Content uploaded by David Edmund Allen
Author content
All content in this area was uploaded by David Edmund Allen
Content may be subject to copyright.
Minimizing Loss at Times of Financial Crisis: Quantile Regression
as a Tool for Portfolio Investment Decisions
By
David E. Allen and Abhay Kumar Singh
School of Accounting, Finance and Economics, Edith Cowan University
School of Accounting, Finance and Economics & FEMARC Working Paper Series
Edith Cowan University
October 2009
Working Paper 0912
Correspondence author:
David E. Allen
School of Accounting, Finance and Economics
Faculty of Business and Law
Edith Cowan University
Joondalup, WA 6027
Australia
Phone: +618 6304 5471
Fax: +618 6304 5271
Email: d.allen@ecu.edu.au
ABSTRACT
The worldwide impact of the Global Financial Crisis on stock markets, investors and fund
managers has lead to a renewed interest in tools for robust risk management. Quantile
regression is a suitable candidate and deserves the interest of financial decision makers given
its remarkable capabilities for capturing and explaining the behaviour of financial return
series more effectively than the ordinary least squares regression methods which are the
standard tool. In this paper we present quantile regression estimation as an attractive
additional investment tool, which is more efficient than Ordinary Least Square in analyzing
information across the quantiles of a distribution. This translates into the more accurate
calibration of asset pricing models and subsequent informational gains in portfolio
formation. We present empirical evidence of the effectiveness of quantile regression based
techniques as applied across the quantiles of return distributions to derive information for
portfolio formation. We show, via stocks in Dow Jones Industrial Index, that at times of
financial setbacks such as the Global Financial Crisis, a portfolio of stocks formed using
quantile regression in the context of the Fama-French three factor model, performs better
than the one formed using traditional OLS.
Keywords:
Factor models; Portfolio optimization; Quantile regression
1
1. INTRODUCTION
From the introduction of Modern Portfolio Theory (MPT) by Markowitz, (1952), the
analysis of historical series of stock returns has been extensively used as the basis of
investment decisions. Diversification, as proposed by MPT, has been used for minimizing
risk, which works on the analysis of the covariance matrix of the chosen universe of stock
returns. Prior to the development of modern computing technology, this was
computationally demanding and short cuts were developed, such as Sharpe’s single index
model (1963). A heuristic which focuses on the empirical estimation of systematic risk,
which has a parallel focus in the modern finance’s central paradigm: the capital asset
pricing model (CAPM). Independently developed by Jack Treynor (1961, 1962), William
Sharpe (1964), John Lintner (1965) and Jan Mossin (1966).
Fama and French (1992, 1993) extended the basic CAPM to include two additional factors;
size and book-to-market as explanatory variables in explaining the cross-section of stock
returns. SMB, which stands for Small Minus Big, is designed to measure the additional
return investors have historically received from investing in stocks of companies with
relatively small market capitalizations. This additional return is often referred to as the
"size premium." HML, which is short for High Minus Low, has been constructed to
measure the "value premium" provided to investors for investing in companies with high
book-to-market values (essentially, the book value of the company’s assets as a ratio
relative to the market value reflecting investor’s valuation of the company, commonly
expressed as B/M).
Ordinary Least Squares regression analysis, has been the work-horse for all the regression
forecasting estimates used to model CAPM and its variations; such as the Fama-French
three factor model or other asset pricing models. With the introduction of alternative robust
risk measures such as Value at Risk (VaR) or Conditional Value at Risk (CVaR), which
are now standard in risk management, more emphasis has been laid on the lower tails of
the return distributions. The way in which OLS is constructed requires it to focus on the
means of the covariates. It is unable to account for the boundary values, or to explore
values across the quantiles of the distribution. It is also a Gaussian technique, with an
assumption of normality of the covariates, which does not sit well with the abundant
evidence of fat tails and skewness encountered in financial asset return distributions. This
2
feature of asset returns is even more acute in times of severe financial distress like the
Global Financial Crisis (GFC). Quantile Regression, as introduced by Koenker and Basset
(1978), has gained popularity recently in finance as an alternative to OLS, as this robust
regression technique can account for the lower and also the upper tails of the return
distribution and automatically accounts for outliers, or extreme events in the distribution,
and hence quantifies more efficiently for risk.
In this paper, we introduce quantile regression as a tool for investment decision making
and also show the applicability of this technique to robust risk management. We show the
effectiveness of quantile regression in capturing the risk involved in the tails of the
distributions which is not possible with OLS. We also use a basic portfolio construction
exercise using the Fama-French three factor model, on the components of the Dow Jones
Industrial 30 stocks index from a period running from 2005-2008 and show how quantile
regression based risk estimates can reduce the losses which we can incur when using OLS
based methods as portfolio construction tools.
2. QUANTILE REGRESSION
Linear regression represents the dependent variable, as a linear function of one or more
independent variables, subject to a random ‘disturbance’ or ‘error’ term. It estimates the
mean value of the dependent variable for given levels of the independent variables. For this
type of regression, where we want to understand the central tendency in a dataset, OLS is a
very effective method. OLS loses its effectiveness when we try to go beyond the mean
value or towards the extremes of a data set by exploring the quantiles.
Quantile regression as introduced in Koenker and Bassett (1978) is an extension of
classical least squares estimation of conditional mean models to the estimation of an
ensemble of models for conditional quantile functions. The central special case is the
median regression estimator that minimizes a sum of absolute errors. The remaining
conditional quantile functions are estimated by minimizing an asymmetrically weighted
sum of absolute errors. Taken together the ensemble of estimated conditional quantile
functions offers a much more complete view of the effect of covariates on the location,
scale and shape of the distribution of the response variable.
3
In linear regression, the regression coefficient represents the change in the response
variable produced by a one unit change in the predictor variable associated with that
coefficient. The quantile regression parameter estimates the change in a specified quantile
of the response variable produced by a one unit change in the predictor variable.
The quantiles, or percentiles, or occasionally fractiles, refer to the general case of dividing
a dataset into parts. Quantile regression seeks to extend these ideas to the estimation of
conditional quantile functions - models in which quantiles of the conditional distribution of
the response variable are expressed as functions of observed covariates.
In quantile regression, the median estimator minimizes the symmetrically weighted sum of
absolute errors (where the weight is equal to 0.5) to estimate the conditional median
function, other conditional quantile functions are estimated by minimizing an
asymmetrically weighted sum of absolute errors, where the weights are functions of the
quantile of interest. This makes quantile regression robust to the presence of outliers.
We can define the quantiles through a simple alternative expedient as an optimization
problem. Just as we can define the sample mean as the solution to the problem of
minimizing a sum of squared residuals, we can define the median as the solution to the
problem of minimizing a sum of absolute residuals. The symmetry of the piecewise linear
absolute value function implies that the minimization of the sum of absolute residuals must
equate the number of positive and negative residuals, thus assuring that there are the same
number of observations above and below the median.
The other quantile values can be obtained by minimizing a sum of asymmetrically
weighted absolute residuals, (giving different weights to positive and negative residuals).
Solving
min
కఢℛ
∑ߩ
ఛ
(ݕ
− ߦ) (1)
where ߩ
ఛ
(∙) is the tilted absolute value function as shown in Figure 1, this gives the ߬th
sample quantile with its solution. To see that this problem yields the sample quantiles as its
4
solutions, it is only necessary to compute the directional derivative of the objective
function with respect to ߦ, taken from the left and from the right.
Figure 1: Quantile Regression
࣋
Function
After defining the unconditional quantiles as an optimization problem, it is easy to define
conditional quantiles in an analogous fashion. Least squares regression offers a model for
how to proceed. If, we have a random sample,ሼݕ
ଵ
, ݕ
ଶ
, … , ݕ
ሽ, we solve
min
ఓఢℛ
∑(ݕ
− ߤ)
ଶ
ୀଵ
(2)
we obtain the sample mean, an estimate of the unconditional population mean, EY. If we
now replace the scalar ߤ by a parametric function ߤ(ݔ, ߚ) and solve
min
ఓఢℛ
∑(ݕ
− ߤ(ݔ
, ߚ))
ଶ
ୀଵ
(3)
we obtain an estimate of the conditional expectation function ܧ(ܻ|ݔ).
We proceed exactly the same way in quantile regression. To obtain an estimate of the
conditional median function, we simply replace the scalar ߦ in the first equation by the
parametric function ߦ(ݔ
௧
, ߚ) and set ߬ to
ଵ
ଶ
. To obtain estimates of the other conditional
quantile functions, we replace absolute values by ߩ
ఛ
(∙) and solve
5
min
కఢℛ
∑ߩ
ఛ
(ݕ
− ߦ(ݔ
, ߚ)) (4)
The resulting minimization problem, when ߦ(ݔ, ߚ) is formulated as a linear function of
parameters, can be solved very efficiently by linear programming methods.
This technique has been used widely in the past decade in many areas of applied
econometrics; applications include investigations of wage structure (Buchinsky and Leslie
1997), earnings mobility (Eide and Showalter 1999; Buchinsky and Hunt 1996), and
educational attainment (Eide and Showalter 1998). Financial applications include Engle
and Manganelli (1999) and Morillo (2000) to the problems of Value at Risk and option
pricing respectively. Barnes, Hughes (2002), applied quantile regression to study CAPM,
in their work on the cross section of stock market returns.
3. THE FAMA-FRENCH THREE FACTOR MODEL
Volatility is widely accepted measure of risk, which is the amount an asset's return varies
through successive time periods. Volatility is most commonly quoted in terms of the
standard deviation of returns. There is a greater risk involved for asset whose return
fluctuates more dramatically than another other. The familiar beta from the CAPM
equation is a widely accepted measure of systematic risk; whilst unsystematic risk is
captured by the error term of the OLS application of CAPM. Beta is a measure of the risk
contribution of an individual security to a well diversified portfolio as measured below;
ߚ
=
௩(
ಲ
,
ಾ
)
ఙ
ಾమ
(5)
where
r
A
is the return of the asset
r
M
is the return of the market
ߪ
ெଶ
is the variance of the return of the market, and
cov(r
A
, r
M
) is covariance between the return of the market and the return of the asset.
6
Jack Treynor (1961, 1962), William Sharpe (1964), John Lintner (1965) and Jan Mossin
(1966) independently, proposed Capital Asset Pricing Theory, (CAPM), to quantify the
relationship between beta of an asset and its corresponding return. CAPM stands on a
broad assumption that, that only one risk factor is common to a broad-based market
portfolio, which is beta. Modelling of CAPM using OLS assumes that the relationship
between return and beta is linear, as given in equation (2).
ݎ
= ݎ
+ ߚ
(ݎ
ெ
− ݎ
ி
)+ ߙ + ݁ (6)
where
r
A
is the return of the asset
r
M
is the return of the market
r
f
is the risk free rate of return
ߙ is the intercept of regression
e is the standard error of regression
Fama and French (1992, 1993) extended the basic CAPM to include size and book-to-
market as explanatory factors in explaining the cross-section of stock returns. SMB, which
stands for Small Minus Big, is designed to measure the additional return investors have
historically received from investing in stocks of companies with relatively small market
capitalization. This additional return is often referred to as the "size premium." HML,
which is short for High Minus Low, has been constructed to measure the "value premium"
provided to investors for investing in companies with high book-to-market values
(essentially, the value placed on the company by accountants as a ratio relative to the value
the public markets placed on the company, commonly expressed as B/M).
SMB is a measure of "size risk", and reflects the view that, small companies logically,
should be expected to be more sensitive to many risk factors as a result of their relatively
undiversified nature and their reduced ability to absorb negative financial events. On the
other hand, the HML factor suggests higher risk exposure for typical "value" stocks (high
B/M) versus "growth" stocks (low B/M). This makes sense intuitively because companies
need to reach a minimum size in order to execute an Initial Public Offering; and if we later
observe them in the bucket of high B/M, this is usually an indication that their public
market value has plummeted because of hard times or doubt regarding future earnings.
7
The three factor Fama-French model is written as;
ݎ
= ݎ
+ ߚ
(ݎ
ெ
− ݎ
ி
)+ ݏ
ܵܯܤ + ℎ
ܪܯܮ + ߙ + ݁ (7)
where s
A
and h
A
capture the security's sensitivity to these two additional factors.
Portfolio formation using this model requires the historical analysis of returns based on the
three factors using regression measures, which quantifies estimates of the three risk
variables involved in the model, i.e. ߚ
, s
A
, h
A
, and the usual regression analysis using
OLS gives us the estimates around the means of the distributions of the historical returns
and hence doesn’t efficiently quantify the behaviour around the tails. Modelling the
behaviour of factor models using quantile regression gives us the added advantage of
capturing the tail values as well as efficiently analysing the median values.
4. DATA & METHODOLOGY
The study uses daily prices of the 30 Dow Jones Industrial Average Stocks, for a period
from January 2005-December 2008, along with the Fama-French factors for the same
period, obtained from French’s website to calculate the Fama-French coefficients.
1
Table
1, gives the 30 stocks traded in the Dow Jones Industrial Average and used in this study.
Table 1 : Dow Jones Industrial 30 Stocks used in the study.
3M EI DU PONT DE NEMOURS KRAFT FOODS
ALCOA EXXON MOBILE MCDONALDS
AMERICAN EXPRESS GENERAL ELECTRIC MERCK & CO.
AT&T GENERAL MOTORS MICROSOFT
BANK OF AMERICA HEWLETT-PACKARD PFIZER
BOEING HOME DEPOT PROCTER&GAMBLE
CATERPILLAR INTEL UNITED TECHNOLOGIES
CHEVRON INTERNATIONAL
BUS.MCHS.
VERIZON
COMMUNICATIONS
CITIGROUP JOHNSON & JOHNSON WAL MART STORES
COCA COLA JP MORGAN CHASE & CO. WALT DISNEY
1
(Available at http://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library.html#International)
8
The approach here is to study the behaviour of the return distribution along the quantiles,
using quantile regression. The coefficients for all the three factors of the model are
calculated both by virtue of their means using OLS and in their quantiles applying quantile
regressions. While OLS calculates the coefficients around the mean, quantile regression
calculates the values for the .05, .25, .50, .75 and .95 quantiles, at 95 percentile confidence
levels.
2
After studying the behaviour of the returns along the quantiles of the distribution,
we use the three factor model for portfolio formation. We use a simple Sequential
Quadratic Programming routine with the help of MATLAB, to minimize risk and
mazimise return for portfolio formation. A hold out period of one year is taken to roll over
the weights calculated from the previous year’s returns to the stock returns of next year to
explore the outcomes of portfolios selected using this method and to compare their
effectiveness with portfolios formed using OLS.
5. QUANTILE ANALYSIS OF FAMA-FRENCH FACTORS
We use OLS regression analysis and quantile regression analysis to calculate the three
Fama-French coefficients. Figure 2, gives an example of the Bank of America stock’s
actual and fitted values obtained from the two regression methods for the year 2008.
Exhibit-a from Figure 2 shows how the actual and fitted values run through the mean of the
distribution for OLS and the next two exhibits, b and c shows the use of quantile
regressions in efficiently capturing the lower and upper tails of the return distribution.
Figure-3, Figure-4, and Figure-5 provide a three dimensional area plot for the quantile
estimates for all the stocks for the year 2007, these figures show how the values are non
uniform across the quantiles and the effect can increase in the lower and upper quantiles, a
feature that is ignored by OLS. The figures present the quantile estimates of beta, the size
effect and the value or book to market effect respectively.
This analysis shows that the three-factor model can provide even more useful risk
information, if it is used in combination with quantile regressions, as we display in the next
stage of our analysis in which we form portfolios.
2
GRETL an open source software is used for OLS and Quantile Regression estimates plus STATA.
Figure 2
: OLS and Quantile
Figure 3
: Beta for stocks across quantiles
: OLS and Quantile
Regression Fitted Versus Actual Values
: Beta for stocks across quantiles
9
Regression Fitted Versus Actual Values
Figure 4
: Size effect for stocks across quantiles
Figure 5
: Value(HML) effect for stock
: Size effect for stocks across quantiles
: Value(HML) effect for stock
s across quantiles
10
11
6. PORTFOLIO FORMATION USING THE FAMA-FRENCH THREE
FACTOR MODEL
We now proceed to portfolio analysis using the three factor model and OLS and quantile
regression estimates. As stated earlier; quantile regression provides better estimates along
the tails of the distribution and hence accounts for risk more efficiently than OLS. We now
introduce an additional advantage of quantile regression whereby its estimated coefficients
can be combined by certain weighting schemes to yield more robust measurements of
sensitivity to the factors across the quantiles, as opposed to OLS estimates around the
mean. This approach was originally proposed by Chan and Lakonishok (1992) in a paper
which featured simulations to establish the facility of quantile regressions in equity beta
estimations. Their results show that the weighted average of quantile beta coefficients is
more robust than the OLS beta estimates. We will test two weighting schemes for robust
measurement of size and book to market effects based on the quantile regression
coefficients. The resulting estimators have weights which are the linear combination of
quantile regression coefficients.
We will use Tukey’s trimean as our first estimator:
ߚ
௧
= 0.25ߚ
.ଶହ,௧
+ 0.5ߚ
.ହ,௧
+ 0.25ߚ
.ହ,௧
, (8)
ݏ
௧
= 0.25ݏ
.ଶହ,௧
+ 0.5ݏ
.ହ,௧
+ 0.25ݏ
.ହ,௧
, (9)
ℎ
௧
= 0.25ℎ
.ଶହ,௧
+ 0.5ℎ
.ହ,௧
+ 0.25ℎ
.ହ,௧
, (10)
ߙ
௧
= 0.25ߙ
.ଶହ,௧
+ 0.5ߙ
.ହ,௧
+ 0.25ߙ
.ହ,௧
, (11)
These are the weighted average of the three quantile estimates. We will test this along with
another robust estimator with symmetric weights covering all the quantile estimates, i.e.
0.05, 0.25, 0.5, 0.75, 0.95.
ߚ
௧
= 0.05ߚ
.ହ,௧
+ 0.2ߚ
.ଶହ,௧
+ 0.5ߚ
.ହ,௧
+ 0.2ߚ
.ହ,௧
+ 0.05ߚ
.ଽହ,௧
(12)
ݏ
௧
= 0.05ݏ
.ହ,௧
+ 0.2ݏ
.ଶହ,௧
+ 0.5ݏ
.ହ,௧
+ 0.2ݏ
.ହ,௧
+ 0.05ݏ
.ଽହ,௧
(13)
ℎ
௧
= 0.05ℎ
.ହ,௧
+ 0.2ℎ
.ଶହ,௧
+ 0.5ℎ
.ହ,௧
+ 0.2ℎ
.ହ,௧
+ 0.05ℎ
.ଽହ,௧
(14)
ߙ
௧
= 0.05ߙ
.ହ,௧
+ 0.2ߙ
.ଶହ,௧
+ 0.5ߙ
.ହ,௧
+ 0.2ߙ
.ହ,௧
+ 0.05ߙ
.ଽହ,௧
(15)
12
The portfolio problem using the Fama-French three factor model, requires a solution for
minimum risk and maximum return. The return and risk of the portfolio is as presented in
equations 16 and equation 17.
ܴ݁ݐݑݎ݊
=∑ߚ
ܣݒ݃(ݎ
ெ
− ݎ
ி
)ݓ
+ ݏ
ܣݒ݃(ܵܯܤ)ݓ
+ ℎ
(ܪܯܮ)ݓ
+ ߙ
ݓ
ୀଵ
(16)
ܴ݅ݏ݇
=∑ߚ
ଶ
ୀଵ
ܸܽݎ(ݎ
ெ
− ݎ
ி
)ݓ
ଶ
+ ݏ
ଶ
ܸܽݎ(ܵܯܤ)ݓ
ଶ
+ ℎ
ଶ
ܸܽݎ(ܪܯܮ)ݓ
ଶ
(17)
This forms a classical portfolio optimization problem of minimizing risk (equation 17) and
maximising the return (equation 16). We apply here sequential quadratic programming
which is also referred to as recursive quadratic programming and is used for solving
general non linear programming problems. (For details refer to Robust portfolio
optimization and management, Frank J. Fabozzi, Petter N. Kolm, Dessislava
Pachamanova, page 284-285). We leave the mathematical details of the algorithm for the
sake of brevity. MATLAB’s optimization toolbox is used to execute the algorithm, with,
additional constrain of maximum 10% weight per asset for well diversified portfolio and
minimum of 0% daily return on the portfolio formed (to prevent optimization from
generating optimized weights for negative portfolio returns).
We generate portfolios using historical data for three consecutive years, 2005, 2006 and
2007 with a following hold out period of one year in each case. We use OLS and quantile
risk measures with Tukey’s trimean and symmetric weights to generate three different
portfolios. We then roll over the weights as calculated by these respective routines to the
next year and calculate the realized return and risk for the next year for each of the three
portfolios. Risk for the rolled over period used as a hold-out sample is the actual
diversifiable risk calculated using the covariance of the daily returns of the stocks and the
weights of the selected portfolios.
We then compare the realized return and risk for the next year obtained from maintaining
the portfolio through the hold out period using the Sharpe Index so as to analyse which
portfolio performs better in times of severe financial distress.
13
Table 2, Table 3, and Table 4, give the weights generated from the historical data of the
years 2005, 2006 and 2007 respectively. W1, W2, W3 represent the weights for quantile
regression coefficients using Tukey’s trimean, the quantile regression coefficients with
symmetric weights and the OLS coefficients respectively.
Table 2: Portfolio Weights from Year 2005 Data
Stocks
3m
ALCOA
American EX
AT&T
BOA
Boeing
Caterpillar
Chevron
Citi Grp
Coca Cola
Ei Du PONT
Exxon
General
Electric
GM
HP
Home Depot
Intel
IBM
J&J
JP MORGAN
Kraft Food
MacD
Merk & Co
Microsoft
Pfizer
P&G
United Tech
Verizon
WalMart
Walt Disney
W1
0.06229200
0.01290900
0.02474200
0.07121300
0.04663300
0.06837600
0.03617300
0.02718700
0.07690000
0.05264500
0.00240340
0.01976700
0.01643300
0.00000000
0.07385500
0.01057300
0.03736100
0.02116300
0.02146600
0.02578700
0.03597500
0.02907300
0.07557900
0.03363300
0.00000000
0.06333700
0.03444800
0.00000000
0.01631000
0.00376770
W2
0.05610500
0.01543800
0.02767300
0.05595500
0.04683900
0.06724900
0.03536900
0.02274000
0.06416700
0.05184300
0.00237950
0.01873900
0.02068200
0.00000000
0.06914300
0.00832410
0.03732700
0.02294500
0.03809600
0.02795800
0.05065400
0.03247500
0.06570100
0.03609500
0.00000000
0.06208600
0.03508400
0.00000000
0.02151200
0.00742090
W3
0.04195300
0.02106300
0.02846100
0.04779500
0.03813200
0.05689900
0.02395100
0.01319100
0.04120500
0.06206800
0.02023100
0.01069700
0.03077100
0.00000000
0.07248900
0.01841600
0.02088100
0.02155600
0.05243100
0.03209500
0.06278900
0.02674400
0.04616000
0.03442900
0.01821700
0.03596300
0.02665100
0.01833700
0.04983900
0.02658600
Table 3: Portfolio Weights from Year 2006 Data
Stocks
3m
ALCOA
American EX
AT&T
BOA
Boeing
Caterpillar
Chevron
Citi Grp
Coca Cola
Ei Du PONT
Exxon
General
Electric
GM
HP
Home Depot
Intel
IBM
J&J
JP MORGAN
Kraft Food
MacD
Merk & Co
Microsoft
Pfizer
P&G
United Tech
Verizon
WalMart
Walt Disney
W1
0.03284900
0.01351000
0.02335900
0.03976500
0.02918900
0.02010400
0.01649400
0.01055100
0.02870300
0.04776700
0.03455100
0.01253800
0.04406400
0.02521600
0.02096300
0.02944500
0.01605000
0.03349600
0.04796600
0.01543100
0.10000000
0.04972700
0.02373600
0.03001600
0.02322900
0.07478600
0.04301200
0.02917800
0.04295900
0.04134800
W2
0.03362400
0.01440800
0.02438800
0.03841900
0.02966100
0.02069000
0.01700400
0.01124700
0.02892000
0.04909600
0.03553700
0.01309600
0.04298500
0.02807200
0.01968800
0.03056700
0.01569300
0.03418700
0.04595400
0.01546500
0.10000000
0.04953500
0.02367400
0.02829300
0.02316100
0.07337400
0.04222300
0.02927200
0.04418600
0.03758400
W3
0.04019300
0.01544900
0.02633800
0.03662300
0.02955800
0.02527500
0.01883900
0.01266900
0.02403100
0.05077000
0.03538400
0.01420500
0.03480700
0.03376900
0.01996100
0.02740400
0.01296800
0.03365800
0.04034600
0.01687000
0.10000000
0.05555300
0.02729900
0.02066000
0.02371600
0.06627500
0.04210000
0.03017800
0.04350900
0.04159400
14
Table 4: Portfolio Weights from Year 2007 Data
Stocks
3m
ALCOA
American EX
AT&T
BOA
Boeing
Caterpillar
Chevron
Citi Grp
Coca Cola
Ei Du PONT
Exxon
General
Electric
GM
HP
Home Depot
Intel
IBM
J&J
JP MORGAN
Kraft Food
MacD
Merk & Co
Microsoft
Pfizer
P&G
United Tech
Verizon
WalMart
Walt Disney
W1
0.054393
0.013114
0.010286
0.021150
0.019031
0.036850
0.023927
0.017666
0.010951
0.071481
0.027253
0.015905
0.026498
0.009273
0.024137
0.020272
0.020009
0.037055
0.100000
0.011823
0.041194
0.060709
0.042629
0.025494
0.033774
0.100000
0.032082
0.034559
0.031722
0.026762
W2
0.100000
0.000000
0.000000
0.004994
0.000000
0.000000
0.037533
0.069037
0.000000
0.100000
0.000000
0.039803
0.014420
0.000000
0.066279
0.000000
0.067342
0.100000
0.027851
0.000000
0.000000
0.100000
0.000000
0.026575
0.000000
0.100000
0.072465
0.073701
0.000000
0.000000
W3
0.040137
0.011691
0.010973
0.023461
0.019723
0.037031
0.023411
0.017536
0.011810
0.078598
0.029683
0.015837
0.027831
0.012645
0.022169
0.024160
0.017283
0.033876
0.100000
0.012573
0.046818
0.076052
0.043153
0.024309
0.035079
0.078537
0.031047
0.033326
0.034201
0.027047
Table 5: Final Risk and Return for all the three types of weights after a roll over
period of a year
Stocks
3m
ALCOA
American EX
AT&T
BOA
Boeing
Caterpillar
Chevron
Citi Grp
Coca Cola
Ei Du PONT
Exxon
General
Electric
GM
HP
Home Depot
Intel
IBM
J&J
JP MORGAN
Kraft Food
MacD
Merk & Co
Microsoft
Pfizer
P&G
United Tech
Verizon
WalMart
Walt Disney
Ret (2006)
0.00553305
0.01477033
0.16464454
0.37828540
0.14572650
0.23491905
0.05979951
0.25868547
0.13779798
0.17979625
0.13638027
0.31060584
0.05980193
0.45861045
0.36374039
-
0.00793655
-
0.20911975
0.16710087
0.09394788
0.19633229
0.23689311
0.27357052
0.31527644
0.13267060
0.10493161
0.10472106
0.11174325
0.24929073
-
0.01333640
0.35746867
Ret (2007)
0.07880810
0.19715053
-0.15382108
0.15058790
-0.25772995
-0.01565545
0.16814469
0.23844811
-0.63762585
0.24052529
-0.09965135
0.20100281
-0.00376953
-0.21044779
0.20338171
-0.39925930
0.27500952
0.10680055
0.01024723
-0.10122828
-0.08991858
0.28434921
0.28728062
0.17582585
-0.13055733
0.13310342
0.20232697
0.15973593
0.02881429
-0.04643890
Ret(2007)
-0.38213873
-1.17742456
-1.03115847
-0.37723408
-1.07513816
-0.71768545
-0.48511167
-0.23248199
-1.47875534
-0.30427655
-0.5554286
-0.16009209
-0.82779677
-2.05131531
-0.330035
-0.15724877
-0.59804162
-0.25033698
-0.10869775
-0.32526381
-0.19496629
0.05418336
-0.64789516
-0.60501284
-0.24955619
-0.17196944
-0.35626441
-0.24901024
0.165061455
-0.35252354
2006
Quantile Regression (Trimean) Quantile Regression (Symmetric Weights
)
OLS
Return Risk Return Risk Return Risk
0.17707517 0.00623073 0.17178290 0.00613572 0.17189773 0.00603117
Sharpe Ratio 25.20975199 24.73760491 25.18545442
2007
Quantile Regression (Trimean) Quantile Regression (Symmetric Weights
)
OLS
Return Risk Return Risk Return Risk
0.02816570 0.00880364 0.02651735 0.00883844 0.03048237 0.00886899
15
Sharpe Ratio 1.72266282 1.52938198 1.97117930
2008
Quantile Regression (Trimean) Quantile Regression (Symmetric Weights
)
OLS
Return Risk Return Risk Return Risk
-0.35877338 0.02189977 -0.28532662 0.02118938 -0.35993622 0.02202203
Sharpe Ratio -16.56516722 -13.65431927 -16.52600367
Table 5, provides the final risk and returns after a hold out period of a year. The risk
(standard deviation), is the total portfolio risk calculated using the covariance of daily
returns of the stocks and the relevant weights. Return is calculated using the first and the
last day’s prices for the stocks for the particular year; the annualized rate of return. The
Sharpe ratio values indicate the efficiency of the portfolios formed through the three
different regression estimates. We can quickly analyse the effectiveness of the portfolios
based on the Sharpe ratio, which is the excess return of a portfolio divided by its risk.
We analyse the return and risk profiles of the portfolios based on the Sharpe Index and also
on the basis of their risk. For the years 2006, and 2007 we can see that the portfolios
formed using OLS do well, as these periods coincide with at time when market was stable
and there were no major losses of the scale that occurred in the year 2008 as a result of the
GFC, yet even so, during these periods the portfolios formed using quantile regressions
performed reasonably well.
Figure 6 shows the returns for all three portfolios for the three observation years, (the
return lines for portfolio 1 and portfolio 3 are almost overlapping due to similar returns).
These years range in period from pre GFC to the onset and establishment of the GFC. The
returns of the portfolios present a rational picture consistent with these varying
circumstances. We can see from Figure 6 that the three test portfolios performed almost
equally well in the year 2006, as the distribution of the returns in the prior historical
analysis period, in which the weights were formed, i.e. the year 2005 were less skewed
towards the lower tails; as they were in years prior to the financial crisis period. We can
further conclude from Figure 6 that as we approach closer to the financial crisis period, our
symmetrically weighted quantile regression coefficient portfolio begins to perform better
than the other two methods; given that during the time of financial distress the return
distributions are more skewed towards the lower tails and portfolio selection methods
based on OLS and the Tukey’s trimean quantiles are unable to capture these extreme
characteristics of the return distributions, and hence unable to give a proper measure of the
risks involved. Our portfolio analyses show the usefu
analysis, as a tool for the quantification of the tail risks involved with the return
distributions of financial assets.
Figure 6: Portfolio Returns across Years
based on OLS and the Tukey’s trimean quantiles are unable to capture these extreme
characteristics of the return distributions, and hence unable to give a proper measure of the
risks involved. Our portfolio analyses show the usefu
l applicability of quantile regression
analysis, as a tool for the quantification of the tail risks involved with the return
distributions of financial assets.
Figure 6: Portfolio Returns across Years
16
based on OLS and the Tukey’s trimean quantiles are unable to capture these extreme
characteristics of the return distributions, and hence unable to give a proper measure of the
l applicability of quantile regression
analysis, as a tool for the quantification of the tail risks involved with the return
17
Our main focus is the period of immense financial distress and downturn in equity markets.
We are testing here, whether quantile regression was able to predict the heavy risks and
whether its application helps to reduce the losses that occurred during this particularly
extreme hold out period. The analysis of portfolios held during the year 2008 clearly shows
that the portfolio formed with symmetric weights from the quantile regression coefficients,
which automatically covered both the extreme lower and upper bounds of the return
distributions performed better than the other two methods. This portfolio saved around 2%
of the relative potential losses to the investor.
The analysis shows that a well distributed quantile regression analysis of historical returns
can give better estimates of the inherent risks than standard OLS analysis. We also show
that the weighting scheme tested here proves more effective in capturing information from
the extreme quantile coefficients that receive more emphasis than that given in the other
two methods considered.
7. CONCLUSION
In this paper we have introduced quantile regression as a tool for investment analysis and
portfolio management. Our study shows that quantile regression can provide more
effective use of information in the entire distribution than is the case with estimates from
the customarily used OLS. We can achieve more efficient risk measures using this robust
regression technique. The technique becomes particularly useful when we want to analyse
the behaviour in the tails of the distributions of returns or to capture a more complete
picture of the risk of a financial instrument. Our analysis suggests that further research
using quantile regression in the context of the application of linear asset pricing models
and their empirical effectiveness in extreme market conditions for portfolio formation is
likely to be fruitful.
18
REFERENCES
Black, Fischer, (1993)"Beta and Return," Journal of Portfolio Management 20 pp: 8-18.
Barnes, Michelle and Hughes, Anthony (Tony) W.,A Quantile Regression Analysis of the
Cross Section of Stock Market Returns(November 2002). Available at SSRN:
http://ssrn.com/abstract=458522
Buchinsky, M., Leslie, P., (1997), Educational attainment and the changing U.S. wage
structure: Some dynamic implications. Working Paper no. 97-13, Department 36 of
Economics, Brown University.
Buchinsky, M., Hahn, J., (1998), “An alternative estimator for the censored quantile
regression model”, Econometrica 66, 653-671.
Chan, L.K. C. and J. Lakonishok, (1992) "Robust Measurement of Beta Risk", The Journal
of Financial and Quantitative Analysis, 27, 2, pp:265-282
Robert F. Engle and Simone Manganelli, (1999) "CAViaR: Conditional Value at Risk by
Quantile Regression," NBER Working Papers 7341, National Bureau of Economic
Research, Inc.
Fama, E.F. and K.R. French (1992) The Cross-section of Expected Stock Returns, Journal
of Finance, 47, 427-486.
Fama, E.F. and K.R. French (1993) Common Risk Factors in the Returns on Stocks and
Bonds, Journal of Financial Economics, 33, 3-56.
Gowlland, C. Z. Xiao, Q. Zeng, (2009) “Beyond the Central Tendency: Quantile
Regression as a Tool in Quantitative Investing”, The Journal of Portfolio
Management, pp:106-119
Hsu, Chih-Chiang and Chou, Robin K.,Robust Measurement of Size and Book-to-Market
Premia. EFMA 2003 Helsinki Meetings. Available at SSRN:
http://ssrn.com/abstract=407726 or DOI: 10.2139/ssrn.407726
Koenker, Roger W and Bassett, Gilbert, Jr, (1978), "Regression Quantiles," Econometrica,
Econometric Society, vol. 46(1), pages 33-50
Koenker aRoger W and Graham Hallock, (2001), “Quantile Regression”, Journal of
Economic Perspectives, 15, 4, pp:143–156
Lintner, John (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.
19
Ma, Lingjie and Pohlman, Larry,Return Forecasts and Optimal Portfolio Construction: A
Quantile Regression Approach(December 19, 2005). Available at SSRN:
http://ssrn.com/abstract=880478
Mossin, Jan, "Equilibrium in a Capital Asset Market", Econometrica, 34, 1966, pp. 768–
783.
Treynor, Jack L. (1961). "Market Value, Time, and Risk". Unpublished manuscript dated
8/8/61, No. 95-209.
Treynor, Jack L. (1962). "Toward a Theory of Market Value of Risky Assets".
Unpublished manuscript. Subsequently published as Chapter 2 of Korajczyk (1999).
Sharpe, W., 1963, A simplified model for portfolio analysis, Management Science 9, 277-
293.
Sharpe, William F. (1964), “Capital asset prices: A theory of market equilibrium under
conditions of risk”, Journal of Finance, 19 (3), 425-442.
