A deep dive into the Mean/Variance Efficiency of the Market Portfolio Master Thesis Supervised by Prof. Dr. Karl Schmedders Chair for Quantitative Business Administration at UZH Zurich Co-supervised by Prof. Dr. Didier Sornette Chair of Entrepreneurial Risks at ETH Zurich Theodoros Giannakopoulos Zurich 2013
210
Embed
A deep dive into the Mean/Variance Efficiency of the ......A deep dive into the Mean/Variance Efficiency of the Market Portfolio Master Thesis Supervised by Prof. Dr. Karl Schmedders
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
A deep dive into the Mean/Variance Efficiency of the
Market Portfolio
Master Thesis
Supervised by Prof. Dr. Karl Schmedders
Chair for Quantitative Business Administration at UZH Zurich
Co-supervised by Prof. Dr. Didier Sornette
Chair of Entrepreneurial Risks at ETH Zurich
Theodoros Giannakopoulos
Zurich
2013
2
3
Abstract
Ever since the appearance of the Capital Asset Pricing Model (CAPM), the mean/variance
efficiency of various market proxies has been theoretically and empirically tested, by a
significant number of previous studies. The findings of almost all of these studies have
shown that the CAPM cannot work in practice; a motion generally accepted amongst the
economists worldwide. The mean/variance efficiency of the Market Portfolio was first
suggested and supported quantitatively by Levy and Roll (2010),1 and afterwards by Ni,
Malevergne, Sornette and Woehrmann, (2011)2. The above mentioned authors attempted
to calculate the return parameters and standard deviations of a sample portfolio by using an
approach different (rather reverse) from the previous researchers, so that they would satisfy
the mean/variance efficiency constraints; by comparing (statistically) their values with the
actual ones for each specific portfolio, they concluded that the difference is within the
statistical error margins, and therefore the CAPM model could hold (at least it cannot be
rejected). In this Thesis, we attempt to stress the robustness of the results supporting the
CAPM, by examining whether the calculated parameters are indeed so close to the actual
ones in a variety of portfolios, in different model implementations and market conditions.
After detailed presentation of the results from analyzing this method, we reached the
conclusion that it has limited power, and by no means can be universally used across
different markets and/or portfolios. Subsequently, an alternative theory was investigated,
which was first presented by Malevergne, Santa-Clara and Sornette (2009)3. In this paper,
the authors employ the idea discovered by Professor George Kingsley Zipf in 19494, in order
to manifest that the heavy-tailed distribution of firm sizes could be used in the form of an
additional risk factor in a time series regression model. The new factor is named the Zipf
factor, and together with the market factor they form a two-factor model that performs
equally well, and sometimes even better, than the three-factor Fama French model5.
1 Levy, M. and Roll, R. (2010). The Market Portfolio May Be Mean/Variance Efficient After All. Oxford
Journals, Review of Financial Studies, 23(6), 2464-2491. 2 Ni, X., Malevergne, Y., Sornette, D. and Woehrmann, P. (2011). Robust Reverse Engineering of Cross
Sectional Returns and Improved Portfolio Allocation Performance using the CAPM. Journal of Portfolio Management, 37 (4), 76-85. 3 Malevergne, Y., Santa-Clara, P. and Sornette, D. (2009). Professor Zipf goes to Wall Street. NBER
Working Paper No. 15295. 4 Zipf, G. K. (1949). Human behavior and the principle of least effort. Oxford, England: Addison-Wesley
Press. xi 573 pp. 5 Fama, E. F. and French, K. R. (1993). Common risk factors in the returns on stocks and bonds.
Journal of Financial Economics. 33 (1), 3–56.
4
5
Statement regarding plagiarism when submitting written work at the University of Zurich By signing this statement, I affirm that I have read the information notice on plagiarism, independently produced this Thesis, and adhered to the general practice of source citation in this subject-area. Der Verfasser erklärt an Eides statt, dass er die vorliegende Arbeit selbständig, ohne fremde Hilfe und ohne Benutzung anderer als die angegebenen Hilfsmiffel angefertigt hat. Die aus fremden Quellen (einschliesslich elektronischer Quellen) direkt oder indirekt übernommenen Gedanken sind ausnahmslos als solche kenntlich gemacht. Die Arbeit ist in gleicher oder ähnlicher Form oder auszugsweise im Rahmen einer anderen Prüfung noch nicht vorgelegt worden. Matriculation number: 11-746-823 _______________________ ___________________________________ place and date signature
6
7
Acknowledgements
Every project is a learning voyage, and this Thesis proved to be the most fascinating research
endeavor I have undergone in my entire academic career. During the time required, working
on this Thesis, a significant number of people contributed to the final success of my work.
Regardless of whether that was directly or indirectly, the final result would have not been
the same without the valuable input from my Professors, friends and colleagues.
From the ETH Zurich and specifically the Chair of Entrepreneurial Risks, I need to primarily
thank Professor Dr. Didier Sornette and Dr. Peter Cauwels for their pioneering ideas that
initially brought this project into life, and subsequently guided me through the research
process with their key remarks. I also need to thank them for inspiring me initially to engage
this project, even though I could foresee that working with the financial markets in such
depth, would pose a serious challenge to my engineering background. I also feel compelled
to thank Dr. Qunzhi Zhang from the same Chair, with whose valuable suggestions I managed
to optimize significantly my Matlab code, a fact that allowed me to perform all the
optimization tests required for my analysis within the deadlines.
In the University of Zurich, I am grateful to Professor Dr. Karl Schmedders from the Chair of
Quantitative Business Administration, for making the cooperation with ETH Zurich feasible
and for his assistance, whenever that was required. I would also like to thank Professor Dr.
Uschi Backes-Gellner and Simone Balestra, as the Management and Economics Program
Director and Coordinator, respectively, for all their valuable efforts that resulted in my
successful graduation. Furthermore, the guidance provided by Dr. Ioannis Akkizidis proved
to be valuable in my early steps in Financial Risk and Portfolio Management.
In personal level, I am deeply grateful to my family for supporting me for all this time.
Finally, I want to thank all my friends and colleagues in Portfolio Investments (Dow Chemical
Europe), for the insightful discussions and comments that were exchanged.
Picture 36: The official Wilshire calculator. ........................................................................... 113
15
Introduction
Levy and Roll, in their 2010 paper6 rejuvenated in full power one of the most debated issues
in modern finance; their results imply that there is still a possibility that the original Capital
Asset Pricing Model (CAPM) presented by Sharpe (1964)7, could describe real market
conditions, a notion which is widely believed to be incorrect, amongst economists.
The motivation of this Thesis is divided in two main parts; first, to fully challenge the Levy
and Roll approach of reverse engineering, by reconstructing the entire code in Matlab and
applying it in a variety of different portfolios, under different market conditions. The goal of
this process is to acquire a deep understanding of the behavior of this model, used by Levy
and Roll, and ultimately to conclude whether it indeed stands true or not. Second, a
different model is tested; the one proposed by the authors of the paper “Professor Zipf goes
to Wall Street”8. Based on the original idea discovered by Zipf (1949)9, we evaluate and test
further the financial use that was first proposed in this paper. As we will see in Chapter 6,
the results are remarkable.
In Chapter 1, an introduction to the financial markets as of the second half of 2013 is given.
Right after, in Chapter 2, a second introductory chapter follows with an insight to notions
such as the CAPM, the efficient frontier and the work of Levy and Roll. Hopefully, these two
chapters would provide enough justification in regard to whether portfolio management is
of such importance, given the special nature of the financial markets.
In Chapter 3, a short introduction to the Levy and Roll model is provided, along with the
fmincon function used (in Matlab) and the general concept of nonlinear optimization. The
purpose of this chapter is to prepare the ground for a more detailed analysis that will follow
in Chapter 4. In this chapter, the entire analysis of the Optimization Problem 1 is analyzed,
starting from its definition, continuing with its implementation and the data preparation,
and completing with a presentation of the results and conclusions. Similarly, Chapter 5 is
devoted to the Optimization Problem 2, by following the same structure as in Chapter 4.
In Chapter 6, the two-factor Zipf model is analyzed; again, from the perspective of the
Matlab process, as well as the data preparation. Finally, in Chapter 7 we draw the final
conclusions of the entire Thesis for both methods (Levy/Roll and Zipf). Chapter 9 includes
some suggestions of the author for further research on some issues that were not
exhausted, scientifically, in this Thesis. That could be either because they deviated
significantly from the main goals of this Thesis, or because there was simply not enough time
to explore all different ideas for testing the different models.
6 Levy, M. and Roll, R. (2010). The Market Portfolio May Be Mean/Variance Efficient After All. Oxford
Journals, Review of Financial Studies, 23(6), 2464-2491. 7 Sharpe, W. F. (1964). Capital Asset Prices: A Theory of Market Equilibrium under Conditions of Risk.
The Journal of Finance, 19 (3), 425-442. 8 Malevergne, Y., Santa-Clara, P. and Sornette, D. (2009). Professor Zipf goes to Wall Street. NBER
Working Paper No. 15295. 9 Zipf, G. K. (1949). Human behavior and the principle of least effort. Oxford, England: Addison-Wesley
Press. xi 573 pp.
16
In Chapter 9, the references on which this work is based on are provided, in order to
facilitate future research. Due to the nature of this project, the largest number of references
provided refers to website pages, especially from Mathworks.
Finally, a complete Appendix is included, divided in three sectors: first, the full description of
the data acquisition procedure from Bloomberg is provided. Second (sector B), the main
modules of the Matlab code are documented, along with the necessary comments for the
understanding of their basic functionality. In the third part of the Appendix, the lists with the
firm names are provided, which are the constituents of the portfolios used throughout the
tests of the optimization problems.
17
Chapter 1
1 Financial Markets
It has been well more than a century, since the first time that an economist tried to explain
the market moves and predict the next one, in order to consequently draft a –hopefully-
lucrative strategy. Such analysis was initially carried out by simply studying the
fundamentals, as well as basic economic notions such as supply and demand, in almost
isolated (from one another) markets. In today’s globalized financial system, the situation is
completely different. Enormous volumes of data are exchanged and analyzed every second
of each day by using the most sophisticated computer systems, in an attempt to evaluate
the current market position, and as such forecast the next move of the highly intercorrelated
markets; a zero-sum game, in theory, which sometimes creates a tremendous amount of
social turbulence.
In a system extremely volatile, which is often characterized as a “random walk” due to its
unpredictable nature, the procedure is usually the same: a new model is tested by using
(existing) historical data. If the model’s explanatory power is consistently high enough, in a
series of different time periods under different macroeconomic conditions, then it is
(considered to be) a good candidate as a predictor of future market moves. For that
purpose, the model is usually fed with a large number of data from one of the known data
providers (Bloomberg, Thomson Reuters, Wharton), analyzes them, and subsequently
produces some quant signals that the analysts of portfolio managers need to evaluate.
But the market “state” belongs usually in one of two different modes: the distressed or the
normal one. While there is no official definition for either state, it is broadly accepted that a
market is in distress (crisis) when we have “a situation in which the value of financial
institutions or assets drops rapidly“10. Usually, the subsequent financial event that takes
place, often called as an “extreme event”, has (or used to have) a possibility to happen so
small, that it was considered almost impossible. Regarding the cause of such a market
downturn, one could argue by studying historical data that sometimes a market crash is
based on an actual event such as a war, while other times it can be based on mere financial
conditions and practices, such as the subprime crisis of 2008.
In the graph below, we can see the most market crashes -by using as market proxy the S&P
500 Index- for the last approximately 100 years, where they are explained as the reaction of
As we can imagine, in general it is very difficult for any quantitative model, to be able to
predict accurately the outcome of the markets or a specific market move in the occurrence
of such extreme events. In addition, it is not always desirable for a model to include modules
able to measure the effect of a crisis on the assets, since that would make the model
severely biased and probably not performing well in “normal” market conditions. On top of
that, there are many factors affecting the performance of an asset or portfolio that cannot
be modeled in a quantitative way.
For all the above reasons, in most cases portfolio managers prefer to have their models
reflecting pure fundamental qualities of the firms they analyze, so that they will be able to
add an overlay analysis regarding for example possible political risks or sustainability - social
responsibility targets that need to be taken into account. By using both approaches
together, one is able to have the overall picture, which is the valuation of the assets based
on their own value, but co-calculating the macroeconomic conditions as well.
Last summer, the S&P 500 Index managed to surpass the peak levels of 2007 for the first
time since the occurrence of the subprime crisis. It took several years for the system to
recover from a crash, that for the first time it was widely accepted as a purely systemic one,
based only on incorrect valuations of assets and as such, without any reflection to shift of
real value.
11
185.Bond vigilantes. Index GPL. http://www.bondvigilantes.com/blog_files/UserFiles/Image/stock_market_crashes2.jpg, 19.04.2013.
19
Picture 2: The rally of S&P 500 Index in 2013.12
In the same spirit, the rally of most equity indices in year-to-date 2013, which broke the
upper trend line several times (even with the correction in May 2013), is considered to be
the result of the third Quantitative Easing (QE3) program of the United States Federal
Reserve Bank (FED) and does not necessarily proves that the macroeconomic conditions are
considerably more favorable than when the subprime crisis emerged.
Picture 3: The effect of the Quantitative Easing (QE) program on the S&P 500 Index.13
12
Source: Russell Investment Group, Standard & Poor’s, FactSet, J.P. Morgan Asset Management.
20
At the same time the economic conditions for most of Europe remain depressed, with the
Eurozone’s first-quarter GDP retreating for the sixth consecutive quarter. In Japan, another
heavy-weighted pole of the global financial system, the effects of a first time in history (with
regards to its size) Quantitative Easing program has been launched (Bank of Japan, BoJ) on
the country’s monetary policy are still not clear. In the meanwhile, “Abenomics” (the
measures introduced by the Japanese Prime Minister Shinzo Abe) have already created a
tsunami of consequences to all the peripheral economies, varying from drops in FX rates and
shifts in the imports/exports trade balance, to the capacity of core industries and actual
people’s jobs. Finally, in China, the first quarter growth rate of 2013 did not reach the
expectations for the first time in the last 10 years, and decelerated to approximately 7.5%,
instead of the double-digit numbers that were promised to the investors and the “Western
World”. This resulted in a sizable decrease in the demand of commodities, which was –by
many economists- the main driver of the drop of the prices of many basic commodities in
stock markets all over the world from the second quarter of 2013 onwards.
These developments resulted in the end of the broadly accepted perception of the financial
world during the recent years, that emerging markets are the main contributors of growth in
the global system, bringing back the economy of the United States of America under the
spotlight. Finally, in order to demonstrate the influence of the FED’s strategy in the domestic
market, as well as the influence of the U.S. economy to the global market, we will refer to
the effect of one statement made by FED's chairman, Ben S. Bernanke (May 2013): the
chairman announced that the FED intends to reduce its asset-buying program (of
approximately $80 billion per month) by the end of the year, putting an end in abundant
liquidity and therefore extremely low interest rates. The fear created by this “tapering of
QE” statement pushed the markets into a very nervous reaction, with an example being
provided below:
13
Financial Sense. Clues to watch for the End of QE “Infinity”. http://www.financialsense.com/contributors/lance-roberts/clues-watch-end-qe-infinity, 23.08.2013.
21
Picture 4: Increase of the U.S. 30 Year Mortgage Rate, as an effect of the FED Chairman’s statement.
14
The yield on U.S. 10-year Treasuries increased 1% during the first 7 weeks after the
announcement by Mr. Bernanke, the sharpest increase in such a short time period for the
last 50 years.
The above stated events and their consequences in asset pricing all over the world should
make clear that active portfolio management is key to whoever wishes to achieve the
highest possible performance of their portfolio of assets. Naturally, there are a lot more
factors that affect the assets’ prices, varying from regulatory changes to behavioral finance.
The focus of this Thesis is concentrated into two competitive theories; first, to add new
evidence in one of the most fundamental issues affiliated with the Modern Portfolio
Management Theory, the Mean/Variance Efficiency of the Market Portfolio. Second, to
demonstrate the explanatory power of the two-factor Zipf Model.
14
The Motley Fool. What “Tapering” Means for Middle America. http://www.fool.com/investing/general/2013/08/17/what-tapering-means-for-middle-america.aspx, 02.09.2013.
22
Chapter 2
2 Capital Asset Pricing Model (CAPM) Theory
In 1952, Markowitz presented the mean/variance approach15. Sharpe16 (1964) and Lintner17
(1964) presented in their work what would afterwards be known as the Capital Asset Pricing
Model (CAPM).
The whole idea can be described by the following equality:
ri,t = Rf,t + βi ( rm,t – Rf,t)
Where ri,t corresponds to the portfolio i’s (expected) returns, Rf,t is the risk free ratio for
each time period (usually monthly), and rm,t is the (expected) market return for the same
period t.
In words, that equality means that the return of a certain portfolio (which could be
constituted by only one security) can be explained by the risk free return plus a risk premium
(market excess return) multiplied by a factor called beta, which captures the sensitivity of
the return of the particular portfolio towards the market returns.
Therefore, somebody could apply this formula in known past returns of a certain portfolio,
obtain its beta, and under the assumption that its sensitivity to the market returns will not
change, use the above formula to calculate the expected returns of that portfolio. The idea
of a linear relationship between risk and returns is also introduced, meaning that the riskier
an asset is considered to be, the higher its returns will be.
The value of this theory, as it is used today, lies basically at the notion that a certain portfolio
is mean/variance efficient under two interchangeable conditions:
It provides with the maximum possible returns, given a certain amount of risk taken.
It bears the minimum possible amount of risk, given a certain return.
This theory turned to be the corner stone of Modern Portfolio Theory, which brings us to the
Efficient Frontier: the set of optimal portfolios that provide with the highest expected return
given a certain level of risk, or the minimum possible risk for a given level of expected
return. The portfolios that reside below the efficient frontier are called “sub-optimal”, since
15
Markowitz, H. (1952). Portfolio selection. The Journal of Finance, 7, 77-91. 16
Sharpe, W. F. (1964). Capital Asset Prices: A Theory of Market Equilibrium under Conditions of Risk. The Journal of Finance, 19 (3), 425-442. 17
Lintner, L. (1965). The valuation of risk assets and the selection of risky investments in stock portfolios and capital budgets. Review of Economics and Statistics, 47, 13-37.
23
they are characterized by lower return for any given level of risk; portfolios that reside on
the right of the efficient frontier are also called sub-optimal, since they bear more risk for
any given level of return.
The curved shape of the efficient frontier, in the space of the returns-risk graph of all
attainable portfolio combinations, pointed out the precious benefit of the diversification
that can be achieved, by choosing the right combination of portfolio constituents.
Picture 5: The Efficient Frontier.18
In order to achieve this most desirable diversification property for a portfolio, and therefore
reduce (ideally eliminate) the systematic risk19, the portfolio constituents need to be assets
with negative correlation, or at least very small (correlation). This can be observed in the
following simplified graph, where we can see that in the case of negative correlation, it is
possible to increase the portfolio (constituted by these two assets) returns while reducing its
risk (standard deviation of return) at the same time. In real market conditions it is rather
Systematic risk or undiversifiable risk is called the amount of risk that is considered to be “systemic”, i.e. cannot be eliminated by further diversification.
24
rare to obtain two assets with such a perfectly desired correlation, but it is still possible to
minimize the systematic risk of a portfolio.
Picture 6: Risk-Return graphs for varying correlation between two assets.20
Ever since, there has been an extensive research activity testing whether these theories
(which basically constitute the mean/variance efficiency of a portfolio) actually apply in real
market data. The efficiency of the market portfolio, except from being the corner stone of
modern finance, is a key subject regarding the debate of active versus passive investing and
whether the riskiness of various assets can be represented by their “beta” (β). Especially
since the last financial crisis, where many professionals in the industry underperformed the
various benchmark indices, there has been a tremendous amount of discussion regarding
whether they “deserve” their management fees, and if they actually create some “alpha”
(α).
Most of the studies have tested various market proxies, which have turned out to be far
from the efficient frontier; in addition, it has been proven that the portfolios on the efficient
frontier usually include short positions. That verifies the opinion than the market portfolio
(in which there cannot be such thing as a “short position”) is not efficient by definition.
To be noted, in the previous studies an extensive collection of sample/modified parameters
has been employed, by using various shrinkage techniques (we will come back to this at the
implementation of the model).
All these could ultimately mean that the portfolio managers and investment consultants
could be dismissed, and the investors could simply invest in an index like the S&P 500, or
even a blend between an index and some risk free assets (like T-bills or German Bunds).
20
IndUS Business Journal. Asset Allocation and the Efficient Frontier. http://www.indusbusinessjournal.com/ME2/dirmod.asp?sid=CA41EB12BCD14CBBB683703875E204A5&nm=&type=Blog&mod=BlogTopics&mid=1AE0C859CE3C43E0B50B7615AAC20015&tier=7&id=958C2D62F6014598A76C3D623A997EE6, 22.08.2013.
25
Amongst the professionals in the industry, it has been accepted for years now that the
CAPM does not really work, therefore there is a series of alternative models, whose base is
the CAPM.
Ross’s (1976) Arbitrage Pricing Theory (APT) is one of them. In general, this model supports
the theory that “the expected return of a financial asset can be modeled as a linear
function“21 of a chosen number of factors, which can be macroeconomic ones and/or market
indices. The coefficient of each factor represents the sensitivity, or “beta” (β), of the asset’s
return to the changes of the value of this factor. After these betas have been calculated, the
model can then be used for pricing this asset.
APT is considered to be one of the basic after-CAPM models, but the most prominent one is
the three-factor Fama-French model (1993). This model expands the CAPM, by having two
additional factors, the “SMB- small (capitalization) minus big” and the “HML- high (book-to-
market) minus low”. More on this model will be provided in Chapter 6.
More recently, in 1993, another model was presented by Cahart, which introduces the
momentum effect; the inclusion of the momentum effect that gains more and more ground
against competitive models, especially in the explanation of returns of hedge funds and
funds of funds.
The first core area of interest of this Thesis lies in the examination of the explaining power of
CAPM, or better, in the determination of whether the market portfolio is mean/variance
efficient; a notion which is considered by the majority of the economists worldwide as not
true. In that case, CAPM would be just a mere pedagogical exercise for finance students,
that does not reflect any actual economic conditions.
In order to accomplish that, a rather reverse approach is being employed. This approach was
first introduced by Levy and Roll in 201022, and was further expanded by Ni, Malevergne,
Sornette and Woehrmann (2011)23.
Their approach tackles the problem from the opposite perspective; most researchers prior
to Levy and Roll implemented the “usual” technique of gathering returns from different
portfolios, market proxies, etc. and feeding them into the CAPM model, in order to
determine whether it holds true. The brilliant idea that basically constitutes the paper of
Levy and Roll is the calculation of the return parameters and standard deviations of a sample
portfolio, so that they would satisfy the mean/variance efficiency constraints.
More precisely put, that means that for a certain portfolio of equities, the authors
determined what the return and standard deviation of each individual stock should be, so
that the entire portfolio would be mean/variant efficient. In order to estimate these returns
21
Arbitrage pricing theory - Wikipedia, the free encyclopedia. http://en.wikipedia.org/wiki/Arbitrage_pricing_theory, 21.08.2012. 22
Levy, M. and Roll, R. (2010). The Market Portfolio May Be Mean/Variance Efficient After All. Oxford Journals, Review of Financial Studies, 23(6), 2464-2491. 23
Ni, X., Malevergne, Y., Sornette, D. and Woehrmann, P. (2011). Robust Reverse Engineering of Cross Sectional Returns and Improved Portfolio Allocation Performance using the CAPM. Journal of Portfolio Management, 37 (4), 76-85.
26
and standard deviations, they used a quantitative model in Matlab (more on this model
under the “implementation” Chapters 4 and 5). That way, they acquired a second set of
returns and standard deviations for each portfolio (by “first set” we mean the actual values,
or sample values of the portfolio), which they compared statistically with the first set of
values. The results were astonishing for the scientific community: the difference between
the values of each set, the sample and the calculated (mean/variance efficient) one, was
lying within the statistical error margins; that would mean that the CAPM theory in terms of
explaining the returns of the particular portfolio (and therefore, market proxy) cannot be
rejected. Details of how this second set of values is calculated, and how the statistical
comparison is being made, will be presented in the respective chapter.
To be noted, a great debate regarding the validity of any testing of the CAPM theory has as
its main subject the validity of any market proxy, from the market proxies that are available
to the researchers. That is because, as many economists support, for a definite test of the
respective theory, one should include all assets, such as human capital, real estate, etc. So
far, there has not been a satisfying approach capable of incorporating equivalent factors in a
model.
27
Chapter 3
3 The Levy/ Roll approach
As stated in the previous chapter, Levy and Roll applied a reverse-engineering model in
order to calculate the values of a pair of parameters (return and standard deviation) of each
stock for a given portfolio, and afterwards compare it with the actual (sample) values of this
stock. Two optimization problems were used by the authors for this procedure, where the
second one is significantly more constrained than the first.
In their approach, the main issue is the calculation of the parameters (returns and standard
deviations) which, given a sample portfolio, will satisfy the mean/variance efficiency
conditions. In this chapter, we will attempt to provide a short introduction of the two
Optimization Problems (Optimization Problem 1 and Optimization Problem 2), so that it will
be possible later in our analysis to elaborate further on them, and build on their
characteristics. The mathematical and programming details for these two Problems will be
analyzed in detail in the two chapters devoted to them.
Regarding Optimization Problem 1, its sole objective is to find a set of μ , σ vectors that
satisfy the nonlinear equation constraint and at the same time make sure that the objective
function f(x) (which is the function D((μ,σ),(μ,σ)sample) in our case) has a value as low as
possible.
We can see the mathematical description of the Optimization Problem 1 right below (all the
parameters are explained after the presentation of Optimization Problem 1):
Minimize (objective function):
Subject to (nonlinear mean/variance constraint):
28
As we will see in Chapter 5, the Optimization Problem 2 is a more constraint version of the
Optimization Problem 1. The objective function and the first constraint remain the same as
in Optimization Problem 1; the additional constraints correspond to the notion that there
might be a predefined (mean) return μ0 and standard deviation σ0 of the estimated
portfolio’s values, which we impose as extra constraints to the initial optimization problem.
For that purpose, the Optimization Problem 1 is equipped with two more constraints, a
linear equation for the returns, and a set of nonlinear equations for the standard deviations.
The Optimization Problem 2 can be formulated as:
Minimize (objective function):
Subject to:
a) Nonlinear mean/variance constraint:
b) Liner constraint for estimated portfolio returns (μ):
c) Nonlinear constraint for estimated portfolio standard deviation (σ):
29
For the rest of this Thesis, we will try to use a coherent naming of the various basic
parameters. There are constituted by two main sets, the sample parameters and the
adjusted (calculated from Matlab) ones.
For the sample parameters, we have μsample and σsample which are vectors constituted
from the return and standard deviation, respectively, of each stock of the sample
portfolio.
For the adjusted parameters, we have μ and σ which are vectors constituted from
the return and standard deviation, respectively, of each stock that have been
calculated via the Matlab procedure.
Furthermore, the variable N corresponds to the number of stocks in the portfolio being
used, rf (or Rf) the risk free ratio, q the constant of proportionality and α (alpha) a parameter
defining the relative weight given to the deviations of the mean returns (versus the weight
given to the deviations of the standard deviations).
By following the most popular approach, the risk free ratio used is the 3-month T-bill rate.
Given the fact that the Levy and Roll Models do not have a time dimension (the naïve
averages are used for the returns of each stock), the same naïve average was calculated and
used for the risk free ratio. As we will see later on, for the period 2003 to 2013 for example,
that was calculated as approximately 1.5% (precise calculations are following, right before
the presentation of the tests results).
The formula for calculating q, reads:
Where σmarket and Rmarket are the market standard deviation and (average) return,
respectively.
Regarding the selection of α (for which holds 0 ≤ α ≤ 1), it has been proven that the model
returns the most robust results when α varies between 0.5 and 0.75. Indeed, the value of α
used in this Thesis is within this range, quantified by 0.6 (its effect is examined in several
optimization tests).
In addition to the previous defined parameters, we also have the vector xmi which represents
the weight of stock i, for each i from 1 until N, of the portfolio. The weights of the stocks,
following the previous work by Levy/Roll24 (as well as almost all available relevant studies),
are calculated based on the market capitalization of each firm. There are several different
approaches as to the market capitalization of which date should be used. In the first tests
(several portfolios constituted by different number of top market capitalization S&P 500
INDEX stocks) the simple average of the market capitalization of each stock was used; later
on, in the first 100 from S&P 500 Index and the S&P 1200 Index, as well as in all the Fama
24
Levy, M. and Roll, R. (2010). The Market Portfolio May Be Mean/Variance Efficient After All. Oxford Journals, Review of Financial Studies, 23(6), 2464-2491.
30
French sets, the market capitalization of each firm as of the last month of the time period
under examination was used.
The formula that gives us the weight xi for every stock i is simply:
The intuition behind D((μ,σ),(μ,σ)sample), where “D” stands for “Distance”, is a mathematical
way to represent how close the sample values are, in relation to the ones calculated by the
Matlab procedure. As expected, the closer the estimated values are to the sample ones, the
lower is the value of D. As we know from statistics, the largest the standard deviation (of a
stock’s returns), the largest the statistical error will be from calculating the estimated
parameter for the same stock, and the larger the confidence interval as well. This is why
they are divided by the sample standard deviation, as an attempt for normalization.
The set of nonlinear equations represents the constraints necessary so that the set of stocks
will satisfy the mean/variance efficiency conditions. For this optimization problem, the
overall number of nonlinear constraint equations is equal to the number of stocks in our
stock set.
Both optimization problems and their constraints will be discussed in detail in the respective
chapters (Chapters 4 and 5 for Optimization Problem 1 and 2, respectively).
Having somewhat described the parameters used, we will see in paragraph 4.6 what
“reasonable close” parameters means, when it comes to comparing the sample and adjusted
parameters, for the stocks of each portfolio. After the calculation of the mean/variance
efficient parameters for each set of stocks, we would accept them as “reasonably close” to
the sample set’s, if at least 95% of the parameters calculated are within the 95% confidence
intervals of the sample parameters. Analytical tables with the statistical results of these
comparison tests and their thorough explanations will be provided at the respective sections
with the results of the optimization tests.
3.1 Implementation in Matlab
For the needs of this Thesis, the entire Matlab code was hardcoded from scratch25, for two
main reasons. First, we needed to make sure that previous coding attempts of the same –or
similar- problems will not influence our approach, which tries to be as accurate to the
original mathematical problem description as possible. Second, and as we will see in the
chapters to come, there are a lot of calibration issues and assumptions that need to be
applied when executing the several functions of the Matlab code, for each specific portfolio
under examination. Hence, only by reconstructing the entire model (in terms of
programming) it would be possible to gain a deep understanding of all these subtle issues
25
With the exception of the Ledoit/Wolf shrinkage technique, as we will see in the respective paragraph.
31
that affect the performance of the model, and result in a stronger conclusion regarding its
validity.
The code is viewed as a tool which would be used to calculate the estimated parameters, so
that we would test the model fit. As such, not much effort was spent in the optimization of
the code itself, except from one very important aspect: the speed of the optimization for
each portfolio, and more specifically, the execution time required for each run. Initially, as
the code was not optimized at all in terms of execution speed, even an execution over a
small portfolio required a tremendous amount of time. To put things into perspective, the
(continuous) execution time in a personal pc for each optimization test with just 10 stocks
was approximately 2 hours, with 20 stocks it was more than 5 hours, and with 50 stocks it
was almost 30 hours; all these, even when the initial guess (x0) was very close to the actual
result. Since it was not possible to conduct all required tests within the time schedule of the
Thesis given these execution times, with the significant help of Dr. Qunzhi Zhang (ETH Zurich,
Chair of Entrepreneurial Risks), the code was optimized (time-wise) and the time frame of 5
hours became approximately 10 seconds; mainly because of eliminating the import-export
process of excel spreadsheets into Matlab, for each iteration. That was the main action
undertaken regarding the performance of the code. Of course, it could be further enhanced,
but since such actions would have absolutely no impact on its functionality, and almost none
on its execution speed, no further action was taken.
We will continue with a short description of the main Matlab function that was used (called
fmincon) throughout the evaluation of the entire Levy/Roll approach, before we come back
to the specifics of the code.
3.2 The “fmincon” Function
For the calculation of the adjusted parameters sets, all the work has been implemented in
Matlab environment, and more specifically by using the fmincon function as the key module.
The descriptive definition of this function, as given by the Matlab documentation, is to find
the minimum value of a certain function f(x) provided by the user, under certain linear
and/or nonlinear constraints (also provided by the user):
c(x) ≤ 0
ceq(x) = 0
A·x ≤ 0
Aeq·x = beq
lb ≤ x ≤ ub
32
Where x, b, beq, lb, and ub are vectors, A and Aeq are matrices, c(x) and ceq(x) are functions
that return vectors, and f(x) is a function that returns a scalar. The functions f(x), c(x), and
ceq(x) can be nonlinear.26
In our case, f(x) is our objective “Distance” function and the vector x corresponds to the
returns and standard deviations of each individual stock for each portfolio. As we will see in
detail further down this Thesis, for the Optimization Problem 1’s constraints, we have only
one set of nonlinear equations, with the actual number of equations being equal to the
number of the stocks of the portfolio; for the Optimization Problem 2, we employ an
additional set of nonlinear equation constraints (with the number of equations again equal
to the number of stocks in the portfolio), as well as a linear one (a single equation,
regardless of the number of stocks).
3.3 Nonlinear Optimization
In order to gain a basic understanding of how the function fmincon produces the results, we
will attempt to describe shortly the basic framework around it, as well as the principles of
nonlinear optimization in similar problems in general.
Based on the documentation available by Matlab, fmincon begins the optimization
procedure by calculating the f(x) function (which must be minimized) for the values of an
initial vector x0 provided by the user. For these initial values, fmincon calculates the value of
f(x), as well as the maximum constraint violation. The maximum constraint violation is
basically the maximum violation that has occurred amongst the constraint functions, given
the values of x0 (at the first run of the function).
For example, an execution of the code implemented for the Optimization Problem 1, for the
portfolio from the 25 Fama-French portfolios (of the entire U.S. equities market-
NYSE/AMEX/NASDAQ), is demonstrated below:
26
Description of the variables taken from the official Mathworks page for the function fmincon: http://www.mathworks.com/help/optim/ug/fmincon.html, 21.08.2012.
Picture 8: The Maximum Constraint Violation being minimized, during the optimization process.
We observe a similar pattern as with f(x), although with less fluctuations (more
monotonous). What is particularly interesting, from the intuitive point of view, is that in
several cases that one of the two graphs has a steep jump (upwards), the other one has the
opposite first derivative: it plunges. That reveals the way fmincon works, since it is not
possible for the function to minimize simultaneously the objective function and the
maximum constraint violation. Therefore, it takes turns in minimizing the one or the other,
and then comes back and evaluates them to determine in which direction it should continue
(in terms of configuring the x vector, i.e. the stocks’ returns and standard deviations).
Now, let us assume that we want to find the (global) minimum of the Peaks27 function:
z = 3*(1-x).^2.*exp(-(x.^2) - (y+1).^2) ...
- 10*(x/5 - x.^3 - y.^5).*exp(-x.^2-y.^2) ...
- 1/3*exp(-(x+1).^2 - y.^2)
27
Mathworks. Peaks Minimization with MultiStart. http://www.mathworks.com/matlabcentral/fileexchange/27178-global-optimization-with-matlab/content/html/msPeaksExample.html, 25.10.2012.
38
The plot of this function is displayed in the following graph:
Picture 9: The plot of the Matlab function Peaks.
We can see multiple local minimums in this plot and only one global minimum. In order to
have a more clear visualization of the problem that fmincon is called to resolve, we can see
the different mountain-valley tops in the following picture.
39
Picture 10: Different starting points resulting in different end points.28
From the different isoheight lines, we can see that if fmincon does not start from the
appropriate initial point, it is possible that it will not return the global minimum, but a local
one. Similarly, if the first derivative of f(x) gets a very small value (below the given TolFun),
then fmincon considers that point as a minimum, even if the function continues further (but
slowly) to a lower value point.
As a last attempt of making absolutely clear how the fmincon function works, even to
somebody that has no previous experience on the whole subject, I find it of use, to cite here
a creative description of how the function works. Even though it is not a pure scientific
explanation, but it feels like the beginning of a novel, after studying its behavior through
hundreds of experiments with it, I believe that the following passage captures the
fundamental workflow of the algorithm with great success:
“Think of fmincon as a blind man trying to find the bottom of a valley. He can stop at any
point and determine the local gradient. From there, he will choose where to search next, as
he also has some idea of the local shape (curvature) of the surface given his previous
meanderings. This is an iteration.
Within each iteration, he will try a few new points before updating his estimate of the local
gradient, which will start a new iteration.”29
At this point, we will continue our analysis with Chapter 4, where the detailed investigation
of Optimization Problem 1 begins.
28
Mathworks. Peaks Minimization with MultiStart. http://www.mathworks.com/matlabcentral/fileexchange/27178-global-optimization-with-matlab/content/html/msPeaksExample.html, 25.10.2012. 29
Mathworks. Thread Subject: Questions about fmincon options. http://www.mathworks.com/matlabcentral/newsreader/view_thread/237130, 26.10.2012.
40
Chapter 4
4 Optimization Problem 1
4.1 Methodology
As already mentioned in this Thesis, while previous researchers tried to determine if the
Levy/Roll model works, which translates to whether the market portfolio can be
mean/variance efficient, the goal is to further examine:
a) If the model indeed works, but more specifically;
b) Under which exact conditions the model works, as well as;
c) Why it works.
For that purpose, while going through the presentation of the results for each optimization
test, we will abandon the previous practice, in which the Levy/Roll process was viewed as a
“black box” which just produced results (but there was no effort spent in describing the
specifics of the Matlab models themselves).
The formulas for Optimization Problem 1 are repeated here, for the kind convenience of the
reader.
Minimize (objective function):
Subject to (nonlinear mean/variance constraint):
The objective function D is constructed in Matlab by basically using the “sum” function.
Initially, a script (script v2.4m in the Appendix) was created with all necessary commands for
the full implantation of Optimization Problem 1, but by using the Matlab Symbolic
(variables) Toolbox. The purpose of this exercise was to fully understand the code and verify
that it functions exactly as expected, without having to use as inputs actually stocks’ returns.
41
Every time the fmincon function is called, in every iteration it evaluates (in the sense of
calculating the value) the D, and for the same values of the vector x, it also evaluates each of
the constraint equations. The nonlinear constraints have been set in a form (all the right
hand side to the left one) so that they will result in a right hand side equation, which must
be equal to zero, and this is what the fmincon checks (in each iteration). At the “max
constraint violation” field, it returns the largest value of these equations (largest distance
from 0), which is named maximum constraint violation.
As the program is running, in each new iteration, if that maximum violation has a value less
than the default (10-6 or defined by the user), for three or more consecutive iterations, and
the predicted change in the objective function D is less than the respective threshold (10-6 or
defined by the user), then the fmincon stops the optimization and returns the result (the
adjusted parameters that it estimated, the value of the objective function, the stopping
criteria, etc.).
The criterion by which fmincon decides if D can be minimized even more, is the TolFun
threshold. Also 10-6 by default, if by each iteration (and therefore change of the objective
function’s input parameters) the difference (towards a smaller value) of the objective
function is less than this threshold, then this value of D is considered to be the minimum
one. After fmincon reaches the condition of the minimization of the objective function, it
would try to minimize even more the maximum constraint violation (without increasing D),
and once this is satisfied as well, it will stop.
We can see a simple representation of this in the following graph:
Picture 11: Tolfun, Tolx.30
If the maximum violation has a value above the TolCon threshold, then fmincon will return
the values that correspond to the minimum constraint violation, unless if the steps of
changing the value of the objective function are very small, but not small enough so that the
program will terminate. In this case, we can observe a very large number of iterations (when
30
Mathworks. Matlab Central. Fmincon, Tolx and Tolfun: How they work. http://www.mathworks.com/matlabcentral/newsreader/view_thread/33662, 21.03.2013.
42
the relevant ‘MaxIter’ setting is set to “infinite”), where fmincon basically goes back and
forward with the values of the estimated parameters, without being able to terminate.
At this point, we understand how important is the selection of the initial point, from which
fmincon begins the optimization, and therefore the evaluation of the objective function.
In similar problems faced in the industry, a linearization of the entire optimization problem
is often attempted, so that the whole procedure will give more robust results. In the context
of this Thesis, a lot of effort was consumed in order to make sure that the best possible
approach is being used, in order to avoid the local-global minimum problem.
Matlab provides the built-in options of Multistart and Global Search31 in order to address
such problems, but in our case we have a great advantage: based on our approach, we know
that the values for the returns and standard deviations that Matlab will return, should be
very close to their sample counterparts. That means, that if we start the optimization with
initial points the vectors of the sample returns and standard deviations, the optimization will
most likely begin from a point in this 2N-D dimensions space, which is really close to the
actual global minimum for our objective function. In other words, it is impossible for our
given problem, to identify a vector x’ of 2N variables that is “far away” from x0, and at the
same time if it is given as input to the objective function, D(x’) will result in a value lower
than the D(x0).
Therefore, in the data sets used to run the optimization problems, the following tests were
conducted, in order to make sure that the minimum acquired is actually (or better, most
likely) the global minimum one.
In a number of stock sets, first we ran the optimization with the sample vectors indeed as x0.
From doing so, we got a vector x*, with the estimated parameters for the same stocks,
returns and standard deviations. We fed these estimated parameters as input in the
objective D, and calculated the “distance” (D(x*)), overall, between these estimated
parameters and the sample ones. That value was saved, and we ran again the optimization
for x0 larger, than the sample values. Then again, we fed the new set of estimated
parameters as input in the objective D, and calculated the “distance”. This procedure was
repeated several times, also for smaller values than the sample ones, as well as for mixed
(some smaller, some larger) ones.
By comparing all the different D(x*), with x* the set of returns and standard deviations from
each run of the optimization mentioned above, it appears that the estimated parameters x*
that we estimated from the first time we ran the model, give us the lowest value for the
objective function. In other words, it always holds that:
D(x0) ≤ D(x*), ∀ x* ≠ x0
With this simple exercise, we consider that the area around the values of x0 was scanned for
resulting in possible lower values of the objective function, and therefore we concluded that
31
Mathworks. Global Optimization Toolbox. Global Search and Multistart Solvers. http://www.mathworks.com/products/global-optimization/description3.html, 20.03.2013.
43
our D(x0) with x0 the sample parameters, was indeed the minimum one. Of course, all the
x*’s that were returned satisfy all constraints. In fact, although some different local
minimums were found, in some occasions, a general multi-dimensional monotonicity was
observed. Most of the times, the problem would converge to the exact same set of values as
the first results (with the sample parameters as x0, and as expected, the further away the
starting vector from it, the more time and iterations would take for fmincon, to “return” to
the global minimum (or to stop to another, local minimum, which would give a larger
distance). Therefore, from now on, for each market proxy we use, we will consider the
parameters that are returned by fmincon with x0 as initial point, the ones that lead to a
global minimum for the objective (distance) function.
The conclusion is that, since the final estimated parameters should be “close” to their
sample counterparts, there should not be a vector with values “far away” from the x0’s ones,
which will return a smaller distance function value.
4.2 Covariance Matrix Shrinkage Methods
There is one particular problem that needs to be addressed, by anybody that is doing
research that includes simulating portfolios with large numbers of stocks; if the number of
observations (in our case, monthly returns) is smaller than the number of free variables (in
our case, the returns and standard deviations for each stock) then the covariance matrix
faces the issue of singularity.
Across literature, several methods have been implemented in order to deal with this, with
most prominent being the following ones32:
a) Shrinkage approach of Jagannathan and Ma (2003)
b) 1/N Portfolio by DeMiguel, Garlappi and Uppal (2007)
c) Shrinkage method of Ledoit and Wolf (2003, 2004)
d) Factor Portfolios (several authors)
In this Thesis, following the originally mentioned paper33 we will use the shrinkage method
presented by Ledoit and Wolf (2003)34. Their code was modified slightly, only for the
purposes of fitting in the original model hardcoded for the Levy/Roll approach; its
functionality was not changed:
(The code is provided here as implemented in all the funCeq.m functions; no separate
function for the shrinking of the covariance matrix was necessary.)
32
The papers in which details of all methods mentioned here can be found, are provided in the Chapter 9 (References). 33
Ni, X., Malevergne, Y., Sornette, D. and Woehrmann, P. (2011). Robust Reverse Engineering of Cross Sectional Returns and Improved Portfolio Allocation Performance using the CAPM. Journal of Portfolio Management, 37 (4), 76-85. 34
Ledoit, O. and Wolf, M. (2003). Honey, I Shrunk the Sample Covariance Matrix. UPF Economics and Business, Working Paper No. 691.
As we can see, we could simplify the explanation of the effect of the Ledoit/Wolf procedure
on the covariance matrix, by stating that while it leaves the variances unchanged, it
increases the lower covariances and decreases the higher ones, in an attempt to reduce
their difference from the mean covariance in a non-proportional way. As mentioned by the
authors of the original paper, the biggest statistical challenge of the entire process is to
identify the optimal shrinkage intensity; details of this method can be found in the original
paper35.
To highlight the importance of such a procedure, based on the 1/N method by DeMiguel,
Garlappi and Uppal36, a portfolio of 25 assets would require (without shrinkage) at least
3000 months of returns for any statistical use of the stocks’ covariance matrix, and a
portfolio of 50 stocks would require 6000 months of returns. As we will see in the relevant
chapter in the Appendix regarding the data acquisition procedure from Bloomberg, such a
case would make any research attempt in past returns of stocks ex ante completely
infeasible.
4.3 Definition of the Dependent Variable in all Matlab Functions
The first issue that had to be solved is the fact that fmincon accepts only a vector x as
variable, while in the optimization problems we have at least37 two sets of vector-variables,
one for the returns and one for the standard deviations. As such the x used be fmincon as
free variable, and of course the very important x0, is a vector of 2·N values, where the first
35
Ledoit, O. and Wolf, M. (2003). Honey, I Shrunk the Sample Covariance Matrix. UPF Economics and Business, Working Paper No. 691. 36
DeMiguel, V., Garlappi, L., Nogales, J. and Uppal, R. (2007). Optimal Versus Naive Diversification: How Inefficient is the 1/N Portfolio Strategy? Rev. Financ. Stud, 22(5), 1915-1953. 37
Across literature, one can see different approaches. In the first part of this Thesis, our x used for fmincon is constituted only be returns and standard deviations.
46
N (1, 2, .., N) values are the returns of the proxy’s stocks, and the second N (N+1, N+2, .., 2N)
represent the standard deviations of these stocks. A very important notice given the nature
of the algorithm, is that the two parameters (return and standard deviation) of each stock,
have the same index in both sets. That means that if the return of a specific stock has the
index 45, for example, then its standard deviation will also have the index 45. Therefore, at
the vector x of fmincon, the same stock’s return and standard deviation will be located at
the places 45 and N+45, respectively.
4.4 Flowchart of the Matlab code
In this paragraph we will provide a simple diagram of the Matlab code, along with a short
description of the main functions.
Picture 12: Matlab code flowchart.
As we can see in the simplified diagram above, for every optimization test (Optimization
Problems 1 & 2) we need to define an objective function to be minimized (named funD.m in
all our tests) and one function that includes all the nonlinear constraints. Depending on the
implementation, that can be one of the funCeq, funCeq2 or funCeq3. In addition, some
variations of funCeq2 were used for Optimization Problem 2, all of which are explained in
detail in Chapter 5. Finally, the entire Matlab code can be found in the relevant Appendix
chapter.
4.5 Construction of the 25 Fama French Portfolios
The ultimate challenge in regard to the power of the Levy/Roll theory is to test it on the 25
Fama French portfolios structure. These portfolios are known throughout the literature, and
have the name of the researchers that first formulated them, because of their ability to
capture, or even better reveal, crucial inner characteristics of any given model that is being
tested on them. In this Thesis, they are considered to be a very important test for the
validity of the model (although, as we will see further on, the objectivity of the results of
fmincon
funCeq.m funCeq2.m funCeq3.m
funD.m
47
such a test can be challenged), and as such, great weight is given to the performance of the
model when testing it on these portfolios.
These 25 Fama French portfolios were formed from data downloaded from
Bloomberg.38There are a lot of issues with missing data from this database, especially when
one tries to download several thousands of stocks for which continuous monthly returns and
other Bloomberg fields are required, over a large number of years (for example, from 1991
until 2009). For that reason, the processing of the raw data downloaded from Bloomberg,
until they were in appropriate format and ready to be used as input to the Matlab
optimization problem, is being described as follows:
38
The entire process of data acquisition from Bloomberg is described in detail in the Appendix.
48
Picture 13: Data acquisition and analysis flowchart.
Downloading of raw data from Bloomberg.
Preparation in excel, with returns, Market Capitalization and Market to Book
Bloomberg fields.
Execution of a small VBA script to get only the stocks with full continuous values for
the given time period.
Execution of a C# script to sort the remaining stocks in two dimensions,
Market Capitalization and Book to Market.
Splitting the whole set of stocks left into the 25 Fama French portfolios, based on
the output of the previous step.
Running the MATLAB optimization procedures for every one of the 25 sets, in
order to get the estimated parameters.
Analyzing the "statistical equality" of the estimated parameters with the sample
ones, for each of the 25 sets.
49
After the acquisition of the raw data from Bloomberg, the form of the data is shown below.
Here the monthly returns (in percentage points, %) of a random selection of alphabetically
concequitive stocks from the NASDAQ Index; the columns correspond to the time series of
the stocks’ monthly returns. The empty cells correspond to missing values, as downloaded
from Bloomberg:
OBAF US EQUITY
OBCI US EQUITY
OPTT US EQUITY
ORIG US EQUITY
OSHC US EQUITY
OCFC US EQUITY
OCLR US EQUITY
OFED US EQUITY
OCLS US EQUITY
OCZ US EQUITY
-9.09 3.42 -10.91 13.04 -23.23 0.39
-15.00 6.75 0.93 4.95 7.62 -22.73
-23.53 -6.83 -2.11 2.09 -9.87 3.03
0.00 -15.25 -6.20 5.64 -26.24 19.90
-15.38 25.76 5.03 -0.97 -40.06 10.95
18.18 12.15 1.98 11.76 49.07 -3.01
0.00 -17.70 8.32 6.58 -5.96 -12.01
7.69 11.85 0.78 -8.85 -6.67 7.76
-21.43 -17.68 -4.11 6.77 13.21 1.34
9.09 -6.59 -2.14 8.08 0.32 -34.21
0.00 -2.08 2.19 11.02 31.60 -18.00
37.50 -20.38 2.86 -0.14 16.61 5.85
39.39 -6.27 -0.17 -3.93 5.53 30.41
4.35 -0.66 1.30 5.58 -8.74 -10.60
-8.33 -2.87 5.41 6.20 21.70 -17.39
-4.55 -4.19 -0.65 -1.00 16.26 -4.55
-14.29 -0.96 3.20 0.34 4.96 -39.60
22.22 -34.25 0.08 -4.83 36.68 19.50
9.09 10.38 3.17 0.00 -36.16 -0.69
12.50 10.40 -0.23 4.96 -36.78 -33.57
-18.52 -2.45 7.41 -12.73 -47.37
24.09 0.77 -2.77 -11.31 -54.00
-15.38 3.53 -0.41 6.71 210.87
0.00 4.17 -2.31 1.26
0.00 -1.82 -25.25 -2.80
-9.09 -0.07 8.11 28.75
15.00 -2.89 -2.53 0.99
8.70 -1.15 9.05 -28.26
-8.00 1.85 5.14 -10.27
-8.70 -2.96 -6.02 -13.36
-4.76 -14.06 -4.03 0.88
20.00 1.82 -3.36 -6.11
2.81 1.25 10.61 4.65
10.64 -3.44 12.30 16.89
Table 4.1: Step one, raw data downloaded from Bloomberg (before deleting the columns with
incomplete data).
50
Each row represents one month; not the full time period is depicted here. In our analysis we
can use the stocks for which we have returns (for example) for all months within the time
period under consideration, therefore by using the following VBA script we delete all the
incomplete ones (parts of the code were found in the forum39 cited):
The sole purpose of this small script is to delete the columns (each column represents a
different stock) for which there is even one value missing (for each row we have the value
for one month, where the oldest date is on top).
Now that we are left only with the stocks for which we have continuous monthly returns for
the desired time period, in order to populate the 25 Fama French portfolios we need to
perform a “two dimensional sorting”, so that we will realize which stock belongs to which
quintile, based on the methodology given by Fama and French.40 The break values for
market capitalization and book-to-market used in this Thesis are the very popular ones in
the literature and previous research work, taken from the database41 maintained by
Professor Kenneth French42.
The procedure followed for their calculation, as described by Professor Kenneth French,
started with the gathering of data from NYSE stocks, for the end of June of every year T.
These stocks were initially sorted by market equity (ME-size, the number of stocks
outstanding times the current stock price). Afterwards, they were also sorted by book-to-
market equity (BE/ME, with BE being book value and ME the market value); the market
The exact same procedure was followed in the paper on which we have been referring before: Ni, X., Malevergne, Y., Sornette, D. and Woehrmann, P. (2011). Robust Reverse Engineering of Cross Sectional Returns and Improved Portfolio Allocation Performance using the CAPM. Journal of Portfolio Management, 37 (4), 76-85. 41
Dartmouth College. Current Research Returns. http://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library.html, 13.03.2013. 42
Kenneth R. French is the Roth Family Distinguished Professor of Finance at the Tuck School of Business at Dartmouth College. http://mba.tuck.dartmouth.edu/pages/faculty/ken.french/curriculum_vitae.html, 20.09.2012.
Sub MaxRows()
For i = 1 To 6000
If Sheets(1).Columns(i).End(xlDown).Row <
Sheets(1).Cells.SpecialCells(xlCellTypeLastCell).Row Then
Columns(i).EntireColumn.Delete
i = i - 1
End If
Next i
End Sub
51
value of a stock is defined as the ME at the end of December of the year T-1, and the book
value is the BE for the fiscal year ending in calendar year T-1.
All the possible values for market value and book-to-market are then split into five quintiles
each, the intersection of which provides the 25 Fama French portfolios.
The new script, which will sort the stocks into these 25 Fame French portfolios, was
implemented in C# and it is presented as follows:
using System;
using System.Collections.Generic;
using System.Linq;
using System.Text;
using Excel = Microsoft.Office.Interop.Excel;
namespace Excel_Ted
{
class Program
{
static void Main(string[] args)
{
Excel.Application excelApp = new Excel.Application();
Console.WriteLine("The BtoM " + i_str_TEDs + " value is " + fl);
if (fl <= 0.358)
excelWorksheet.get_Range('E' +
i_TEDs.ToString()).Value = 1;
52
else if (fl <= 0.562)
excelWorksheet.get_Range('E' +
i_TEDs.ToString()).Value = 2;
else if (fl <= 0.798)
excelWorksheet.get_Range('E' +
i_TEDs.ToString()).Value = 3;
else if (fl <= 1.102)
excelWorksheet.get_Range('E' +
i_TEDs.ToString()).Value = 4;
else if (fl <= 6.905)
excelWorksheet.get_Range('E' +
i_TEDs.ToString()).Value = 5;
j_str_TEDs = 'C' + i_TEDs.ToString();
Excel.Range excelCell2 =
(Excel.Range)excelWorksheet.get_Range(j_str_TEDs); // MKT CAP
fl = excelCell2.Cells.Value;
Console.WriteLine("The MKT CAP " + j_str_TEDs + "
value is " + fl);
if (fl <= 568.09)
excelWorksheet.get_Range('D' +
i_TEDs.ToString()).Value = 1;
else if (fl <= 1480.28)
excelWorksheet.get_Range('D' +
i_TEDs.ToString()).Value = 2;
else if (fl <= 3172.82)
excelWorksheet.get_Range('D' +
i_TEDs.ToString()).Value = 3;
else if (fl <= 8557.61)
excelWorksheet.get_Range('D' +
i_TEDs.ToString()).Value = 4;
else if (fl <= 394611.12)
excelWorksheet.get_Range('D' +
i_TEDs.ToString()).Value = 5;
}
// string ans = Console.ReadLine();
excelWorkbook.SaveAs(workbookPath, true);
excelApp.Quit();
}
}
}
With each command like the one following, we compare the book-to-market or market capitalization value of the stock with the break points (here, the book-to-market value of the stock (BtoM) is compared to the lowest breakpoint (0.358): Console.WriteLine("The BtoM " + i_str_TEDs + " value is " + fl);
if (fl <= 0.358)
The output of this script is two additional columns in the spreadsheet which we give as
input, where in the first one it returns the classification of the particular stock in one of the
five quintiles in respect to its market capitalization, and the second the classification of the
stock in respect to its book-to-market ratio. The structure of the output for the first 20
stocks can be seen in the following table:
53
Firm Name Book-to -market
Market Cap.
Book-to -market Rank
Market Cap. Rank
Mean Return
DDD US EQUITY 0.57 177.58 3 1 1.84
MMM US EQUITY 0.27 39906.48 1 5 0.87
AIR US EQUITY 0.90 656.20 4 2 1.26
ABT US EQUITY 0.20 82853.35 1 5 1.04
ABM US EQUITY 0.85 759.51 4 2 1.23
ATU US EQUITY 0.61 1011.49 3 2 2.23
AMD US EQUITY 0.09 1327.62 1 2 1.45
AFL US EQUITY 0.65 21389.63 3 5 1.66
GAS US EQUITY 0.64 2410.82 3 3 0.86
APD US EQUITY 0.41 10539.27 2 5 0.97
ARG US EQUITY 0.48 3162.83 2 3 1.97
ALK US EQUITY 0.63 1061.04 3 2 0.68
AIN US EQUITY 1.34 384.67 5 1 0.41
AA US EQUITY 1.74 9011.57 5 5 0.79
ALX US EQUITY 0.18 1297.85 1 2 1.52
Y US EQUITY 0.97 2332.92 4 3 1.02
AGN US EQUITY 0.32 12260.83 1 5 1.40
ALE US EQUITY 0.80 1052.00 4 2 0.93
AB US EQUITY 1.05 1877.83 4 3 1.46
LNT US EQUITY 0.88 3222.90 4 4 0.67
Table 4.2: Data from Bloomberg sorted. (Market capitalization in USD Million, throughout the
whole Thesis.)
Afterwards, it comes down to a simple exercise in excel, to create the 25 different portfolios
that we need from this original spreadsheet.
4.6 Statistical Analysis of the test results
Before we proceed with the various tests of the Optimization Problems 1 & 2, it is necessary
to provide with the procedure with which the sample values are compared with the
generated from Matlab ones. Hence, for every portfolio, the table below is constructed for
each one of the tests:
54
1 2 3 4 5 6 7 8 9
μ sample
μ Matlab
σ sample
σ Matlab
t statistic |t| < 1.96 satisfied
lower bound satisfied
variances ratio
upper bound satisfied
0.014 0.017 0.109 0.108 0.29 yes yes 0.982 yes
0.001 0.017 0.110 0.110 1.56 yes yes 0.991 yes
0.041 0.017 0.115 0.107 -2.37 no yes 0.871 yes
0.012 0.016 0.062 0.060 0.61 yes yes 0.956 yes
0.009 0.016 0.050 0.049 1.42 yes yes 0.959 yes
0.012 0.016 0.059 0.057 0.71 yes yes 0.952 yes
0.016 0.016 0.066 0.064 0.08 yes yes 0.941 yes
0.011 0.016 0.073 0.072 0.71 yes yes 0.974 yes
0.014 0.017 0.097 0.096 0.34 yes yes 0.977 yes
0.013 0.016 0.088 0.087 0.33 yes yes 0.976 yes
0.021 0.017 0.122 0.121 -0.32 yes yes 0.975 yes
0.005 0.016 0.052 0.052 2.34 no yes 0.982 yes
0.011 0.016 0.050 0.049 1.09 yes yes 0.949 yes
0.016 0.017 0.115 0.114 0.08 yes yes 0.980 yes
0.017 0.016 0.097 0.095 -0.09 yes yes 0.968 yes
0.011 0.017 0.100 0.099 0.69 yes yes 0.984 yes
0.013 0.016 0.065 0.064 0.52 yes yes 0.957 yes
0.009 0.020 0.301 0.299 0.40 yes yes 0.988 yes
0.013 0.017 0.113 0.112 0.37 yes yes 0.984 yes
0.042 0.017 0.189 0.185 -1.50 yes yes 0.954 yes
0.008 0.016 0.083 0.083 1.06 yes yes 0.983 yes
0.015 0.016 0.096 0.094 0.22 yes yes 0.975 yes
0.033 0.016 0.114 0.109 -1.72 yes yes 0.919 yes
0.006 0.017 0.090 0.089 1.25 yes yes 0.988 yes
0.010 0.018 0.182 0.182 0.48 yes yes 0.995 yes
0.008 0.016 0.072 0.071 1.28 yes yes 0.982 yes
0.026 0.016 0.078 0.074 -1.39 yes yes 0.897 yes
0.028 0.017 0.125 0.122 -1.01 yes yes 0.948 yes
0.016 0.016 0.101 0.099 0.06 yes yes 0.974 yes
0.015 0.017 0.136 0.135 0.18 yes yes 0.987 yes
0.014 0.016 0.073 0.071 0.33 yes yes 0.962 yes
0.012 0.016 0.090 0.089 0.54 yes yes 0.978 yes
0.016 0.017 0.098 0.097 0.10 yes yes 0.974 yes
0.010 0.016 0.065 0.064 1.04 yes yes 0.970 yes
0.020 0.016 0.083 0.080 -0.42 yes yes 0.945 yes
-0.001 0.016 0.120 0.120 1.54 yes yes 0.996 yes
0.019 0.016 0.091 0.089 -0.36 yes yes 0.956 yes
0.031 0.018 0.176 0.174 -0.84 yes yes 0.972 yes
0.017 0.016 0.078 0.076 -0.06 yes yes 0.953 yes
0.004 0.016 0.089 0.089 1.48 yes yes 0.990 yes
55
Table 4.3: Statistical tests carried out for each optimization sets’ results.
At the table above we can see a sample of the results of the statistical test. We have picked
40 stocks out of a sample portfolio constituted by 100 stocks of the S&P 500 Index, sorted by
alphabetical order.
Each row corresponds to one specific stock, for which as has already explained there are two
sets of parameters: the estimated ones μ and σ, and the sample ones μsample and σsample.
More specifically (explanation of the variables):
μ sample: the (simple) average return of each stock of the portfolio.
μ Matlab: the return of each stock estimated by fmincon.
σ sample: the standard deviation of each stock of the portfolio.
σ Matlab: the standard deviation of each stock estimated by fmincon.
The objective of this statistical exercise is to determine how “close” the sample and the
estimated parameters are; so we compare the μ with the μsample, and the σ with the σsample.
Two simple statistical tests are employed in order to test how “close” the estimated
parameters are to the real ones:
a) T-test for the returns. The t-values for the estimated returns are given in column 5
of the above table. As we can see, for each and every one of the stocks of this
sample, the t-value states that the difference between the adjusted and the sample
parameter is not significant at the 95% level of confidence level. This can be seen at
column 6 of our table, where an if-function in excel informs us whether the null
hypothesis (H0: μisample – μi = 0, i = 1, 2, .., 121 at the 5% significance level) holds true.
b) Confidence interval for the standard deviations. In order to test the standard
deviations, we test whether the confidence interval of the ratio (σi)2/(σi
sample)2 for
each stock is within range; the range differs depending on the number of stocks, for
the above sample it is [0.7873-1.27016]. This is calculated as follows: We use 121
monthly return observations, therefore n=121. Since we want are results to be in
line with the 95% confidence interval for (σi)2/(σi
sample)2, and under the null
hypothesis that the estimated standard deviations returned by Matlab follow the
same distribution as the sample parameters, we are looking for the critical values c1
and c2, such that P(x2120-1 > c1 ) = 0.025 and P(x2
120-1 < c2 ) = 0.025. For the first
equation we have (2·c1)^(0.5)-(2·121-1)^(0.5) = 1.96, and for the second
(2·c2)^(0.5)-(2·121-1)^(0.5) = 1.96. From these two equations, we get c1 = 152.42
and c2 = 90.58. In order to calculate the confidence interval, we employ the standard
formula c1 < (121-1)·(σi)2/(σi
sample)2 < c2, which gives us the above numbers for the
specific stock set.
Another candidate for testing the null hypothesis is the Bonferroni test (1945), which states
that a multiple comparison null hypothesis like this one should be rejected at the 5%
significance level if any one of the estimated return parameters is significantly different from
its sample counterpart at the 2.5% significance level. In this particular case two of the
56
estimated parameters are significant at the 5% significance level, therefore even by
Bonferroni’s test, the multiple null hypothesis cannot (marginally) be rejected.
In the chapters devoted to the tests results for the optimization problems under the
Levy/Roll procedure (Chapter 4 for Optimization Problem 1 and Chapter 5 for the
Optimization Problem 2), a large number of tests will be presented. They were picked out of
a much larger pool of all the relevant optimization tests conducted, for the purposes of this
Thesis, in an attempt to familiarize with the behavior of these models under real market
conditions.
4.7 Investigation of correlations and other varying attributes
Picture 14: Correlations are significantly higher when in crisis.43
The rapid change of correlations, even within a short period of only few years, poses a
challenge to the statistical analysis of stock attributes. We believe that using the same (or
“average”) correlation for the entire period, like in the implementation by Levy and Roll
(2010), corresponds to a serious over-generalization and represents a very strong
assumption for projections of future returns. In addition, a lot of questions have been raised
regarding the structural robustness of the model -that is, whether it is robust because of its
mathematical nature- and it does not necessarily reflect any characteristics of the actual
stocks.
For that reason, two distinct series of tests with different correlation matrices were
implemented. In order to accomplish that, we will replace the matrix “sample” at line 52 of
the constraint functions (funCeq.m), which represents the covariance matrix of the sample
set of stocks (the “real one”) with the artificial covariance matrix, for which we want to
stress the model. This artificial covariance matrix is constructed by using the original
43 Empirical Research Partners LLC, Standard & Poor’s, J.P. Morgan Asset Management. Capitalization
weighted correlation of top 750 stocks by market capitalization, daily returns, 30.06.2013.
57
standard deviations of the stocks, but a different correlation matrix, for which we will create
for different versions (two sets). The portfolio chosen for these tests is the most robust one
from all the optimizations ran so far, the first 100 stocks from the S&P 500 Index. The set-up
that we will use for testing the correlation matrices is the “proper” one, based on market
conditions and risk free rate; the one that, as we will see in the subsequent chapters, is
proven to be robust. This way we believe that the necessary conditions are met, in order to
achieve a reliable conclusion regarding the use of different correlation matrices in a known
problem.
1. For the first test we will create a random correlation matrix. The following simple
Matlab script is being used for that purpose:
% generate a pseudo random definite correlation matrix % positive definite NxN matrix SPD = randn(N); % N the number of stocks in our portfolio %covariance matrix: SPD = SPD'*SPD; % Convert S into a correlation matrix: CMATR = sqrt(diag(SPD)); C = diag(1./ CMATR)* SPD *diag(1./ CMATR);
2. For the second test we will create a random positive correlation matrix, given that
most stocks have positive correlations. The following simple Matlab script is being
used for that purpose:
% generate a pseudo random definite correlation matrix % positive definite NxN matrix constituted only by positive
values SPD = randn(N); % N the number of stocks in our portfolio %covariance matrix: SPD = SPD'*SPD; % Convert S into a correlation matrix: CMATR = sqrt(diag(SPD)); C = diag(1./ CMATR)* SPD *diag(1./ CMATR);
3. For the third test, we employ the Cholesky factor for sampling, for the generation of
% Compute the standard deviations from the original sample covariance
matrix var=diag(sample); sqrtvar=sqrt(var);
% calculate the artificial correlation matrix (procedure demonstrated % above)
ArtCOR; % Use the new correlation matrix in the Final expression for non the
linear restriction ceqshrunkinv = S* ArtCOR*S*xma - q*mr2n;
The reason why we refer to the “original” sample covariance matrix is that as we know, the covariance matrix used in all tests is the shrunk one (based on the cited paper45), although the standard deviations remain unchanged after shrinking; there is no apparent reason why the random covariance matrices should be shrunk, therefore we substitute the shrunk ones with them, at the final formulas in the Matlab code. The results from running the model with these correlation matrices will indicate whether the
mathematical model is robust by itself, disregarding the actual relationships between the
stocks, or if it actually reveals some intuition behind these relationships. In the following
table we have summarized the execution trials of Optimization Problem 1:
Model name
Rejected μ's
Rejected σ's
MaxFunEvals Levy/Roll complied
1 44% 87% 40000 N
2 43% 91% 40000 N
3 85% 82% 40000 N
4 86% 52% 84690 N
Table 4.4: Results of the optimizations conducted by using the generated random correlation
matrices (the explanation of the parameters is provided in page 66).
As we can see from the results, as well from the behavior of the model while running in
Matlab, the different correlation matrices caused the complete breakdown of the whole
optimization procedure (even though the maximum –objective- function evaluations were
several times above the usual number).
44
Stackoverflow. How to generate pseudo-random positive definite matrix with constraints on the off-diagonal elements? http://stackoverflow.com/questions/1037340/how-to-generate-pseudo-random-positive-definite-matrix-with-constraints-on-the-o, 20.08.2013. 45
Ledoit, O. and Wolf, M. (2003). Honey, I Shrunk the Sample Covariance Matrix. UPF Economics and Business, Working Paper No. 691.
59
Our results indicate that there is a significant amount of information included in the
correlation matrices. Therefore, in the final chapter we will propose an idea of repeating the
tests with random correlation matrices constrained by certain conditions, so that the largest
part of information will not be lost.
For that reason, it is considered reasonable to proceed with tests in actual market data, right
after we analyze how to deal with the final calibration of the input parameters.
4.8 Definition of Market Proxy – Indices
Extensive research was carried out in order to determine the best set of market parameters
that need to be fed into the model for each different set of tests. Therefore, it was
considered absolutely necessary to devote a subchapter of this Thesis on this issue, in order
for the reader to gain a better insight of the historical return and standard deviation values,
for different indices which are considered to be market proxies, throughout the literature.
For the needs of the entire Thesis, the following Indices were used:
1. S&P 500
2. S&P 500 Equally Weighted
3. S&P Global 1200
4. Willishire 5000
5. Willishire 5000 Equally Weighted
6. Russell 3000
The equally weighted indices were used at the third main part of the Thesis, in order to
calculate the Zipf factor. For that particular part, the Equally Weighted Market Returns from
CRSP Database were used as well, but as it is not officially considered an Index, it is not
stated here.
From the analysis of the below provided graphs, as well as the tables downloaded from
Bloomberg, we are trying to demonstrate two major characteristics: the returns and the
standard deviation of each index, and of course the degree of correlation between the
different indices for the same market, as well as for overlapping markets. The reason for
undertaking so much effort to analyze the behavior of the main (U.S.) market Indices, is that,
as we will see in the respective chapters, we need to calculate accurately (sometimes up to
10-4) the proxy market’s returns and standard deviations, in order to feed them into the
models. Since the backbone of the stress testing of the Levy/Roll approach (Optimization
Problems 1 and 2) is based on whether they perform well under the specific market
conditions for each, we attempted to specify for each optimization test, based on the
constituents of the portfolio used, and the specific time period, the most appropriate mean
market returns and standard deviation.
60
Zephyr StyleADVISORZephyr StyleADVISOR: Dow Chemical Company
Trailing Market ReturnsJanuary 2003 - February 2013 (not annualized if less than 1 year)
Retu
rn
0
5
10
15
20
YTD 1 quarter 1 year 3 years 5 years 10 years
Citigroup 3-month T-bill
Russell 3000
MSCI EAFE Index
S&P 500
S&P 500 Equal Weight
S&P 1000 Pure Growth
Wilshire 5000 (Full Cap)
Wilshire 5000 - Equal Weight
Trailing Market ReturnsJanuary 2003 - February 2013 (not annualized if less than 1 year)
S&P 500 MKT VL, a = 0.7 0.087 0.1807 0.0012 0.7 4 4% 8 8% N
S&P 500 MKT VL, a = 0.6 0.087 0.1807 0.0012 0.6 4 4% 8 8% N
x0 (sample parameters)
Table 4.12: Full cluster of sets for market values based on the S&P 1200 Global Index.
Picture 17: Testing a number of parameter combinations: none of them works, including the ones with correct market values.
-1%
0%
1%
2%
3%
4%
1 15 29 43 57 71 85 99
Return
Stock #S&P 500 MKT VL BL TOP 100 orig shouldn’t work S&P 500 MKT VL BL TOP 100 orig s=0.1 shouldn’t work S&P 500 MKT VL BL TOP 100 orig m=0.05
S&P 500 MKT VL, Rf = 0.0012 shouldn’t work S&P 500 MKT VL, Rf = 0 shouldn’t work S&P 500 MKT VL, a = 0.7
S&P 500 MKT VL, a = 0.6 x0 (sample parameters)
71
Rf test market
μ market
σ Rf α
Rejected μ's (#)
Rejected μ's (%)
Rejected σ's (#)
Rejected σ's (%)
Levy/Roll complied
BL TOP 100 original 0.0012 Rf 0.052 0.104 0.0012 0.75 2 2% 0 0% Y
BL TOP 100 original 0.02 Rf 0.052 0.104 0.02 0.75 44 44% 8 8% N
BL TOP 100 original 0.015 Rf 0.052 0.104 0.015 0.75 17 17% 6 6% N
x0 (sample parameters)
Table 4.13: Testing the model for different values of Rf.
Picture 18: Testing the Rf: the model should not work for 0.12%, but for 1.5%, which is the historical average. The results indicate abnormal behavior.
-1,0%
-0,5%
0,0%
0,5%
1,0%
1,5%
2,0%
2,5%
3,0%
3,5%
4,0%
1 15 29 43 57 71 85 99
Return
Stock #
BL TOP 100 original 0.0012 Rf BL TOP 100 original 0.02 Rf BL TOP 100 original 0.015Rf x0 (sample parameters)
72
Alpha (α) test market
μ market
σ Rf α
Rejected μ's (#)
Rejected μ's (%)
Rejected σ's (#)
Rejected σ's (%)
Levy/Roll complied
S&P 500 MKT VL, a = 0.75 0.087 0.1807 0.0012 0.75 1 1% 8 8% N
S&P 500 MKT VL, a = 0.7 0.087 0.1807 0.0012 0.7 4 4% 8 8% N
S&P 500 MKT VL, a = 0.6 0.087 0.1807 0.0012 0.6 4 4% 8 8% N
x0 (sample parameters)
Table 4.14: Testing the model for different values of α.
Picture 19: Alpha’s effect on the results: insignificant.
-1,0%
-0,5%
0,0%
0,5%
1,0%
1,5%
2,0%
2,5%
3,0%
3,5%
4,0%
1 15 29 43 57 71 85 99
Return
Stock #
S&P 500 MKT VL, a = 0.75 S&P 500 MKT VL, a = 0.7 S&P 500 MKT VL, a = 0.6 x0 (sample parameters)
73
Alternative market proxies market
μ market
σ Rf α
Rejected μ's (#)
Rejected μ's (%)
Rejected σ's (#)
Rejected σ's (%)
Levy/Roll complied
S&P 500 MKT VL BL TOP 100 orig 0.087 0.1807 0.015 0.75 12 12% 4 4% N
S&P 1200GL MKT VL BL TOP 100 original 0.084 0.047 0.015 0.6 22 22% 10 10% N
x0 (sample parameters)
Table 4.15: Testing the model output when using as calibration parameters the ones from the Indices S&P 500, Wilshire 5000 and S&P 1200 Global.
-1%
0%
1%
2%
3%
4%
5%
6%
7%
1 15 29 43 57 71 85 99
Return
Stock #S&P 500 MKT VL BL TOP 100 orig Wilshire 5000 MKT VL Wilshire 5000 MKT VL a=0.6 S&P 1200GL MKT VL BL TOP 100 original x0 (sample parameters)
Picture 20: The model was also tested with the parameters (return and standard deviation) of the Wilshire 5000, as market parameters. As we can see, it failed the test.
74
Table 4.16: Finally, the model was tested for the actual market parameters, for this market proxy: the returns and standard deviation of S&P 1200 Global Index.
-1%
0%
1%
2%
3%
4%
5%
6%
7%
1 15 29 43 57 71 85 99
Return
Stock #SP1200GL MKT VL BL TOP 100 original SP1200GL MKT VL Rf=0.001 SP1200GL MKT VL Rf=0.001_s=0.1
Table 4.21: Small model sensitivity to market μ. (Actual values for S&P 500 Index are 8.7% for mean returns and 18.07% standard deviation. The average risk free ratio
Table 4.22: We notice a small but sufficient sensitivity to market μ (the middle test perhaps should have been rejected, but on the other hand, 0.15 can be considered
not that “far” from 0.087, which is the correct value).
Picture 27: Testing variations in the risk free ratio, as well as to the market standard deviation (by keeping all others constant). The model works only for the correct Rf.
84
Test Works Should work Levy/Roll complied
S&P 500 1920-2000 MKT VL N N Y
Correct market values # 1 Y Y Y
Correct market values #2 N N Y
Rf and σ test Rf =0.001 N N Y
Rf and σ test b Rf =0.002 N N Y
Rf and σ test b σ=0.1 N N Y
Rf and σ test b σ=0.25 Y Y Y
Table 4.24: Overview of tests for Optimization Problem 1, first 100 stocks from the S&P 500
Index.
Finally, just like with the previous set of 100 stocks (from the S&P 1200 Index), we aggregate
all the results in the above table. Surprisingly, by following the exact same methodology as
before, but only for U.S. stocks this time, all the results comply with the Levy/Roll theory.
Therefore, the final conclusion of the cluster of tests on the second portfolio in our analysis
proved to be 100% supportive to the Levy/Roll theory, and the overall approach.
In order to further investigate the validity of the proposed approach, we will continue by
testing it on the 25 Fama French portfolios, the construction of which was described in
previous paragraphs.
4.9.3 The 25 Fama French NYSE Portfolios
In this paragraph, the performance of the Levy/Roll approach on the 25 Fama French NYSE
portfolios will be presented.
The number of firms in each Fama French portfolio is presented in the next table:
Book to market
Quantile low 2 3 4 high
small 15 16 19 14 63
Size 2 15 20 30 29 32
3 10 19 40 20 20
4 25 14 19 19 14
big 36 24 26 15 10
Table 4.25: Stocks constituting the 25 Fama French NYSE portfolios. Total number of stocks: 564.
For the next table, we have that:
Time period: 1.1991-12.2008.
BtMx, x=1, 2, 3, 4, and 5: Book-to-Market quintile; low (1) to high (2).
85
Second column x=1, 2, 3, 4, and 5: market capitalization quintile; small (1) to large
(2).
The entire Levy/Roll analysis’ results are given in the table:
Book to Market
Size # stocks in
the Portfolio Rejected
μ's (#) Rejected μ's (%)
Rejected σ's (#)
Rejected σ's (%)
Levy/Roll complied
NYSE BtM1 small 15 1 6.67% 0 0.00% N
NYSE BtM1 2 15 0 0.00% 0 0.00% Y
NYSE BtM1 3 10 0 0.00% 0 0.00% Y
NYSE BtM1 4 25 0 0.00% 0 0.00% Y
NYSE BtM1 large 36 0 0.00% 0 0.00% Y
NYSE BtM2 small 16 0 0.00% 0 0.00% Y
NYSE BtM2 2 20 0 0.00% 0 0.00% Y
NYSE BtM2 3 19 1 5.26% 0 0.00% N
NYSE BtM2 4 14 0 0.00% 0 0.00% Y
NYSE BtM2 large 24 2 8.33% 0 0.00% N
NYSE BtM3 small 19 0 0.00% 0 0.00% Y
NYSE BtM3 2 30 0 0.00% 0 0.00% Y
NYSE BtM3 3 40 1 2.50% 0 0.00% Y
NYSE BtM3 4 19 0 0.00% 0 0.00% Y
NYSE BtM3 large 26 1 3.85% 0 0.00% Y
NYSE BtM4 small 14 0 0.00% 0 0.00% Y
NYSE BtM4 2 29 0 0.00% 0 0.00% Y
NYSE BtM4 3 20 1 5.00% 0 0.00% Y
NYSE BtM4 4 19 0 0.00% 0 0.00% Y
NYSE BtM4 large 15 0 0.00% 0 0.00% Y
NYSE BtM5 small 63 3 4.76% 0 0.00% Y
NYSE BtM5 2 32 1 3.13% 0 0.00% Y
NYSE BtM5 3 20 0 0.00% 0 0.00% Y
NYSE BtM5 4 14 2 14.29% 0 0.00% N
NYSE BtM5 large 10 0 0.00% 1 10.00% N
Table 4.26: Performance of the 25 Fama French NYSE portfolios.
Overall, 20 out of the 25 portfolios’ estimated parameters are fully in line with the Levy/Roll
method, while from the five that are not, only two have rejected values (either in returns or
standard deviations) more than 10%. The last column is set at “N” (No) if the particular
portfolio’s returns or standard deviations tests are rejected at more than a 5% significance
level. Overall, we consider that the behavior of these 25 portfolios is consistent with the
approach of Levy and Roll.
86
4.9.4 The 25 Fama French NASDAQ/NYSE/AMEX Portfolios
Now the results of the Levy/Roll approach on the 25 Fama French NYSE/AMEX/NASDAQ
(entire U.S. Equity Market) portfolios will be presented.
The number of firms in each Fama French portfolio is presented in the next table:
Book to market
Quantile low 2 3 4 high
small 35 46 60 89 118
Size 2 22 36 55 45 25
3 17 36 39 31 20
4 48 41 27 28 19
big 74 60 35 24 4
Table 4.27: Stocks constituting the 25 Fama French NYSE/AMEX/NASDAQ portfolios. Total number
of stocks: 1034.
For the next table, we have that:
Time period: 1.1991-12.2008.
BtMx, x=1, 2, 3, 4, and 5: Book-to-Market quintile; low (1) to high (2).
Second column x=1, 2, 3, 4, and 5: market capitalization quintile; small (1) to large
(2).
The entire Levy/Roll analysis’ results are given in the table:
87
Book to Market Size
# stocks in the Portfolio
Rejected μ's (#)
Rejected μ's (%)
Rejected σ's (#)
Rejected σ's (%)
Levy/Roll complied
U.S. BtM1 small 35 0 0.00% 0 0.00% Y
U.S. BtM1 2 22 0 0.00% 0 0.00% Y
U.S. BtM1 3 17 1 5.88% 1 5.88% N
U.S. BtM1 4 48 0 0.00% 0 0.00% Y
U.S. BtM1 large 74 0 0.00% 0 0.00% Y
U.S. BtM2 small 47 2 4.26% 0 0.00% Y
U.S. BtM2 2 36 0 0.00% 0 0.00% Y
U.S. BtM2 3 36 1 2.78% 0 0.00% Y
U.S. BtM2 4 41 0 0.00% 0 0.00% Y
U.S. BtM2 large 60 3 5.00% 0 0.00% Y
U.S. BtM3 small 60 0 0.00% 0 0.00% Y
U.S. BtM3 2 55 4 7.27% 0 0.00% N
U.S. BtM3 3 39 1 2.56% 0 0.00% Y
U.S. BtM3 4 27 0 0.00% 0 0.00% Y
U.S. BtM3 large 35 1 2.86% 0 0.00% Y
U.S. BtM4 small 89 0 0.00% 0 0.00% Y
U.S. BtM4 2 45 0 0.00% 0 0.00% Y
U.S. BtM4 3 31 0 0.00% 0 0.00% Y
U.S. BtM4 4 28 0 0.00% 0 0.00% Y
U.S. BtM4 large 24 1 4.17% 0 0.00% Y
U.S. BtM5 small 118 2 1.69% 0 0.00% Y
U.S. BtM5 2 25 3 12.00% 0 0.00% N
U.S. BtM5 3 20 1 5.00% 0 0.00% Y
U.S. BtM5 4 19 1 5.26% 0 0.00% N
U.S. BtM5 large 4 0 0.00% 2 50.00% N
Table 4.28: 26 Performance of the 25 Fama French NYSE/AMEX/NASDAQ portfolios.
Even though the NYSE stocks represent only approximately half of the U.S. Equity Market,
and that was the case in the portfolio used after the data cleaning (after the downloading of
the raw data from Bloomberg) as well, the results of the same procedure onto the entire
U.S. market are almost identical with the ones for the NYSE stocks only.
The results produced for the 25 Fama French portfolios for the purposes of this Thesis are
slightly different than the ones in previous papers. That could very well be because of the
data preparation process, where it is impossible to guarantee that every researcher uses the
exact same set of stocks (after the “cleaning” of raw data).
As a conclusion of the 25 Fama French portfolios tests with our code, it could be stated that
their performance is in line with the Levy/Roll theory- a significantly strong result.
88
4.9.5 Portfolios of variable size from the top Market Capitalization Firms of S&P
500 Index
Finally, we chose to present in this chapter the results of the top market capitalization (as of
February 2013) stocks from the S&P 500 Index, with a variable number of stocks (5, 10, 20,
50, 100, 146 and 200), in an attempt to manifest the behavior of the Levy/Roll model onto
portfolios constituted by top market capitalization stocks. This practice (constructing
portfolios by the top market capitalization firms of a market proxy) holds a dominant
position in the literature, like for example in the original paper by Levy and Roll (there, the
authors did all their tests by using the first 100 top capitalization stocks).
Portfolio Stocks
# Rejected
μ's (#) Rejected μ's (%)
Rejected σ's (#)
Rejected σ's (%)
Levy/Roll complied
top 5 S&P 500 5 0 0.00% 0 0.00% Y
top 10 S&P 500 10 0 0.00% 0 0.00% Y
top 20 S&P 500 20 2 10.00% 0 0.00% N
top 50 S&P 500 50 3 6.00% 0 0.00% N
top 146 S&P 500 146 4 2.74% 0 0.00% Y
top 200 S&P 500 200 5 2.50% 0 0.00% Y
Table 4.29: Performance summary for the variable size top market capitalization portfolios.
The results of these tests are mixed. But there is an important bias: the very top market
capitalization stocks have completely different attributes from the rest of the stocks, even in
a portfolio formulated by the largest market capitalization firms. Therefore, every time the
number of stocks changes, so does the distribution (although not that much), but the same
huge market capitalization stocks remain in the portfolio. That results in the phenomenon
that, after the portfolio size of 20, it is the exact same stocks that fail to meet the Levy/Roll
statistical tests. This is recognized as a possible idea for further research, and is discussed in
more detail in the final chapter.
4.10 Conclusions for the Optimization Problem 1
In this chapter, our goal was to determine whether the model performs as it is supposed to,
given the inputs. That means that there were tests when the inputs were correct, therefore
the model should work, and there are tests were the inputs were partially or completely
wrong, in which case the model should fail.
As a conclusion for this first major cluster of tests, the model did not perform in line with the
theory at the portfolio constructed by the S&P 1200 Index, but it performed perfectly well
when tested on the portfolio constructed by the S&P 500 Index. With regard to the 25 Fama
French portfolios, the results inclined towards the validity of the model, while the tests on
the various size portfolios were rather ambiguous.
89
Overall, it seems that the model could possibly reveal some financial attributes of a certain
portfolio, but not under all conditions. Given that the results are not conclusive, we will test
the portfolio that performed fully in line with the Levy/Roll approach under the much more
constrained Optimization Problem 2 in the next Chapter.
90
Chapter 5
5 Optimization Problem 2
The analysis so far, of Optimization Problem 1, has provided evidence that the model can
work under certain circumstances, but the results are not robust enough in order for us to
be absolutely positive towards whether it worked due to the fact that it is mathematically
correct, or because it reflects some meaningful, intuitive financial attributes.
In order to strengthen our methodology, and by following the original Levy and Roll
approach, we apply two extra constraints to the Optimization Problem 1, which is now
named Optimization Problem 2. The idea is that the estimated parameters should, on top of
the mean/variance constraints of Optimization Problem 1, satisfy two additional constraints:
the returns and standard deviation of the estimated portfolio should be ex ante identical to
the sample portfolio’s ones (or to another, also prespecified pair of returns and standard
deviation values.
5.1 Mathematical Problem Description
The formulas for Optimization Problem 2 are repeated here, for the kind convenience of the
reader.
Minimize (objective function):
Subject to:
a) Nonlinear mean/variance constraint:
91
b) Liner constraint for estimated portfolio returns (μ):
c) Nonlinear constraint for estimated portfolio standard deviation (σ):
The first extra constraint (b) corresponds to a single equation (for the portfolio return, μ). It
states that the overall average (value-weighted) return, calculated from the estimated
parameters for our portfolio, should be equal to the actual average (value-weighted) return
of the sample. This is a very strong condition, given that the constraint violation tolerance is
just 10-6. This equation is linear, whereas already described, xm is the value-weighted vector
of the returns, μ the estimated returns and μ 0 the actual sample portfolio return.
The second extra constraint (c) is constituted by a set of nonlinear equations, similarly with
the first constraint (and only one, for Optimization Problem 1). It states that the estimated
portfolio’s standard deviation (σ) is equal to the sample portfolio’s one (σ0). More
specifically, by assuming that the correlations between the stocks should remain the same,
only the standard deviations are allowed to change, and that can be seen in the respective
mathematical formula above.
The σ0 and μ0 variables are being calculated as follows, from the sample stocks:
The variable xm (xm) corresponds to the weights derived from the market capitalization of the firms, sigma (σ) to the shrunk covariance matrix of the sample portfolio, and the variable x0 (x0) to the vector comprised by the N returns and N standard deviations, as it has been explained before. The variable x0 is multiplied with the vector [xm, zeros (1,100)], in order to get μ 0 because as we have stated in the Matlab variables paragraph, only the first half of the elements of every x0 corresponds to stocks’ returns – and these are the ones we need here; the second half elements of x0 correspond to the respective standard deviations (which are multiplied by 0, here). For our particular set of stocks, by using the above formulas, we have that
σ0 = sqrt(xm*sigma*xm'); σ0 = 0.0137
92
μ0 = [xm,zeros(1,100)]*x0'; μ0 = 0.0044
The mean return might appear particularly low (0.44%), but we should co-calculate the fact that this portfolio also bears very little risk (1.37%) and it is basically a random portfolio. Therefore, no techniques have been applied in terms of stock picking that would possibly maximize returns and/or minimize the standard deviation; the goal was to have a portfolio as random as possible. In the following paragraphs of this chapter, we will always use the correct calibration parameters, that were proven to work (with this specific portfolio) in the previous chapter. These would be the following values:
market μ market σ Rf α
0,087 0,18 0,015 0,6
Table 5.1: Calibration parameters and their values for this portfolio.
According to the theory and the empirical data provided by Levy and Roll, for the above values of parameters, the Optimization Problem 2 procedure should provide us with a range of σ0 and μ0 values, in the two-dimensional σ0 - μ0 space, for which the estimated parameters would be close enough to their sample counterparts. Since the actual values of σ0 and μ0 are equal to 1.37% and 0.44%, respectively, we expect the Optimization Problem 2 to work perfectly well for these exact values, as well as for small deviations of either one or both of them, around the original values. But in contradiction of the expected outcome, as we will see in the following paragraphs, the Optimization Problem 2 as provided with all its constraints, failed systematically to provide with proper convergence vectors (returns and standard deviations) in at least twenty different portfolios tested, whereas the Optimization Problem 1 worked for all of them. Based on these unexpected results, we decided to break down the different sets of extra constraints of the Optimization Problem 2 (one set for the returns and another for the standard deviations) and examine the behavior of the model in each of these individual subproblems. In the following paragraph, the specifics of these subproblems are presented, by presenting their mathematical formulas, as well the details of adjusting the Matlab implementation to each case. It is of high importance to state the exact specifications of the Matlab code used, when calling the fmincon function, since, as we will see, there are a lot of different combinations and misconceptions that can alter the final result. The detailed construction of the subproblems could also serve as a reference for future research.
5.2 Methodology
As stated in the previous paragraph, in order to dive deeper into the analysis, we impose the
new sets of constraints to the problem one by one; the reason for this careful approach was
that based on existing research, the model gets extremely sensitive to minor changes of the
93
μ0 and σ0 extra constraints, once all of them are in place. Therefore, we continue with the
construction of the Optimization subproblems ceq2, ceq2b, ceq and linear equality, ceq2
and linear equality, ceq3 and finally, ceq2 with linear inequality for the m0 constraint. Below
we explain the formation and the reasoning behind each different combination of the
constraints.
In the fmincon function of Matlab there is a possibility, structural, to define a linear equation
either as linear or nonlinear. More specifically, from the definition48 of the function:
c(x) ≤ 0
ceq(x) = 0
A·x ≤ 0
Aeq·x = beq
lb ≤ x ≤ ub
We have the possibility of only two sets (it can be more than just one equation in each set)
equations: the nonlinear set, which is defined by the ceq(x) =0 equation, and the linear one,
which is the Aeq·x = beq. Equivalently, we can impose nonlinear and linear inequality
constraints, by using c(x) ≤ 0 and A·x ≤ b. The linear (in)equalities are implemented
differently from the nonlinear (in)equalities, when calling the fmincon function; the
parameters of the linear ones are given as input parameters when calling the fmincon, while
the nonlinear ones are implemented in the ceq.m function. In this Thesis, the fmincon is
% Final expression for the σ squared nonlinear restriction of OPT 2:
% prsq: variance
ceq2= xm*S*RSIGMA*S*xma-prsq;
% Final nonlinear constraints used:
ceq = [ceqshrunkinv; ceq2];
a) Nonlinear mean/variance constraint:
b) Nonlinear constraint for estimated portfolio standard deviation (σ):
This problem corresponds to the Optimization Problem 2 of Levy and Roll, but without the
linear equation for the estimated portfolio returns (which, as we will see in the following
pages, is the strongest constraint).
95
5.2.2 Constraints Function funCeq2b
Calling the fmincon function by: [x,fval,exitflag,output] = fmincon(@funD,x0,[],[],[],[],[],[],@funCeq2b,options)
Linear constraints: none
ceq (nonlinear constraints) formulation used: % Final expression for nonlinear restriction with LW ceqshrunkinv = S*RSIGMA*S*xma - q*mr2n;
% Final expression for the μ linear restriction of OPT 2: % m: estimated returns vector
ceq3 = xm*m'-0.0137;
% Final nonlinear constraints used:
ceq = [ceqshrunkinv; ceq3];
a) Nonlinear mean/variance constraint:
b) Liner constraint for estimated portfolio returns (μ):
The set of nonlinear equations from Optimization Problem 1, as well as the linear equation
for the estimated portfolio returns, are implemented as nonlinear equation (in the ceq.m
function, not with the Aeq, beq parameters when calling fmincon).
5.2.3 Constraints Function funCeq (+linear equality)
Calling the fmincon function by: [x,fval,exitflag,output] = fmincon(@funD,x0,[],[],Aeq,beq,[],[],@funCeq,options);
Linear constraints:
Aeq=[xm,zeros(1,100)];
beq=[xm,zeros(1,100)]*x0';
ceq (nonlinear constraints) formulation used: % Final expression for nonlinear restriction with LW
ceqshrunkinv = S*RSIGMA*S*xma - q*mr2n;
96
% Final nonlinear constraints used:
ceq = ceqshrunkinv.';
The set of nonlinear equations from Optimization Problem 1, as well as the linear equation
for the estimated returns, but with the linear equation for the estimated portfolio returns is
implemented as a linear equation by fmincon.
5.2.4 Constraints Function funCeq2 (+linear equality)
Calling the fmincon function by: [x,fval,exitflag,output] = fmincon(@funD,x0,[],[],Aeq,beq,[],[],@funCeq2,options);
Linear constraints:
Aeq=[xm,zeros(1,100)];
beq=[xm,zeros(1,100)]*x0';
ceq (nonlinear constraints) formulation used: % Final expression for nonlinear restriction with LW
ceqshrunkinv = S*RSIGMA*S*xma - q*mr2n;
% Final expression for the σ squared nonlinear restriction of OPT 2:
ceq2= xm*S*RSIGMA*S*xma-prsq;
% Final nonlinear constraints used:
ceq = [ceqshrunkinv; ceq2];
a) Nonlinear mean/variance constraint:
b) Nonlinear constraint for estimated portfolio standard deviation (σ):
c) Liner constraint for estimated portfolio returns (μ):
97
In this subproblem all the sets of constraints that describe the Optimization Problem 2 have
been included. The linear equation for the estimated portfolio returns is implemented as a
linear equation and the subproblem corresponds to the “original” Optimization Problem 2.
5.2.5 Constraints Function funCeq3
Calling the fmincon function by: [x,fval,exitflag,output] = fmincon(@funD,x0,[],[],[],[],[],[],@funCeq3,options);
Linear constraints: none
ceq (nonlinear constraints) formulation used: % Final expression for nonlinear restriction with LW ceqshrunkinv = S*RSIGMA*S*xma - q*mr2n;
% Final expression for the σ squared nonlinear restriction of OPT 2:
ceq2= xm*S*RSIGMA*S*xma-prsq;
% Final expression for the μ linear restriction of OPT 2: ceq3 = xm*m'-0.016;
% Final nonlinear constraints used:
ceq = [ceqshrunkinv; ceq2; ceq3];
In this subproblem all the sets of constraints that describe the Optimization Problem 2 have
been included, but with the linear equation for the estimated portfolio returns is
implemented as a nonlinear equation. In fact, all the constraints are given to the program by
a single Matlab file called funCeq3.m.
5.2.6 Constraints Function funCeq2 (+linear INequality)
Calling the fmincon function by: [x,fval,exitflag,output] = fmincon(@funD,x0,A,b,[],[],[],[],@funCeq2,options);
Linear constraints: none A=-[xm,zeros(1,100)];
b=-[xm,zeros(1,100)]*x0';
ceq (nonlinear constraints) formulation used: % Final expression for nonlinear restriction with LW ceqshrunkinv = S*RSIGMA*S*xma - q*mr2n;
% Final expression for the σ squared nonlinear restriction of OPT 2:
ceq2= xm*S*RSIGMA*S*xma-prsq;
% Final nonlinear constraints used:
ceq = [ceqshrunkinv; ceq2];
98
The final test mimics a practice of previous research work:50 transforming the linear equality
for μ0, into a linear inequality. The reason for changing the problem by replacing the equality
with an inequality, is that we are giving fmincon an much less constraint problem to resolve,
so that we will collect the estimated values (xIN) and feed them as inputs to the sets of
constraints (ceq, ceq2, ceq2b and ceq3) and validate whether they hold true.
Because Matlab accepts the inequality constraint only in the form A·x ≤ b, in our case this
translates into:
But since we want the overall returns of the estimated portfolio to be larger (ideally) or at
least equal to the sample ones, we make the following transformation:
Overall, as mentioned before, the model “ceq2 and linear equality” is the one described as
Optimization Problem 2 in the literature. Nevertheless, we find that by imposing each
additional set of constraints to Optimization Problem 1 separately, we could gain a better
insight of the behavior of the Levy/Roll approach; the results are demonstrated below (in all
the following graphs, the first one hundred points correspond to the returns for the 100
stocks, while the last 100 points correspond to the standard deviations of the same stocks).
5.3 Test results: Portfolio of the 100 first stocks from S&P 500 Index
50
Ni, X., Malevergne, Y., Sornette, D. and Woehrmann, P. (2011). Robust Reverse Engineering of Cross Sectional Returns and Improved Portfolio Allocation Performance using the CAPM. Journal of Portfolio Management, 37 (4), 76-85.
99
Constraints Function Short description μ0 σ0 market
μ market
σ Rf α
Rejected μ's (#)
Rejected μ's (%)
Rejected σ's (#)
Rejected σ's (%)
Levy/Roll complied
ceq non liner eq. OP1 - - 0.087 0.18 0.015 0.6 3 3% 0 0% Y
ceq2b m0=0.0252 ceq2b m0=0.01 ceq2b m0=0.0044 ceq+linear eq ceq2 with linear INeq
101
ceq μ0 σ0 market
μ market
σ Rf α
Rejected μ's (#)
Rejected μ's (%)
Rejected σ's (#)
Rejected σ's (%)
Levy/Roll complied
ceq - - 0.087 0.18 0.015 0.6 3 3% 0 0% Y
x0 (sample parameters)
Table 5.3: For illustration purposes, before we move on with the Optimization Problem 2, we picture once again the results for Optimization Problem 1, for the same
Picture 29: Optimization Problem 1. As we can see, the results demonstrate almost perfect compliance with the Levy/Roll model.
102
Table 5.4: By imposing only the second set of constraints, for σ0, the model is incapable of producing parameters close enough to their sample counterparts.
Table 5.5: The model seems to work well with the extra constraint for μ0 (0.0044 is the correct value). It is extremely sensitive to this constraint, since we can see that
for the “wrong” values (0.044, 0.0252 and 0.01), the rejection rates are high.
ceq2 with linear INeq 0.0044 0.0137 0.087 0.18 0.015 0.6 3 3% 0 0% Y
x0
Table 5.6: When the linear inequality is imposed, it has a minor effect on the results, in comparison to the exact same problem but without the linear inequality. In both
cases, the model seems to work in line with the Levy/Roll theory (2% and 3% rejection rate, respectively).
Picture 32: In this plot we can see perhaps the most important test: Ceq2with linear inequality. The estimated parameters are statistically close to the sample ones.
105
After a series of tests with all the different subproblems presented in paragraph 5.2, we
came to the conclusion that, while the model can work (most of the times) for different
combinations of two out of the three sets of constraints, it is impossible for fmincon to solve
the Optimization Problem 2 with all three sets of constraints for this portfolio.
This is a very strong result, since in the previous chapter, this particular portfolio, with the
exact same market parameters, managed to pass all tests with almost 100% success rate.
Regarding the inability of the model to return a “proper” vector of estimated returns and
standard deviations, we could state that the biggest difficulty seems to lie when including
the linear equality constraint for the mean portfolio returns (μ0).
As we did in paragraph 3.3, we present here the graph of the objective function and the
maximum constraint violation. We can see that as fmincon tries to satisfy all the constraints
(the particular graphs correspond to the ceq3 model which is exactly equivalent to the
Optimization Problem 2), the value of the objective function increases, up to an extreme
value. For the record, the value of D for this set should be approximately 0.07, when
optimized; in this situation it goes well above 10.
Picture 33: Objective Function in the case of a model break-down.
The behavior of the objective function in combination with the behavior of the maximum
constraint violation, shown in the graph right below, depicts clearly the inability of the
model to return a “proper” vector of returns and standard deviations. Of course, by
“proper” we mean a vector of estimated parameters, which will be close enough,
statistically, with the sample parameters. The largest accepted value of a constraint
violation, with which we would assume that the constraint is satisfied, is 10-6. As we can see
below, in this case the maximum constraint violation could not go below approximately 10-2,
0
2
4
6
8
10
12
14
1
15
29
43
57
71
85
99
11
3
12
7
14
1
15
5
16
9
18
3
19
7
21
1
22
5
23
9
25
3
26
7
28
1
29
5
30
9
32
3
33
7
35
1
36
5
37
9
39
3
Ob
ject
ive
Fun
ctio
n V
alu
e
Fmincon Iteration Number
106
even though the number of iterations was almost eight times more than the number
required in the Optimization Problem 1, with the exact same setting (without the two extra
sets of constraints for Optimization Problem 2).
Picture 34: Maximum Constraint Violation in the case of a model break-down.
As a way around this problem, we followed a practice that has been used before in the
literature;51 we employed a linear inequality instead of the linear equality (for μ0), in which
as shown in paragraph 5.2.6 the returns of the estimated parameters can be larger or equal
than the returns of the sample ones. Indeed, Matlab returned a vector xIN (by executing the
subproblem named ceq2 with linear inequality) that is in line with the Levy Ross approach:
the estimated parameters (xIN) are not statistically different from the sample parameters
(xsample), at the 5% significance level.
Before we accept this solution, as a valid solution for the Optimization Problem 2, we need
to verify that the estimated vector of 200 elements (the 100 returns and 100 standard
deviations of the 100 stocks) satisfies all the constraints of the problem, but as they are
defined originally; with an equality for the estimated portfolio returns equation. Based on
the generally accepted standards, which are in line with Matlab’s default thresholds, that
would mean that all constraints should be satisfied with a maximum deviation of 10-6.
In order to accomplish that, we used the vector xIN that was produced by the last test (ceq2
with linear inequality set) as input to all the constraint functions demonstrated above. If xIN
is indeed a solution to the original problem, the following inequalities should hold:
51
Ni, X., Malevergne, Y., Sornette, D. and Woehrmann, P. (2011). Robust Reverse Engineering of Cross Sectional Returns and Improved Portfolio Allocation Performance using the CAPM. Journal of Portfolio Management, 37 (4), 76-85.
0.009
0.0095
0.01
0.0105
0.011
0.0115
1
16
31
46
61
76
91
10
6
12
1
13
6
15
1
16
6
18
1
19
6
21
1
22
6
24
1
25
6
27
1
28
6
30
1
31
6
33
1
34
6
36
1
37
6
39
1
Max
imu
m C
on
stra
int
Vio
lati
on
Val
ue
Fmincon Iteration Number
107
ceq(xIN) ≤ 10-6
ceq2(xIN) ≤ 10-6
ceq2b(xIN) ≤ 10-6
ceq3(xIN) ≤ 10-6
That would mean that xIN, the solution found by employing an inequality instead of an
equality in the original Optimization Problem 2 can serve as a solution to the whole problem.
As it has been mentioned before, the nonlinear constraint that corresponds to the
mean/variance efficiency condition for the Optimization Problem 1, is a set of N equalities,
where N=100 in this case of a 100-stock portfolio. That means that ceq(xIN) will be a vector
of also 100 values, which will correspond to the constraint violation for each stock of the
portfolio.
On the other hand, the extra constraints of Optimization 2, i.e. the linear equality for μ0 and
the non linear equation for σ0 correspond to one equation each; that means that ceq2(xIN)
and ceq2b(xIN) will include N+1 (101 in this case) values, where the 100 first will be the
constraints violations of ceq, and the 101st the constraint violation for σ0 and μ0,
respectively.
Finally, the vector resulting from ceq3(xIN) will include N+2 (102 in this case) values, where
the first 100 will be the constraints violations of ceq, the 101st the constraint violation for μ0
and the 102nd the constraint violation for μ0.
Overall, the constraint violations resulting by feeding the estimated (from the Optimization
problem ceq2 and inequality) vector xIN, to the actual constraints of the full Optimization
problem 2 (and its subproblems), are depicted in the following graph:
108
Picture 35: The constraint violations resulting by feeding the estimated (from the Optimization problem ceq2 and inequality) vector x IN into the subproblems constraints functions.
-0.00014
-0.00012
-0.0001
-0.00008
-0.00006
-0.00004
-0.00002
0
0.00002
13 5
79
11
13
15
17
19
21
23
25
27
29
31
33
35
37
39
41
4345
4749515355
5759
61
63
65
67
69
71
73
75
77
79
81
83
85
87
89
91
9395
9799 101
ceq ceq2 ceq2b ceq3
109
As we can see, the equality constraint for the returns is satisfied with a deviation of 1.4·10-4,
which is approximately 100 times larger than the maximum allowed threshold (10-6). This is
observed for both ceq2b and ceq3 subproblems, which identifies that the divergence comes
from the linear constraint for estimated portfolio returns (μ0).
5.4 Conclusions for the Optimization Problem 2
After the step-by-step analysis of the Optimization Problem 2 that we demonstrated in this
chapter, we concluded that it is not possible to produce results in accordance with the
Levy/Roll approach, when the two additional constraints are applied simultaneously. More
specifically, the addition of the linear equality constraint for the estimated portfolio returns
(μ0) which corresponds to the first extra constraint imposed on the Optimization Problem 1
by Levy and Roll for the Optimization Problem 2, cannot be satisfied for a given portfolio,
having the rest of the constraints implemented in the Matlab code. Even by replacing it by
an inequality so that we will relax the constraint and produce smoother results, backtesting
of the Matlab function revealed that not all constraints are satisfied.
110
Chapter 6
6 Zipf Approach
All the details regarding the Zipf distribution can be found in the relevant pages in the
References sector of this Thesis. In the Introduction of this Chapter, a brief description of the
Zipf Theory will be presented; the following work is based mainly on the paper “Professor
Zipf goes to Wallstreet”52.
6.1 Introduction
The key idea of this approach, an idea first stated by George Kingsley Zipf (1949)53, is that in
many systems in nature (including human-related activities) the size of an object or network
is dependent to its ranking amongst its fellow objects or networks of the same type. In fact,
the size is found to be inversely proportional to the rank of the object studied. This
remarkable relationship has been tested and proved to be true in a lot of different contexts,
such as English words, cities, distribution of wealth and others.
Zipf’s law, in its original form, it is written as
, where x is the size of the object, i is
its rank and xs is the maximum size in the specific system (or set) of objects.
In our case we employ the Zipf law in the context of firm sizes, stating the idea from the
paper “Professor Zipf goes to Wallstreet” cited previously, that because the distribution of
the firms’ sizes are heavy tailed, the market portfolio is not diversified enough; since the top
few firms in reality represent a huge portion of the total market capitalization. That
relationship has been proven to be valid in real life when finding that approximately 20% of
the firms in a given market correspond to the 80% of the market capitalization.
The authors of the paper54 mentioned in the previous paragraph were inspired by this fact,
and supported the opinion that a factor representing this effect could have significant
explanatory power in a time series regression model. Indeed, as we will see in the rest of this
chapter, by adding the “Zipf factor” to the plain market model, we were able to accomplish
results even better than the three-factor Fama French model.
52
Malevergne, Y., Santa-Clara, P. and Sornette, D. (2009). Professor Zipf goes to Wall Street. NBER Working Paper No. 15295. 53
Zipf, G. K. (1949). Human behavior and the principle of least effort. Oxford, England: Addison-Wesley Press. xi 573 pp. 54
Malevergne, Y., Santa-Clara, P. and Sornette, D. (2009). Professor Zipf goes to Wall Street. NBER Working Paper No. 15295.
111
6.2 Matlab Implementation
The third and last Matlab test set of this Thesis (as the first and second Matlab set tests we
consider the Optimization Problem 1, and the Optimization Problem 2, respectively) is solely
a number of linear regressions of portfolios’ returns on different factors. Having as reference
the examples of Mathworks55, the code used is the minimum one, and it is presented below
as such:
Dataset_1= dataset('XLSFile','1.xlsx');
With the above command, we import an excel file (the file '1.xlsx') into Matlab which is
being converted to a data set (named Dataset_1 in our example) with all the variables’
names.
Finally, with the following command we run a linear regression on the previously created
data set:
M_MARKET_MODEL_1= LinearModel.fit(Dataset_1)
The most important part of this procedure is that the excel spreadsheet imported should be
constructed such that the first row has the variables names, and the factors are one by one
ordered in a different column (starting from the first factor), with the last column being the
regressand (portfolio returns).
An example of the Matlab output for the three-factor Fama French model is given as follows:
Kenneth_mkt_factor_MODEL =
Linear regression model:
smallLow_Rf ~ 1 + Mkt_RF + SMB + HML + RF
Estimated Coefficients:
Estimate SE tStat pValue
(Intercept) -1.0238 0.29969 -3.4161 0.00073691
Mkt_RF 1.112 0.038951 28.547 4.1672e-82
SMB 1.3527 0.052288 25.871 7.1059e-74
HML -0.34482 0.05453 -6.3235 1.1088e-09
RF 1.1994 0.95988 1.2495 0.2126
Number of observations: 265, Error degrees of freedom: 260
Root Mean Squared Error: 2.66
R-squared: 0.901, Adjusted R-Squared 0.899
F-statistic vs. constant model: 590, p-value = 4.02e-129
As we can see, from such an output we can gather all the information that we need in order
to evaluate the performance of each model (each model’s evaluation will be analyzed in the
subsequent paragraphs).
55
MathWorks. Documentation Center. Time Series Regression I: Linear Models. http://www.mathworks.com/help/econ/examples/time-series-regression-i-linear-models.html, 20.12.2012.
112
6.3 Data
The data used for this chapter can be divided in two separate data sets: the portfolios’
returns which we will regress on a number of factors, and the factors themselves.
The portfolios returns used, are the ones that Professor Kenneth French has calculated and
published at his website56; the first set that has been tested is the 25 Fama French portfolios,
always by size and book-to-market. Afterwards, data downloaded directly from the CRSP
Database (Wharton interface) are used, namely the entire universe of NYSE/AMEX/NASDAQ
stocks divided in 10 deciles, then 4 quadrilles, and also just 2 sets (halves), by market
capitalization (with the largest companies in portfolio 1 and the smallest in portfolio 10).
For the factors, we initially applied 5 different tests, overall. First, the only factor was the
market one, but there were two different versions: one with the market factor Rm from the
database of Professor French, and a second one by using the returns of the Wilshire 5000
Index. This index is constituted by all the stocks of NYSE, NASDAQ and AMEX Indices,
therefore it could be considered as the stocks “universe”, being a representative proxy of
the market.
6.4 Time Series Regression Models
While in the literature it is always the case that very long time periods are used for such
kinds of tests, initially we wanted to test this model for the exact same period employed for
testing the 100 stocks (for the S&P500 Index and S&P 1200 Index) the Optimization Problem
1. Having already seen how the previous models cope with this specific data set, it is
interesting to examine the performance of this new approach as well.
In this subchapter, we will present the different time series regression models used in the
rest of the chapter. For reasons of completeness, we will demonstrate a table of results for
the R2 statistical parameter for each different model, so that the reader will acquire a
“feeling” of the different models’ fit and as such have certain expectations from their
performance. Thereafter, we will go on with the presentation of the main results.
6.4.1 Market Model
The first time series regression model employed is the “Market Model”: only one
explanatory factor, the Market Factor. The first set of regressions is described by the
following simple formula:
ri,t – Rf,t = αi + βi (rm,t – Rf,t) + εi,t
56
Dartmouth College. Research Returns. http://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library.html, 13.03.2013.
113
Where ri,t corresponds to each different portfolio’s returns, for i = [1,25], Rf,t is the risk free
ratio for each month, and rm,t is the market return time series. As in all regressions, εi,t
represents the residual (or error term).
At the first set, as mentioned above, the market return rm,t is the one given at the website of
Professor Kenneth French, with the following results:
Book-to-market
low 2 3 4 high
small 0.56 0.55 0.61 0.58 0.59
2 0.69 0.71 0.69 0.66 0.63
Size 3 0.72 0.79 0.74 0.68 0.63
4 0.79 0.81 0.75 0.72 0.64
big 0.89 0.82 0.75 0.62 0.57
Table 6.1: R
2 for the Market model, CRSP Data.
The average R2 equals to 68.7%.
At the second set, as mentioned above, the market return rm,t is calculated by the official
Wilshire website,57 and it was retrieved from the official website of the Wilshire 5000 Index:
Picture 36: The official Wilshire calculator.
The results were the following:
57
Wilshire. Wilshire Index Calculator. http://www.wilshire.com/Indexes/calculator, 31.07.2013.
114
Book-to-market
low 2 3 4 high
small 0.55 0.54 0.54 0.58 0.59
2 0.68 0.71 0.70 0.66 0.63
Size 3 0.71 0.79 0.75 0.69 0.65
4 0.78 0.82 0.77 0.74 0.66
big 0.89 0.83 0.77 0.63 0.58
Table 6.2: R
2 for the Market model, Wilshire 5000 Data.
The average R2 equals to 68.8%: just 0.1% higher in comparison to the market data from the
website of Professor Kenneth French; that indicates that the Wilshire 5000 Index could be
used as a market proxy for such experiments, with a performance almost identical to the
one of the widely accepted database of Professor Kenneth French.
6.4.2 Zipf Model
For the second round of regressions, we will add another factor to the market model, the
Zipf Factor.
Based on the theory presented in the beginning of this chapter, the addition of this factor to
the market model is expected to provide with considerable better fitting (or explanation
power) to the portfolios’ returns, and not much worse than the fitting by the three-factor
Fama French model.
The time series for the Zipf factor is calculated as the difference between the returns of the
equally weighted market portfolio (rm_EQ.W,t), and the value (market capitalization) market
portfolio (rm_V.W,t):
rz,t = rm_EQ.W,t – rm_V.W,t
Given the Zipf factor, the new regression model will be:
Fama French 3.36 2.61 2.67 2.89 3.00 2.48 3.06 2.70 4.01 12.40 3.92
Table 6.11: Statistical results: U.S. Equities Market divided in ten portfolios (January 1991 - December 2012).
126
Chapter 7
7 Conclusion
The process of evaluating quantitative models by applying them on real-market financial
data is known to be very data dependent. For that reason, to make our case as robust as
possible, we ran the models constructed for the purpose of this Thesis numerous times, with
their parameters varying in each execution in order to make sure that we have a
representative pool of results for each distinct case.
Regarding the Levy/Roll approach, the results for the optimizations are very sensitive to the
choice of the portfolio used, the market returns and standard deviation, as well as to the
choice of the risk free ratio. In the personal opinion of the author of this Thesis, it is possible
to manipulate these results, up to a certain point, by demonstrating specific cases/
combinations of the above stated parameters, in order to accomplish a better outcome and
improve the robustness of the model. Of course, such results would be severely biased,
since they would basically demonstrate the performance of a model for a certain subcase
and not for the universe of stocks, corresponding to the real market conditions. Our own
analysis showed that if we control for the “calibration parameters” and feed the models
with their real market values, the performance of the models is not robust enough in order
to justify global acceptance.
More specifically, having already drawn the detailed conclusions regarding each
optimization problem in the respective chapters (Chapters 4 and 5), we will attempt here to
sum up the main points and conclude about the entire approach as proposed by Levy and
Roll in 2011.
Regarding the Optimization Problem 1, we showed that the Levy/Roll approach can work for
certain portfolios, but not universally. It appears that it has increased explanatory power
over portfolios comprised exclusively by U.S. equities, versus portfolios that include stocks
from other markets. Previous research work has focused its efforts solely in U.S. equities.
For that reason, we proceeded with the step-by-step evaluation of the Optimization
Problem 2. In order to make our case more robust, we used only the portfolio for which the
Optimization Problem 1 produced results fully in line with the Levy/Roll approach (we refer
to the portfolio composed of the first 100 stocks from the S&P 500 Index). But even after
relaxing the most difficult constraint for the model to meet, the inability of the process to
produce results in line with the Levy/Roll approach is considered to be the final hit to the
economists’ theory (at least as it has been manifested). Therefore, the value of the whole
approach diminishes significantly, with the Levy/Roll approach being dismissed, conclusively.
127
In our quest for better “fitting” the returns of a portfolio, an idea coming from the paper
“Professor Zipf goes to Wall Street”58 seemed that it was worth exploring. Chapter 6 of this
Thesis is devoted to a series of tests, aiming to prove whether this new theory holds.
After conducting numerous time series regressions with three different regression models
(the market model, the Zipf model and the three-factor Fama French model), the results
verify at full power the initial motion introduced in the paper.
Two very large data sets were used, from which six different test sets were produced. Three
of them were constructed in the known format of the 25 Fama French portfolios, and the
other three were based on portfolios constructed by deciles, in which the stocks were
ranked by market capitalization. As we saw in the respective chapter, Zipf’s model
performance in all three 25 Fama French portfolios test sets was fully competitive, by always
overtaking the simple market model, and always lacking slightly (with some fluctuation)
from the three-factor Fama French model. But the results of the last three portfolio settings,
that were constituted by stocks based on (decreasing) market capitalization were even more
impressive. When the market was divided in only two parts, the Zipf model lagged slightly
over the Fama French one (but still has an R2 of 96%) and when the market was divided in
four parts (by market capitalization), both models had the exact same performance. But in
our last test, when the market was divided in 10 deciles based on the firms’ market
capitalization, the Zipf model surpassed the Fama French one (in terms of performance), by
boasting an R2 of 91% versus 83% (of the three-factor Fama French one). That is because of
the theory behind the Zipf model; in this case it can deploy its full potential.
As a final comment on the two approaches, we believe that the Levy Roll procedure could be
employed in order to achieve superior returns, in certain portfolios that their constituents
belong to well-defined markets. Our tests have shown that when we used representative
(for the market) values for the calibration parameters, and most importantly it was possible
to calculate the risk free ratio accurately, the model had a high probability to work well. On
the other hand, the enormous explanatory power of the Zipf factor in portfolios composed
of equities sorted by market capitalization, could be the foundation for further research on
achieving higher portfolio returns by using the Zipf model.
58
Malevergne, Y., Santa-Clara, P. and Sornette, D. (2009). Professor Zipf goes to Wall Street. NBER Working Paper No. 15295.
128
Chapter 8
8 Suggestions for continuing this work
As with every research process, several issues emerged along the way of this Thesis, which
were not “visible” or were not considered as important enough from the beginning. Most of
them were addressed until they were resolved, or at least until their mechanism was fully
understood by the author, so that later they would be presented in a decent way in this
Thesis. But not each and every idea that comes along while working on a certain research
topic can always be explored in its full detail, given the several constraints (with time being
the most important one). Therefore, in this section we will shortly present some issues that
were considered of secondary importance while working on this Thesis, but at the same
time are worth mentioned, for future research. These points are:
1) During the various tests conducted while examining the behavior of Optimization
Problem 1, it became evident that the estimated by Matlab parameters for certain
stocks were always not “close enough” to their sample counterparts. After a brief
look, it appeared that these stocks did not move along with the others in the same
portfolio, meaning that the correlation of their returns to the portfolio’s returns was
low. Therefore, further research could attempt to determine under which exact
conditions, a specific stock in a given portfolio will most probably not comply with
the Levy/Roll theory.
2) It has also been documented that the behavior of a portfolio (again, in the
optimization problems by Levy and Roll) changes, with regard to the order of the
stocks that it is being constituted. That could pose, for example, an idea for
investigating what the effects of market capitalization sorting versus sorting by
volatility are, in the same stocks of a given portfolio.
3) In the same spirit, it could be possible to find a critical number of stocks per
portfolio under examination with the Levy/Roll procedure, which would achieve the
maximum compliance to the theory. Again, it is difficult to control for such an effect
while keeping everything else constant (ceteris paribus), therefore such an approach
would require a tremendous amount of tests, covering all different cases (depending
on given markets/indices, sorting method of stocks in the portfolio, time-periods
used, etc.).
4) From a different perspective, it is possible that the number of monthly returns used
in any test (optimization for the Levy/Roll approach or time series regression
analysis for the Zipf factor) also has its implications on the results. Extensive
research has been already conducted on a similar subject, which is how many
monthly returns should be used in a model so that it would be able to forecast
market moves with a certain accuracy rate. This point was taken under
consideration in this Thesis, up to a certain extend.
129
5) As an extension of the previous point, but from the perspective of macroeconomics,
it is almost certain that even if the number of months is the same, it matters which
exact time period is covered. Even intuitively, we can understand that if the period
that is being used to test a model contains an “extreme event”, or it is generally
considered as a distressed period for the financial sector, it is expected to have
completely different characteristics versus a “normal” period. Examples of attributes
that are key to a financial analysis and vary significantly in such a distressed period
are the correlations between assets and the liquidity shortages caused after such
events (that affect asset pricing). Therefore, all tests could demonstrate significantly
different results, depending on which crisis they coincide with.
6) Referring to Paragraph 4.7, a different –and perhaps more intuitive- way of testing
the model with random correlation matrices could be applied, by using the view of
Malevergne and Sornette in their 2004 paper.59 In their work, they discovered that
the largest part of the information included in the correlation matrix of (in our case)
a portfolio, is located in the highest eigenvalues, while the rest of the correlation
matrix is white noise. Therefore, we could repeat the random correlation model
tests by imposing the necessary constraints on the (average) correlations of the
random correlation matrices, so that the information of the original correlation
matrix will not be lost (as it is the case with completely random correlation
matrices).
7) Regarding the optimization algorithm (solver) itself, there is a popular alternative
module that gains ground, Tomlab.60 It is possible that we would have different
results, by implementing Tomlab in the existing optimization problems for the
Levy/Roll procedure. That is because, in nonlinear optimization problems, the choice
of the solver employed is very important, and can –in certain cases of high
nonlinearity as in our case- lead into different results.
8) Finally, an issue that is more of a statistical nature: while the paper61 on which this
Thesis is based for the Levy/Roll approach part employs “naïve” averages for the
stocks’ returns, there was recently a paper62 in which the authors had similar results
but by using annualized mean returns. Since in this Thesis in none of the tests
annualized mean returns were used, it is unknown how all the different models
would behave in that case; but it is certainly an issue that could be addressed in
future work.
59
Malevergne, Y. and Sornette, D. (2004). Collective origin of the coexistence of apparent random matrix theory noise and of factors in large sample correlation matrices. Physica A: Statistical Mechanics and its Applications. 331(3–4), 660–668. 60
Levy, M. and Roll, R. (2010). The Market Portfolio May Be Mean/Variance Efficient After All. Oxford Journals, Review of Financial Studies, 23(6), 2464-2491. Ni, X., Malevergne, Y., Sornette, D. and Woehrmann, P. (2011). Robust Reverse Engineering of Cross Sectional Returns and Improved Portfolio Allocation Performance using the CAPM. Journal of Portfolio Management, 37(4), 76-85. 62
Brière, M., Drut, B., Mignon, V., Oosterlinck, K. and Szafarz, A. (2011). Is the Market Portfolio Efficient? A New Test to Revisit the Roll (1977) versus Levy and Roll (2010) Controversy. EconomiX Working Papers, 2011-20.
130
Chapter 9
9 References
General note on the references for this entire Thesis: each time there is need for a direct
reference (for a graph, for example) throughout the text, the source is given in a footnote in
the same page. All the sources used for this work, including the ones mentioned in the
footnotes, are provided in this Chapter, organized in three categories: Books, Papers and
Websites.
9.1 Books
1. Brammertz, W., Akkizidis, I., Breymann, W., Entin, R., Rüstmann, M. (2011). Unified
Financial Analysis: The Missing Links of Finance. John Wiley & Sons, The Wiley
Finance Series.
2. Brealey, R. A., Myers, S. T., Franklin, A. (2008). Principles of Corporate Finance.
Mcgraw-Hill, Irwin, 9.
3. Hull, J. C. (2011). Options, Futures and Other Derivatives. Prentice Hall, 9.
9.2 Papers
1. Brière, M., Drut, B., Mignon, V., Oosterlinck, K. and Szafarz, A. (2011). Is the Market
Portfolio Efficient? A New Test to Revisit the Roll (1977) versus Levy and Roll (2010)
Controversy. EconomiX Working Papers, 2011-20.
2. Chesnay, F. and Jondeau, E. (2000). Does Correlation between Stock Returns Really
Increase during Turbulent Period? Economic Notes, 30(1), 53-80.
3. Cristelli, M., Batty, M. and Pietronero, L. (2012). There is More than a Power Law in
Zipf. Scientific Reports 2, 812.
4. DeMiguel, V., Garlappi, L., Nogales, J. and Uppal, R. (2007). Optimal versus Naive
Diversification: How Inefficient is the 1/N Portfolio Strategy? Rev. Financ. Stud,
22(5), 1915-1953.
5. DeMiguel, V., Garlappi, L., Nogales, J. and Uppal, R. (2009). A Generalized Approach
to Portfolio Optimization: Improving Performance by Constraining Portfolio Norms.
Management Science, 55(5), 798-812.
6. Disatnik, D. and Benninga, S. (2006). Estimating the Covariance Matrix for Portfolio
Optimization. Available at SSRN: http://ssrn.com/abstract=873125 or
http://dx.doi.org/10.2139/ssrn.873125.
131
7. Fama, E. F. and French, K. R. (1993). Common risk factors in the returns on stocks
and bonds. Journal of Financial Economics. 33 (1), 3–56.
8. Flavin, T. J., Hurley, M. J., Rousseau, F. (2001). Explaining Stock Market Correlation: A
Gravity Model Approach. The Manchester School, 70(S1), 87-106.
9. Forbes, K. J. and Rigobon, R. (2002). No Contagion, Only Interdependence:
Measuring Stock Market Co-Movements. The Journal of Finance, 57(5), 2223-2261.
10. Jagannathan, R. and Ma, T. (2003). Risk Reduction in Large Portfolios: Why Imposing
the Wrong Constraints Helps. The Journal of Finance, 58)4), 1651–1684.
11. Ledoit, O. and Wolf, M. (2003). Honey, I Shrunk the Sample Covariance Matrix. UPF
Economics and Business, Working Paper No. 691.
12. Levy, M. and Roll, R. (2010). The Market Portfolio May Be Mean/Variance Efficient
After All. Oxford Journals, Review of Financial Studies, 23(6), 2464-2491.
13. Lintner, L. (1965). The valuation of risk assets and the selection of risky investments
in stock portfolios and capital budgets. Review of Economics and Statistics, 47, 13-
37.
14. Malevergne, Y. and Sornette, D. (2004). Collective origin of the coexistence of
apparent random matrix theory noise and of factors in large sample correlation
matrices. Physica A: Statistical Mechanics and its Applications. 331(3–4), 660–668.
15. Malevergne, Y., Santa-Clara, P., and Sornette, D. (2009). Professor Zipf goes to Wall
Street. NBER Working Paper No. 15295.
16. Markowitz, H. (1952). Portfolio selection. The Journal of Finance, 7, 77-91.
17. Ni, X., Malevergne, Y., Sornette, D. and Woehrmann, P. (2011). Robust Reverse
Engineering of Cross Sectional Returns and Improved Portfolio Allocation
Performance using the CAPM. Journal of Portfolio Management, 37 (4), 76-85.
18. Sharpe, W. F. (1964). Capital Asset Prices: A Theory of Market Equilibrium under
Conditions of Risk. The Journal of Finance, 19 (3), 425-442.
19. Zipf, G. K. (1949). Human behavior and the principle of least effort. Oxford, England:
Addison-Wesley Press. xi 573 pp.
9.3 Websites
Note: all websites cited in this section were visited from August 2012 until September 2013,
for the purposes of this Thesis. For each website, the date is stated individually at the end of
each link, of which the general website origin/Institution is also provided as a title.
1.4 Guys from Rolla. Sorting a Two-Dimensional Array using the Bubble Sort.
There was a large amount of peculiarities that had to be dealt with, when attempting to
download data with the Bloomberg excel add-in. The most important one is the way that the
add-in uses, in order to fill in the columns (it is always the case, that each column represents
the field of a different stock), when it comes to large numbers of columns (above several
hundred). In several cases the add-in would fill in every second column the intermediates
were completely blank) and we would have to click “enter” in the formula of each separate
cell, in order to be filled (when the number was small). In other cases, especially above 1500
columns, after a certain point (the last 3-5 hundred, for example) were also completely
blank.
For these reasons and a few more, in order to acquire all the available data for each major
index, a massive amount of time was consumed. This would include the break-up of the
large sets in several smaller excel files (so that the add-in would not jam) and then checking
that the columns that were still blank are because there are actually no data available for
the particular stock for the particular period and not because of any other bug. After the
check was concluded, all the partial excel sheets would be accumulated back into a master
file, in which we would run the VBA script, in order to keep only the stocks for which we
have full data over the desire period (even if one out of 215 months of returns was missing,
the entire column would be deleted).
1.2 Bloomberg Formulas
Below we can see all the different Bloomberg formulas used for the data acquisition, as well
as the explanation of the Bloomberg fields. For reasons of space, every different formula is
presented only for a particular set/time period.
1. Index Members
=BDS("Security ID","Field Mnemonic") In order to download the members of the Wilshire 5000 Index, we would use the previous formula as:
=BDS("W5000 Index","INDX_MEMBERS")
146
2. Bloomberg Security Names
After the previous command, we need to make the index members names compatible with the Bloomberg security names. For that reason, we employ the following excel formula:
=A$1&" EQUITY"
3. General formula (historical fields)
For downloading any historical field from the Bloomberg database, we use the formula: =BDH("Security ID","Field mnemonic","Start Date","End Date")
4. Monthly returns (using day-to-day field with the option “per=cm”):
Field mnemonic: "DAY_TO_DAY_TOT_RETURN_GROSS_DVDS"