
(Σ𝑥)𝑖
As a consequence, 𝑅𝐶𝑖 = 𝑥𝑖 ⋅ . (2)
√𝑥 𝑇 Σ𝑥
19
(P1) 𝑥 ∗ = {𝑥 ∈ [0,1]𝑛 , ∑𝑛𝑖=1 𝑥𝑖 = 1, ∑𝑛𝑖=1 𝑏𝑖 = 1, 𝑥𝑖 ⋅ 𝜕𝑥𝑖 𝑅(𝑥) = 𝑏𝑖 ⋅ 𝑅(𝑥), 𝑏 ∈ (0,1]𝑛 , ∑𝑛𝑖=1 𝑏𝑖 =
1}
It is not possible to find a closed-form solution to (P1) but more of a numerical. We can try to
use Broyden or SQRF algorithms to solve (P1) but the solution doesn’t always converge.
In order to make (P1) more tractable in terms of implementation we will consider an equivalent
version, as an optimization problem:
So 𝑓(𝑥; 𝑏) = 𝐴(𝑥) + 𝐵(𝑥) − 𝐶(𝑥) and the problem (P2) can be rewritten as follows:
(P2’): 𝑥 ∗ = 𝑎𝑟𝑔𝑚𝑖𝑛𝑥∈𝑅𝑛 𝐴(𝑥) + 𝐵(𝑥) − 𝐶(𝑥) subject to:
1𝑇 𝑥 = 1
{
𝑥 ≥ 0, 𝑥 ∈ 𝑅 𝑛
Particular case: ERC portfolio.
1 1 1
The ERC portfolio is the portfolio where 𝑅𝐵 = (𝑛 , 𝑛 , … , 𝑛). In the APPENDIX you have
attached the MATLAB functions for both the ERC portfolio in particular and RB portfolio in
general.
3.2.2 Forecasting returns
The index data on which I’ll work will be consisting of weekly data for 877 weeks starting from
5 of January 2001 to 20th October 2017 (for purposes of homogeneity of data – some indexes lack data
th
before).
20
Myopic forecast
1. The index data is 𝐼 ∈ 𝑀877,8 (𝑅+ ), based on weekly data, the eight columns representing the
indexes mentioned in section 2.2. I am working on rolling bases. That means, the weekly
relative returns form a matrix, 𝐼𝑤 ∈ 𝑀876,7 (𝑅), the monthly relative returns form a matrix
𝐼𝑚 ∈ 𝑀873,7 (R), the quarterly 𝐼𝑞 ∈ 𝑀865,7 (𝑅), and so on.
2. We take rolling periods of 1 year, and the myopic forecasted return at time 𝑡 will simply be
𝐼𝑥 (𝑡 − 1, : ) where 𝑥 ∈ {𝑤, 𝑚, 𝑞, 𝑠𝑎, 𝑦}.
3. The myopic estimates for weekly relative returns will form a matrix 𝑅1𝑊 ∈ 𝑀825,7 (𝑅), the
monthly relative returns 𝑅1𝑀 ∈ 𝑀822,7 (𝑅),…, 𝑅1𝑌 ∈ 𝑀774,7 (𝑅)
Risk-adjusted forecasts
REMARK: The allocation results of myopic and risk-adjusted approach will prove to be almost
identical, but the forecasts are not.
REMINDER:
𝑋𝑡1
2
We say that the vector 𝑋𝑡 = 𝑋𝑡 follows a VAR (p) model if 𝑋𝑡 = μ + Φ1 𝑋𝑡−1 + Φ2 𝑋𝑡−2 + ⋯ +
…
(𝑋𝑡𝑘 )
Φ𝑝 𝑋𝑡−𝑝 + 𝜖𝑡 where Φ1 , Φ2 , … , Φ𝑝 ∈ 𝑀𝑘 (𝑅), 𝜖𝑡 ∼ 𝑁(𝑂𝑘 , Ω), 𝜇 = (𝜇1 , 𝜇2 , … , 𝜇𝑘 )𝑇
A VAR(p) can be rewritten as a VAR(1) as follows:
𝑋𝑡 𝜇 Φ1 Φ2 … Φ𝑝−1 Φ𝑝 𝑋𝑡−1 𝑒𝑡
𝑋𝑡−1 0𝑘1 𝐼 𝑂𝑘 … 𝑂𝑘 𝑂𝑘 𝑋 𝑡−2 𝑒𝑡−1
( … )=( … )+( 𝑘 )( … ) + ( … ) = 𝜇 + Φ ⋅ 𝑋𝑡 + 𝑒𝑡
… … … … …
𝑋𝑡−𝑝+1 0𝑘1 𝑂𝑘 𝑂𝑘 … 𝐼𝑘 𝑂𝑘 𝑋𝑡−𝑝 𝑒𝑡−𝑝+1
The model is said to be stable if det(𝐼𝑘 − Φ1 𝑧 − Φ2 𝑧 2 − ⋯ − Φ𝑝 𝑧 𝑝 ) ≠ 0, ∀|𝑧| ≤ 1.
METHODOLOGY:
Here I will use two vector auto-regression models, VAR (1) and VAR (2) models.
The rolling period is 52 weeks and the forecasted to be analyzed will start with the 53th week. The
forecasted returns are treated in blocks (for all indexes at once).
We say that a univariate random process 𝑿𝒕 follows a GARCH (1,1) model if 𝑿𝒕 = 𝝁 + 𝚽𝑿𝒕−𝟏 + 𝝐𝒕 ,
where 𝝐𝒕 = 𝝈𝒕 𝒛𝒕 and 𝝈𝟐𝒕 = 𝜶𝟎 + 𝜶𝟏 𝝈𝟐𝒕−𝟏 .
21
The same methodology is applied as above, but the forecasted relative returns are treated
independently.
3.3.1 𝜎 − problem
Methodology
We will consider 𝝈 = annualized standard deviation of the portfolio starting from the covariance
matrices of relative returns of MSCI indexes with respect to MSCI Europe Index.
In order to choose proper values for 𝝈 we need a clear idea about the annualized standard deviation
𝟏 𝟏 𝟏
of weekly relative returns of indexes and for the Equally Weighted portfolio (𝒙 = (𝟕 , 𝟕 , … , 𝟕)).
So I will show the evolution of the historical standard deviations of relative returns starting from January
2002 and ending to October 2017.
22
As you can see, the standard deviation of the reference EW portfolio lies around 2%.
Therefore I will choose 3 targets for
Average weight for each index in the 𝝈 − problem, for each 𝝈 ∈ {𝟎. 𝟎𝟎𝟏, 𝟎. 𝟎𝟎𝟓, 𝟎. 0𝟏}
23
Allocation coefficient evolution for weekly returns, myopic estimates (January 2001 – October 2017)
24
Allocation coefficient evolution for monthly returns, myopic estimates (January 2001 – October 2017)
25
Allocation coefficient evolution for myopic estimates of quarterly returns (Jan 2001 – Oct 2017)
REMARK:
1. One can remark the decreasing evolution of the allocation % for the growth index, the higher
is the return periods.
2. For low volatility targets, e.g. 𝝈 = 𝟎, 𝟎𝟎𝟏 or 𝟎. 𝟏%, the allocation results are stable for Growth
and Index factor (between 45% and 55%). For higher volatility targets the results are less
accurate to determine a precise decision.
3. The results for 𝝈 − 𝒑roblem where 𝝈 = 𝟎. 𝟎𝟎𝟏 are similar to these of the corresponding MV
portfolio, because the MV portfolio’s standard deviation lies around that value.
26
3.3.2 MV allocation
Average allocation results across 2001 – 2017 period.
Myopic Price
ret/𝑴𝑽 −pb MMT Growth Value Quality Low vol High dvd Small cap
1W return 1.3% 45.5% 48.1% 1.7% 1.0% 1.4% 1.0%
1M return 1.6% 45.0% 47.1% 1.9% 1.3% 1.9% 1.2%
3M return 2.7% 41.9% 46.2% 2.8% 1.9% 2.8% 1.7%
6M return 3.3% 40.3% 44.6% 3.5% 2.1% 3.9% 2.3%
1Y return 3.8% 39.8% 41.2% 5.1% 2.5% 4.8% 2.9%
27
Allocation % evolution for MV portfolio (January 2001 – October 2017)
REMARK:
1. The standard deviation of the MV portfolios obtained are as indicated in the figure below:
2. The minimum standard deviation of portfolio attained is around 0.1%, and this is why in the 𝜎-
problem we should consider targets above 0.1%.
28
3.3.3 ERC allocation
Average allocation results for the ERC portfolio.
REMARKS:
1. Inside the strategy defined above, an allocation of 7% to the low-vol index would have a 14.28%
risk-contribution while a 23.3 % allocation to growth index would contribute the same to the
portfolio volatility.
2. The ERC allocations along each factor behave this time very discontinuously and it’s harder to
set up a pattern.
29
Allocation % for Growth and Value Index
30
3.4 RISK Adjusted Returns
3.4.1 𝜎-problem
As in section 3.3.1 𝜎 – represents the standard deviation of the portfolio derived from annualized
standard deviations of relative returns of indexes.
31
REMARK:
The difference between the risk-adjusted approach and the myopic forecast is very small.
On the value index there is the greatest difference between allocations, on average (on the 6M horizon
and 1Y orizon) (for all 𝜎-problems considered).
3.4.2 MV allocation
Risk-adj Price MMT Growth Value Quality Low vol High dvd Small cap
1W returns 1.3% 45.5% 48.1% 1.7% 1.0% 1.4% 1.0%
1M returns 1.6% 45.0% 47.1% 1.9% 1.3% 1.9% 1.2%
3M returns 2.7% 41.9% 46.2% 2.8% 1.9% 2.8% 1.7%
6M returns 3.3% 40.3% 44.6% 3.5% 2.1% 3.9% 2.3%
1Y returns 3.8% 39.8% 41.2% 5.1% 2.5% 4.8% 2.9%
32
Comparison between Myopic forecast MV problem and Risk-weighted forecast MV problem results.
33
3.4.3 ERC allocation
Price MMT
Growth Value Quality Low vol High dvd Small cap
1W returns 7.9% 28.7% 18.8% 9.0% 6.9% 17.6% 11.1%
1M returns 7.7% 28.1% 22.1% 9.0% 6.7% 16.2% 10.2%
3M returns 9.4% 23.0% 23.6% 11.6% 7.0% 15.5% 10.0%
6M returns 10.5% 23.8% 22.1% 11.1% 6.9% 14.4% 11.2%
1Y returns 10.8% 23.3% 20.2% 14.1% 7.0% 13.4% 11.2%
REMARKS:
1. We see an increasing allocation weight on Price Momentum and Quality Index, the higher is the
investment horizon, given equal risk weights.
2. In terms of risk, we see that in case of weekly horizon, 28.7% of capital allocated on growth
contributes to 14.28% of the total risk of portfolio while only 6.9% of total capital allocated on
low-vol index contributes with the same amount of risk to the total risk of the portfolio.
CONCLUSION:
Low vol index does not mean a less risky index in terms of risk contribution to a portfolio of 7 indexes.
Actually, among all 7 indexes, the low-vol index has the relative returns contributing the strongest to the
portfolios risk. It presents a relative risk to the MSCI Europe benchmark 4 times higher than the growth’s
relative risk.
34
35
3.5 Vector Auto-regression Returns
3.5.1 𝜎-problem
VAR (1) average coefficients
36
VAR (2) average coefficients
REMARKS:
1. The Growth index weight decreases both with frequency of returns and with targeted volatility
in case of VAR (1) forecasting model. In the same time, an increasing allocation weight should be
considered for Small Cap Index, the higher the horizon time of investment.
2. In case of the VAR (2) forecasting model the results behave in a similar way, but the weights do
not depend on the targeted volatility.
3.5.2 MV allocation
VAR(1) – allocation results
Table: VAR (1) and VAR (2) forecast average allocation results.
Price High Small
MMT Growth Value Quality Low vol dvd cap
1W return 1.3% 45.5% 48.1% 1.7% 1.0% 1.4% 1.0%
1M return 1.6% 45.0% 47.1% 1.9% 1.3% 1.9% 1.2%
3M return 2.7% 41.9% 46.2% 2.8% 1.9% 2.8% 1.7%
6M return 3.3% 40.3% 44.6% 3.5% 2.1% 3.9% 2.3%
1Y return 3.8% 39.8% 41.2% 5.1% 2.5% 4.8% 2.9%
37
REMARK:
The forecasting results of VAR(1) and VAR(2) models are slightly different but the allocation results are
identical. (See below, the forecasts of relative returns of Growth and Value index using VAR(1) and
VAR(2) models).
For more details about the back-tests of the model see Section 3.5.4 about Back-testing and stability
issues of the models.
38
3.5.3 ERC allocation
VAR(1) and VAR(2) allocation results:
High Small
Price MMT Growth Value Quality Low vol yield cap
1W return 7.9% 28.7% 18.8% 9.0% 6.9% 17.6% 11.1%
1M return 7.7% 28.1% 22.1% 9.0% 6.7% 16.2% 10.2%
3M return 2.7% 41.9% 46.2% 2.8% 1.9% 2.8% 1.7%
6M return 3.3% 40.3% 44.6% 3.5% 2.1% 3.9% 2.3%
1Y return 3.8% 39.8% 41.2% 5.1% 2.5% 4.8% 2.9%
REMARKS:
1. In each choice of the time horizon, a significant portion of capital should be allocated along
Growth and Value Indexes, but especially if one considers investing on 3 month and 6 month
horizon.
2. On short term investment, one should consider the high yield stocks (index) for equally risk
budget portfolios.
39
3.5.4 Remarks regarding VAR stability and back-testing
1. The estimated VAR models for relative returns are stationary and stable, that is, the coefficient Φ
from the model 𝑋𝑡 = 𝜇 + Φ𝑋𝑡−1 + 𝜖𝑡 has the property that the roots of |Φ − 𝜆𝐼𝑁 | = 0 are inside
the unit circle where 𝑁 = 7 in our case.
2. According to the following back-test, between VAR (1) and VAR (2), the 2nd model is more accurate.
BACKTESTING METHODOLOGY:
1. We consider 𝑅𝑖𝑗 =realized relative return of index i at time j, and 𝑅𝑖𝑗 =forecasted VAR(1)
′
relative return of index i at time j, and 𝑅𝑖𝑗 = forecasted VAR(2) relative return of the same
quantity.
𝑅11 − 𝑅11 𝑅12 − 𝑅12 ⋯ 𝑅17 − 𝑅17
2. We consider the error matrices 𝐴 = ( ⋮ … ⋱ ⋮ ) and 𝐵 =
𝑅𝑛1 − 𝑅𝑛1 𝑅𝑛2 − 𝑅𝑛2 ⋯ 𝑅𝑛7 − 𝑅𝑛7
′ ′ ′
𝑅11 − 𝑅11 𝑅12 − 𝑅12 ⋯ 𝑅17 − 𝑅17
|𝑅𝑖𝑗 −𝑅𝑖𝑗 |
( ⋮ … ⋱ ⋮ ) and we compute|𝐶| = ( ′
) .
′ ′ ′ |𝑅𝑖𝑗 −𝑅𝑖𝑗 |
𝑅𝑛1 − 𝑅𝑛1 𝑅𝑛2 − 𝑅𝑛2 ⋯ 𝑅𝑛7 − 𝑅𝑛7 𝑖𝑗
3. The decision is made depending on the average and median obtained on each column. If the
average error ratio is significantly greater than 1 and the medians are close to or greater than
1 then we choose VAR (2) model rather than VAR (1) model.
We can extend the analysis to VAR (p) models as well. The results will give VAR (2) as the best model
among these.
The cumulative distribution functions (except the upper tail of 95% level) for the error ratios of some
MSCI indexes are given in the figure below:
40
DECISION:
We choose VAR(2) as benchmark model among autoregressive models for the relative performances
of the indexes.
To check the stability of VAR(2) along the forecasting history we will compute for the model
The spectral radius of the companion matrices 𝐹 𝑖 for both VAR (1) and VAR (2) models of weekly
relative returns can be seen in the figure below:
41
42
3.6 Risk-weighted returns (GARCH(1,1))
We obtain the following results for weekly returns:
REMARK:
43
3.7: Information ratio
3.7.1: Methodology
Instead of considering Sharpe ratio for our portfolios, we will consider an alternative measure, called
information ratio which measures the risk/return performance or risk-adjusted performance.
Definition:
For a portfolio 𝒙 = (𝒙𝟏 , 𝒙𝟐 , … , 𝒙𝒏 ) and the benchmark 𝒃 = (𝒃𝟏 , 𝒃𝟐 , … , 𝒃𝒏 ), if 𝝁 = (𝝁𝟏 , … , 𝝁𝒏 )is the
((𝒙−𝒃)𝑻 𝝁
vector of expected returns, 𝚺 =covariance matrix of matrix returns, then 𝑰𝑹(𝒙|𝒃) = .
√(𝒙−𝒃)𝑻 𝚺(𝒙−𝒃)
𝒙𝑻 𝝁
In the absence of benchmark, 𝑰𝑹(𝒙): = .
√𝒙𝑻 𝚺𝒙
In our case, we do not consider any benchmark since we are working with relative returns.
We consider the allocations, returns and standard deviations obtained dynamically at the previous
steps.
44
45
REMARKS:
1. A peak information ratio is remarked around the end of 2008 followed by
minimum information ratio immediately after that.
2. Also I remark a seasonality of high/low levels succession, along 6 months
periods.
46
3.7.3: VAR(1) forecasted returns results
47
48
3.7.4 REMARKS for VAR allocation results:
1. In case of VAR forecast models, given a certain time horizon, we obtain similar patterns for
information ratios, no matter which allocation strategy we choose.
2. In all investment horizon considered previously (1 week, 1 month, 1 quarter, 6 months, 1 year), if we
target 1% volatility, we obtain most of the time, positive information ratios, (that is, positive expected
returns). The back-test gives negative information ratio on 𝜎 = 1% target problem only around the
period of 2008-2009 given high volatilities of the markets.
49
3.7.5 GARCH(1,1) forecasted returns results
There are a few outliers so I will build instead a quantile graph to see the evolution of the sharpe ratio
until the 97% quantile.
Quantile plot for sharpe ratios obtained by dynamic allocation throughout 2001-2017 period
50
3.7.6 Remarks
1. For all the returns considered on the VAR (1) econometric model one can remark a sharpe ratio
peak around 2008 and a positive information ratio for all 𝜎 − 𝑝𝑏 except the one for weekly and
monthly returns where we see a minimum achieved at the beginning of the financial crisis.
2. On the GARCH(1,1) model there is a peak information ratio of 3500 touched around April 2008.
That is because the historical volatility attained before the financial crisis is very small.
Chapter 4: APPENDIX:
Matlab code for the functions ERC, MV, RPB, 𝜎 −problem for allocation strategies
RPB and ERC portfolio:
function [str,ret,vol] = risk_budget_portfolio(returns,bench,budgets,varargin)
% returns -- expected returns of the assets / indices
% budgets -- risk budgets: the max amount of vol allowed for each component
% bench -- can be a vector of the same length as the returns/empty
% varargin -- can contain 1 parameter, the covariance matrix or
% 2 parameters -- the correlation matrix and the individual standard deviations
if nargin==4
cov_mat = varargin{1};
elseif nargin==5
cov_mat = covariance(varargin{1},varargin{2});
end
[A,B,C] = auxiliary(cov_mat,bench,budgets);
f = @(x) A(x)-B(x)+C(x);
k = size(cov_mat,1);
x0 = ones(1,k)/k;
ub = ones(1,k);lb = zeros(1,k);
Aeq = ones(1,k); beq = 1;
[str,~] = fmincon(f,x0,[],[],Aeq,beq,lb,ub);
if isempty(bench)==0
if isempty(returns)==0
ret = (str-bench)*returns';
vol = sqrt((str-bench)*cov_mat*(str-bench)');
else
ret = [];
vol = sqrt((str-bench)*cov_mat*(str-bench)');
end
else
if isempty(returns)==0
51
ret = str*returns';
vol = sqrt(str*cov_mat*str');
else
ret = [];
vol = sqrt(str*cov_mat*str');
end
end
end
function y = covariance(sig,correl)
sigs = repmat(sig',1,length(sig));
y = (sigs.*sigs').*correl;
end
52
end
end
function y = covariance(sig,correl)
sigs = repmat(sig',1,length(sig));
y = (sigs.*sigs').*correl;
end
To check the consistency and correctness of the two functions you can choose example 15 of Thierry
Roncalli’s book, Risk Parity and Budgeting, by taking the risk budgets given there and then compare with
ERC function by using equal weights.
MV portfolio
function [x,ret,vol] = MV_port(bench,lb,ub,returns,varargin)
% Min variance portfolio -- x = the structure, ret = the portfolio expected
% return and vol = the volatility attained
% bench = structure of benchmark (the sum of the components = 1)
% lb = lower bound for these components; ub = upper bound of these
% components
if nargin==5
cov_mat = varargin{1};
elseif nargin==6
cov_mat = covariance(varargin{1},varargin{2});
end
[x,ret,vol]= MV(cov_mat,bench,lb,ub,returns);
end
function y = covariance(sig,correl)
sigs = repmat(sig',1,length(sig));
y = (sigs.*sigs').*correl;
end
53
function [x,ret,vol] = MV(cov_mat,bench,lb,ub,returns)
if isempty(bench)==1
fun = @(x)1/2*x'*cov_mat*x;
Aeq = ones(1,size(cov_mat,1));
beq = 1;
x0 = ones(1,size(cov_mat,1))'*1/size(cov_mat,1);
[x,~] = fmincon(fun,x0,[],[],Aeq,beq,lb,ub);
vol = sqrt(x'*cov_mat*x);
if isempty(returns)==1
ret = [];
else
ret = returns*x;
end
else
fun = @(x)1/2*(x-bench)'*cov_mat*(x-bench);
Aeq = ones(1,size(cov_mat,1));
beq = 1;
x0 = ones(1,size(cov_mat,1))'*1/size(cov_mat,1);
[x,~] = fmincon(fun,x0,[],[],Aeq,beq,lb,ub);
vol = sqrt((x-bench)'*cov_mat*(x-bench));
if isempty(returns)==1
ret = [];
else
ret = returns*(x-bench);
end
end
end
You can confront this code against example 1 provided in Roncalli’s book, Risk Parity and Budgeting.
SOURCES:
1. https://www.msci.com/eqb/methodology/meth_docs/MSCI_Quality_Indices_Methodology.pdf
2.
https://www.msci.com/eqb/methodology/meth_docs/MSCI_Momentum_Indexes_Methodology_Sep20
14.pdf
54