This action might not be possible to undo. Are you sure you want to continue?
https://www.scribd.com/doc/65683506/CrossSectionalPredictabilityofstockreturnsPreWorldWarIEvidence
08/09/2013
text
original
DEPARTMENT OF ACCOUNTING AND FINANCE
DISSERTATION
CrossSectional Predictability of Stock
Returns:
PreWorld War I evidence
THESIS SUBMITTED IN ORDER TO OBTAIN THE DEGREE
OF
DOCTOR IN APPLIED ECONOMICS
Author: Supervisor:
Lord Mensah Prof. Dr. Jan Annaert
September, 2011
This research was funded by the Faculty of Applied Economics, from the 2007 to 2011
research grant for the Accounting and Finance Research group.
CrossSectional Predictability of Stock return: PreWorld War I Evidence
Lord Mensah, Antwerpen, Belgium, 2011
ISBN: 9789089940445
Printed by: Universitas Antwerpen
© Copyright 2011, Lord Mensah
All rights reserved. No part of this book may be reproduced or transmitted in any form
by any electronic or mechanical means (including photocopying, recording or
information storage and retrieval) without permission in writing from the author.
Doctoral Jury
Internal members
Prof. Dr. Jan Annaert (Supervisor)
Universiteit Antwerpen, Belgium
Department of Accounting and Finance
Jan.annaert@ua.ac.abe
Prof. Dr. Marc De Ceuster (chair/PhD committee chair)
Universiteit Antwerpen, Belgium
Department of Accounting and Finance
marc.deceuster@ua.ac.be
Dr. Frans Buelens (PhD committee member)
Universiteit Antwerpen, Belgium
Department of Accounting and Finance
Fran.buelens@ua.ac.be
Prof. Dr. Joseph Plasmans
Universiteit Antwerpen, Belgium
Department of Economics
joseph.plasmans@ua.ac.be
External Members
Prof. Dr. Wim Janssens (PhD committee member)
Universiteit Hasselt, Belgium
Department of Marketing
wim.janssens@uhasselt.be
Prof. Dr. Laurens Swinkels
Erasmus School of Economics, Rotterdam, Netherlands
Vice President, Robeco Investment Solutions, Rotterdam
lswinkels@ese.eur.nl
Prof. Dr. Patrice Fontaine
EUROFIDAI and University of Grenoble 2, France
patrice.fontaine@eurofidai.org
ACKNOWLEDGEMENTS
In as much as writing a PhD thesis requires independent study and scientific research, the
success marked in this project depend on the environment I worked in and the people
around me. As a result, I may like to express my profound gratitude to everyone who
contributed to the success of this project both professionally and personally.
Primarily, I would like to thank my supervisor Prof. Dr. Jan Annaert for the confidence
pose in me, and giving me the opportunity to be part of the research project. You guided
me through the research and thought me how to select the literature to support my
arguments. In a motherchild scenario, I will say you know the right food combination
you will give to your child in order to mature quickly. I also appreciate how quick you
read my chapters and direct me for changes here and there. In short, this dissertation
would not be what it is now without your suggestions, encouragement and guidance.
Secondly, I would like to express my sincere gratitude to my PhD committee members
Prof. Dr. Mark De‟Ceuster, Prof. Dr. Wim Janssens and Dr. Frans Buelens. Your
comments and suggestions have been of immense value in completing this dissertation.
Dr. Frans Buelens, thank you very much for the updates you gave me on the data. In
addition, I am also thankful to the external doctoral jury members Prof. Dr. Joseph
Plasmans, Prof. Dr. Laurens Swinkels and Prof. Dr. Patrice Fontaine for taking your time to read
and discuss my work. Your comments and suggestions have been of great value to improve and
polish the chapters in this thesis.
Genuinely, I have enjoyed every moment at work for the past four years. On this note, I
want to thank all my colleagues from the Accounting and Finance department of the
University of Antwerp. Specifically, my sincere gratitude goes to the secretariat office
for their beaming smiles and friendly services. I want to thank the department head for
the annual outing, lunches and visit to museums in Antwerp. This allows me to release
stress at work for the past four years. I hope that we will stay in touch for the future.
Finally, the period of writing PhD thesis comes with moments of joy and doubt.
However, how to handle the stress depends on the people around you after office work. I
have been blessed with supportive family. Firstly, I owe an immense debt to my wife
Sefam Ekua LordMensah. What you have done for the past four years goes beyond your
call to duty. I have come to understand that only few people are fortunate to have caring
and supporting wife like you. Completion of my PhD, gives me the opportunity to thank
you for the unconditional love and support you have given me, since we got married.
Second words go to my lovely kids Lakeisha LordMensah, Zita LordMensah and
Nhyira LordMensah. Thank you for making my home a place of endless happiness.
Thank you for your patience and not been on your sight ten to eleven hours every day.
My final words go to all those who made my education a success by contributing
financially. Special thanks to my one and only Uncle Maxford Kwadwo Kwakye who
contributed to my senior high school education. I do appreciate your support, and I will
always remember you.
Lord Mensah
Table of Contents
CHAPTER 1 .................................................................................................................................. 1
1 INTRODUCTION, DATA DESCRIPTION AND SUMMARY ................................... 1
Introduction .................................................................................................................... 1 1.1
1.1.1 The aim of this Dissertation .................................................................................... 6
Data Description and Methodology ................................................................................ 7 1.2
1.2.1 The Riskfree rate and the Market Index .............................................................. 11
1.2.2 Methodology ......................................................................................................... 14
Summary of the chapters .............................................................................................. 20 1.3
1.3.1 Chapter 2 ............................................................................................................... 20
1.3.2 Chapter 3 ............................................................................................................... 21
1.3.3 Chapter 4 ............................................................................................................... 22
1.3.4 Chapter 5 ............................................................................................................... 23
CHAPTER 2 ................................................................................................................................ 25
2 ASSESSMENT OF BETA IN THE 19
th
CENTURY BSE ........................................... 25
Introduction and Literature Review .............................................................................. 25 2.1
Beta Coefficient Descriptive Statistics .......................................................................... 33 2.2
Beta Stability ................................................................................................................. 37 2.3
2.3.1 Blume and Vasicek stability adjustment techniques............................................. 39
Beta Bias ........................................................................................................................ 44 2.4
Impact of outlying observations on Beta ...................................................................... 47 2.5
Conclusion..................................................................................................................... 51 2.6
CHAPTER 3 ................................................................................................................................ 52
3 THE TEST OF CAPITAL ASSET PRICING MODEL (CAPM) AND THE SIZE
EFFECT IN 19
th
CENTURY BSE ............................................................................................. 52
Introduction and Literature Review .............................................................................. 52 3.1
Expected returns of portfolios sorted on betas ............................................................ 54 3.2
3.2.1 The CrossSectional Regressions ........................................................................... 59
Expected Returns, Beta, and the Size Effect ................................................................. 66 3.3
3.3.1 FamaMacBeth CrossSectional Regressions to Test the Size Effect..................... 71
Conclusion..................................................................................................................... 79 3.4
CHAPTER 4 ................................................................................................................................ 80
4 DOES THE MOMENTUM EFFECT EXIST IN THE 19
TH
CENTRURY? ............. 80
Introduction and Literature Review .............................................................................. 80 4.1
Momentum Trading Strategies and their Returns ......................................................... 85 4.2
4.2.1 Expected Returns and Average Size of Quintile Portfolios ................................... 90
4.2.2 Momentum Profit within Size and Betabased Subsamples ................................. 91
Seasonality and Subperiod Analysis of the Momentum Profit .................................... 94 4.3
Post holding Period Momentum Profits ........................................................................ 98 4.4
The Momentum profit and the Market State ............................................................. 103 4.5
Conclusion................................................................................................................... 106 4.6
CHAPTER 5 .............................................................................................................................. 109
5 THE COMBINED EFFECT OF DIVIDEND YIELD, SIZE, TOTAL RISK
AND MOMENTUM (18681913 EVIDENCE)....................................................................... 109
Introduction and Literature Review ............................................................................ 109 5.1
Measures of Characteristics ........................................................................................ 115 5.2
Descriptive Summary Statistics of the Characteristics ................................................ 118 5.3
5.3.1 SizeDividend yield double sorts ......................................................................... 118
5.3.2 MomentumDividend yield double sorts ............................................................ 121
5.3.3 Total riskDividend yield double sorts ................................................................. 125
Average Excess Returns on Portfolio Sorts ................................................................. 130 5.4
The CrossSectional regressions .................................................................................. 136 5.5
5.5.1 Pervasiveness of the CrossSectional Relationships ........................................... 140
Conclusion ................................................................................................................... 143 5.6
6 CONCLUSION .............................................................................................................. 144
References .................................................................................................................................. 150
NEDERLANDSTALIGE SAMENVATTING ........................................................................ 157
TABLES
Table 1.1: Summary Statistics for Riskfree rate and the ValueWeighted Index ......................... 13
Table 2.1: Beta coefficient descriptive statistics for the 15 estimated periods ............................ 34
Table 2.2: Average beta and average coefficient of determination of the size based subsamples
....................................................................................................................................................... 36
Table 2.3: Weighted average of correlation and Spearman rank order correlation across
successive periods ......................................................................................................................... 38
Table 2.4: Measurement of regression tendency of estimated beta coefficient for individual
stocks ............................................................................................................................................. 40
Table 2.5: Predictive performance of Blume and Vasicek (Bayesian) procedures of estimating
beta ............................................................................................................................................... 42
Table 2.6: Modified DieboldMariano test statistics (pvalue in parentheses) ............................. 44
Table 2.7: Dimson Aggregate Coefficient (AC) beta Adjustment .................................................. 45
Table 2.8: Comparison of the market model betas and the iterative reweighted least square
betas .............................................................................................................................................. 48
Table 2.9: Test of equal predictive accuracy between MM and IRLS models ............................... 50
Table 3.1: Time Series Mean (%), Standard Deviation (%), and Postranking Betas of Decile
portfolios formed from preranking betas in Jan. 1868Dec. 1913 .............................................. 58
Table 3.2: Average time series slopes from the FamaMacBeth CrossSectional Regressions in
Jan. 1868Dec. 1913 ...................................................................................................................... 61
Table 3.3: Average Time Series Slopes from FamaFrench CrossSectional Regression in Jan.
1868Dec. 1913 ............................................................................................................................. 64
Table 3.4: Subperiod look into estimated slopes and excess market returns ............................. 65
Table 3.5: Beta Estimate and Mean Excess Return for the BSE equally weighted size portfolios,
Jan. 1868 Dec. 1913 ..................................................................................................................... 68
Table 3.6: Equally weighted portfolios excess returns without the firstsize decile group .......... 69
Table 3.7: Average time series slopes and intercept from the FamaMacBeth crosssectional
regression: Jan 1868Dec. 1913 .................................................................................................... 73
Table 3.8: Average Time Series Slopes and Intercepts from the FamaFrench CrossSectional
Regressions: Jan. 1868Dec. 1913 ................................................................................................. 76
Table 3.9: Average Time Series Slopes and Intercepts from the FamaFrench CrossSectional
Regressions without the FirstSize Decile: Jan. 1868Dec.1913 .................................................... 78
Table 4.1: Profitability of momentum Strategies on BSE (Jan.1868Dec. 1913) ........................... 89
Table 4.2: Average Returns and Average Size of Quintile Momentum Portfolios ........................ 91
Table 4.3: Portfolio Returns of the Momentum Strategies with Size and Beta Subsamples ......... 92
Table 4.4: Seasonality in momentum profits ................................................................................ 95
Table 4.5: Subperiod Analysis of Momentum Profit .................................................................... 97
Table 4.6: Long Horizon Momentum Profits ............................................................................... 100
Table 4.7: The Momentum profit and the Market State ............................................................ 104
Table 5.1: Summary statistics for SizeDividend doublesorts .................................................... 119
Table 5.2: Summary statistics for MomentumDividend yield double sorts .............................. 123
Table 5.3: Summary statistics of total risk dividend yield double sorts .................................... 126
Table 5.4: Annual Time Series Average of the correlation between the entire characteristic and
Average return ............................................................................................................................ 129
Table 5.5: Equal and Valueweighted portfolios excess returns (%) of doublesorted
characteristics ............................................................................................................................. 131
Table 5.6: CrossSectional Regression of Excess Returns on Dividend Yield, Size, Total Risk and
Momentum ................................................................................................................................. 138
Table 5.7: CrossSectional regression of Excess Returns on Dividend Yield, Size, Total Risk and
Momentum of Size subsamples .................................................................................................. 141
FIGURES
Figure 1.1: Number of listings on the 19
th
Century Brussels Stock Exchange ................................. 9
Figure 1.2: Total market capitalization of the BSE ........................................................................ 10
Figure 1.3: Evolution of the value weighted market index in the two subperiods of our study . 12
Figure 2.1: The graph of the average beta of each period for large stocks and small stocks ....... 36
Figure 2.2: Plot of average market model betas and IRLS betas for stocks with outlier
observation less than 4 ................................................................................................................. 49
Figure 3.1: Number of stocks in our selection criteria for the entire period of the preworld war I
SCOB data ...................................................................................................................................... 55
Figure 3.2: Sixty months moving average of the crosssectional slopes and excess market returns
using Dimson beta estimates ........................................................................................................ 65
Figure 3.3: Size Portfolio betas ...................................................................................................... 70
Figure 4.1: Number of common stocks in our sample for the momentum studies...................... 86
Figure 4.2: Time line of sample periods ........................................................................................ 86
Figure 4.3: Average returns of the momentum profit in all calendar months ............................. 95
Figure 4.4: Cumulative Returns for Five years after portfolio formation ................................... 101
Figure 5.5.1: Percentage of Zerodividend paying stocks and their Relative Market Capital: 1868
1913 ............................................................................................................................................. 116
1
CHAPTER 1
1 INTRODUCTION, DATA DESCRIPTION AND SUMMARY
More than fifty variables have been used to predict returns. The overall picture remains murky, because
more needs to be done to consider the correlation structure among the variables, use a comprehensive set
of controls and discern whether the results survive simple variations in methodology
1
.
Introduction 1.1
The issue about why stock returns differ in the crosssection from one another at a
particular time has been a hot topic of financial research for the past decades. The capital
asset pricing model (CAPM) proposed by Sharpe (1964), Lintner (1965) and Mossin
(1966) seems to provide an adequate description of the crosssection of stock returns
until the 1980s. The model postulates a linear relationship between expected returns and
the covariance between the market portfolio returns and the returns of an asset (beta).
This implies that the asset beta with respect to the market portfolio is sufficient to
determine its expected returns. However, since the 1980s, documentation on deviations
from the model (anomalies) has been an extremely active area of research. Either the
anomalies represent inadequacies in the CAPM, inefficiencies in the market (profit
opportunities) or a data snooping bias. If the presence of an anomaly truly indicates the
inadequacy of the CAPM, then factors other than beta can predict stock returns. This
implies that anomalies would continue to exist before and after their discovery. While
many have accepted the factors other than beta to predict returns, others believe that they
were discovered out of luck and are due to data snooping bias (Lo and MacKinlay
1
Subrahmanyam, Avanidhar, 2010, The crosssection of expected stock returns: What have we learnt from
the past twentyfive years of research?, European Financial Management 16, 2742.
2
(1990)). The possible ways that data snooping bias can occur are (i) when researchers
continuously test the properties of a data set or the outcomes of other studies on data set
(ii) form predictive models based on the characteristics of the previous results and (iii)
test the power of their models on the same data set. As a result, any anomaly found
might appear to be valid within the data set, but they would have no statistical
significance outside the sample from which it was discovered. Indeed, Schwert (2003)
documents that anomalies often seem to disappear, reverse, or attenuate after they have
been documented and analyzed in the academic literature. The problem can be addressed
by using data from markets that have not been searched exhaustively, or by making
predictions using periods that are new to asset pricing research. The situation remains
murky, as more needs to be done to distinguish between the data snooping hypothesis
and the persistence of the characteristics other than beta. Many studies have tried to
differentiate between the two possibilities, but most of them concentrate on the post
World War I data, usually limited to the USA. On the international front, Haugen and
Baker (1996) find some degree of commonality in the characteristics that are most
important in determining comparative expected returns among different stocks. Because
of this, testing the asset pricing models on a new data set provides an obvious way to
distinguish between the data snooping hypothesis and the persistence of the
characteristics identified to predict returns. This is to test whether the anomaly exists in a
new and independent sample. This dissertation fills these gaps in the literature by
introducing an independent data set and a different set of characteristics to test the cross
sectional predictability of stock returns.
3
Several characteristics have been studied in the literature to predict stock returns. Among
them are size (market capitalization: measured as price times shares outstanding), book
to market value ratio, momentum, return reversal, price earnings ratio, dividend yield,
accruals, illiquidity, net issues, asset growth, profitability, total risk (idiosyncratic risk),
etc. Subrahmanyam (2010) categorizes the origins of the predictive characteristics based
on four principles. These are:
 Theoretical motivation based on riskreturn model variants.
 Informal Wall Street wisdom (such as value and size investing)
 Predictors originated from the behavioral biases of investors.
 Models that include market frictions (illiquidity).
The theoretically motivated risk/return models conduct a test to see whether a higher
return has been associated with higher risk (measured as beta). Black, Jensen and Scholes
(1972), Sharpe and Cooper (1972) and Fama and MacBeth (1973) find support for the
CAPM in the 70s. On the contrary, Fama and French (1992) find that their data do not
appear to support the CAPM. The traditional CAPM argues that only market risk should
be incorporated into asset prices and command compensation. However, Malkiel and Xu
(2006) document that the CAPM may not hold when some investors are not able to hold
the market portfolio, due to various reasons such as transaction costs. Investor‟s inability
to hold market portfolios will force them to care about the total risk, not simply the
market risk, as implied by the CAPM. Hence, idiosyncratic risk would be priced in the
market. In effect, there is a positive relationship between idiosyncratic risk and average
returns. On the contrary, Ang, Hodrick, Xing and Zhang (2006) find a negative
relationship between idiosyncratic risk and expected returns. The negative relationship
4
between idiosyncratic risk and expected returns remains debatable, since no theoretical
framework has been established to determine the source. Idiosyncratic risk may also
limit arbitrageurs from exploiting mispricing opportunities on the market. That is, for
high idiosyncratic risk stocks, it is a challenge to execute arbitrage activity that is free
from idiosyncratic risk; hence, such stocks are more likely to trade at a price far from
their fundamental values. Doukas, Kim and Christos (2010) find a positive relationship
between mispricing and idiosyncratic risk, which is consistent with the limited arbitrage
argument. The arbitrage theory posits that mispricing can persist whenever the cost of
arbitrage exceeds the benefit. In addition, as a specific example, McLean (2010) finds a
strong reversal mispricing in high idiosyncratic risk firms.
The Wall Street wisdom characteristics are just found by chance or motivated by
informally appealing to the knowledge of finance professionals. The premise of these
characteristics is not based on any prior theoretical reasoning. For example, Basu (1977)
documents the negative relationship between price/earnings ratio and abnormal returns.
He partially based his assertion on the notion that recommending stocks based on
price/earnings ratio is common on the Wall Street. Similarly, Banz (1981) documents
that stocks with low market capitalization outperforms those with high market
capitalization. The literature on the informal Wall Street characteristics is given a
magnificent boost by Fama and French (1992). They document the importance of size
and book/market value in the crosssection of expected stock returns. In addition, they
show the CAPM is not supported in their data. Fama and French (1993) build on the role
of size and value to postulate that a three factor model based on factors formed on size,
book/market value characteristics and the market returns can explain expected returns.
5
On the return predictions based on past performance, Jegadeesh and Titman (1993)
document the prediction of three to twelve months of past returns. They find that stocks
or portfolios that have performed well in the past 3 to 12 months will continue to do so in
the next 3 to 12 months (momentum). However, the source of the momentum effect is
subject to debate. Conrad and Kaul (1998) and Bulkley and Nawosah (2009) argue that
the effect is mainly due to the crosssectional variation in expected returns. On the
contrary, Jegadeesh and Titman (2001) do not only address the datamining critique in
explaining momentum effect but also document that models of investor behavioral bias
offer a good explanation of the momentum effect. Their argument is because momentum
profit is due to delay in overreactions that are eventually reversed. However, they
indicate the support for the behavioral model should be tempered with caution. Others
believe momentum effect can be explained by the state of the market (Cooper, Gutierrez
and Hameed (2004) and Chabot, Ghysels and Jagannathan (2009)).
The characteristics derived from behavioral biases or cognitive challenges are based on
informal arguments about investor overreaction and underreaction to information. The
premise of the behavioral notion is that the conventional financial theory ignores how
investors take decisions. Daniel, Hirshleifer and Subrahmanyam (1998) argues that if an
investor places more weight on his private information signal, it causes stock price
overreaction. As information that is more public arrives, the prices move closer to their
fundamental values. The overreaction correction pattern is consistent with a long run
negative autocorrelation in stock returns (longterm reversal in returns). On the other
hand, if the investor begins with unbiased beliefs about his private information signal, he
tends to see a new public information signal as a confirmation of his private signal. This
6
suggests that the public can also trigger further overreaction to the preceding private
signal, thereby causing momentum in security prices. In effect, the biased selfattribution
implies the shortrun momentum and longterm reversal. With the notion of investor
reactions to value, Cooper, Gulen and Schill (2008) show that growth in book assets are
crosssectionally related to future returns, and the implication is that investors underreact
to information in the time series of balance sheets.
Investors require compensation for market friction (illiquidity). Hence, market friction is
a predictor of stock returns. The most difficult issue in relating market illiquidity to
expected returns is the measure of illiquidity. There is a pack of measures in the literature
related to illiquidity. Some of these include the bidask spread, absolute return to the
dollar trading ratio, relationship between price changes and order flows, share turnover,
the proportion of zero returns, market capitalization, etc. Amihud and Mendelson (1986)
find a significant premium for the bidask spread measure. In a reaction to the possible
problem associated with providing a consistent liquidity measure for all markets, Amihud
(2002) proposes price impact measures, defined as the absolute value of stock returns
divided by the dollar volume. He finds a positive relationship between his illiquidity
measure and average return. From emerging markets, Bekaert, Harvey and Lundblad
(2007) use the proportion of zero returns as a measure of illiquidity to establish a positive
relationship with expected returns.
1.1.1 The aim of this Dissertation
In this dissertation, we study the robustness of the crosssectional patterns in stock
returns, using the completely independent database of the Brussels Stock Exchange (BSE
from here onwards) in the 19
th
century and the first few years of the 20
th
century. Apart
7
from the quality of the data set, it allows us to study the influence of strong varying
conditions in the economic and institutional environment on stock returns. The period of
the data set avoids the data mining critique. With this data,
 We will test the validity of the CAPM.
 In addition, we test whether other characteristics can explain the crosssectional
variation in stock returns.
As of the time of writing this dissertation, accounting and transaction data has not been
digitalized for the 19
th
century BSE, so we will not be able to investigate the
predictability of accountingrelated characteristics on stock returns. However, we
investigate whether size, momentum (short run past returns), total risk (firm specific risk)
and dividend yield (an asset value' indicator like the booktomarket value ratio) can
crosssectionally predict returns.
With the importance of illiquidity, in the absence of volume and transaction data, we
follow Bekaert et al. (2007) to measure illiquidity as a proportion of zero returns in the
last 30 months. However, our unreported results indicate that zeroreturn illiquidity
measure is strongly correlated with market capitalization (also a proxy for liquidity) and
stock price level.
Data Description and Methodology 1.2
It is worthy to reiterate that empirical research on the stock market return predictability
has a long tradition. The large body of the financial literature in this area uses data from
the Center for Research in Security Prices (CRSP). The CRSP data consist of
comprehensive and accurate historical returns for all stocks listed on the NYSE, the
8
Amex and the NASDAQ stock markets in the USA. Researchers make use of the long
time series (NYSE from 1926 to date, Amex and NASDAQ from 1962 and 1973,
respectively to date) and high quality financial data to carry out research about general
equilibrium asset pricing models and predictable patterns in returns, among others.
However, the obsessed and continuous search for predictable patterns in a single data set
will likely reveal an interesting (spurious) pattern (Lo and MacKinlay (1990)). Since one
of the most studied quantities on the market to date is the stock return, the tests on
financial asset pricing models seem especially at risk. Recent anomalies detected in
empirical studies call for researchers to suggest a modification to standard economic
theories about asset prices. However, the CRSP data are mostly used in asset pricing
research due to the nonavailability of reliable and independent data sources. There is a
considerable danger of inducing data mining/snooping bias when a single data set is used
repeatedly. This may attenuate the reliability of statistical analysis on the data. Data
from financial markets of other countries are normally used to investigate asset return
predictability. However, common characteristics have been identified that can predict
returns on contemporary markets.
To reduce the doubts regarding inferences of the test of asset pricing predictability, we
analyze a completely new and unique historical data set of stock returns from the BSE
during the 19
th
century and the beginning of the 20
th
century. The data set was
constructed at the University of Antwerp in Belgium (Studiecentrum voor Onderneming
en Beurs (SCOB)). The data for our study start from 1832 and ends at 1914, just before
the outbreak of World War I. During the World War I period, the BSE was closed and
this can be regarded as a natural breaking point of the long time series of the stock
9
returns. The BSE was considered one of the biggest markets in the world at that time,
because Belgium was one of the first nations on the European continent to become
industrialized (see Van der Wee (1996) and Neymarck (1911)). On the industrial output
per head ladder, Belgium stood second after Britain in 1860, and third in 1913, after the
UK and the USA (see Bairoch (1982)). During this period, highly developed banking
system coupled with liberal stock market regulations attracted a great deal of domestic
and foreign capital in Belgium. In confirmation, Van Nieuwerburgh, Buelens and
Cuyvers (2006) document that the development of the financial sector, accompanied with
the stock marketbased financing of firms, played an important role in the economic
growth of 19
th
century Belgium.
Figure 1.1: Number of listings on the 19
th
Century Brussels Stock Exchange
Source: Annaert, Buelens and De Ceuster (2004)
Figure 1 indicates the number of listings on the 19
th
century BSE. The figure depicts the
rising popularity of the BSE. The bold line shows the number of common stocks, bonds
10
and preferred stocks. The evolution of common stock listings is represented with the
dashed line. It is clear from the figure that the BSE almost exclusively lists common
stocks until the mid1850s. The importance of the common stocks declines to about 65%
just before the World War I. To illustrate the international attraction of the BSE, the
faint line shows the common stock listings of foreign companies. About one fifth of the
common stock listings from the late 1870s are for foreign companies. The number of
common stock listings on the BSE is quite important, even if we limit our attention to
Belgian common stocks. From 1868 onwards, at least 100 Belgian common stocks are
listed, gradually increasing to about 600 just before the World War I.
Figure 1.2: Total market capitalization of the BSE
Source: Annaert, Buelens and De Ceuster (2004)
It is obvious from Figure 2 that increasing number of common stock listings increases the
market capitalization of the BSE. Notably, the financial liberalization from the early
1970s onwards reflects in the growth of the total market capitalization. Annaert, Buelens
and De Ceuster (2004) shows that in the 19
th
century stock market capitalization was
11
27% of gross national product in 1846, 57% in 1880 and steadily growing to 80% by the
end of the period of our study.
Annaert et al. (2004) classified all companies listed on the BSE into five categories,
based on the geographical location of the major production facilities and the company‟s
country of residence. However, in this research, we restrict ourselves to the analysis of
data from the securities of Belgium owned companies. The stocks on the BSE are well
diversified across industries such as transportation, mining and extraction, financials,
utilities and industrials. The BSE provides monthly data on stock prices, dividends and
the number of shares outstanding for more than 1000 different companies officially
quoted on the exchange. This enables the computation of the market capitalization and
total returns for individual common stocks on the exchange. The total returns of the
common stocks on the BSE were computed by considering dividends, stocks splits,
mergers, delisting and other capital operations. The data permit us to test the robustness
of the predictive power of characteristics on common stock returns, and the analysis of
the anomalies identified in the existing literature on asset pricing. All available
information begins in January and ends in December of each year.
1.2.1 The Riskfree rate and the Market Index
We use the annualized short rate converted to a monthly rate as a proxy for the riskfree
rate. The short rate is based on the commercial rate from the official quotation list of the
Antwerp Stock Exchange. The rate is extracted from the newspapers in the period of our
study.
2
For the market index, we employ the all share monthly value weighted index
constructed by Annaert et al. (2004). We concentrate on the valueweighted index as it
2
Journal du Commerce, d‟Anvers, L‟Avenir, Moniteur des Intérêts Matériels and Het Handelsblad.
12
mirrors the return evolutions of the investable assets' universe. The valueweighted index
is constructed by considering the market capital of the individual stocks in the index
portfolios. The index represents the total returns of the stock market investable assets.
Figure 1.3: Evolution of the value weighted market index in the two subperiods of our study
It is clear from Figure 1.3 that, 100 Belgian Franc invested in the index would have
grown to about 500 Belgian Franc at the end of the period of strict regulation and
industrial revolution (Jan.1832Dec.1867). Evidence from the Figure depicts a market
0
50
100
150
200
250
300
350
400
450
500
1830 1835 1840 1845 1850 1855 1860 1865 1870
Valueweighted index evolution in the period of industrial
revolution: Jan. 1832Dec.1867
0
200
400
600
800
1000
1860 1870 1880 1890 1900 1910 1920
Valueweighted index evolution in the period of
deregulation and expansion: Jan. 1868Dec.1913
13
drop in 1948, the year of industrial revolution across Europe. In the period of the
deregulation and expansion period, 100 Belgian Franc invested in the beginning of the
period would have grown to about 1000 Belgian Franc at the end of the period (Jan.1868
Dec.1913).
Table 1.1: Summary Statistics for Riskfree rate and the ValueWeighted Index
Table 1.1 shows the time series summary statistics of the riskfree rate and the value
weighted index. As indicated in Annaert et al. (2004), the period from 1832 to 1913 is
divided into two subperiods based on the environment in which the BSE was operating.
These were the periods of the industrial revolution and high regulation (18321867) and
the period of deregulation and expansion (18681913). We show the summary statistics
in the entire period and both subperiods. The average riskfree rate, as well as the
average value weighted return for the entire period, is almost the same in the two sub
periods. On the other hand, the standard deviation of the valueweighted index is high in
the first period as compared to the second period. The high standard deviation and
Standard 1st Order
Mean Deviation Skewness Kurtosis Min Median Max Autocorr.
Riskfree 0.26% 0.08% 0.59 3.13 0.12% 0.25% 0.60%
ValueWeighted Index 0.40% 2.42% 0.02 20.92 20.68% 0.27% 15.78% 0.28*
Riskfree 0.28% 0.07% 0.19 2.54 0.17% 0.29% 0.53%
ValueWeighted Index 0.41% 3.01% 0.15 18.61 20.68% 0.21% 15.78% 0.30*
Riskfree 0.24% 0.08% 0.99 4.26 0.12% 0.23% 0.60%
ValueWeighted Index 0.40% 1.86% 0.55 9.28 8.16% 0.34% 11.99% 0.25*
autocorrelation of the valueweighted index series of the riskfree rate and the valueweighted market index.The Ta
ble also displays the descriptive statistics for two subperiods.That is the period of the industrial revolution (18321867)
and the period of deregulation and expansion (18681913). * = significantly different from zero at 5% (two sided).
Jan. 1832Dec. 1913
Jan. 1868Dec. 1913
Jan. 1832Dec. 1867
Note: In this Table, we compute the descriptive statistics of the riskfree rate and the valueweighted market index.
We calculate the time series mean, standard deviation, skewness, minimum, median, maximum and the first order
14
kurtosis in the first period are not surprising, as the period is characterized with (i) a large
market drop in 1848, a year of revolution across Europe (although not in Belgium) and
(ii) strong restrictions on joint stock holdings that may affect the volatility of prices
(Annaert et al. (2004)). The standard deviation recorded on the 19
th
century BSE is low
compared to the high values recorded on the USA market. However, it is close to the
values recorded on the UK market in the period 18701913 (see Grossman and Shore
(2006) page 281, Table 2).
The average difference between the value weighted index and the riskfree rate is
between 0.13% and 0.16% for the overall period and the two subperiods. This implies a
risk premium of 1.92% per annum. This might be too low compared to the annual risk
premium of 4.61% recorded on the USA markets. However, the low risk premium
recorded on the 19
th
century BSE is not surprising as Grossman and Shore (2006) record
far too low annual risk premium on the UK market between 1870 and 1913 ( risk
premium of 0.53% and 1.93% for value weighted and equally weighted respectively).
1.2.2 Methodology
As shown in the diagram below, basically, there are two methods for testing the cross
sectional predictability of stock returns. These are the sorting and the regression method.
The sorting method is subdivided into the univariate, independent and conditional
double sorts. The univariate sort method ranks stocks on beta, or their characteristic, and
groups them to form portfolios. The average returns of the portfolios should be a
monotonic function of the characteristic. The method tests the significance of the
difference in average returns between the high and low ranked portfolios. A univariate
sort has the disadvantage of not being able to disentangle the effect of two (or more)
15
characteristics. To distinguish between the effects of two characteristics on the cross
section of stock returns, the conditional double sort is applied. In the conditional double
sort method, stocks are first ranked on a single characteristic and split into groups. The
stocks in each group are then sorted on the second characteristic. The conditional double
sorting method keeps one characteristic constant and tests the effect of the other
characteristic on the crosssection of expected returns.
Fama and French (1992) adopted the conditional double sort to disentangle the effects of
beta and size on average returns. The independent double sort method simultaneously
tests the effect of two characteristics on the crosssection of stock returns. In such a case,
one characteristic is expected to predict stock returns better, holding the other constant
and vice versa. Empirically, stocks are separately ranked on each characteristic and
grouped. Portfolios are formed from the stocks that exist at the intersections of the
groups. Asness (1997) used the independent double sort method to test the interactive
effect of value and momentum on the crosssection of stock returns. The advantage of the
CrossSectional
Predictability of
Stock Returns
Sorting
Regression
Independent
Double Sort
Conditional
Double Sort
Time Series
CrossSectional
Univariate
sort
16
sorting methods is that they do not impose any functional relationship between
characteristics and expected returns. The method answers the question of whether the
highranked characteristic portfolio outperforms the lowranked characteristic portfolio.
As stated before, the method test is based on whether the difference in the average
returns of the high and low ranked portfolios are significantly different from zero. The
sign and the significance of the difference in average returns determine whether the
relationship is negative or positive. However, it is weak to use the information on the
high and low ranked portfolios to represent the crosssectional effect of a characteristic
on expected returns. In addition, it is difficult to disentangle the marginal effect of
several characteristics because the sorting method cannot sort on more than three
characteristics. It also ignores the noisy nature of the stock or portfolio returns.
The regression method addresses some of the shortcomings of the sorting methods. The
two types of regressions used to test the crosssectional predictability of stock returns are
the time series and the crosssectional regression method. Black et al. (1972) initiated the
time series method, which was later developed by Gibbons, Ross and Shanken (1989).
The method is based on time series regression, for which the intercepts are tested to
determine whether they are significantly different from zero. The method runs the
regression of each asset base on the equation
,
jt j j t jt
r b F o c = + + (1)
where r
jt
is the time series excess returns of asset j (usually portfolio of assets). F
t
is the
factor portfolio, usually comprised of the market excess returns and portfolio returns,
formed from the difference in the high and lowranked characteristic portfolios. α
j
, b
j
and ɛ
jt
are the regression coefficients and the regression error. The estimated intercepts
17
are grouped together to form a vector. The intercepts represent pricing errors. If the
model can explain the crosssectional variation in returns, the intercepts should not be
jointly significantly different from zero. Gibbons et al. (1989) suggest computing the
statistic
( )
1
' 1 2
1
1 ~ ( , 1) ,
T N
SR F N T N
N
o o
÷
÷
÷ ÷
¿ + ÷ ÷
where T is the length of the time series, N is the number of assets or portfolio of assets,
o is vector of intercepts from the regression equation above, ∑ is the covariance matrix
of the residuals from the regression above and SR is the Sharpe ratio of the factor. Under
the null hypothesis that the intercept equals zero, the method compares the statistic to the
F distribution with degrees of freedom N and TN1 (F (N, TN1)).
Unlike the time series method, which requires regressors that are also returns, the two
pass crosssectional regression method can take on factors that are not returns (see Fama
and MacBeth (1973)). The method often involves two regression stages (hence twopass
regression) because of the test of CAPM. The first pass regression (time series) estimates
betas, which then serve as input for the second pass crosssectional regression. In the
second pass regression, several characteristics can be combined as independent variables
or regressors. That is, for each time t, the crosssectional regression method runs a
regression of the form
1 1 , 1
2
,
C
jt ot t jt vt v jt jt
v
r ¸ ¸  ¸ c
÷ ÷
=
= + + I +
¿
(2)
where r
jt
is the crosssectional excess returns of all assets at the time t, and β
jt1
are the
timevarying beta estimates. Γ
v,jt1
can be any characteristic determined prior to t. ɛ
jt
is the
18
error term. ( )
0
, ,
t Ct
¸ ¸
is a vector of regression coefficients. Using the time series of the
regression coefficients, we estimate their expected values and test if they are significantly
different from zero. If the CAPM holds, the expected value of
0t
¸ should be zero and the
expected value of
1t
¸ , which is the risk premium on the market should be significantly
positive. Any characteristic that can explain the crosssectional variation in stock returns
has an average coefficient significantly different from zero. We now form the ttest
( )
( )
v
v
v
t
¸
¸
o ¸
= (3)
where
1
1
T
v vt
t
T
¸ ¸
=
=
¿
and
( )
v
o ¸ is the standard error normally adjusted for the Newey
West heteroskedastic and autocorrelation consistency (HAC). The NeweyWest standard
error is estimated as
( ) ( )
'
0
1
1
1
q
v w w
w
w
q
o ¸
=
  (
= u + ÷ u +u

(
+
¸ ¸ \ .
¿
where
( )( )
w vt vt vt w vt
E ¸ ¸ ¸ ¸
÷
u = ÷ ÷ (
¸ ¸
and q is the number of lags.
Betas can be estimated with a rolling window or the full window regressions. The rolling
window regressions estimate betas with information in the period prior to t. The betas are
then used in the crosssectional regression for the subsequent period. There is a tendency
to measure betas with error in the first regression. Using estimated betas in the second
pass regressions therefore induces an errorinvariable problem, biasing the slope
coefficient in small samples. Estimating betas over the full sample period mitigates the
measurement error, as the time series sample size T increases. In this case, the second
crosssectional regression estimate is Tconsistent. In the full window regression method,
portfolios are formed by sorting stocks on betas estimated in the period prior to t.
19
Portfolios are then rebalanced by repeating the process of beta estimation and portfolio
formation for each time t in the period studied. Finally, portfolio returns are computed.
Full sample betas are estimated using these time series of portfolio returns. Based on the
assumption that betas are constant throughout the sample period, the full sample betas
can be used in the crosssectional regression (see Ibbotson, Kaplan and Peterson (1997)).
On the other hand, asset beta varies with time, so each time, the full sample betas can be
assigned to individual stocks in the portfolios. The main assumption underlying this
approach is that stocks in a portfolio have the same exposure to longterm systematic
risk. This approach will yield N and T consistent estimates, since it uses the full
information in the sample period and the crosssection. Fama and French (1992) utilized
the fullsample beta estimate in the test of the CAPM.
A second approach to minimize the errorinvariable problem induced by beta or any
other estimated variable is to multiply the denominator in equation 3 by a correction
factor. Shanken (1992) suggests an adjustment factor
( )
2
0
2
1
m
m
r
r ¸
o
 
÷
+ 

\ .
. Where
2
m
r
o is the
standard deviation of the market returns in excess of the riskfree rate,
m
r is the average
excess market returns and
0
¸ is the average intercept series from the crosssectional
regressions.
In this dissertation, we will use all the possible sorting methods depending on the task.
For instance, we will use the double sorting methods to disentangle the marginal effect of
each characteristic on average returns. The sorting method provides the advantage of not
placing any linear restrictions on the return/characteristic relationship. To test the effect
20
of several characteristics on returns, we will adopt the crosssectional regression method.
This will serve as a confirmation of the result in the sorting method.
For all portfolio formations we adopt the Fama and MacBeth (1973) breakpoint method.
That is, if N is the number of stocks in year t and nis the number of portfolios required,
stocks are allocated to int( / ) N n portfolios, where ( ) int / N n is the nearest integer less or
equal to / N n. The middle portfolios have ( ) int / N n stocks each. If N is even,
( )
( )
int / 1/ 2 int / N n N n N n ( + ÷ ×
¸ ¸
stocks will be allocated to the first and the last
portfolio. If N is odd, one stock will be added to the last portfolio. The method
sometimes allocates more stocks to the extreme portfolios, which are of a great deal of
interest because of the formation of the hedge portfolio (topranked portfolio returns,
minus the bottomranked portfolio returns). It also ensures that no stock is lost in the
portfolio formation process.
Summary of the chapters 1.3
In this doctoral research, we conduct four empirical studies with the BSE data from the
19
th
and the beginning of the 20
th
century. We cover the assessment of beta in chapter 2.
In chapter 3, we test the validity of the CAPM and the size effect. The fourth chapter
tests the momentum effect. The combined effects of Size, Momentum, Total Risk and
Dividend yield on the crosssection of stock returns is studied in chapter 5. The last
chapter concludes the dissertation.
1.3.1 Chapter 2
In the second chapter, we focus on the assessment of beta. Beta serves as the main input
of the CAPM. It is an estimated variable that can be measured with error, thereby biasing
21
the test results of the CAPM. Beta instability, bias and nonrobustness to outliers have
become the center stage of research since the development of the CAPM. The chapter
studies the relative performance of different methods of estimating beta, based on their
ability to predict the subsequent beta. Specifically, we compare the market model betas
to the betas estimated with the Blume (1971) and Vasicek (1973) autoregressive
techniques.
We show that the individual stock market model betas are not stable. The predictability
of the market model betas can be improved by forming portfolios with at least ten or
more stocks. Most strikingly, there is no significant difference in the predictive accuracy
of Blume and Vasicek‟s adjusted betas. Using Dimson‟s method of estimating betas,
small numbers of stocks prove to have returns that lead or lag the market returns. To
account for outliers, we use iterative reweighted least square techniques (IRLS) to
estimate betas. The betas from the IRLS are small in magnitude compared to the market
model betas, but they have the same predictive accuracy as the market model. Based on
the study in chapter two, we determine how to estimate beta for the study of the CAPM.
1.3.2 Chapter 3
In the third chapter, we use the sorting and the Fama and MacBeth (1973) (from here on
FM) crosssectional regression methods to investigate whether the CAPM is valid before
World War I. We will also test if the size effect (the propensity of small size to have
higher returns than large size stocks) exists in the 19
th
century BSE.
With the sorting and the FM crosssectional regression methods, we find no empirical
validity for the CAPM. We also show that the relationship between beta and returns
varies with time. If anything, the null hypothesis of the equality of the estimated slopes
22
from the crosssectional regression and the market excess returns is not rejected in the
period 1868 through to 1893. On the contrary, using sizesorted portfolios (equally
weighted) in the crosssectional regression establishes the betaexcess return relationship
for beta as the lonely regressor (the CAPM is valid). However, the conditional double
sorting method adopted by Fama and French (1992) separates the effect of beta and size
on expected returns. The method makes the average crosssectional slope of the beta
insignificant, whether placed alone or together with size in the regression.
We find size to be negatively related to excess returns. Size is negatively significantly
related to excess returns (size effect), but beta does not relate to excess returns when
placed together in the cross sectional regressions. Detailed analysis of the data reveals
that the size effect is mainly due to small size stocks, which accounts for about 0.35% of
the total market size. Eliminating these small stocks destroys the relationship between
excess returns and beta or size. Both sorting and crosssectional regression reveals that
the size effect disappears when stocks are value weighted to form portfolios. In
summary, the CAPM did not work in the 19
th
century BSE. Estimating betas with the
market model, the Dimson and Vasicek method will not establish the model. The size
effect exists, but it is mainly due to a small group of stocks that represent a very small
portion of the total market capitalization.
1.3.3 Chapter 4
In chapter 4, we investigate whether the momentum strategy can generate abnormal
profit in the 19
th
century BSE. The momentum strategy buys stocks that have performed
well in the past 312 months, and sells stocks that have performed poorly in the past 312
months. There is convincing evidence that the strategy is profitable on the 19
th
century
23
BSE. Finding momentum in this era confirms the assertion that the momentum profit
found on the postWorld War II markets is not mainly due to datasnooping biases.
Detailed analysis reveals that the momentum effect does not exist in the small size group
of stocks. Additional investigation into the momentum profit in each calendar month
shows that the profit was positive for all months. The January reversal effect, found in
the postWorld War II USA markets, cannot be found in the 19
th
century BSE. In fact,
January records the fourth highest momentum profit relative to the other months of the
year. We find that momentum profit is not strong in the first twenty years of our study
period.
In order to investigate the source of the momentum profit, we use the approach of
Jegadeesh and Titman (2001) to study the returns of the momentum portfolios in the
postholding period. The momentum returns reverse in the second to fifth years after the
holding period. Further study reveals the reversal is mainly due to small size stocks. We
also test whether the momentum profit and the long run reversal in the crosssection of
stock returns depends on the state of the market. The 6month formation and the 6 to 12
month holding period strategies are profitable solely in periods of market gains.
1.3.4 Chapter 5
In this chapter, we investigate whether size, momentum, total risk and dividend yield can
explain the crosssection of stock returns between the years 1868 and 1913. We repeat
size and momentum characteristics in the analysis to test the pervasiveness and the
combined effect of these characteristics on the crosssection of excess returns. The size
sort also tests whether the effect of the other characteristics are not confined to the small,
illiquid group of stocks. Total risk for each stock is measured as the standard deviation of
24
the past 24 to 60 months excess returns. Each year, dividend yield for a stock is the sum
of all dividends paid within the last 12 months divided by the current month price. We
investigate whether the relationships are pervasive across all dividend yield, size, total
risk and momentum groups. Sorting and FM crosssectional regression methods are
adopted in the analysis. We confirm that size has a negative significant relationship with
excess returns. However, repeating the analysis on different size groups reveals that the
negative relationship is completely driven by microsize stocks accounting for less than
3.67% of the market capital. This confirms the finding in Chapter 3 that the size effect is
mainly due to the first decile size portfolio, which forms part of the micro size group. We
did not find a consistent relationship for total risk. Momentum shows a consistent
positive relationship with excess returns in stocks, which accounts for more than 96% of
the market capital. We find a negative relationship between dividend yield and
momentum among dividendpaying stocks, but each of them is positively related to
average excess returns. Dividend yield and momentum are positively related to the
average excess returns in our large stocks, which account for about 96% of the market
capital.
25
CHAPTER 2
2 ASSESSMENT OF BETA IN THE 19
th
CENTURY BSE
3
Beta risk estimation and the testing of asset pricing models have a long tradition in
financial literature. Beta serves as the main input of the CAPM, and it is assumed stable
in empirical applications. However, the instability, bias and nonrobustness (to outliers)
of the beta have raised concerns in the literature. The notion of its estimation is of
fundamental importance to the testing of CAPM. The objective of this chapter is to
introduce the new data set from the 19
th
century BSE to test the performance of the
alternative techniques of estimating beta. Studying the various techniques of estimating
beta, we may determine relatively unbiased and stable beta for the test of CAPM, which
posits a positive relationship between the beta and expected returns.
Introduction and Literature Review 2.1
Beta stability, bias and robustness evaluations have become the center stage of research
in finance since the development of the CAPM by Sharpe (1964), Lintner (1965) and
Mossin (1966). Beta, one of the parameters of a time series' regression, plays a key role
in CAPM applications. Beta is usually estimated with the standard market model (MM).
The MM is a statistical model that relates the return of any given stock to the return of
the market index. Therefore, for returns of a stock j in period t we write,
,
jt j j mt jt
R R o  c = + + (4)
3
This chapter was presented as a paper at the First World Finance Conference on the 27
th
May 2010 at
Viana do Castelo, Portugal (blind reviewed)
26
where
( )
2
jt j
Var
c
c o =
4
,
jt
R and
mt
R are the period t returns on the stock j and the
market index return, respectively. The parameters
j j
 o , and
2
j c
o are the parameters to
be estimated from the MM peculiar to stock j . The MM assumes that
j
 (beta) is
constant over the estimation period. However, substantial evidence in the financial
literature has established that security betas are not stable over time. Blume (1971) and
(1975) pioneered the research on beta stability based on the crosssectional correlation
between beta estimates for successive periods. He finds portfolio betas having stronger
correlation across successive periods than individual security betas. This indicates that
portfolio betas are more stable than individual stock betas. Blume (1971) and Vasicek
(1973) show that betas are not stable, but they revert towards their crosssectional mean.
They propose autoregressive models to capture this variation in beta. The autoregressive
method adopted by Blume is the crosssectional regression of betas in period t on the
betas in period t1. That is,
1
,
jt jt jt
a b   q
÷
= + + for j=1,…,N (5)
N is the number of stocks in the crosssection. a and b are the intercept and the slope of
the crosssectional regression of betas in the period t on betas in the period t1. The
jt
q is
the error of the crosssectional regression of betas in successive periods. The objective of
Blume‟s autoregressive model is based on the tendency of betas from successive periods
to revert to the mean. In effect, the method adjusts historical betas towards their cross
sectional mean. The Vasicek autoregressive method not only adjusts betas toward their
crosssectional mean, but the adjustment also depends on the uncertainty (standard error)
4
( ) x E is the expected value of x , ) (x Var is the variance of x.
27
about beta. Betas estimated with high standard errors have a greater tendency to deviate
from the crosssectional mean of the betas. Therefore, high uncertainties call for greater
adjustment. Vasicek (1973) applied a Bayesian correction method to capture the
differences in standard errors. The Bayesian correction method places weights on the
crosssectional average beta and the asset‟s beta estimate. The weights sum up to one,
and the more the uncertainty surrounding either estimate of beta, the lower the weight
placed on it. He used crosssectional information from the previous period betas. That is,
beta is estimated with the model
( )
( )
( )
2
1
1 1
2 2
1 1
var
for 1, 2, , ,
var var
jt
jt jt jt
jt jt
j N

 

o
  
 o  o
÷
÷ ÷
÷ ÷
= + =
+ +
(6)
where
jt
 is the mean of the posterior distribution of beta for stock , j which serves as
the beta forecast.
2

o is the standard error from the estimation of the market model
regression coefficient,
1 jt

÷
.
1 jt

÷
is the crosssectional mean of betas in period 1 t ÷ , and
( )
1
var
jt

÷
is the variance of the crosssection of betas. Faff and John (1992) use
Australian data to confirm the evidence in beta nonstationarity. They show that there is a
strong connection between beta instability and the market index used to estimate beta. On
the contrary, GregoryAllen, Impson and Karafiath (1994) use daily data to document
that portfolio betas are not more stable than individual security betas on the US market.
They show that, accounting for the variance around beta estimation will not make
portfolio betas more stable than individual stock betas.
Also, the literature has shown that nonsynchronous trading biases betas estimated with
the MM. When stocks are ranked based on their market capital (size), beta is biased
28
downward for small size stocks and upward for large size stocks. Scholes and Williams
(1977), Dimson (1979) and Fowler and Rorke (1983) developed methods to correct the
bias. Dimson (1979) adopted the aggregate coefficient method, which computes the lead
and the lag betas in a multiple regression of the stock return on a number of lead/lag
market returns. The sum of the estimated beta coefficients from the multiple regressions
is Dimson‟s estimate of beta. Dimson‟s technique involves estimating a multiple
regression of the form
,
l
jt j j mt i jt
i l
R R o  c
+
=÷
= + +
¿
(7)
Dimson‟s beta estimate is then given by
dim
l
i
i l
 
=÷
=
¿
. The error term
jt
c follows the
assumptions of the classical linear regression model. On the other hand, Scholes and
Williams (1977) compute the lead and lag beta coefficients by univariate regressions.
The literature has found that the bias in beta estimates is due to autocorrelation in returns
caused by infrequent trading. Ibbotson et al. (1997) found that beta estimates for small
stocks are severely biased downwards. This implies that the returns of small stocks are
more capable of exhibiting autocorrelation than large stocks. They recommend the
inclusion of the lagged market returns in the estimation of beta when securities are traded
infrequently. On the other hand, Bartholdy and Riding (1994) found that the MM betas
are less biased, more efficient and consistent on the New Zealand Market than the
Dimson, Scholes and Williams biased correction model betas.
The literature also highlights the impact of outliers (extreme observations) on the beta
estimates. Outlier observations in returns, depending on their location in the stock return/
29
market return plane, can have an extreme influence on beta when the MM is used to
estimate beta. Under idealized conditions, the MM beta produces the best estimate
because of the ordinary least squares method. The idealized conditions are that the paired
returns of the stock and the market conform to the linear model relationship with zero
mean error not correlated with the market returns. Unfortunately, these conditions often
fail in empirical settings. One explanation of the failure of the condition is the occurrence
of outlier observations in returns. Nevertheless, the MM does not consider the effect of
outlying observations on beta. Chatterjee and Jacques (1994) document the effect of
outlying “observations” on the beta parameter. Their detailed study compares the betas
estimated from the MM and the outlier resistance methods. They indicate that the MM
cannot detect a large number of outlying returns in their data. Identifying the outlying
observations and estimating betas with the outlying resistance method yield beta
estimates that are lower than the MM betas. Chan and Lakonishok (1992) said the robust
method (outlier detecting method) of estimating beta is good for stocks susceptible to
stock splits, dividend cuttings, and initial public offerings (IPO). Stocks that exhibit this
behavior are mostly small stocks. Martin and Simin (2003) confirmed that outlier
resistance beta is a better predictor of future beta than the market model beta. They also
reveal that small stocks betas are most vulnerable to outliers.
The literature proposes various techniques of minimizing the impact of outliers on beta
estimates. The weighted least square estimation technique is one of the methods used to
minimize the impact of outliers on beta estimates. Based on the assumption that outliers
have a lower probability of occurring in the future, less weight is placed on them (see
Martin and Simin (2003)). An alternative technique would be to remove the extreme
30
observations and perform the regression with fewer observations. The least squares
approach adopted by the MM minimizes the sum square residuals with respect to the
model parameters
j
o and
j
 :
( )
2
,
min ,
j j
jt j j mt
R R
o 
o  ÷ ÷
The squaring of the residuals magnifies the effect of the outliers on the estimated
parameters. For instance, expected return is likely to shift toward the outliers while the
covariance matrix will be inflated toward the outliers. To reduce the influence of outliers,
the statistics literature emphasizes the use of iterative reweighted leastsquares (IRLS)
method. This method estimates beta by iteratively minimizing a weighted function that
depends on standardized return residuals:
,
min ,
j j
j
jt j j mt
R R
W
o 
c
o 
o
 
÷ ÷


\ .
(8)
with weight function W and
j
c
o the standard deviation of the return residual. We
estimate the regression parameters from the least squares regressions and use these
parameters as initial input for iteration. A weight function is applied to the standardized
residuals. We apply Huber or the bisquare weight function, defined as:
( )
2
2
1
,
0
jt
jt
jt
jt
s
for s T
T
W s
for s T
¦
(
 
¦
÷ s
(
 ¦
= \ .
( ´
¸ ¸
¦
>
¦
¹
31
with the standardized residual
j
jt j j mt
jt
R R
s
c
o 
o
÷ ÷
= and T , a tuning constant. Huber
(2004) sets the value of T to 4.685. Small values of T
introduce more resistance to
outliers, but at the detriment of efficiency when returns are outlier free. He chose the
value of T so the method will have less influence on beta when there is no extreme
observation and it still provides protection against outliers. For the iterations, the initial
input is the coefficients from the MM. The residual from the MM is standardized and the
weight function defined above is used to transform them. Estimate the parameters by the
weighted least squares:
 
j
R W W b
0 '
1
0 ' 1
9 9 9 =
÷
where  
,
mt
R i 9=
with i a vector of ones with the same size as
mt
R (market index returns).
0
W is an initial
standardized weight diagonal matrix and
1
b the estimated beta parameter. The new set
of parameters serves as input for the next iteration. The procedure above is repeated until
the parameter of interest (beta) converges. A standardized residual observation that
exceeds the tuning constant is assigned zero weight (outliers).
The literature also suggests that estimated betas vary with time and macroeconomic
conditions. Therefore, static beta CAPM should replace with time varying beta
conditional CAPM (see Jagannathan and Wang (1996)). In this dissertation, we did not
pursue the timevarying beta because, Ghysels (1998) used monthly NYSE data to
document that pricing errors with static beta models are smaller than those with time
varying beta models. He indicated that timevarying beta model might be mispecified.
32
To the best of my knowledge, the literature has not traced the stability, bias and
robustness of beta using data before 1926 in any country. Luoma, Martikainen, Perttunen
and Pynnönen (1994) simulated returns data to generate artificial markets. They found
that various beta estimation techniques behave uniquely in different markets.
In this chapter, individual and portfolio betas from the various estimation techniques will
be compared based on their predictive accuracy. The root mean square error (RMSE)
criterion adopted by Blume (1971) and Klemkosky and Martin (1975) is used to
determine the predictive accuracy of the betas estimated from the various techniques.
The 19
th
century data provide a very good platform to determine whether betas before
1914 exhibit a similar pattern as betas after 1914.
This chapter reveals that for individual stocks, the market model betas are weak in
predicting their future. Predictability can be improved when a portfolio of 10 or more
stocks components is formed. The study also shows no significant difference between the
Blume and Vasicek adjusted betas in terms of their predictive accuracy. Estimating betas
with the Dimson method reveals a very small number of stocks showing a lead and lag
relationship with the market returns. The iterative reweighted least squares method
produces betas that have the same predictive power as the MM betas, but are lower in
their magnitude on average.
The subsequent sections of the chapter are organized as follows: Section 2.2 provides
descriptive statistics of beta when estimated with the least squares regressions of the
MM. Section 2.3 examine the Blume correlation technique of detecting the stability of
beta. The autoregressive technique of Blume and Vasicek is used to adjust betas in
33
Subsection 2.3.1. The predictive accuracy of the adjustment techniques based on the
RMSE criterion is tested in the same section. The modified Diebold and Mariano test
proposed by Harvey, Leybourne and Newbold (1997) is used to test for equal predictive
accuracy of the models. In Section 2.4, we determine how the stocks in each period lead
or lag the market index using the Dimson‟s model. In Section 2.5, we compare outlier
resistant betas to the benchmark MM betas. The final section presents the conclusion.
Beta Coefficient Descriptive Statistics 2.2
The data sample is divided into 15 fiveyear nonoverlapping periods
i
t for 1,...,15. i =
Stocks with complete returns data for five years are considered in each subperiod. For
the proxies of the market portfolio, we consider the valueweighted market index. Table
2.1 presents the number of stocks in each period under study, the statistical
characteristics of the betas estimated in each period and displays the 15 periods studied,
starting from January 1837 to December 1911. The number of stocks with full returns
data in each period ranges from 21 to 424 as shown in Column 2. It can be seen in Table
2.1 that the number of stocks that have five years of data does not exceed 100 before
1877. The beta coefficient of each stock in each period is estimated by simply regressing
the monthly returns of the stock in each period on the corresponding monthly value
weighted index of the market by using the MM above. Crosssectional statistics of the
betas in each period are computed and the result is displayed. The equally weighted
average beta and the valueweighted average beta are displayed in Columns 3 and 4,
respectively. We compute the valueweighted mean beta by considering individual stock
relative to market capitalization at the beginning of the period. From the high values of
equally weighted mean beta and low values of valueweighted mean beta in the periods
34
Table 2.1: Beta coefficient descriptive statistics for the 15 estimated periods
Note: This table displays crosssectional descriptive statistics of betas estimated by market model (1) of returns on their market counterparts over the 15 fiveyear
subperiods between January 1832 and December 1911. The number of stocks in the various periods ranges from 21 to 424 for stocks with full return data within
a period. The table also reports the percentage of beta sample in a period that is less than zero. The maximum and minimum beta of each period is also recorded.
The last column reports the average coefficient of determination ( )
2
R . Equally weighted (EW) and valueweighted (VW) mean betas are also displayed in the
table. The valueweighted mean is the average of the betas weighted by their market capital at the beginning of the period.
Percentage
Number of Average Average Standard of BETAS Mean
Period Stocks beta(EW) beta(VW) Deviation <Zero Minimun 0.10 0.25 0.50 0.75 0.90 Maximum R
2
(%)
1/183712/1841 21 1.37 1.24 0.67 0 0.38 0.50 0.77 1.35 1.83 2.29 2.90 29
1/184212/1846 33 1.10 1.01 1.84 12 1.18 0.04 0.27 0.79 1.39 2.21 10.19 18
1/184712/1851 32 0.69 1.05 0.72 0 0.04 0.15 0.33 0.48 0.83 1.24 3.92 25
1/185212/1856 33 0.76 0.98 0.55 3 0.43 0.11 0.35 0.75 1.09 1.56 1.98 19
1/185712/1861 54 0.98 1.15 0.55 2 0.19 0.28 0.71 0.99 1.25 1.55 2.74 27
1/186212/1866 72 0.83 1.01 0.73 4 0.86 0.20 0.40 0.70 1.07 1.57 3.70 12
1/186712/1871 76 0.74 0.76 0.58 8 0.91 0.09 0.42 0.67 1.02 1.35 2.36 11
1/187212/1877 82 1.05 0.96 1.01 10 0.71 0.01 0.30 0.85 1.70 2.59 3.55 15
1/187712/1881 113 1.29 0.96 1.50 12 3.54 0.03 0.32 0.93 1.99 3.14 7.80 17
1/188212/1886 154 1.03 1.03 1.44 18 1.33 0.35 0.14 0.66 1.66 2.90 7.94 6
1/188712/1891 161 1.52 0.74 1.73 14 0.76 0.05 0.10 0.94 2.61 4.03 7.81 22
1/189212/1896 196 1.24 0.95 1.32 12 1.74 0.02 0.34 1.04 1.71 2.87 7.23 8
1/189712/1901 252 1.13 0.93 1.11 8 3.03 0.02 0.27 1.09 1.71 2.38 5.47 14
1/190212/1906 374 1.24 1.03 1.36 11 1.33 0.08 0.33 1.15 1.74 2.63 8.92 10
1/190712/1911 424 1.16 1.05 1.18 9 2.14 0.01 0.34 1.06 1.64 2.45 7.33 12
Fractiles
35
after 1873, we can conclude that small stocks have high betas in these periods. The
values in the last column show the average coefficient of determination ( )
2
R in
percentages, which is a measure of explanatory power of the MM. It is worthy to note the
percentage of negative betas in nearly all the subperiods. Surprisingly, the literature
reports negative betas on different markets after World War I. For example, Altman,
Jacquillat and Levasseur (1974), Dimson and Marsh (1983) recorded a large number of
negative betas in their weekly returns interval estimation of beta on the French and the
American markets. It is also important to note the very high and low betas in the 19
th
century BSE compared to the post1926 market betas recorded in the literature. The
possible explanation for these extreme values of the beta and low coefficients of
determination might be linked to the infrequent trading effect and the influence of
extreme observations (outliers) in the returns series of the stocks. The removal of stocks
that did not trade fully within the fiveyear estimation periods does not introduce
survivorship bias since there is no significant difference in the beta coefficient
descriptive statistics when stocks with at least 24 observations are added to the sample
(not reported).
In order to capture the effect of size (market capitalization) on the average beta and
2
R
values, we computed the same crosssectional statistics for betas in each period.
However, the stocks in each period are subdivided into three mutually exclusive size
based subsamples. The composition of the subsamples is as follows: the sample of
stocks in each period is sorted in descending order based on their market capitalization
(size) at the beginning of the period.
36
Table 2.2: Average beta and average coefficient of determination of the size based subsamples
Note: In this table, stocks in each period were sorted in descending order based on their market capital.
The first 30 percent of stocks in the periods are classified as large stocks, the next 40 percent as medium
stocks and last 30 percent as small stocks. Crosssectional statistics of the betas of the subsamples were
computed and their equally weighted averages of betas and coefficient of determination ( )
2
R are
displayed. The L, M and S subscripts on betas indicate large, medium and small stocks, respectively.
Figure 2.1: The graph of the average beta of each period for large stocks and small stocks
Note: The average beta estimates for large and small stocks are plotted against their corresponding periods
of estimation. The diamond points indicate the averages of large stocks and the square points for small
stocks. This depicts the extreme values of beta estimate recorded for the small stocks relative to the
benchmark beta of one.
Period mean(β
iL
) mean R
2
(%) mean(β
iM
) mean R
2
(%) mean(β
iS
) meanR
2
(%)
1/183712/1841 1.29 30.1 1.73 32.8 1.07 23.0
1/184212/1846 1.19 36.5 0.79 10.5 1.40 10.1
1/184712/1851 0.94 43.2 0.49 21.6 0.67 10.1
1/185212/1856 1.05 35.4 0.66 14.2 0.61 10.2
1/185712/1861 1.15 40.5 1.08 27.0 0.69 12.9
1/186212/1866 1.04 22.6 0.76 8.0 0.70 5.5
1/186712/1871 0.77 17.2 0.59 9.4 0.90 7.2
1/187212/1877 1.06 19.2 1.02 16.4 1.10 10.0
1/187712/1881 1.03 20.2 1.51 18.0 1.30 12.3
1/188212/1886 1.01 9.4 1.16 6.1 0.86 3.4
1/188712/1891 0.84 18.2 1.73 27.6 1.95 17.7
1/189212/1896 1.00 9.5 1.29 7.4 1.40 5.7
1/189712/1901 0.98 17.6 1.22 14.6 1.16 9.3
1/190212/1906 1.06 15.5 1.27 9.1 1.39 4.3
1/190712/1911 1.18 20.1 1.17 11.0 1.13 5.5
Large Medium Small
37
The first 30 percent of stocks are classified as the large stocks portfolio, next 40 percent
as medium stocks portfolio, and the last 30 percent as the small stocks portfolio. Table
2.2 reports the average beta and the
2
R values for each subsample in each period.
Generally, the high values of
2
R recorded for large stocks are striking because on the
UK market, Dimson (1979) recorded a similar pattern with 15 years of monthly data.
This might be attributed to the valueweighted index used in the computation, since the
index will have more explanatory power for largesized stocks. Figure 2.1 depicts the
crosssectional average beta of each period for large stocks and small stocks. It is clear
from Figure 2.1 that small stocks record comparatively low betas in the periods before
1867 and high betas thereafter. The result in the periods before 1867 corroborates the
assertion of Dimson (1979), Scholes and Williams (1977) and Beer (1997). They said
that if trading frequency is highly correlated with the market capitalization of the stock,
betas of small stocks (infrequently traded stocks) are lower when estimated with the
market model. On the contrary, Ibbotson et al. (1997), based on NYSE data between
1926 and 1994, found that small stock returns due to infrequent trading show a high
degree of autocorrelation and that they are capable of recording high betas. On the 19
th
century BSE, we confirm their result in the periods after 1867. The only anomaly is the
period between 1882 and 1886, where the small stock average beta lies slightly below the
average beta of large stocks.
Beta Stability 2.3
Here, we adopt the Blume (1971) and Altman et al. (1974) correlation method of
investigating the stability of beta estimates. Blume (1975) shows that the beta coefficient
38
between two successive periods is stationary if
5
( ) ( )
1 i i
t t
E E  
+
= ,
( ) ( )
1 i i
t t
Var Var  
+
= ,
( )
1
, 1
i i
t t
corr  
+
= where
i
t
 are the betas in period
i
t . Betas in period
i
t are used to rank
stocks existing in periods
i
t and
1 i
t
+
in ascending order. In period
i
t equally weighted
portfolios of ,... 3 , 2 , 1 = s stocks are formed as follows: the first portfolio consists of stocks
with s smallest beta estimates. The next portfolio consists of stocks with the next
smallest beta estimates. This process of portfolio formation is repeated until the number
of stocks left is less than s. For each s, the betas of all portfolios in period
1 i
t
+
are also
computed. We compute the correlation and Spearman rank order correlation of the betas
between each two adjacent periods. Table 2.3 reports the weighted average correlation
across all the subperiods studied. The weighted average correlation takes into account
the number of stocks or portfolios in each adjacent tenyear period.
Table 2.3: Weighted average of correlation and Spearman rank order correlation across successive periods
Note: This table shows the weighted average correlation and Spearman‟s rank correlation of betas of
individual stocks and portfolios in successive periods across the 15 periods studied. For a stock to be
included in this analysis, it has data for two complete consecutive periods. Betas in a period (estimation
period) are used to rank the betas in that period and the next adjacent period (prediction period) in
ascending order. Portfolios are formed with their constituency as follows: the first portfolio is the first s
stocks for s = (1,2,4,7,10,20). The second portfolio contains the following s stocks and so on until
available stocks is less than s . Assuming equal amounts are invested in each stock, then the portfolio beta
will be the mean of the betas of stocks included in the portfolio. We computed the weighted average
correlations by considering the number of portfolios in each tenyear period.
5
) , ( y x corr is the correlation between x and y .
per Portfolio Correlation
Spearman's
rank Correlation
No. of Stocks
0.92
0.95
0.56
0.64
0.72
0.77
0.91
0.96
0.54
0.62
0.72
0.76
1
2
4
7
10
20
39
The weighted average correlation across the successive periods ranges from 0.54 for
individual stocks to 0.95 for portfolios of 20 stocks. These values indicate that the beta of
individual stocks have, on average, less information about their future beta than the
portfolio beta. Blume (1971) found a similar result on the US market. Due to the limited
number of stocks, we were not able form up to 50sized portfolios on the BSE. Blume
finds 0.62(0.67) and 0.91(0.93) mean correlation (rank correlation) for 1 and 10sized
portfolios using 84month estimation periods respectively. Furthermore, based on 52
week estimation periods, Levy ( 1971) records 0.44 and 0.82 mean correlation for 1 and
10sized portfolios on the same exchange. On the UK market between 1955 and 1979,
Dimson and Marsh (1983) used valueweighted market index, monthly returns interval
measurement and sixtymonth estimation periods to obtain an average correlation of 0.56
and 0.91 for 1 or 10sized portfolios respectively. The correlations in Table 2.3 show
that the betas from the 19
th
century BSE are stable compared to the betas in post1926
US and UK markets.
2.3.1 Blume and Vasicek stability adjustment techniques
Individual stock betas estimated from the MM are noted as unstable in the previous
section (also in Blume (1971), Collins, Ledolter and Rayburn (1987), Faff and John
(1992), GregoryAllen et al. (1994), Eisenbeiss, Kauermann and Semmler (2007)). There
is a tendency for a high beta estimate to overstate its true value and vice versa. Therefore,
we use the Blume (1971) autoregressive adjustment model to improve the stability of
beta estimates for both individual stocks and portfolios. Table 2.4 presents regression
tendencies implied between adjacent periods, where a and b are the constant term and
slope coefficients, respectively. The values of the coefficients in the periods between
40
1837 and 1867 are striking. It is not consistent with the Blume assertion that all the
coefficients lie between zero and one. The last two Columns show the tstatistics of the
test of a hypothesis of the slope coefficient equal to zero or one. The tstatistics of the
slope coefficients show that the null hypothesis of the slope coefficient equal to zero is
rejected in the adjacent periods after 1872. In addition, the null hypothesis of the slope
coefficient equal to one is rejected in all the adjacent periods (except the first two
periods). The R
2
values also show that the betas in the periods after 1872 have more
explanations for their prior betas than those before 1872. As can be seen in Table 2.4, the
coefficients change over time, but there are extreme coefficients outside the interval
between zero and one in the first two periods.
Table 2.4: Measurement of regression tendency of estimated beta coefficient for individual stocks
The extreme values may be attributed to the number of stocks in a period as we record
less than 50 stocks in our first two fiveyear periods. A result not reported shows that
increasing the length of the estimation period (such as seven years in Blume (1971))
Regression Tendency
Implied Between Periods a b
R
2
(%) H
0
:b=0 H
0
:b=1
1/1842 12/1846 and 1/1837 12/1841 1.40 0.00 0 0.00 1.31
1/1847 12/1851 and 1/1842 12/1846 0.74 0.07 3 0.96 14.11
1/1852 12/1856 and 1/1847 12/1851 0.48 0.33 20 2.58 7.60
1/1857 12/1861 and 1/1852 12/1856 0.53 0.54 26 3.22 5.41
1/1862 12/1866 and 1/1857 12/1861 0.27 0.55 18 3.32 5.43
1/1867 12/1871 and 1/1862 12/1866 0.63 0.13 3 1.38 10.34
1/1872 12/1876 and 1/1867 12/1871 0.77 0.43 5 1.94 4.04
1/1877 12/1881 and 1/1872 12/1876 0.65 0.70 20 3.99 5.03
1/1882 12/1886 and 1/1877 12/1881 0.44 0.67 39 7.80 10.91
1/1887 12/1891 and 1/1882 12/1886 0.86 0.69 31 7.66 10.49
1/1892 12/1896 and 1/1887 12/1891 0.58 0.39 27 7.42 18.75
1/1897 12/1901 and 1/1892 12/1896 0.60 0.38 27 7.81 19.99
1/1902 12/1906 and 1/1897 12/1902 0.49 0.70 28 8.94 11.98
1/1907 12/1912 and 1/1902 12/1906 0.78 0.39 15 7.67 19.21
0.67 0.46 20 18.75 41.46
null hypothesis of the slope coefficient (b ) equal to zero or one is also reported in the last two columns.
All periods
β
jt
=a+bβ
jt 1
+η
jt
Note: In this table beta of stock existing in a period are regressed on the betas of the same stocks in a prior adjacent peri
od. R
2
(%) is the percentage of the variance of betas in the period t explained by betas in the period t 1. The tstatistic of the
41
improves the
2
R values and the tstatistics but at the expense of losing more stocks, since
fewer stocks have complete returns' data for longer periods.
With the regression tendencies, suppose we want to forecast the beta for any stock or
portfolio in the period 18421846. We compute its beta in 18371841. The forecast of the
beta is obtained by substituting it for β
t1
in equation (5) with the coefficients in the first
row of Table 2.4. β
t
is then computed from the equation and used as the forecast. The
adjustment process is repeated for stocks in the subsequent 13 adjacent periods using
their respective coefficients. We also introduce the Vasicek adjustment model (6) to
adjust betas in successive adjacent periods. We test the predictive performance of the
various adjusted betas by using the root mean square error
6
(RMSE) criterion. The
RMSE tests the performance of the autoregressive methods based on variation and
unbiasedness of their beta forecast. The adjustment method is repeated for all adjacent
periods on equally weighted portfolios of size 2, 4, 7, 10 and 20. In order to compare the
predictive performance of the MM betas and the autoregressiveadjusted betas, we
compute the RMSE of betas estimated with MM in adjacent periods. Table 2.5 displays
the average RMSE of the adjusted betas and MM betas across the 15 periods studied. It is
clear from the table that the average RMSE of the adjusted betas is lower than that of the
market model betas. Table 2.5 also shows that the predictive performance improves as
the number of stocks in a portfolio increases for both adjusted and the MM betas. For
individual stocks, the Bayesian adjustment technique proposed by Vasicek is superior to
Blume‟s adjustment as reflected in the small average RMSE. Blume (1971) and
6
The root mean square error was calculated by
( )
1
2
i i
jt jt
N
 
÷
÷
¿
for N j ,..., 1 =
42
Klemkosky and Martin (1975) recorded similar patterns of the predictive performance on
the NYSE market. Their adjusted betas mean square errors were smaller than their MM
betas. For portfolios of size 7 or more, one cannot see much difference between the
Blume adjustment method and the Bayesian approach.
Table 2.5: Predictive performance of Blume and Vasicek (Bayesian) procedures of estimating beta
Note: Average RMSE across the 15 periods studied were used to compare the predictive performance of
the various adjusted betas and the market model (MM) betas in successive periods. The average RMSE
across the successive adjacent periods for the various equally weighted portfolio formations are displayed.
The conclusion is that on the 19
th
century BSE, the predictive accuracy of betas estimated
by the MM can be improved by adjusting betas using either the Blume or Bayesian
adjustment methods and a portfolio with a sizeable number of stocks.
The reliability of the conclusion above can be confirmed by performing an additional test
on the root mean squared error values. The possible method is to test whether the
differences in the values of the RMSE‟s are statistically significant. Harvey et al. (1997)
presented a modified Diebold and Mariano test statistic that will be used for this purpose.
Therefore, suppose we want to compare the forecasts of Blume (BL) and Vasicek (VA)
models.
{ }
BL
c and
{ }
VA
c are the forecasting errors from the Blume and Vasicek models,
respectively. In our case, we consider the root mean square error function,
( )
BL
f c = root
mean square error of the Blume adjusted betas. The test is based on the loss differential
Bayesian
0.48
0.38
0.31
0.27
0.27
0.23
No.of Stock
per Portfolio
1
2
0.40
1.08
4
7
10
0.57
MM Blume
Average RMSE
0.23
0.19 20
0.45
0.57
0.44
0.35
0.29
0.76
0.25
43
function
( ) ( )
BL VA
j j j
d f f c c = ÷ for 1,..., j H = . The null hypothesis of expected equal
predictive performance is H
0
:
( )
0
j
E d = and the alternative hypothesis of the Blume
model predicting worse than the Vasicek model is
a
H :
( )
0.
j
E d > The Modified
DieboldMariano (MDM) statistic is:
( )
( )
1
1 2
, 1 1
2
0
1
1 2 1
. ~ 0,1 ,
2
H
i
i
H h H h h
d
MDM t
H
v
ì ì
÷
÷
·
=
( + ÷ + ÷
=
(
¸ ¸  
+

\ .
¿
where
1
1
H
j
j
d H d
÷
=
=
¿
,
( )
cov ,
i j j i
d d ì
÷
= , h is the horizon of forecast and
, 1
(0,1)
H
t
v ÷
a
student‟s t distribution with 1 H ÷ degrees of freedom and v is the significant level
usually set at 5%. The test compares the DieboldMariano test statistics to critical values
from the student‟s t distribution. We reject the null hypothesis of equal predictive
accuracy when the test statistic is greater than the critical value at v level. In order to
apply this test, betas estimated with the Market, Blume and Vasicek models in all periods
are pooled together to form three series of length H. Then we perform the test on the
three series.
Table 2.6 reports the modified DieboldMariano test statistics between the various
models under study. Betas estimated from the various models are considered across the
entire period studied. The period studied is divided into two subperiods based on the
environment in which the BSE operated, a period of strict regulation and a period of
deregulation and expansion. The values in the first row of the table reveal that we can
confidently reject the null hypothesis of onestep ahead equal predictive accuracy of the
44
Blume and MM betas. For instance, in the overall period of our sample the null
hypothesis can be rejected at the 5 percent level. During the strictly regulated period, we
find that Blumeadjusted beta significantly outperforms the market model beta. Between
the Vasicek adjusted betas and the MM betas, we can reject the null hypothesis of equal
predictive accuracy for the deregulated and expansion period.
Table 2.6: Modified DieboldMariano test statistics (pvalue in parentheses)
The significance level of the rejection becomes weak in the strictly regulated period. The
equality in the predictive performance of the Vasicek and the MM betas is strongly
rejected in the entire period. The values from the bottom row of Table 2.6 show that there
is no significant difference between the Vasicek betas and the Blume betas in terms of
their onestep ahead forecast. Therefore, we cannot reject the null hypothesis of equal
predictive accuracy between the two models.
Beta Bias 2.4
Considering the period of the study and the trading frequency of the market, we might
expect that some stocks may not trade every month for economic reasons or because of
Models
β
MM
vs. β
BL
β
MM
vs. β
VA
β
BL
vs. β
VA
Overall period
1/183712/1911
2.18
period
Strict Regulatory
1/183712/1871
Deregulation and
expansion period
1/187212/1911
(0.01)
1.99
(0.02)
2.87
(0.00)
3.75
(0.00)
4.11
(0.00)
1.44
(0.08)
H
a
= E(d
j
) >0
Note: This table reports the modified DieboldMariano test statistics for one step ahead equal forecast
accuracy between the market model, Blume and the Vasicek's adjusted betas. β
MM
=market model betas,
β
BL
= Blume's adjusted betas, β
VA
= Vasicek's adjusted betas. The hypothesis test H
0
: E(d
j
) = 0 and
0.56
(0.29)
0.78
(0.22)
0.19
(0.42)
45
regulatory conditions. These stocks may systematically lead or lag the market movement,
producing biased betas when beta is estimated with the MM. In order to expose the
presence of possible lead or lag effects, we test the significance of the coefficient of the
returns on the lagged or lead market index.
Table 2.7: Dimson Aggregate Coefficient (AC) beta Adjustment
Number of AC
Period stocks Beta
β
3
β
2
β
1
β
0
β
1
β
2
β
3
1/183712/1841 21 1.37 0.03 0.05 0.12 1.42 0.07 0.02 0.07
(10) (5) (5) (95) (0) (0) (5)
1/184212/1846 33 1.47 0.11 0.37 0.10 1.06 0.12 0.10 0.03
(3) (12) (18) (58) (15) (0) (6)
1/184712/1851 32 0.69 0.00 0.07 0.16 0.76 0.03 0.07 0.02
(6) (13) (3) (72) (13) (13) (3)
1/185212/1856 33 0.80 0.06 0.03 0.00 0.76 0.01 0.07 0.01
(3) (6) (0) (73) (6) (6) (0)
1/185712/1861 54 0.97 0.00 0.02 0.07 0.99 0.10 0.00 0.04
(2) (7) (6) (78) (4) (2) (11)
1/186212/1866 72 0.56 0.21 0.19 0.09 0.90 0.16 0.05 0.12
(4) (10) (8) (42) (0) (3) (0)
1/186712/1871 76 0.49 0.09 0.15 0.05 0.73 0.04 0.00 0.08
(1) (3) (5) (67) (3) (3) (3)
1/187212/1877 82 1.17 0.07 0.08 0.24 1.02 0.17 0.12 0.02
(10) (5) (11) (60) (7) (7) (7)
1/187712/1881 112 1.67 0.12 0.05 0.10 1.19 0.07 0.09 0.15
(6) (7) (13) (55) (6) (11) (11)
1/188212/1886 154 0.74 0.03 0.03 0.01 0.97 0.11 0.10 0.18
(6) (3) (6) (38) (9) (5) (1)
1/188712/1891 160 1.68 0.03 0.12 0.09 1.58 0.04 0.00 0.08
(4) (7) (5) (52) (7) (4) (6)
1/189212/1896 196 1.42 0.07 0.07 0.12 1.19 0.08 0.08 0.04
(6) (6) (6) (42) (8) (5) (4)
1/189712/1901 252 0.96 0.07 0.05 0.05 1.17 0.18 0.06 0.02
(9) (5) (6) (59) (4) (8) (7)
1/190212/1906 374 1.11 0.04 0.09 0.22 1.18 0.03 0.16 0.15
(4) (7) (10) (47) (3) (3) (4)
1/190712/1911 424 1.23 0.03 0.04 0.12 1.13 0.04 0.05 0.01
(2) (3) (10) (56) (5) (6) (5)
Overall periods 2075 1.09 0.01 0.03 0.04 1.07 0.02 0.00 0.02
(5) (6) (8) (53) (5) (5) (5)
Mean Lag,Match and the Lead beta estimates
Note: This table reports Dimson's aggregate coefficient adjusted beta in each five year period. The numbers in parenthesis are
the percentage of stocks in a period that reject the null hypothesis of the coefficient been zero at 5% significant level.
46
The Dimson bias adjustment equation with maximum lag or lead of three months is
considered, that is 3 ,..., 3 ÷ = i . We use only three months lag and lead because beta bias
has been documented as not prevalent in monthly returns data (see Cohen, Hawawini,
Maier, Schwartz and Whitcomb (1983)). For each stock, the estimates of the parameters
i
 indicate the lagged, matched and lead beta coefficients. We test the hypothesis
0 :
0
=
i
H  against the alternative 0 :
1
=
i
H  for each stock.
Table 2.7 reports the crosssectional average of the lag and lead betas in each period. The
numbers in parentheses are the percentage of stocks that reject the null hypothesis.
Evidence from this table indicates that beta coefficients
i
 for 0 = i are not significantly
different from zero for the majority of the stocks. This shows that the explanatory power
of the model for 0 = i
is approximately zero for most of the stocks. Unsurprisingly, there
are some stocks with lead and lag coefficients that are statistically significant, but their
numbers does not exceed the coefficients corresponding to the match.
This indicates that
there is no severe timing problem in the 19
th
century data.
As most of the lead and lagged coefficients are significantly equal to zero, we can
interpret this as evidence of the market model (MM) producing statistically reliable beta
estimates in relation to the other models, which incorporate the lagged and lead market
indexes. These results can be compared to the results from the postWorld War I markets.
For example, Hawawini and Michel (1979) found a similar pattern of results on the
Belgium stock exchange by using weekly interval returns data between 1963 and 1976.
The result also follows Cohen et al. (1983) hypothesis that there is a strong relationship
between beta estimates and the length of the interval over which returns are measured.
47
They established that beta bias mostly shows up in the short length interval (daily) of
returns, and the bias disappears when the difference of the interval is lengthened
(monthly). Similarly, on the New Zealand market, Bartholdy and Riding (1994) used
monthly data to establish that betas estimated from MM are less biased. On the contrary,
Ibbotson et al. (1997) reports that lagged coefficients should be considered when
estimating beta.
Impact of outlying observations on Beta 2.5
The extreme (maximum/minimum) beta estimates recorded by some stocks in Table 2.1
for the fiveyear periods studied might be due to the influence of outliers or unexpected
movement by the stock or the market returns. The literature shows that outliers have a
tendency to reduce or increase the magnitude of the beta when it is estimated with the
MM (Chatterjee and Jacques (1994)). In such cases, reducing the impact of outliers in the
estimation of the beta can significantly change the value of beta. We apply the IRLS
(outlier resistant), which minimizes a weighted sum of squares of residuals. The weights
given to each return pair observation depends on the distance between the observation
and the fitted line (Martin and Simin (2003)).
Table 2.8 reports how the presence of outliers affects the beta value. The number of
stocks in each fiveyear period is grouped into two. As explained in Section 2.2, stocks
with identified outliers less than or equal to 4 are grouped into Category A and those with
identified outliers greater than 4 are grouped in category B. We compute the average beta
of each category. In each period, we compare the crosssectional average betas of the
MM and IRLS for each category. For example, in the first period out of the 21 stocks,
three fall in Category A with an average market model (MM) beta estimate of 1.29 and
48
IRLS beta of 0.46. In Category A, the difference between the average MM beta and the
average IRLS beta is 0.83. Looking across periods, except for periods 1 to 3, the rest of
the periods have more stocks in Category A than Category B.
Table 2.8: Comparison of the market model betas and the iterative reweighted least square betas
Total Number of Number of detected Number of Average Average
Period stocks outlier observations Stocks MM IRLS β
MM
β
IRLS
1/183712/1841 21 A 3 1.29 0.46 0.83
B (18) 1.38 0.19 1.19
1/184212/1846 33 A 10 0.72 0.49 0.23
B (23) 1.26 0.03 1.24
1/184712/1851 32 A 9 0.60 0.52 0.08
B (23) 0.72 0.26 0.46
1/185212/1856 33 A 23 0.97 0.81 0.16
B (10) 0.28 0.07 0.21
1/185712/1861 54 A 41 1.10 1.00 0.10
B (13) 0.60 0.10 0.51
1/186212/1866 72 A 48 0.88 0.60 0.28
B 24 0.72 0.15 0.57
1/186712/1871 76 A 42 0.79 0.56 0.23
B (34) 0.67 0.06 0.61
1/187212/1877 82 A 54 1.21 0.80 0.42
B (28) 0.75 0.14 0.61
1/187712/1881 112 A 78 1.47 1.05 0.42
B (34) 0.91 0.20 0.71
1/188212/1886 154 A 110 1.19 1.13 0.06
B (44) 0.62 0.05 0.57
1/188712/1891 160 A 120 1.77 1.41 0.36
B (40) 0.80 0.16 0.63
1/189212/1896 196 A 146 1.45 1.06 0.39
B (50) 0.62 0.12 0.50
1/189712/1901 252 A 198 1.33 1.07 0.26
B (54) 0.40 0.07 0.34
1/190212/1906 374 A 290 1.36 1.07 0.29
B (84) 0.83 0.19 0.64
1/190712/1911 424 A 346 1.32 1.12 0.20
B (78) 0.47 0.07 0.40
Note: In this Table, stocks with outlier observations less or equal to 4 are grouped in Category A and those with outlying observ
ations greater than 4 are grouped in category B. In each period, we estimate betas in each category by using the Market model and
iterative reweighted least square (IRLS) model. We also show the number of stocks that falls in each category. We compute the cross
sectional average of the betas estimated with the market model and the IRLS model.
49
In each period, the difference between the MM betas and the IRLS betas for Category B
is greater than the difference in Category A (last column). This implies that the more the
outlier observations in the return series, the higher the market model overestimates beta.
Figure 2.2: Plot of average market model betas and IRLS betas for stocks with outlier observation less than 4
Notes: This figure depicts the average market model and iterative reweighted least square betas for stocks with at most
4 detected outliers across the 15 periods of study.
The MM beta is always greater than the IRLS beta in Category A (across periods in
Figure 2.2). It confirms the result by Chatterjee and Jacques (1994) that the weighted
least squares estimation reduces the MM betas by a certain percentage. As in Subsection
2.3.1, we apply the modified DieboldMariano test to compare onestep ahead predictive
accuracy of the MM and IRLS estimated betas. The modified Diebold Mariano test
statistics proposed by Harvey et al. (1997) are employed to test the null hypothesis of
equal predictive accuracy against the alternative of IRLS betas forecasting better than the
MM betas. A pooled sample of betas within the period studied is considered. The
modified Diebold Mariano's test statistic between the two models is 1.37 with pvalue of
0.09 for onestep forecasts in the period of our study.
50
Table 2.9: Test of equal predictive accuracy between MM and IRLS models
This shows that we cannot reject the null hypothesis of equal predictive accuracy at the 5
percent significance level in the overall period. The null hypothesis can be rejected only
at the 10 percent level. From the deregulation and expansion period, the null hypothesis
of equal predictive accuracy is not rejected. From Table 2.9, we conclude that on the 19
th
century BSE, the IRLS method can help to curb the influence of outliers on estimated
betas, but it does not significantly outperform the standard MM in terms of their ability to
predict onestep ahead in the period of deregulation and expansion.
Models
β
MM
vs. β
IRLS
betas, β
IRLS
= iterative resistive least squares method betas. The hypothesis test H
0
: E(d
j
) = 0 and
H
a
= E(d
j
) >0. Pvalues are in parenthesis.
(0.01) (0.08) (0.09)
Note: This table reports the modified DieboldMariano test statistics for one step ahead equal forecast
accuracy between the market model and the iterative resistive least squares model. β
MM
=market model
1/183712/1871 1/187212/1911 1/183712/1911
2.29 1.43 1.37
Strict Regulatory Deregulation and
period expansion period Overall period
51
Conclusion 2.6
This chapter evaluates the relative performance of different methods of estimating beta
based on their ability to predict subsequent beta on the 19
th
century BSE. The analysis of
the different beta techniques reveals that beta estimated with the market model is not
stable. Specifically, the study reveals that for individual stocks, the market model beta is
weak in its ability to predict the future beta. The predictability can be improved by
grouping 10 or more stocks to form a portfolio or adjusting betas with the Vasicek and
Blume autoregressive techniques. The study also shows no significant difference
between the Blume and Vasicek adjusted betas in terms of their predictive accuracy.
Applying the Dimson method, correcting nonsynchronous trading effect reveals that
returns of few stocks have a significant relationship with the lead and lag market returns.
There is no significant difference in the predictive accuracy of the betas estimated with
the IRLS method and the market model in the deregulation and expansion period.
In the next chapter, we study the ability of beta to explain returns in the crosssection of
stocks, which is the primary implication of the CAPM.
52
CHAPTER 3
3 THE TEST OF CAPITAL ASSET PRICING MODEL (CAPM)
AND THE SIZE EFFECT IN 19
th
CENTURY BSE
The CAPM posits a positive relationship between the systematic risk (beta) and the
expected return of an asset. Despite the dominance of the CAPM in empirical finance, a
number of researchers have found evidence against the model. The purpose of this
chapter is to investigate if CAPM is valid before World War I, using the Brussels Stock
Exchange (BSE) data. In addition, we will investigate whether a company‟s market
capitalization (size) is related to its average excess returns (size effect). The use of the
preWorld War I (18681914) data will provide very good grounds for an outofsample
test of the CAPM and the size effect. This will minimize the data mining critique if the
CAPM is valid, and the size effect exists.
Introduction and Literature Review 3.1
This section investigates the crosssectional relationship between stock returns and beta.
We use preWorld War I Belgium data. Since the development of the CAPM in the
1960s by Sharpe (1964), Lintner (1965), and Mossin (1966), the literature has questioned
the validity of the model and suggest other characteristics than beta to explain expected
returns. The empirical study that supports the CAPM model in the 1970s is from Fama
and MacBeth (1973). It investigates whether there is a positive linear relationship
between expected returns and beta. In addition, it also examines whether other
parameters such as beta square and idiosyncratic risk can explain expected returns.
However, Banz (1981) finds a size effect in stock returns. The effect implies the
propensity for stocks with low market capitalization to outperform those with high
53
market capitalization. In addition, Lakonishok and Shapiro (1986) and Ritter and Chopra
(1989) do not detect any significant relationship between beta and expected returns. With
the debate on the validity of the CAPM still ongoing, Fama and French (1992) find no
association between betas and average returns, even when beta is the only explanatory
variable in their crosssectional regressions. Instead, they conclude that size and the
booktomarket value ratio can explain the variation in expected returns when placed
together in a crosssectional regression. In contrast, Kothari, Shanken and Sloan (1995)
use annual portfolio returns and equally weighted market index to document evidence in
support of the CAPM. The above literature on beta and size focuses on postWorld War I
return data and sometimes mainly on data from the US. Another view is that the
characteristic may have been discovered out of luck through data snooping bias (see Lo
and MacKinlay (1990)). In this case, the effect should not be found in other periods.
Dimson and Marsh (1999) test the presence of the size effect by using FTSE all share
monthly returns data from the period 1955 to 1998 and document that the effect has
disappeared after 1979. In addition, Schwert (2003) documents that the size effect
disappears after 1981 in the US market using monthly data for the period 1962 to 2002.
Grossman and Shore (2006) use preWorld War I UK data to present evidence against
the size effect. They find a size effect among extremely small stocks, which account for
about 0.2% of market capitalization, but the size effect disappears when these stocks are
eliminated.
To distinguish between data snooping and the persistence of size effect, we investigate
the effect on another dataset. The anomaly is initially discovered outside the period of
our study. To add to the existing literature on asset pricing, this chapter introduces the
54
19
th
century BSE data to test the validity of CAPM and the presence of the size effect. To
this end, we resort to the sorting and the FM crosssectional regression methods
discussed in Chapter one.
We find no relationship between beta and expected returns. We also find a size effect on
the 19
th
century BSE, but it disappears when stocks are value weighted to form
portfolios. Detailed investigation reveals that the size effect in our data is confined to
small stocks, which represent on average 0.35% of the total market capitalization.
The remainder of the chapter is organized as follows: In Section 3.2, we show the
expected returns of portfolios sorted on MM betas ( )
MM
 , Dimson betas ( )
dim
 , and
Vasicek betas ( )
V
 . The FM crosssectional regressions are used to test the relationship
between beta and expected returns (CAPM) in Subsection 3.2.1. In Section 3.3, we
investigate the effect of size and beta on excess returns by using the sorting method. In
Section 3.4, we use FM crosssectional regression analysis to confirm the above sorting
results. Section 3.4 concludes the chapter.
Expected returns of portfolios sorted on betas 3.2
In the sorting method, we rank stocks based on beta and group them to form portfolios.
As stated before, the question answered by this method is whether highbeta stocks
outperform lowbeta stocks. As the aim of this chapter is to test the validity of the
CAPM, the method needed to estimate its input is worth considering. In testing the
CAPM, one needs to form portfolios in order to improve on the precisions of individual
betas. In the previous chapter, Table 2.3 shows that a betasorted portfolio should contain
at least seven stocks in order to have a reliably stable portfolio beta estimate. Figure 3.1
55
shows the number of stocks included in our portfolio formation every year. Evidence
from this figure shows that until 1868 decile portfolios will not have the minimum of
seven stocks. The changes in legislation in 1867 ease the establishment of a company,
which is reflected in the number of stocks listed on the BSE. Furthermore, Van
Nieuwerburgh et al. (2006) indicate the importance of the longterm relationship between
the development of the BSE and economic growth in Belgium after legal liberalization.
In addition, as shown in Chapter 2 (Table 2.4), individual betas before 1868 do not
predict well their subsequent fiveyear beta. Based on these reasons, this chapter and the
subsequent ones will focus on the data between 1868 and 1914.
Figure 3.1: Number of stocks in our selection criteria for the entire period of the preworld war I SCOB data
For a stock to be included in the portfolio formation, it must have a minimum of 24
months of observations out of the 60 months required to estimate beta before the
56
portfolio formation year. In this chapter, we do not restrict our analysis to stocks with a
complete fiveyear returns data as in chapter two. This enables us to capture more stocks
in the crosssection. Including stocks with at least 24 months of returns does not change
the descriptive statistics of the prior betas (from here on preranking betas).
For comparison purposes, we study three beta estimates in this chapter. The first is the
MM( )
MM

beta, which is the traditional beta. It is the slope coefficient from the
regression equation
( )
,
jt ft MM mt ft jt
R R R R o  c ÷ = + ÷ + (9)
where
jt
R is the return on a portfolio or stock for period t ,
ft
R is the riskfree rate for
period t , and
mt
R is the market portfolio for period t . We use the valueweighted market
portfolio constructed by Annaert et al. (2004) as a proxy for the market portfolio. The
annualized money market rate, converted to a monthly rate, is used as a proxy for the
riskfree rate. We compute the second beta estimate using the Vasicek model introduced
in Chapter 2. The model is used to compute the posterior Vasicek betas ( )
V
 using the
crosssectional information on betas estimated by equation (9).
We recall from Chapter 2 (Table 2.7) that some stocks systematically lead or lag behind
the market movement, which may produce biased betas, when we estimate beta by the
MM. Possible explanation for the significant lead (lagged) relationship is because large
(small) firm prices adjust quickly (slowly) to market wide information. Thus, since the
market index used in this analysis is heavily weighted towards large stocks, small stock
returns have the tendency to lead or lag behind in relation to the market wide returns. We
57
adjust for the lag effect by using the Dimson model in Chapter 2 to obtain a third beta
estimate. That is, we run the regression
( ) ( )
,0 , 1 1 1
,
jt ft j j mt ft j mt ft jt
R R R R R R o   o
÷ ÷ ÷
÷ = + ÷ + ÷ + (10)
where
,0 j
 captures the contemporaneous covariation between the returns of a stock
(portfolio) and the market returns.
, 1 j

÷
captures the correlation between stock‟s current
period returns and the lagged market returns. Dimson onemonth lagged beta is estimated
as
,0 , 1 dim j j
  
÷
= + , which captures the correlation between the current period returns of
a stock and current and lagged market returns. For our monthly data, we use only one
month lag because Chapter 2 (Section 2.6) reveals that infrequent trading effect is not a
severe problem. In addition, Dimson (1979) with UK data, documents that the infrequent
trading effect is not severe when monthly returns are used to estimate betas.
Stocks are assigned to decile portfolios using the FM breakpoint method. We estimate
postranking portfolio betas for the sample period (18681914) by using valueweighted
and equally weighted portfolio returns. Specifically, beginning in January 1868, we
compute betas (preranking) for all stocks using the past 24 to 60 months of returns data.
We sort stocks into decile portfolios based on the preranking betas (univariate sort).
Portfolio 1 contains stocks with the lowest betas, while portfolio 10 contains stocks with
the highest betas. The postranking valueweighted and equally weighted return for each
month is calculated for each portfolio. New estimates of preranking betas are calculated
in December each year, and the portfolio formation is repeated. We account for the
possible timevariation in betas by rebalancing stocks in each year. Monthly portfolio
58
formation for each year yields 552 monthly returns for each decile portfolio. This process
is followed for all three beta estimates ( ) , and
MM V dim
   .
Table 3.1: Time Series Mean (%), Standard Deviation (%), and Postranking Betas of Decile portfolios formed from
preranking betas in Jan. 1868Dec. 1913
Table 3.1 reports the average excess return (time series), standard deviation, and the post
ranking betas of the ten portfolios. From Panel A, when both preranking and post
ranking betas are estimated with the market model, beta does not exhibit any relationship
with average returns. The average returns do not show any pattern as beta progressively
increases from low to high beta portfolios. The result does not change when we consider
Low 1 2 3 4 5 6 7 8 9 High10
0.34 0.13 0.18 0.45 0.45 0.25 0.32 0.23 0.09 0.26
3.91 2.14 1.93 2.88 3.14 3.22 3.68 3.82 4.71 6.67
0.68 0.45 0.53 0.82 1.10 1.19 1.49 1.60 1.94 2.68
0.17 0.18 0.09 0.12 0.40 0.36 0.27 0.16 0.06 0.05
3.17 1.22 1.62 2.02 2.82 2.90 3.06 3.34 4.07 5.66
0.62 0.33 0.41 0.67 1.01 1.21 1.33 1.45 1.80 2.48
0.18 0.30 0.28 0.26 0.20 0.30 0.35 0.23 0.26 0.32
3.02 3.12 2.33 2.64 3.00 3.07 3.61 4.12 5.13 6.38
0.71 0.64 0.61 0.95 1.26 1.23 1.48 1.71 2.14 2.49
0.06 0.33 0.17 0.32 0.18 0.26 0.30 0.03 0.14 0.05
1.63 3.35 1.50 1.95 2.79 2.45 3.08 3.56 4.52 5.48
0.39 0.58 0.45 0.75 1.13 0.98 1.32 1.56 1.92 2.32
0.25 0.23 0.26 0.43 0.37 0.25 0.30 0.23 0.20 0.17
3.59 2.32 2.04 3.08 3.09 3.22 3.68 4.70 5.09 5.58
0.92 0.93 0.92 0.97 1.10 1.19 1.45 1.58 1.57 1.55
0.20 0.13 0.12 0.19 0.36 0.36 0.20 0.17 0.06 0.01
2.79 1.24 1.47 2.34 2.71 2.90 3.36 3.39 4.35 4.66
0.80 0.81 0.81 0.86 0.99 1.21 1.36 1.37 1.41 1.39
series portfolio excess returns and the corresponding excess market returns.
Market Model (Value Weighted)
Dimson Betas(Equally Weighted)
Vasicek Betas(Value Weighted)
Mean (%)
Standard Deviation(%)
Beta
Mean (%)
Standard Deviation(%)
Standard Deviation(%)
Beta
At the beginning of each year, stocks are sorted based on preranking betas. The preranking betas are estimated with
Beta
Standard Deviation(%)
Panel A
Mean (%)
Standard Deviation(%)
Beta
Market Model (Equally Weighted)
Mean (%)
Standard Deviation(%)
Beta
Beta
Panel B
Mean (%)
Dimson Betas(Value Weighted)
Vasicek Betas(Equally Weighted)
market model (β
MM
), Vasicek's adjustment (β
V
) model and the Dimson's model with one month lag (β
dim
). The Fama
MacBeth breakpoint technique is used to assign stocks to decile portfolios. Portfolio 1 contains the lowest betas and Portfolio
10 contains the highest betas. Mean (%) is the time series average of the portfolio excess returns for the entire period.We
compute time series Standard Deviation(%) of the postranking excess returns. Betas are estimated by using the long time
Mean (%)
Panel C
59
the valueweighted portfolio excess returns. Estimating betas with the Dimson and
Vasicek methods in Panels B and C does not establish the relationship between beta and
expected returns. The most striking of all is that the postranking betas almost surely
follow the ordering of the preranking betas (except the first, second, and the sixth decile
portfolios). The univariate beta sorting results confirm Fama and French (1992) findings.
They use Dimson adjusted betas to establish a flat relationship between beta and average
return. We can also compare our result to the evidence of Reinganum (1981) who finds
no relationship between beta and average return in the period 19641979.
3.2.1 The CrossSectional Regressions
As stated in chapter one, the standard approach to test the validity of the CAPM is
sorting and FM (1973) crosssectional regression. In this subsection, we use FM cross
sectional regression to test the robustness of the above sorting result. The FM approach
also provides a straightforward procedure to test whether the reward for bearing beta risk
(risk premium) is equal to the excess market returns (the return of the market less the risk
free rate) as implied by Sharpe, Lintner, and Mossin‟s version of the CAPM. The method
also considers the noisy nature of portfolio or stock returns by running monthly cross
sectional regressions of betasorted portfolio returns on betas. That is,
0 1 jt ft t t jt t
R R ¸ ¸  q ÷ = + + (11)
where
0t
¸
and
1t
¸ are the regression intercept and slope for month t respectively.
jt
 is
the beta estimated from the fullsample portfolio returns. The slope coefficient from each
regression is treated as the reward per unit of the beta risk in that month (risk premium).
The time series average of the monthly coefficient is the average reward for bearing the
beta risk. The standard deviation of the monthly time series of slopes is used to perform a
60
ttest, whether the average slope is statistically significant from zero, in other words,
whether the beta risk is priced on average. To mitigate a possible errorinvariable
problem, the result in Chapter 2 (Section 2.4) shows that we can rely on the portfolio
betas. Fama and French (1992) also rely on full window portfolio betas to mitigate the
errorinvariable problem. Moreover, it is common to rely on large sample size statistics
to draw inferences. This curbs the argument that the test can be incorrect if the size of the
sample is not large enough for the asymptotic results to provide a good approximation.
We adopt the method by Fama and French (1992) to estimate full window portfolio
betas. The only difference is that we replicate Ibbotson et al. (1997) method and use the
portfolio betas for the crosssectional regression instead of assigning the portfolio beta to
individual stocks in the portfolio each year. As in the previous section, we sort stocks
based on their estimated preranking betas (Market model betas, Vasicek betas, and
Dimson betas) and form portfolios each year. Portfolio 1 contains the lowest beta stocks
while portfolio 10 contains the highest beta stocks. We form equally weighted and value
weighted portfolios from the betasorted group of stocks each month. We repeat the
process each year to account for time variations in betas. This will produce 552 monthly
returns of decile portfolios (postranking returns). The postranking betas are estimated
by using the postranking long time series' returns of the decile portfolios. We repeat the
process for the various estimates of betas ( ) , ,
MM V dim
   .
The postranking beta serves as the input for equation (11) to perform the crosssectional
regressions. Each month, we regress the postranking excess returns of the decile
portfolios on their corresponding beta (postranking) estimates.
61
Table 3.2: Average time series slopes from the FamaMacBeth CrossSectional Regressions in Jan. 1868Dec. 1913
Eventually, we obtained 552 crosssectional regressions for each estimate of beta. After
performing the monthly crosssectional regressions, the time series mean of the slope
coefficients is tested for statistical significance. The significance of the average slope is
tested by using heteroskedastic and autocorrelation consistent standard errors (Newey
ttest
Intercept β
MM
β
dim
β
V
H
0
:Slope=(R
m
R
f
)
Panel A: Equally Weighted Portfolio
0.30% 0.02% 1.41
(2.49) (0.17)
0.24% 0.02% 0.93
(1.87) (0.14)
0.42% 0.12% 1.05
(1.57) (0.45)
Panel B: Value Weighted Portfolios
0.24% 0.06% 2.16
(2.92) (0.48)
0.30% 0.11% 2.63
(3.41) (0.81)
0.28% 0.10% 0.94
(1.19) (0.34)
Panel C: Individual Stocks
0.29% 0.01% 1.47
(2.37) (0.16)
0.26% 0.02% 1.16
(2.10) (0.29)
0.36% 0.06% 1.55
(2.70) (0.41)
monthly crosssectional regression of postranking portfolio
excess returns on postranking beta estimates. It also shows
the hypothesis test of mean slope (risk premium) equal to the
average excess market returns as implied by the SharpeLintner
β
MM
=Market Model beta, β
V
=Vasicek beta and β
dim
=Dimson's
CAPM. Newey West adjusted tstatistics are in parentheses.
This table reports average time series slopes and intercepts from
beta with one month lag.
62
and West (1987) correction with default lag of int (
1
4
T ), where T
is 552). We also adopt
the Shanken‟s correction factor discussed in Chapter one for the computation of t
statistics. This is to eliminate the possible errorinvariable biased induced by the
estimated betas.
Table 3.2 reports the average intercepts, slopes, and their corresponding tstatistics in
parentheses. As shown by the sorting method, Panel A indicates that the market model
postranking beta estimated with equally weighted portfolio returns does not provide a
significant relationship with returns. Estimating preranking and postranking betas with
the Vasicek and Dimson methods does not establish the betareturn relationship.
Specifically, in Panel A, the mean estimated slope for the market model beta is negative,
and it is only 0.17 standard errors from zero. The negative slope is quite surprising as it
goes against the notion of positive risk premium (CAPM). Fama and French (1992) had a
negative slope for beta when placed together with size in the crosssectional regression.
The average slope using the Dimson beta is 0.02% with a tstatistic of 0.14. The
estimated mean slope with the Vasicek beta is also not significant.
The values in the last column show the tstatistics from the hypothesis test of average
slope (risk premium) equals the average excess market return as implied by the CAPM.
In Panel A, the hypothesis cannot be rejected at the 5% level, regardless of how beta is
estimated. However, it may be possible that the result is influenced by small stocks, since
equally weighted portfolios give undue weight to small stocks. Therefore, in Panel B, we
use valueweighted portfolios for the estimation of postranking betas and in the cross
sectional regression. The average slope of all the beta estimates in the crosssectional
regression is not significantly different from zero. Most strikingly, the hypothesis of
63
equality between the average slope and the average excess market return is rejected at the
5% level for the market model and the Dimson betas. In Panel C, we follow the
traditional FM (1973) rolling window approach by using individual preranking betas in
the crosssectional regression. This is a predictive test since the preranking betas are
estimated over a period prior to the period over which the crosssectional regression is
performed. The results do not support the CAPM for the three beta estimates.
Although, portfolio betas are used for crosssectional regression, others believe that
portfolios may conceal important information contained in the individual stock betas. For
example, Ang, Liu and Schwarz (2008) show that the slope coefficient (risk premium) of
crosssectional regression can be estimated more precisely using individual stocks
instead of portfolios because creating portfolios reduces the crosssectional variation in
betas. As a result, we apply the Fama and French (1992) approach of estimating full
window portfolio beta and assigning the portfolio beta to the individual constituent
stocks of the portfolio in the crosssectional regression. This serves as a robustness check
of the results in Table 3.2.
Table 3.3 reports the average crosssectional regression slopes for both equally weighted
and valueweighted portfolio betas assigned to individual stocks. The market model beta
and the Vasicek beta estimate still maintains the negative nonsignificant relationship
with average returns. A detailed look at Panel A shows that the Dimson beta is weak in
explaining average returns (average slope of 0.02% but with a tstatistic of only 0.17).
Using valueweighted portfolios (Panel B) to estimate postranking betas does not
establish the beta return relationship. This confirms Fama and French (1992) result; they
assert that beta is flat in relationship with average returns for post1960s USA data.
64
Surprisingly, in all cases the hypothesis that the mean slope is equal to the mean excess
market return is not rejected. The positive average slope of the Dimson beta cross
sectional regressions (Table 3.3, Panel A) calls for a detailed look into its time series'
behavior with the excess market returns. In addition, the average intercept is marginally
significant, and it is close to the average riskfree rate as postulated by CAPM. To
investigate the evolution of the slope coefficient and the excess market return through
time,
Table 3.3: Average Time Series Slopes from FamaFrench CrossSectional Regression in Jan. 1868Dec. 1913
ttest
Intercept β
MM
β
dim
β
V
H
0
:Slope=(R
m
R
f
)
Panel A: FamaFrench approach (eq)
0.29% 0.02% 1.40
(2.45) (0.17)
0.24% 0.02% 0.89
(1.81) (0.17)
0.42% 0.12% 1.03
(1.54) (0.44)
Panel B: FamaFrench approach (vw)
0.29% 0.02% 1.39
(2.54) (0.18)
0.24% 0.03% 0.89
(2.04) (0.19)
0.41% 0.12% 0.97
(1.56) (0.42)
enthesis.
vw=value weighted. Newey West adjusted tstatistics are par
excess market returns. β
MM
=Market Model beta, β
V
=Vasicek
beta and β
dim
= Dimson beta with one lag. eq=equally weighted
In this table, we assign the postranking portfolio beta to the
individual stocks in the portfolio. Portfolios are rebalanced
annually. Mean slope and their corresponding ttstatistic is
reported in parenthesis. We also report the tstatistic for the
test of hypothesis of the mean slope equal to the average
65
Figure 3.2: Sixty months moving average of the crosssectional slopes and excess market returns using Dimson beta
estimates
Surprisingly, there seems to be a close correlation between the slopes and the excess
market returns for much of the period except between the years (1880, 1885) and (1907,
1913).
Table 3.4: Subperiod look into estimated slopes and excess market returns
Figure 3.2 presents a fiveyear moving average of the estimated slopes and excess market
returns. The graph shows that the relationship between beta and expected returns varies
with time. In Table 3.4, we report subperiod average slope and intercept from the Fama
1870 1875 1880 1885 1890 1895 1900 1905 1910 1915
0.015
0.01
0.005
0
0.005
0.01
0.015
year
s
l
o
p
e
s
/
e
x
c
e
s
s
m
a
r
k
e
t
r
e
t
u
r
n
s
slope
excess market returns
Intercept Slope
0.15% 0.02%
(0.75) (0.09)
0.32% 0.03%
(3.15) (0.21)
H
0
:Slope=Avg.(R
m
R
f
)
ttest
subperiods
0.09
(Avg. R
m
R
f
=0.04%)
Jan. 1868Dec. 1893
Jan. 1894Dec. 1913
(Avg. R
m
R
f
=0.20%)
2.04
In this table, Dimson's beta estimated from equally weighted
portfolios is used in the crosssectional regressions for the two
subperiods. Avg. =Average. Newey West tstatistic in parenthesis.
66
and French (1992) crosssectional regressions using the Dimson beta. The last column
shows the tstatistics for the test of equality of the average slope and average excess
market return. For the first subperiod, the average excess market return (0.04%) is very
close to the average slope (0.02%). The null hypothesis of the equal average cannot be
rejected. In contrast, the null hypothesis that the average slope is equal to the average
excess market returns is rejected (tstatistic of 2.04) in the second subperiod as the
difference in magnitude confirms (0.03% average slope and 0.20% average excess
market returns). Chan and Lakonishok (1993) document similar results with post1920
Amex and NYSE data and caution researchers and practitioners not to rush into
discarding beta. The average slope is significantly less than the average excess market
return (a difference of about 0.17%).
Expected Returns, Beta, and the Size Effect 3.3
This section examines the wellknown size effect on the 19
th
century BSE. That is, the
propensity for large stocks to have consequent lower returns than small stocks. Early
works of Banz (1981), Reinganum (1981), (1983), Chan, Chen and Hsieh (1985), and
Chan and Chen (1988) first documented the size effect in modern data. Fama and French
(1992) present evidence that size and booktomarket combine to capture the cross
sectional variation in average stock returns in the period 19631990. Subsequently, Fama
and French (1993) build a threefactor model, which uses the excess market returns, size,
and booktomarket factors. The finance literature uses the threefactor model as a
benchmark model to measure longrun abnormal returns and many other factors. This
shows that researchers and practitioners have accepted size as an important characteristic
to explain the crosssectional behavior of longrun stock returns. On the contrary, a
67
recent paper by Horowitz, Loughran and Savin (2000) presents evidence against the size
effect in the USA market. It conjectures the magnitude of size effect is not robust when
the transaction costs and very small stocks (the removal of stocks with market
capitalization less than $5 million) are taken into account. Schwert (2003) uses US
monthly returns data between the year 1962 to 2002 to document that the size effect
disappears after 1981. With historical data, Grossman and Shore (2006) do not find any
presence of the size effect on UK data between the years 1870 to 1913. This would imply
size is not a systematic risk factor. We present similar evidence on the 19
th
century BSE
covering the same period.
Each year, we sort (univariate sort) stocks based on their size (or market capitalization) at
December of the prior year and then split them into decile portfolios. The market
capitalization is measured as the price of stock multiplied by shares outstanding. Again,
FM breakpoint method is employed to group the stocks into decile portfolios. As in the
previous sections, the smallest size stocks are put in decile one, and the largest size
stocks are put in decile ten. Portfolios are rebalanced each year to capture changes in
their constituent stock market capital over time. Monthly portfolio returns are calculated
as the valueweighted and equally weighted averages of the individual stock returns
within each of the ten portfolios. We compute the relative percentage size of a portfolio
as the time series average of the crosssectional sum of the market size of the stocks in
the portfolio divided by the sum of the size of stocks in our sample. That is, if
t
n
is the
number of stocks in a portfolio for the month t ,
t
N
is the number of stocks in the cross
section of our sample for the month t . T is the number of years. The relative percentage
of markets size is computed as
68
1
1
1
1
% Market Size = 100 ,
t
t
n
it T
i
N
t
jt
j
Size
T
Size
=
=
=
×
¿
¿
¿
Table 3.5: Beta Estimate and Mean Excess Return for the BSE equally weighted size portfolios, Jan. 1868 Dec. 1913
Two beta estimates of the size portfolios are calculated using equation (9) and (10) in the
previous section. Table 3.5 reports the beta estimates and the average excess returns of
the tensize portfolios during the period 1868 to 1914. It is well known in empirical
finance that small stocks have both a higher beta and average return than large stocks.
However, this is not the case when size portfolios are value weighted in our sample.
Column 3 reveals that equally weighted portfolio 1 has an extreme average excess return
(1.12%) which is almost three times the next largest excess return (0.38% from portfolio
6). The negative relation between size and returns is concentrated in the first decile
Size Portfolio % Market Size EW(%) VW(%) β
MM
β
dim
β
MM
β
dim
EW(%) VW(%)
1 0.35 1.12 0.01 1.61 1.89 1.43 1.56 6.16 5.26
2 0.94 0.29 0.14 1.37 1.45 1.26 1.42 4.45 3.86
3 1.60 0.10 0.10 1.16 1.30 1.16 1.28 3.43 3.37
4 2.43 0.14 0.16 1.43 1.54 1.41 1.52 3.65 3.57
5 3.56 0.06 0.01 1.09 1.16 1.10 1.16 2.77 2.73
6 5.02 0.38 0.52 1.43 1.46 1.61 1.67 3.93 5.16
7 7.04 0.12 0.10 1.16 1.17 1.16 1.17 2.71 2.67
8 10.06 0.20 0.21 1.35 1.36 1.35 1.38 2.97 2.91
9 15.56 0.16 0.19 1.12 1.13 1.12 1.13 2.45 2.43
10 53.46 0.17 0.15 0.88 0.84 0.79 0.74 1.84 1.68
0.95 0.14
tstatsitics (3.74) (0.63)
Fstatistics with the first decile 4.04
Fstatistics without the first decile 0.79
3 and 4. Relative market size is reported in column 2. We also report the Dimson and market model betas for the
In this table, stocks are ranked each year based on their size at the end of the prior year. They are then grouped
deciles for portfolio formation. Portfolio one contains the smallest size stocks, portfolios ten contains the largest
size stocks. Portfolios are rebalancec each year.Average excess returns of the decile portflios are reported in column
EW V W Standard Deviation R
p
R
f
mean of hedge portfolio (%)
decile portfolios. We also retport the standard deviation of the portfolio return series. The Fstatistic for the test
of hypothesis of equal mean of the porftolio returns is also reported.We test the hypothesis with and without the 1st
decile portfolio. EW=equallyweighted and VW=value weighted.
69
portfolio as the average excess return sharply drops from portfolio 1 to portfolio 2. As the
data source has been well checked for outliers, the extreme excess returns of 1.12%
recorded for the first size decile is not due to data error. Not able to study the events in
1880, 1890 and 1905, the stocks in the lowest size portfolio exhibit very high returns in
these years. In addition, the relative market capitalization (0.35%) of these stocks shows
that they are likely so illiquid that it would have been difficult to profit from buying
them. Looking at the Fstatistics, the null hypothesis of equal average returns is rejected
at the 1% significant level when the first decile is included in the test. Excluding the first
decile portfolio fails to reject the hypothesis. In addition, the effect disappears when
stocks are value weighted in the portfolios. The average excess return of the equally
weighted hedge portfolio (mean excess return of 0.95% and tstatistic 3.74) shows that
the size effect exists in our data. Surprisingly, the value increases to 0.14% (tstatistic of
0.63) for the valueweighted hedge portfolio.
Table 3.6: Equally weighted portfolios excess returns without the firstsize decile group
R
p
R
f
(EW%)
1 0.28
2 0.10
3 0.11
4 0.05
5 0.39
6 0.12
7 0.17
8 0.21
9 0.19
10 0.14
mean of hedge portfolio 0.14
tstatsitic (0.80)
Size Portfolio
70
Recently, Fama and French (2008) used USA data from 1963 to 2004 to document that
the size effect owes much of its power to micro caps and that it is marginal for small and
big caps. As mentioned earlier, Grossman and Shore (2006) find similar results for the
UK market in the same period of our study.
For robustness, we eliminate the stocks in the firstsize decile each year and perform the
size sorting analysis. As shown in Table 3.6, the size effect disappears when we
eliminate the first decile portfolio (portfolio with relative market size of about 0.35%)
before the size portfolio formation every year. This corroborates Horowitz et al. (2000),
who find no size effect in the period from 1963 through to 1981 when they eliminate
firms with less than $5 million in market value on the USA market.
Figure 3.4 plots the MM ( )
MM
 and the Dimson ( )
dim
 betas with the onemonth lag
for each equally weighted size portfolio. Clearly, the difference between the
MM
 and
dim
 progressively gets smaller as stock size gets larger.
Figure 3.3: Size Portfolio betas
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
1 2 3 4 5 6 7 8 9 10
B
e
t
a
Size Portfolio
Market Model Betas
Dimson Betas
71
This shows that small stock betas are underestimated when estimated with the market
model. This might be due to nonsynchronous trading as chapter two reveals that some
stocks show lead or lag relationship with the market returns. Ibbotson et al. (1997) find
similar results on the USA market between the years 1926 and 1994. They recommend
the inclusion of lagged information of market returns in the estimation of beta. We also
recommend the use of the Dimson beta with the onemonth lag when estimating betas for
small stocks in our sample. This is to curb the possible underestimation of smallstock
beta. There is a clear negative correlation (0.79 with a pvalue of 0.0065) between size
and portfolio beta (Figure 3.4).
3.3.1 FamaMacBeth CrossSectional Regressions to Test the Size Effect
In order to support the above evidence on size effect, we resort to the FM crosssectional
regression method adopted by Ibbotson et al. (1997). We regress the crosssection of
excess returns for a given month on the beta estimate (full window beta estimate) and
natural logarithm of size by using an extension of equation (11):
( )
0 1 2
1
ln Size
jt ft t t jt t t
jt
R R ¸ ¸  ¸ q
÷
÷ = + + +
(12)
where
0t
¸ ,
1t
¸ and
2t
¸ are the regression intercept and slopes for month t , respectively.
t
 is the full period estimate of beta for portfolio. In our sample, the previous section
reveals that size is crosssectionally correlated with beta. In addition, Chan and Chen
(1988) argue that as size serves as a proxy for betas, they expect the betas of size
portfolios to be strongly correlated crosssectionally with size. However, when both
characteristics are included in a regression, the correlation will increase the standard
errors of the estimates, and this will make the outcomes murky for interpretation. Fama
72
and French (1992) show that when portfolios are formed on size alone, there is evidence
of a positive relationship between average return and beta (CAPM). The correlation
between size and beta makes the test on size portfolios unable to disentangle the effect of
size and betas on average returns.
We show that when equally weighted portfolios are built on size alone, there is support
for CAPM. However, allowing the variations in beta that are unrelated to size, it breaks
the effect of size and beta even on equally weighted portfolio excess returns. We achieve
this by conditional double characteristics sorting. Specifically, we first sort stocks based
on size and then sort within each size group on preranking beta. We find a strong
relation between size and average excess return but no relation between beta and average
return for equally weighted portfolios. The size effect disappears when stocks are value
weighted in portfolios. The size effect does not exist when we eliminate the firstsize
decile portfolio in the analysis each year.
As in the sorting method, we form decile size portfolios. This is to confirm the effect of
the correlation between size and beta on the betareturn relationship. To separate the
effect, we sort stocks into three size groups each year. Each size group is then sorted into
five groups based on their preranking
MM
 or
dim
 beta estimates. The equally and
valueweighted return for each portfolio is computed for each month of the following
year. The conditional doublesorting portfolio formation is repeated at the end of each
year. The procedure generates fifteen sizebeta portfolios for each beta estimate. For all
portfolio formations, we use the FM breakpoint technique. Postranking betas are
estimated with postranking returns over the entire period from 1868 through to 1914.
Each month, we regress portfolio excess returns on beta and the natural log of size by
73
using equation (12). The full period postranking betas are used in the crosssectional
regressions. Size is determined at the end of the year before the portfolio formation year.
Table 3.7: Average time series slopes and intercept from the FamaMacBeth crosssectional regression: Jan 1868
Dec. 1913
Intercept β
MM
β
dim
ln (Size) Intercept β
MM
β
dim
ln (Size)
1.06% 1.05% 0.23% 0.31%
(3.62) (3.46) (0.83) (1.12)
0.88% 0.86% 0.10% 0.20%
(3.50) (3.43) (0.45) (0.86)
2.12% 0.13% 0.34% 0.03%
(2.56) (2.63) (0.47) (0.79)
0.42% 0.85% 0.03% 2.00% 0.62% 0.09%
(0.40) (2.87) (0.55) (1.72) (1.74) (1.71)
2.05% 1.08% 0.06% 2.45% 0.62% 0.12%
(1.72) (3.50) (0.97) (1.81) (1.76) (1.86)
Panel B1: Sizeβ
mm
Portfolios Panel B2: Sizeβ
mm
Portfolios
0.26% 0.01% 0.29% 0.12%
(2.18) (0.09) (2.72) (0.91)
1.63% 0.09% 0.36% 0.03%
(2.08) (2.05) (0.56) (0.91)
1.69% 0.03% 0.09% 0.03% 0.12% 0.02%
(2.35) (0.20) (2.18) (0.05) (0.89) (0.61)
Panel C1: Sizeβ
dim
Portfolios Panel C2: Sizeβ
dim
Portfolios
0.17% 0.08% 0.27% 0.09%
(1.35) (0.56) (2.51) (0.67)
1.69% 0.10% 0.22% 0.02%
(2.09) (2.06) (0.32) (0.63)
1.64% 0.01% 0.09% 0.10% 0.09% 0.01%
(2.17) (0.07) (2.10) (0.15) (0.64) (0.31)
EQUALLY WEIGHTED VALUE WEIGHTED
Panel A1: Size Portfolios Panel A2: Size Portfolios
ing betas are used in the crosssectional regression. tstatistics are in parenthesis.
This will yield 15 sizebeta equally and value weighted portfolios. In all portfolio formations we use the FM
break point. Estimate postranking betas by using the full period postranking excess returns. Post rank
and value weighted portfolio returns are computed each month in the year. The joint effect of size and
beta is seperated by first forming three size portfolios and splitting each size group into five beta groups.
Each year, we sort stocks into ten portfolios based on their size at the end of the prior year. Equally
74
Table 3.7 reports the time series averages of the slopes and intercept of the regression.
The time series standard deviations of the slopes and the intercepts are used to test
whether the average is significantly different from zero. We use Newey and West (1987)
heteroskedastic autocorrelation corrected standard errors for the computation of the t
statistics (reported in parentheses). The Shanken‟s adjustment factor is also adopted in
the computation of the tstatistics.
The values in Panel A1 show that the CAPM is valid for equally weighted univariate
sizesorted portfolios. Both
MM
 and
dim
 are positively related to excess return when
placed alone in the crosssectional regression. Size is negatively related to excess returns.
When size and any of the beta estimates are placed simultaneously as independent
variables, only the beta estimate is significantly related to excess returns. Interestingly,
size is sometimes positively but insignificantly related to excess returns when placed
simultaneously with beta in the regressions. This is contrary to Ibbotson et al. (1997)
result, where size is significant when placed together with market model betas in the
regression. When equally weighted portfolios are formed on size alone, both the market
model and the Dimson beta with the onemonth lag can predict returns at the expense of
size.
The implication of the CAPM requires that the average crosssectional regression
intercept will not be significantly different from zero. However, from Panel A1, the
average intercept is significantly negative when MM and the Dimson‟s betas are used in
the crosssectional regressions. The negative intercept becomes significantly positive
when size is placed alone as a regressor in the crosssectional regression. The average
75
intercept is not significantly different from zero when the MM and Dimson‟s betas are
placed together with size in the crosssectional regression.
Panels B1 and C1 show the crosssectional regression slope and intercept for conditional
doublesorted sizeβ
MM
and sizeβ
dim
portfolios respectively. Both betas
are no more
significantly related to returns, whether placed alone or with size in the regressions.
There is a statistically significant relationship in size to excess returns, whether placed
alone or with any of the beta estimates. This is in support of Fama and French (1992)
evidence that the conditional doublesort portfolio (sizebeta sort) allows variations in
beta that are unrelated to size and would break the combined of size and beta on
expected returns. Therefore, size will be related to average returns, but beta will not. It is
worthy to note the significance of the intercepts. These show that beta and size cannot
combine to capture the crosssectional variation in stock returns.
Most interestingly, when the valueweighted portfolios are used in the analysis (be it
univariate size sorting or conditional double sizebeta sorting), beta or size is not
significantly related to the average excess return (See Table 3.7, Panels A2, B2, and C2).
This suggests that the result from the equally weighted portfolio is due to the influence of
small stocks since it assigns equal weights to all stocks in portfolio formations and in the
crosssectional regressions. This confirms the sorting result in Table 3.5; size effect does
not exist when stocks are value weighted in portfolios. We repeat the above analysis by
adopting Fama and French (1992) method. At the end of each year, the postranking
betas estimated with the full period postranking returns will be assigned to each stock in
the portfolio. Assigning full period postranking betas to stocks does not mean a stock‟s
76
beta is constant, as stocks can move across portfolios with yearly rebalancing. The
method uses the information available for individual stocks in the cross section.
Table 3.8: Average Time Series Slopes and Intercepts from the FamaFrench CrossSectional Regressions: Jan. 1868
Dec. 1913
Intercept β
mm
β
dim
ln (Size) Intercept β
mm
β
dim
ln (Size)
1.05% 1.04% 0.54% 0.65%
(3.58) (3.44) (2.57) (2.79)
0.87% 0.85% 0.48% 0.58%
(3.46) (3.42) (2.58) (2.76)
2.63% 0.16% 2.63% 0.16%
(3.15) (3.29) (3.15) (3.29)
1.51% 0.37% 0.12% 2.43% 0.06% 0.15%
(1.34) (1.22) (2.07) (2.44) (0.26) (2.88)
1.16% 0.38% 0.10% 2.86% 0.10% 0.17%
(0.86) (1.20) (1.43) (2.62) (0.46) (2.91)
Panel B1: Sizeβ
mm
Portfolios
0.31% 0.03% 0.26% 0.00%
(2.53) (0.27) (2.27) (0.03)
2.63% 0.16% 2.63% 0.16%
(3.15) (3.29) (3.15) (3.29)
2.70% 0.07% 0.16% 2.77% 0.07% 0.16%
(3.54) (0.59) (3.39) (3.57) (0.50) (3.48)
Panel C1: Sizeβ
dim
Portfolios
0.21% 0.04% 0.17% 0.08%
(1.58) (0.28) (1.44) (0.54)
2.63% 0.16% 2.63% 0.16%
(3.15) (3.29) (3.15) (3.29)
2.69% 0.05% 0.16% 2.78% 0.05% 0.16%
(3.52) (0.35) (3.47) (3.55) (0.37) (3.55)
EQUALLY WEIGHTED VALUE WEIGHTED
Panel A2: Size Portfolios Panel A1: Size Portfolios
Each year, we sort stocks into ten portfolios based on their size at the end of the prior year. Equally
and value weighted portfolio returns are computed each month in the year. The joint effect of size and
beta is seperated by first forming three size portfolios and splitting each size group into five beta groups.
This will yield 15 sizebeta equally and value weighted portfolios. In all portfolio formations we use the FM
break point. Estimate postranking betas by using the full period postranking excess returns. We assign
postranking betas to the constituent stocks in the portfolio. Portfolios are rebalanced each year. tstatis
tics are in parenthesis.
Panel B2: Sizeβ
mm
Portfolios
Panel C2: Sizeβ
dim
Portfolios
77
Table 3.8 reports the average slopes and intercepts of the crosssectional regressions
using equally weighted and valueweighted portfolios to estimate the postranking betas.
The values in parentheses are the Newey West adjusted tstatistics for the test of a
hypothesis of the average slope or intercept significantly different from zero. From Panel
A1, when the fullperiod equally weighted portfolio returns used to estimate postranking
betas are formed on size alone, both
MM
 and
dim
 have a strong relation with returns
when placed alone in the regression. They lose their relationship when placed together
with size in the regression. This indicates that beta, which is correlated with size, serves
as a proxy for size when placed alone in the regression. From Panels B1 and C1,
conditional doublesorting returns based on size and betas break the hold up between size
and beta. It can be seen that beta has no relationship with excess return when it is placed
alone or together with size. The result is similar when valueweighted postranking
returns are used to estimate postranking betas (See Panels A2, B2 and C2).
For robustness, we repeat Fama and French (1992) crosssectional analysis, but exclude
stocks in the firstsize decile each year. As in the sorting method, Table 3.9 does not
show the significant relationship between betas and expected returns when placed alone
or combined with size for singlesorted size portfolios in panel A1. In Panels B1 and C1,
double sorting stocks to form portfolios will not establish the relationship between betas,
size, and returns. When valueweighted portfolio returns are used in the analysis, size and
beta have no relationship with return as shown in panels A2, B2, and C2. This shows that
any size effect present in our data is driven by a small group of stocks with an average
relative market size of about 0.35%.
78
Table 3.9: Average Time Series Slopes and Intercepts from the FamaFrench CrossSectional Regressions without
the FirstSize Decile: Jan. 1868Dec.1913
Intercept β
mm
β
dim
ln (Size) Intercept β
mm
β
dim
ln (Size)
0.21% 0.31% 0.17% 0.28%
(0.92) (1.25) (0.85) (1.37)
0.10% 0.21% 0.11% 0.22%
(0.50) (0.98) (0.65) (1.24)
0.20% 0.00% 0.20% 0.00%
(0.28) (0.07) (0.28) (0.07)
1.39% 0.54% 0.06% 0.88% 0.37% 0.04%
(1.27) (1.83) (1.09) (0.89) (1.55) (0.77)
1.65% 0.51% 0.08% 1.22% 0.37% 0.06%
(1.30) (1.67) (1.22) (1.10) (1.61) (1.05)
Panel B1: Sizeβ
mm
Portfolios Panel B2: Sizeβ
mm
Portfolios
0.28% 0.09% 0.28% 0.10%
(2.70) (0.74) (2.82) (0.72)
0.20% 0.00% 0.20% 0.00%
(0.27) (0.06) (0.27) (0.06)
0.34% 0.10% 0.00% 0.52% 0.10% 0.02%
(0.55) (0.79) (0.10) (0.86) (0.76) (0.41)
Panel C1: Sizeβ
dim
Portfolios Panel C2: Sizeβ
dim
Portfolios
0.28% 0.08% 0.27% 0.08%
(2.57) (0.62) (2.73) (0.60)
0.20% 0.00% 0.20% 0.00%
(0.27) (0.06) (0.27) (0.06)
0.41% 0.09% 0.01% 0.55% 0.09% 0.02%
(0.68) (0.67) (0.22) (0.91) (0.67) (0.49)
This will yield 15 sizebeta equally and value weighted portfolios. In all portfolio formations we use the FM
break point. Estimate postranking betas by using the full period postranking excess returns. We assign
postranking betas to the constituent stocks in the portfolio. Portfolios are rebalanced each year. tstatis
tics are in parenthesis.
EQUALLY WEIGHTED VALUE WEIGHTED
Panel A1: Size Portfolios Panel A2: Size Portfolios
Each year, we sort stocks into ten portfolios based on their size at the end of the prior year. Equally
and value weighted portfolio returns are computed each month in the year. The joint effect of size and
beta is seperated by first forming three size portfolios and splitting each size group into five beta groups.
79
Conclusion 3.4
We used sorting and the crosssectional regression method to investigate whether the
CAPM model is valid in the period before World War I. We find no support for the
CAPM in the 19
th
century BSE. Estimating beta with the market model, Dimson model,
and Vasicek model does not establish the crosssectional relationship between expected
returns and the beta.
However, when we use equally weighted size portfolios in the crosssectional
regressions, we find the relationship between excess returns and size or beta. There is a
negatively significant relationship of size to excess returns (size effect), but beta does not
relate to excess returns, when placed at the same time as regressors in the crosssectional
regression. This is due to a strong correlation existing between size and beta. We find
that conditional doublesorting portfolios by size and then by beta breaks the effect
between size and beta on excess returns. As a result, the average slope of the cross
sectional regression of returns on betas becomes insignificant when placed alone in the
regression or combined with size. We recommend researchers estimate betas with the
Dimson method with a onemonth lag since smallstock betas are underestimated when
estimated with the market model with this data.
Further investigation reveals that the size effect in our data is mainly due to small stocks
with relative market size of about 0.35% of the total market size. Eliminating these small
stocks destroys the relationship between excess returns, beta, and size. Both the sorting
and the crosssectional regression methods reveal that the size effect disappears when the
valueweighted portfolios are used in the regression.
80
CHAPTER 4
4 DOES THE MOMENTUM EFFECT EXIST IN THE 19
TH
CENTRURY?
Fama and French (2008) document that the momentum or relative strength strategy
produces an average return that is robust and significant across different size categories
of stocks. The strategy buys stocks that have performed well in the past 6 to 12 months
and sells stocks that have performed poorly in the same period. It produces significant
profit in the following 6 to 12 months. However, most studies testing the returns
momentum effect in 6 to 12month horizons use postWorld War II (WWII) data from
the USA, Europe and Emerging Markets (see Jegadeesh and Titman (1993),
Rouwenhorst (1999), Nijman, Swinkels and Verbeek (2004) and etc.). In this chapter, we
use our preWorld War I BSE data to test the presence and source of momentum effect
between the years 1868 and 1914. This will serve as an outofsample and robustness test
of the postWWII momentum profits documented.
Introduction and Literature Review 4.1
Jegadeesh and Titman (1993) document that portfolios or stocks that have performed
well in the past 3 to 12 months often continue to deliver high returns in the subsequent 3
to 12 months. Their result confirms the pervasive, positively significant, lagpredictable
patterns of stock returns documented by Jegadeesh (1990). Subsequent to this, there has
been extensive literature published confirming the robustness of the momentum effect.
For example, Rouwenhorst (1998) documents the momentum effect in 12 European
equity markets. Though researchers and practitioners have subscribed to the view that the
momentum strategy yields a significant profit, the source of the momentum profit poses a
81
strong challenge in the literature. The literature attributes the source of the profit to cross
sectional variation in expected returns, data mining bias, the state of the market and the
behavioral biases of investors.
To examine the possibility that the profit can be explained by the crosssectional
variation in expected returns, Jegadeesh and Titman (1993) compute momentum profit
within size and betabased subsamples with lower dispersion in expected returns. They
find that the profit is not necessarily smaller in the subsamples. Based on this result,
they conclude that the momentum profit is not due to the crosssectional variation in
expected returns. On the contrary, Conrad and Kaul (1998) and Bulkley and Nawosah
(2009) argue that the momentum profit is due to the crosssectional variation in the
expected returns, but not to autocorrelation in stock returns. Autocorrelation is the
correlation between two returns observations of the same returns series at different times
(time series pattern in stock returns). The crosssectional variation of expected returns
explaining momentum implies that the returns of the momentum portfolio will be
positive on average in any postranking period. However, Jegadeesh and Titman (2002)
attribute the results found by Conrad and Kaul (1998) to a small sample bias in their
empirical test and bootstrap experiments. The bias is due to the likelihood of drawing
with replacement, the same returns observation in the formation and the holding periods.
Jegadeesh and Titman (2002) adopt sampling without replacement to mitigate the bias
selection of a particular observation and realized that the crosssectional variations in
expected returns contribute very little of the momentum profits.
While some argue that the momentum profit is due to crosssectional variations in
expected returns, others believe the profits may be due to data mining. To counter the
82
data mining explanation, Jegadeesh and Titman (2001) extend the data from the period
19651989 to 1998 and perform an outofsample test. Extending the data enables them
to assess whether the result in the period 19651989 is merely due to chance. In addition,
it will allow them to assess whether investors have changed their investment strategies
and if the profit of the momentum strategy does not exist anymore, especially for large
size stocks, which are easier and less expensive to trade. However, they find that the
momentum effect still exists in their extended data, and that the profit found in Jegadeesh
and Titman (1993) is therefore not likely to be a data mining bias. They find momentum
in the first 12 months after portfolio formation. However, the cumulative return in
months 13 to 60 after portfolio formation is negative (return reversal), which is consistent
with the behavioral theories.
On the relationship between the state of the market and momentum, Cooper et al. (2004)
apply the behavioral theory of Daniel et al. (1998) to predict the differences in
momentum profit across different market states, like a bull versus a bear market, as
aggregate overconfidence should be greater following market gains. They find that
momentum profits depend on the state of the market. Likewise, Chabot et al. (2009) use
19
th
century data (Victoria Era) from the UK to document the dependence of momentum
profit on the state of the market, when a three year lag in average return is used to define
market states.
It is imperative that executing the momentum strategy involves frequent trading. Despite
the existence of momentum profits on different markets and in different time periods,
Lesmond, Schill and Zhou (2004) argue that when transaction costs are considered,
momentum profit is just an illusion. This is because the strategy tends to pick stocks that
83
have a high trading cost. Frequent trading in these high cost stocks will prevent gainful
strategy execution.
The literature also shows that momentum is strong in industry portfolios than individual
stocks. For example, Moskowitz and Grinblatt (1999) document a strong and persistent
industry momentum effect that cannot be explained by size, value, individual stock
momentum, the crosssectional dispersion in expected returns and possible
microstructure effects. In contrast, Nijman et al. (2004) documents that the momentum
profits on the European markets are mainly due to individual stock effects. Since the
primary motive of this chapter is to investigate whether momentum profit exists in the
19
th
BSE, future research can be conducted on this data to determine whether the
momentum profit is due to individual stock or industry effect.
In this chapter, we investigate whether a 3 to 12months momentum effect existed in the
19
th
century and early part of the 20
th
century BSE using the methodology similar to
Jegadeesh and Titman (1993). The only difference is that Jegadeesh and Titman (1993)
used decile portfolios in their analysis, while we use quintile portfolios in order to have a
sufficient number of stocks in our portfolios. We find investors can earn significantly
positive returns over 3 to 12month holding periods when they adopt the momentum
strategy. The result is not influenced by forming the portfolios just after the formation
period or by skipping one month after the portfolio formation period. Finding a
momentum effect in the 19
th
century casts additional doubt on the datamining
explanation for momentum.
84
A detailed look into an extensively researched momentum strategy (6 months formation
period and 6 months holding period) on the 19
th
century BSE reveals that the strategy
tends to pick small size stocks. However, further tests also suggest that momentum exists
in beta and large size subsamples. Unlike Jegadeesh and Titman (1993), (2001) who find
negative momentum profit in January (a reversal) each calendar year, we do not record a
January reversal in the 19
th
century BSE. The momentum strategy yields a positive return
in all months throughout the year. Subperiod analysis also shows that momentum profit
is robust in all tenyear subperiods, except the years between 1878 and 1887.
When we replicate the postholding period event analysis documented by Jegadeesh and
Titman (2001), we find a strong shortterm momentum profit in the first 12 months and
long run reversal, two years after portfolio formation. There is no return reversal in the
first month after the portfolio formation. There is evidence of a sharp rise in profit in the
first year, but the profit declines and reverses to become negative in the second to fifth
year after the portfolio formation, as in the US. We can conclude that in the 19
th
century,
the profit does not remain positive after 12month holding periods, as claimed by Conrad
and Kaul (1998). We also investigate the momentum profit across market states, as it
may be consistent with many behavioral explanations. We follow Cooper et al. (2004) to
predict the cyclicality of the momentum profit across different market states. We find
that momentum profit depends on the state of the market when three years lagged value
market weighted returns are used to define market states.
The rest of the chapter is organized as follows: we explain and compute the returns of the
various combinations of momentum strategies in Section 4.2. In Subsection 4.2.1, we
will focus on a sixmonth formation/ sixmonth holding period strategy. This strategy has
85
been extensively researched in the literature to represent the other strategies. We
compute average returns (equally weighted and value weighted), and the average market
capitalization of the quintile momentum portfolios. We document the returns of the
strategy in size and beta based subsamples in Subsection 4.2.2. Section 4.3 investigates
seasonality in the momentum profits across calendar months and tenyear subperiods. In
section 4.4, we investigate the source of the momentum profit by studying the
characteristics of the momentum profit after the 12month holding period. We study the
momentum profit across different market states in section 4.5. Section 4.6 concludes our
findings.
Momentum Trading Strategies and their Returns 4.2
As in the previous chapters, this chapter uses the 19
th
century BSE data, consisting of
monthly prices, total returns (returns adjusted for stock splits and dividends), and
outstanding shares listed in Brussels between the years 1868 and 1913. As can be seen
from Figure 4.1, the number of stocks in our sample varies from 71 in 1868 to 513 in
1913. The number of stocks increases sharply from 1890 to 1913. The momentum
strategy and its profitability are constructed in this section as follows: At every given
month , we rank stocks in ascending order based on their previous
(formation period) compound returns. Based on these rankings, five equally weighted
and value weighted portfolios are formed. In order to obtain sufficiently diversified
portfolios, we form quintile portfolios instead of the decile portfolios formed by
Jegadeesh and Titman (1993) and Bulkley and Nawosah (2009). The weights are
determined by the market capitalization of each stock in the portfolio at the end of the
portfolio formation.
t months ÷ P
86
Figure 4.1: Number of common stocks in our sample for the momentum studies
Figure 4.2: Time line of sample periods
Formation Period Holding Period
tP t t+Q
The bottom quintile portfolio is called “Loser," and the top quintile portfolio is called
“Winner.” The momentum strategy is to buy the winner portfolio and sell the loser
portfolio in each month . The strategy holds the position for
(holding
period). Either the holding period is immediately after the formation period, or we skip
one month after the formation period. Figure 4.2 shows the various periods we
considered in the strategy. We update our portfolio formation every month. Various
combinations of P and Q are considered, where P and Q equal three, six, nine and twelve
months. This yields 16 strategies each for equally weighted and valueweighted portfolio
formation. We calculate holding period returns from the same month in which the stocks
were ranked in order to form portfolios. Specifically, the strategy that selects stocks
0
100
200
300
400
500
600
1860 1870 1880 1890 1900 1910 1920
N
u
m
b
e
r
o
f
S
t
o
c
k
s
year
t months ÷ Q
87
based on the past Pmonths return and holds it for Qmonths is known as a P/Q strategy.
For a comparative purpose, we also calculate returns for skipping one month between the
formation period and the holding period. This will be referred to as P/Q/1 strategy.
Specifically, we construct 6/6 strategy as follows: at the end of each month we sort
stocks in our sample based on their past six month returns (month 6 to month 1) and
group the stocks to form five portfolios based on the ranks. Portfolios are held for six
months (month 0 to month 5) following the ranking months. In order to improve the
power of our test, in all the portfolio formations, we replicate the overlapping portfolio
method adopted by Jegadeesh and Titman (2001). That is, for the 6/6 strategy, the
winning portfolio in month t will contain the top quintile of stocks ranked over the
previous t5 to t, t6 to t1, t7 to t2, and it will continue until t11 to t6. To illustrate
this on the calendar months, the return in December of the winning portfolio will contain
the top quintile of stocks ranked over the previous June to November, the previous May
to October, and it will continue over the period from January to June. We form equally
weighted and valueweighted quintile portfolios from the returns of stocks that coexist in
the same ranking. We resort to the FamaMacBeth method for the portfolio break points.
Table 4.1 presents the monthly average returns of the equally weighted and value
weighted portfolios of the various strategies, over the period from January 1868 to
December 1913 on the BSE. In Panel A, we report both the equally weighted and value
weighted average returns for strategies with holding periods starting immediately after
portfolio formation. In Panel B, we report equally weighted and valueweighted average
portfolio returns for strategies with holding periods of one month after the portfolio
formation.
88
For each cell in Table 4.1, we report the average returns of the winner, the loser and the
zero cost (winner minus loser) portfolios of the various strategies. The tstatistics based
on NeweyWest heteroskedasticity and autocorrelationadjusted standard errors are
shown in parentheses. For all strategies, the average return of the value weighted zero
cost portfolio is higher than the equally weighted zerocost average return. This suggests
that momentum on the 19
th
century BSE is mainly due to large size (market capital)
stocks. In Panel A, the average return of the equally and valueweighted zerocost (WL)
portfolios for all strategies are positive and significant, except the equally weighted
average return of the 3/3 strategy, which is marginally significant. In Panel B, the
average returns of the equally and value weighted zerocost portfolio for all strategies are
positive and significant, except the equally weighted 12/12/1 strategy, which is
marginally significant.
The most profitable zerocost strategy selects stocks based on their previous 6 month
returns and holds the portfolio for 6 months. It yields 0.71% and 0.75% per month for the
equally weighted zerocost portfolio, when portfolios are formed immediately and one
month after the formation period respectively. For the valueweighted portfolio
formations, it yields 1.06% each for both 6/6 and 6/6/1 strategies. This is contrary to the
results found by Jegadeesh and Titman (1993), where the most profitable strategy was a
12/3 strategy. The 9/3 strategy, with an equally weighted portfolio formation, yields
almost the same profit as the 6/6 equally weighted strategy (0.70% per month). However,
the profit for the latter strategy declines (0.69 % per month) when there is a time lag
between the portfolio formation period and the holding period.
89
Table 4.1: Profitability of momentum Strategies on BSE (Jan.1868Dec. 1913)
Formation Period Portfolio
EQ VW EQ VW EQ VW EQ VW EQ VW EQ VW EQ VW EQ VW
3 Winner 0.81% 0.75% 0.87% 0.79% 0.89% 0.82% 0.90% 0.80% 0.87% 0.79% 0.90% 0.81% 0.92% 0.83% 0.88% 0.78%
(4.32) (4.76) (5.44) (6.10) (6.41) (7.31) (7.10) (7.72) (4.69) (5.12) (5.82) (6.60) (6.83) (7.67) (7.09) (7.57)
Loser
0.53% 0.15% 0.44% 0.11% 0.39% 0.06% 0.37% 0.06% 0.48% 0.12% 0.40% 0.08% 0.35% 0.03% 0.36% 0.06%
(2.54) (0.95) (2.31) (0.71) (2.13) (0.42) (2.20) (0.48) (2.29) (0.71) (2.04) (0.51) (1.88) (0.22) (2.12) (0.48)
WL 0.29% 0.60% 0.44% 0.68% 0.51% 0.76% 0.52% 0.74% 0.39% 0.67% 0.51% 0.73% 0.57% 0.80% 0.52% 0.72%
(1.73) (4.08) (3.40) (5.40) (4.58) (7.56) (5.56) (8.69) (2.37) (4.32) (4.09) (5.93) (5.22) (8.16) (5.78) (8.53)
6 Winner 1.04% 0.98% 1.04% 0.98% 1.00% 0.94% 0.91% 0.85% 1.07% 0.99% 1.04% 0.98% 0.97% 0.91% 0.88% 0.81%
(6.61) (6.86) (7.50) (7.95) (7.72) (8.27) (7.43) (7.79) (6.99) (7.13) (7.58) (7.98) (7.49) (7.92) (7.12) (7.33)
Loser 0.40% 0.05% 0.33% 0.07% 0.33% 0.04% 0.38% 0.04% 0.33% 0.08% 0.30% 0.08% 0.33% 0.02% 0.40% 0.07%
(1.68) (0.26) (1.52) (0.41) (1.69) (0.26) (2.16) (0.25) (1.35) 0.38 (1.35) (0.44) (1.68) (0.16) (2.29) (0.50)
WL 0.64% 1.04% 0.71% 1.06% 0.67% 0.98% 0.53% 0.82% 0.74% 1.06% 0.75% 1.06% 0.64% 0.93% 0.48% 0.74%
(3.75) (5.99) (4.84) (7.44) (5.73) (8.89) (5.37) (8.63) (4.28) (6.44) (5.32) (8.02) (5.78) (9.05) (5.10) (8.29)
9 Winner 1.04% 0.99% 0.99% 0.94% 0.92% 0.87% 0.86% 0.80% 1.03% 0.97% 0.95% 0.90% 0.88% 0.83% 0.83% 0.77%
(7.02) (7.18) (7.25) (7.51) (7.20) (7.46) (7.05) (7.23) (6.89) (7.12) (6.92) (7.22) (6.81) (7.04) (6.72) (6.86)
Loser 0.35% 0.12% 0.37% 0.02% 0.42% 0.07% 0.46% 0.14% 0.33% 0.08% 0.38% 0.03% 0.44% 0.12% 0.48% 0.17%
(1.49) (0.62) (1.77) 0.135 (2.25) (0.44) (2.76) (0.97) (1.43) 0.414 (1.85) (0.19) (2.42) (0.77) (2.91) (1.23)
WL 0.70% 1.11% 0.62% 0.96% 0.50% 0.80% 0.40% 0.67% 0.69% 1.05% 0.57% 0.87% 0.44% 0.71% 0.35% 0.60%
(4.36) (7.25) (4.66) (7.72) (4.63) (7.92) (4.31) (7.41) (4.46) (7.12) (4.43) (7.29) (4.19) (7.25) (3.88) (6.72)
12 Winner 0.94% 0.90% 0.89% 0.84% 0.84% 0.79% 0.80% 0.75% 0.91% 0.87% 0.86% 0.81% 0.81% 0.76% 0.78% 0.72%
(6.17) (6.49) (6.35) (6.57) (6.44) (6.64) (6.44) (6.59) (5.95) (6.26) (6.06) (6.23) (6.16) (6.32) (6.17) (6.28)
Loser 0.47% 0.06% 0.52% 0.15% 0.55% 0.20% 0.58% 0.23% 0.49% 0.12% 0.55% 0.19% 0.58% 0.23% 0.60% 0.25%
(2.05) (0.33) (2.56) (0.87) (3.07) (1.31) (3.49) (1.67) (2.17) (0.63) (2.72) (1.15) (3.23) (1.57) (3.63) (1.88)
WL 0.47% 0.84% 0.37% 0.70% 0.29% 0.60% 0.23% 0.52% 0.42% 0.75% 0.31% 0.62% 0.24% 0.53% 0.18% 0.47%
(2.92) (5.42) (2.81) (5.50) (2.66) (5.65) (2.35) (5.46) (2.71) (5.02) (2.46) (5.04) (2.25) (5.14) (1.89) (4.97)
weighed by their market capital, one month before the holding period. EW= Equally Weighted and VW= Value Weighted.
Panel A: Holding Period Panel B: Holding Period
3 6 9 12 3 6 9 12
This table reports the returns from momentum strategy based on BSE returns data. Each month t, stocks a ranked based on Pmonths compound returns. Qmonths return are
calculated based on the ranking. Equally weighted and value weighted quintile portfolios are formed. The top quintile is called "Winner" portfolio and the bottom quintile
is called "Loser" portfolio. The zerocost (WinnerLoser) portfolio is the portfolio formed by going long on the winner and short on loser portfolios. In Panel A, portfolios are
formed immediately after the portfolio formation and in Panel B they are formed one month after the portfolio formation. Different values of P and Q are shown
in the second row and first column respectively. The monthly average of the three portfolios are reported with their NeweyWest heteroskedastic and autocorrelation adjusted
standard error tstatistics. We adopt the Fama MacBeth break point method in the portfolio formation. W=Winner and L=Loser. For the value weighted portfolios, stocks are
90
All equally weighted strategies yield profits around 0.50%, regardless of the holding
period and of the onemonth lag between the formation period and holding period. It is
out of the ordinary to note that the positive average returns of the equally weighted loser
portfolios are usually significant (except the 6/3, 6/6, 6/9, 9/3, 9/6, 6/3/1, 6/6/1, 6/9/1,
9/3/1 and 9/6/1 strategies, which are marginally significant). This shows that the positive
returns of the zero cost portfolios are mostly due to the positive significant returns of the
winner portfolios and positive significant returns of the loser momentum portfolios. In
effect, 19
th
century investors who would find it difficult to go short on loser portfolios
could have profited from the momentum effect by going long on the winner portfolios. In
contrast, the average return of the value weighted loser portfolio for all the strategies is
not significant, may even be negative sometimes.
The strategy that has been extensively researched in the literature is the 6/6 strategy.
After confirming that momentum strategies of all combinations would yield significantly
positive returns on the 19
th
century BSE, the rest of the chapter will focus on the 6/6
strategy (this strategy will be analyzed, and the result will be used to represent the other
strategies that comprise formation and holding periods ranging from three to twelve
months).
4.2.1 Expected Returns and Average Size of Quintile Portfolios
Table 4.2 shows the portfolios‟ average returns (equally weighted and value weighted)
and average market capitalization (price multiplied by the shares outstanding) of the
quintile portfolios. As before, we construct portfolios by adopting the overlapping
method. For equally weighted portfolios, the average returns increase from the losing
91
portfolio to the winning portfolio, but there is a marginal drop in average returns from the
loser portfolio to the second portfolio.
Table 4.2: Average Returns and Average Size of Quintile Momentum Portfolios
In contrast, the average return of the valueweighted portfolios monotonically increases
from the losing portfolio to the winning portfolio. The values in the last column of the
Table indicate that past losing and winning portfolios pick stocks that have low market
capital on the average. This result is consistent with the findings of Jegadeesh and
Titman (1993) and Chabot et al. (2009).
4.2.2 Momentum Profit within Size and Betabased Subsamples
In order to investigate whether the profit from the momentum strategy is not confined to
any particular group of stocks, we follow Jegadeesh and Titman (1993) to examine the
profitability of the strategy on subsamples. We group stocks into subsamples based on
the beta and size. Size may serve as a proxy for liquidity and beta for volatility. The
subsample analysis of the momentum profit also provides evidence about the source of
the momentum profit. It tests whether the momentum profit is due to the crosssectional
Average Market
Capitalization
Portfolio EQ VW (Million Belgium Franc)
Loser 0.33% 0.07% 3.80
2 0.26% 0.11% 9.23
3 0.51% 0.43% 11.76
4 0.76% 0.70% 11.28
Winner 1.04% 0.98% 7.50
WL 0.71% 1.06% 
Average returns of the values weighted quintile portfolios.
weighted. The sample period is January1868 to December 1913. Market capital is in millions.
W=Winner , L=loser, EQ=Average returns of the equally weighted portfolios and VW is the
This table reports the average returns and average market capital of the quintile portfolios
of 6 /6 months momentum strategy. The market capital is the price times the number of
shares outstanding. Average market capital is the average of the holding period market
capitals of the stocks in each portfolio. Quintile portfolios are equally weighted and value
92
variations in expected returns or predictable patterns in the time series of stock or
portfolio returns. If the profit is due to the crosssectional variations in expected returns,
the profit will be reduced in the subsamples, as the dispersion of the crosssection of
expected returns is lower in the subsamples than in the full sample. If the momentum
profit is due to predictable patterns in individual stock returns, then the subsamples will
yield positive profits, as will the entire sample.
Table 4.3: Portfolio Returns of the Momentum Strategies with Size and Beta Subsamples
Table 4.3 presents the average return of momentum portfolios, momentum profits and the
difference in the mean of the full sample profit and each subsample profit. Portfolios are
Portfolio All Micro Small Big β
1
β
2
β
3
Loser 0.33% 1.01% 0.04% 0.09% 0.31% 0.42% 0.53%
(1.52) (3.19) (0.19) (0.69) (1.72) (2.14) (1.80)
2 0.26% 0.47% 0.35% 0.41% 0.32% 0.41% 0.39%
(2.03) (2.17) (3.01) (3.96) (3.89) (3.37) (1.89)
3 0.51% 0.58% 0.60% 0.54% 0.48% 0.64% 0.58%
(4.58) (3.61) (5.22) (5.48) (8.63) (5.56) (3.16)
4 0.76% 0.92% 0.75% 0.70% 0.74% 0.80% 0.77%
(6.85) (5.92) (6.34) (7.61) (8.05) (7.23) (4.36)
Wi nner 1.04% 1.30% 0.98% 0.82% 0.98% 1.06% 1.06%
(7.50) (7.12) (7.07) (7.02) (7.93) (7.69) (5.69)
WL 0.71% 0.29% 1.01% 0.73% 0.67% 0.64% 0.53%
(4.84) (1.17) (7.55) (6.26) (5.09) (4.16) (2.84)
Difference  0.42% 0.30% 0.02% 0.04% 0.07% 0.18%
 (2.86) (3.35) (0.18) (0.32) (0.68) (1.77)
months. The overlapping method is employed in the portfolio formation. Loser/winner
portfolio is the equally weighted portfolio of stocks in the lowest/top quintile when
stocks are ranked over the previous six months returns. We adopt Fama Macbeth method
portfolio formation to estimate betas. The last two rows show the difference in mean
This table reports the average monthly returns of portfolios formed based on size and
beta subsamples. Portfolios are formed based on six month returns and held for six
Average Monthly Returns
sample period is January 1968 to December 1913. W=Winner and L=Loser
subsamples. Size is determine at the beginning of each formation period. We estimate
betas for stock returns prior to the portfolio formation period. Dimson's method one
month lag is applied to stocks with at least 24 month returns within the five years prior to
and their Newey West tstatistics between full sample profit and each subsample profit.The
for all portfolio breakpoints. "Micro" contains the smallest stocks, "Small" contains the next
smallest and "Big" contains the largest stocks. β
1
, β
2
and β
3
the lowest to the highest beta
93
formed based on the 6/6 strategy. The momentum strategy is applied on beta and size
subsamples. Size is determined at the end of each formation period. We estimate beta by
using the Dimson method with a onemonth lag. The analysis includes stocks with at
least 24month returns within five years before the formation month.
Clearly, the momentum strategy is profitable in all subsamples except the “Micro” size
subsamples. This indicates that the profit does not exist for the “Micro” size subsample
of stocks. The values in the last two rows show the average of the difference in profit
between the full sample momentum profits and the subsample momentum profits. The t
statistics for the test of a hypothesis that the difference is equal to zero is reported in
parentheses. The nonexistence of momentum in the “Micro” is mainly due to the extreme
average return recorded by its loser portfolio (average return of 1.01% with a tstatistic of
3.19). Not finding momentum in our micro size stocks support Hong, Lim and Stein
(2000) results. They used NYSE and AMEX data between 1980 and 1996 to document
that momentum does not exist in their smallest size decile stocks. On the contrary, the
result does not support Jegadeesh and Titman (1993) and Chabot et al. (2009), who find
momentum profit in all the small size subsamples. The result also deviates from Fama
and French (2008) who find significantly positive momentum profit for small stocks. As
in our sample, it is positive but not significant. Like Chabot et al. (2009) we do not find a
strong relationship between size and momentum profit. However, it is worthy to note the
high profit (the profit of 1.01% with tstatistics of 7.55) for the “Small” size subsample.
Chabot et al. (2009) also find the largest momentum profit for the middle size group.
Consistent with the previous chapter, the beta subsamples show no significant difference
in profit from the full sample. The “Big” size subsample also shows no significant
94
difference in the mean with the full sample. The difference in momentum profits between
the full sample and the subsamples (except the “Micro” and “Small” size subsamples)
indicates that the crosssectional difference in the expected returns of the stocks may not
determine the momentum profits. If anything, the predictable patterns in individual stock
returns may contribute to the momentum profits in the beta subsamples, as the profit is
not reduced significantly in these groups. It is important to note that the profit of the
“Big” size subsample is mainly due to the buy side of the transaction rather than the sell
side. The expected return of the past winner portfolio is significantly positive, and the
average return of the past loser portfolio is not statistically significant. This implies that
even when investors are not allowed to short stocks, they can still earn a momentum
profit by buying winners.
Seasonality and Subperiod Analysis of the Momentum Profit 4.3
We are motivated to look into the seasonal patterns in momentum profits, as Jegadeesh
and Titman (1993) document positive momentum profits in all calendar months except
January. Jegadeesh and Titman (2001) also confirm the January reversal (negative)
returns in later data. However, the reversal becomes marginal when they exclude stocks
with a price lower than $5 per share and stocks in the smallest decile in their sample.
Therefore, they infer that most of the negative profits recorded in the month of January
by Jegadeesh and Titman (1993) are due to small and low priced stocks, which are likely
to be difficult to trade. Table 4.4 reports the momentum profits in all calendar months on
the Brussels stock exchange between January 1868 and December 1913. This will test
the possible seasonal effect of the momentum profits in the 19
th
century.
95
Table 4.4: Seasonality in momentum profits
Figure 4.3: Average returns of the momentum profit in all calendar months
It also shows the percentage of positive profits in all the calendar months. Again, we
form momentum portfolios based on the 6/6 strategy. The zerocost portfolio is the
winning portfolio minus the losing portfolio return in each month. We also report the
Months Jan. Feb. Mar. Apr. May June July Aug. Sep. Oct. Nov. Dec. FebDec Fstats
All 0.87% 0.94% 0.97% 0.70% 0.76% 0.72% 0.52% 0.49% 0.50% 0.63% 0.65% 0.76% 0.70% 0.41
(4.02) (4.61) (3.95) (2.72) (2.95) (2.73) (1.39) (1.46) (1.73) (2.80) (2.69) (3.96) (8.84)
76 78 74 74 78 80 78 76 72 67 65 67 74
Micro 0.66% 0.39% 0.47% 0.37% 0.52% 0.35% 0.21% 0.17% 0.38% 0.34% 0.16% 0.34% 0.34% 0.11
(2.08) (1.15) (1.18) (0.85) (1.36) (0.84) (0.46) (0.42) (0.96) (0.96) (0.42) (1.03) (3.11)
65 65 63 59 61 63 67 61 59 54 52 50 59
Small 1.07% 1.29% 1.30% 0.89% 0.96% 0.96% 0.87% 0.88% 0.66% 0.75% 0.93% 0.96% 0.95% 0.65
(3.91) (5.45) (4.89) (3.04) (3.42) (3.51) (3.14) (2.97) (2.25) (4.10) (5.37) (4.31) (11.78)
85 85 83 72 85 80 83 80 78 72 87 76 80
Big 0.92% 0.93% 0.90% 0.69% 0.56% 0.56% 0.56% 0.59% 0.62% 0.89% 1.00% 0.94% 0.75% 0.81
(3.81) (4.32) (4.23) (3.28) (2.68) (3.08) (2.68) (3.49) (3.41) (6.40) (5.44) (5.26) (12.18)
85 83 83 76 76 89 85 80 85 89 87 85 83
For all months in the period 18681913, this table reports the average momentum profit, their related tstatistics and the percentage
of momentum profits that are positive in calendar month. The momentum portfolios are formed based on past 6 months returns
and held for 6 months. The equally weighted portfolio is formed from the lowest past return decile and is called Loser. The
equally weighted portfolio returns formed from the past return top rank quintile stocks is called Winner. The zero cost portfolio is
the Winner minus Loser portfolio. The average return, related tstatistic and the percentage of positive profits is also reported for
each size subsample."Micro" contains the smallest stocks, "Small" contains the next smallest stocks and "Big" contains the largest
size stocks. Size is determined at the end of each portfolio formation period. Newey West standard error adjusted tstatistics are in
parenthesis.The last column report the Fstatistics for the test of hypothesis of equal average profits in the various calendar months.
This figure report s t he average ret urns of t he moment um profit by calender mont h.
0.00%
0.20%
0.40%
0.60%
0.80%
1.00%
1.20%
Jan. Feb. Mar. Apr. May June July Aug. Sep. Oct. Nov. Dec.
M
o
m
e
n
t
u
m
P
r
o
f
i
t
s
96
average return, corresponding tstatistics and the percentage of positive momentum
profits in each calendar month for size subsamples. Size is determined at the end of each
portfolio formation period. “Micro” is the lowest tercile group for size, “Small” is the
middle tercile group for size, and “Big” is the highest tercile group for size. We compute
the Fstatistics in the last column under the null hypothesis that the average returns on the
zero cost portfolios are equal in all calendar months.
The results from the 19
th
century BSE do not support the negative momentum profit
found on the USA market by Jegadeesh and Titman (1993) and Jegadeesh and Titman
(2001). The winner portfolio returns exceed the loser portfolio returns in all months. This
is supported by a significant percentage of positive momentum profits in all calendar
months. Specifically, January records the third highest momentum profit (0.87% average
return, with tstatistic of 4.02 and 76% positive profits) after February and March. The
absence of the January effect on 19
th
century BSE is not surprising, as Chabot et al.
(2009) record similar results in almost the same period (Victorian Era) in the UK. They
argue that investors, as of that time, were not taxed on capital gains and that the tax year
does not end in December. The same reason applies to the BSE. The Fstatistics in the
last column indicate that the null hypothesis of equal profit in all calendar months cannot
be rejected for the entire sample and the size subsamples. This shows that there is no
significant evidence of difference in average momentum profit for the calendar months.
As shown in Figure 4.3, the profit decreases gradually from January to August but picks
up again from September to December. Jegadeesh and Titman (1993) also recorded the
lowest profit in August. Comparing the profits in all calendar months for the various size
subsamples, the “Micro” size group records the lowest profits in all months. Table 4.5
97
documents the zerocost portfolio for the 6/6 month strategy in tenyear subperiods and
the last five years before World War I. This table shows that the momentum strategy
yields positive significant profit in all subperiods except in the period 1878 to 1887,
although still positive.
Table 4.5: Subperiod Analysis of Momentum Profit
This table reports the average monthly returns of the zerocost portfolios in ten year subpe
Sample Month 18681877 18781887 18881897 18981908 19081913
All 0.55% 0.19% 1.05% 1.13% 0.59%
(1.91) (0.48) (5.60) (4.89) (4.03)
All Jan. 0.96% 0.29% 0.60% 1.96% 0.33%
(1.88) (0.84) (3.60) (4.55) (3.99)
Feb.Dec. 0.51% 0.18% 1.09% 1.05% 0.62%
(3.96) (0.81) (11.21) (6.54) (6.87)
All 0.17% 0.16% 1.22% 0.30% 0.26%
(0.29) (0.32) (3.48) (0.68) (1.00)
Micro Jan. 0.50% 0.06% 0.87% 1.68% 0.11%
(0.52) (0.08) (2.18) (2.18) (0.57)
Feb.Dec. 0.14% 0.18% 1.25% 0.17% 0.30%
(0.58) (0.72) (7.02) (0.66) (1.73)
All 0.54% 0.43% 1.17% 1.76% 0.87%
(2.58) (1.29) (6.17) (7.88) (9.08)
Small Jan. 0.67% 0.52% 0.76% 2.42% 0.90%
(1.55) (3.16) (2.46) (3.94) (4.65)
Feb.Dec. 0.53% 0.42% 1.21% 1.70% 0.86%
(4.82) (2.09) (9.56) (11.43) (9.86)
All 0.75% 0.31% 0.72% 1.23% 0.83%
(2.57) (1.26) (6.73) (6.13) (7.10)
Big Jan. 1.28% 0.36% 0.34% 1.86% 0.69%
(3.25) (1.73) (0.95) (3.27) (2.66)
Feb.Dec. 0.70% 0.30% 0.76% 1.17% 0.85%
(6.36) (2.09) (11.75) (8.33) (11.45)
riods and the last five years before World War I. The zerocost portfolios are formed based
on six month past returns and held sixm months. We sorts stocks in ascending order based
on past six months returns and equally weighted portfolios are formed from lowest qui
ntile stocks group. This group is called the sell portfolio and equally weighted portfolios
formed from the top quintile group of stocks is called Winner portfolio. The zero cost portfo
lio is the winner minus loser portfolios.The average return of the zerocost portfolio formed
using sizebased subsamples of stocks within subperiods is also reported."Micro" is the sub
sample which contains the smallest stocks and "Big" contains the largest size stocks. Newey
west standard error adjusted tstatistics is reported in parenthesis. The sample period is
Jan. 1968 Dec. 1913.
98
When the strategy is applied to the middle and largest size subsamples, it produces
significantly positive profits in all subperiods except in the periods 1878 and 1887. It is
important to note the significant average returns in January and outside January for
“Micro” size subsamples in 1888 and 1897 period. This shows that there is a momentum
profit for small stocks in this tenyear subperiod. This may counter the assertion that
momentum profit does not exist in small stocks for the entire period (the assertion is not
robust over time). If we base our research on this tenyear period, we may find
momentum profit for all size subsamples, and small stocks will contribute to the greater
part of the momentum profit across the entire sample period. Except for the periods
18681877, 18781887 and 18881897, there is a significant positive profit for January
and outside January in all the periods for the “Small” and the “Big” size subsamples.
Post holding Period Momentum Profits 4.4
The support for shortterm reversal, intermediate term continuation and long run post
holding period reversal of momentum profits has been extensively documented in the
literature (Jegadeesh and Titman (1993), Conrad and Kaul (1998), Jegadeesh and Titman
(2001) and Chabot et al. (2009)). In effect, there seems to be strong evidence of past
losers exceeding past winners in the first month after portfolio formation (shortterm
reversal). In addition, past winners continue to exceed the past losers for two to twelve
months after portfolio formation (intermediate term continuation), and past losers will
reverse to exceed past winners over three to five years. The behavioral explanation of
momentum indicates that the time series' variation in individual stock or portfolio returns
contributes to momentum profit. That is, the time series‟ abnormal holding period return
(momentum) is due to investor delay in overreaction to information, which pushes the
99
prices of winners (losers) above (below) their fundamental values. Upon returning to
their senses, investors force the prices to revert to their fundamental values. However,
Conrad and Kaul (1998) argue that the crosssectional variation in expected returns
generates momentum profit. Empirical evidence presented by Bulkley and Nawosah
(2009) support the assertion. Conrad and Kaul (1998) hypothesize that stock prices
follow a random path with various drifts, and that every stock has a unique drift. They
show that the difference in unconditional drifts across stocks explains the momentum
profits. Since the predictability under this hypothesis is based on the difference in
unconditional drifts across stocks, but not on the variation in the time series of prices for
individual stocks in any particular period, the profits from the momentum strategy will
continue to remain positive in any post formation period. To differentiate between the
behavioral bias hypothesis and the Conrad and Kaul (1998) hypothesis, we examine the
returns of the momentum portfolios, in the periods following the holding periods
considered in the previous sections. That is, if momentum profits are completely due to
behavioral biases, we expect profit to be reduced to zero overtime, and if possible,
reverse its sign. On the other hand, if the profit is due to crosssectional variation in
expected returns, the profit should continue to increase after the formation period with
time.
Table 4.6 presents the average monthly returns for the first five years after portfolio
formation. From Panel A, across all stocks, the winner portfolio returns drop from 0.91%
in the first year to 0.46% in the fifth year. In contrast, the loser portfolio increases from
0.38% in the first to 0.86% in the fifth year, more than double its value in the first year.
The profit in the second year is approximately the same for the winner and the loser
100
portfolios. From the third year to the fifth year, the zerocost (winner minus loser)
portfolio profit becomes significantly negative, as the loser portfolio returns exceed the
winner portfolio in these periods. The strongly significant average returns of 0.31%
from the second through the fifth year confirms the return reversal.
Table 4.6: Long Horizon Momentum Profits
Month Months Months Months Months Months Months
Portfolios 1 112 1324 2536 3748 4960 1360
Panel A: All
0.50% 0.53% 0.02% 0.27% 0.46% 0.40% 0.31%
(2.66) (5.37) (0.23) (2.90) (4.95) (4.29) (6.83)
0.99% 0.91% 0.64% 0.53% 0.46% 0.46% 0.59%
(5.65) (7.43) (4.64) (4.34) (3.72) (3.97) (8.76)
0.50% 0.38% 0.66% 0.80% 0.92% 0.86% 0.90%
(2.03) (2.16) (4.07) (4.54) (5.35) (5.11) (10.09)
0.02% 0.30% 0.23% 0.52% 0.73% 0.59% 0.53%
(0.06) (1.94) (1.70) (3.62) (5.37) (4.61) (7.83)
1.24% 1.13% 0.81% 0.64% 0.54% 0.52% 0.73%
(5.43) (7.15) (4.18) (4.08) (3.50) (3.25) (8.41)
1.26% 0.83% 1.03% 1.15% 1.26% 1.11% 1.27%
(3.41) (3.46) (4.85) (4.95) (5.88) (5.43) (10.79)
called "loser portfolio". Zerocost portfolio is the winner minus loser portfolios. We
repeat the portfolio formation for size subsamples. "Micro" is the lowest size group,
"Small" is the middle size group and "Big", the largest size group. Size is the market
capitalization of stock and is determined at the end of portfolio formation. Newey
This table reports monthly average momemtum profit for zerocost, winner and
loser portfolios, one to five years after portfolio formation. Portfolios are formed
based on past six months returns. Stock are sorted based based on the pased six
months returns and equally weighted portfolio is formed from the lower and the
upper quintile group of stocks. The portflio formed from the upper group of stocks
is called the "winner portfolio" and the portfolio formed from the lower group is
West standard error adjusted tstatistics are reported in parenthesis. The sample
period is Jan. 1868 to Dec. 1913.
Zerocost Portfolio
Winner Portfolio
Loser Portfolio
Winner Portfolio
Loser Portfolio
Panel B: Micro Zerocost Portfolio
101
Figure 4.4: Cumulative Returns for Five years after portfolio formation
It is important to note that there is no shortterm return reversal in the 19
th
century BSE,
as the momentum profit one month after portfolio formation is positively significant
(0.50% with tstatistic of 2.66). Figure 4.4 depicts the cumulative momentum profits for
Table 4.6 Continued
Month Months Months Months Months Months Months
Portfolios 1 112 1324 2536 3748 4960 1360
0.90% 0.77% 0.27% 0.02% 0.25% 0.18% 0.09%
(4.85) (8.02) (2.76) (0.23) (2.46) (1.59) (1.81)
0.82% 0.78% 0.58% 0.48% 0.46% 0.46% 0.53%
(4.64) (6.56) (4.81) (4.13) (3.37) (4.16) (7.91)
0.08% 0.01% 0.31% 0.51% 0.71% 0.63% 0.62%
(0.36) (0.06) (2.09) (3.11) (4.01) (3.54) (7.42)
0.70% 0.61% 0.17% 0.10% 0.13% 0.21% 0.03%
(4.22) (8.27) (2.31) (1.59) (2.04) (2.76) (0.95)
0.82% 0.76% 0.51% 0.50% 0.43% 0.44% 0.50%
(5.03) (7.29) (4.62) (4.61) (4.30) (4.32) (8.28)
0.12% 0.15% 0.34% 0.40% 0.56% 0.65% 0.54%
(0.83) (1.39) (3.14) (3.74) (4.73) (5.09) (7.74)
Panel D: Large Zerocost Portfolio
Winner Portfolio
Loser Portfolio
Winner Portfolio
Loser Portfolio
Panel C: Small Zerocost Portfolio
0.30%
0.20%
0.10%
0.00%
0.10%
0.20%
0.30%
0.40%
0.50%
0.60%
0.70%
0.80%
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53 55 57 59
C
u
m
u
l
a
t
i
v
e
R
e
t
u
r
n
s
Event Months
Figure 4.4: Cumulative Returns for five years after portfolio
formation
102
five years after portfolio formation. For our 1868 to 1913 periods, the graph shows that
the momentum profit increases monotonically from the first month to the seventh month.
The profit declines from the seventh month to the twelfth month, when it falls below its
first month value. As shown in Table 4.6, the momentum profits become negative around
the second year after portfolio formation, but the negative profit is not strong enough to
offset the positive profit recorded in the first year. This keeps the cumulative profits
above zero until the fourth year, in which the negative profits exceed the positive profits.
Therefore, the cumulative profit falls below zero on the fourth year after portfolio
formation. In effect, there is a strong return reversal from the second to the fifth year
after portfolio formation on the 19
th
century BSE. This result is consistent with the
behavioral explanation found on the USA market by Jegadeesh and Titman (2001).
However, similar analysis on size subsamples (Panels B, C and D) shows that the strong
longterm return reversal recorded for all stocks is mainly due to the lower long run
return reversal in the “Micro” size group of stocks. For the “Micro” size stocks in Panel
B, the first month after portfolio formation records an insignificantly negative
momentum profit. The momentum profit one year after portfolio formation is significant
(0.30% with a tstatistic of 1.94). There is a strong return reversal from the second to the
fifth year (0.53, tstatistic of 7.83). From Panels C and D, the first year records a very
strong momentum profit, but the profit reverses from the second year to the fifth year,
and it is not significant (0.09 and 0.03 with tstatistics of 1.81 and 0.95 respectively).
The evidence that the return reversal draws its power from small size stocks cautions us
to interpret the behavioral model explanation of momentum with care. In general, the
result is still not consistent with the argument by Conrad and Kaul (1998) either.
103
The Momentum profit and the Market State 4.5
In this section, we investigate whether the condition of the state of the market can
determine the profitability of the momentum strategy. As indicated in Cooper et al.
(2004), the behavioral theory can be extended to predict differences in momentum profit
across the state of the market. The confidence of a group of investors is expected to be
higher following market gains (Daniel et al. (1998) and Gervais and Odean (2001)). The
upwards adjustment of market prices will tend to be attributed unduly to investor skill.
This will result in greater aggregate overconfidence, as aggregate investors hold the long
positions in the equity market. The high overconfidence in following up markets will
trigger strong overreactions, and it will eventually lead to shortrun momentum. We
follow Cooper et al. (2004) and Chabot et al. (2009) in defining two market states: for
each month t, the UP (DOWN) market is when the lagged threeyear value weighted
market return is positive (negative). We compute average momentum profits across
market states by taking the momentum profit (winner minus loser quintiles) for each
formation month, and taking the average across all formation months that qualify for a
particular market state. Our sample period is still January 1868 to December 1913. For
each month t, we sort stocks into a quintile portfolio based on the past sixmonth return.
We adopt the portfoliooverlapping method in all portfolio formations. We do not skip
one month between the formation period and the holding period, since the previous
section shows no sign of return reversal in the first month after the formation period. The
Fama and MacBeth (1973) breakpoint method is used to form portfolios. The holding
period profits are computed for three horizons: t to t +5, t to t +11 and t +12 to t +59.
104
Table 4.7: The Momentum profit and the Market State
Table 4.7 reports the mean of the momentum profit for both UP and DOWN markets in
the period January 1868 to December 1913. In order to verify whether the number of
months used to define the market state has any influence on the results, we investigate
the profit of the 6/6 strategy on three (one year, two years, three years, etc.) different
definitions of the market state.
7
For brevity, we report the UP and DOWN market
momentum profit for the threeyear definition of market state. In Panel A, for the 6 and
12 months holding periods, the UP market momentum profits are significant, and are all
above 0.60%. Cooper et al. (2004) and Chabot et al. (2009) find similar results. The
DOWN market in panel B shows that momentum profit for 6 and 12 month holding
periods are not significantly different from zero, when the three year lagged value
weighted returns index is used to define the market state. To corroborate the
7
The result is not robust when less than 3 years return is used to define the market states.
Market States Definition
N
Mean Momentum profit
tstatistics
market. The sample period is January 1868 to December 1913.
424
0.88%
(7.42)
424
0.65%
(7.61)
markets. NeweyWest tstatistics are in parenthesis. N is the number of observation for each state of the
Months t to t +5 Months t to t +11 Month t +12 to t +59
Market States Definition
Mean Momentum profit
N
Holding Periods
This table reports momentum profits for the UP and DOWN market states. The momentum portfolio is formed
based on past six month returns and held for six months. Each month, stocks are sorted in ascending order
based on past six months returns. Equally weighted quintile portfolios are formed. The First quintile portf
olio is called the loser portfolio and the top quintile is called the Winner portfolio. The profit of the mo
mentum portfolio (winner minus loser quintiles) is cumulated across the holding periods : months t to
t +5, t to t +11 and t +12 to t +59 if t 1 is the end month of the formation period. Market is classified as
UP (DOWN) if the value weighted market index over the period t 36 to t 1 is positive (negative).We
report the mean monthly profits, Panel A and B reports the momentum profit following UP and DOWN
Panel A: UP market states
3years
4.28 4.59 3.65
128
0.15%
(0.37) (0.62) (4.85)
Panel B: Down market states
Panel C: Test of Equality of the Mean Momentum profit (UP=DOWN)
128
3years
0.15%
376
0.27%
(6.07)
128
0.43%
105
unconditional results from Lee and Swaminathan (2000), Jegadeesh and Titman (2001)
and Cooper et al. (2004), we find that UP market momentum profit significantly reverses
over the long run. The average momentum profits are significantly below 0.20% over
the holding period of 13 to 60 months. We also find significant longrun reversal for
DOWN markets. Cooper et al. (2004) find similar results, and they assert that long run
reversals are not solely due to corrections of prior momentum.
In Panel C, we test the hypothesis of equals in the mean of the momentum profit between
the UP and DOWN markets. The hypothesis is rejected at 5% for the fiveyear holding
period horizons considered after portfolio formation. Chabot et al. (2009) did not find a
significant difference in the mean between the UP and DOWN momentum profits.
106
Conclusion 4.6
We investigate the trading strategy that buys past return winners and sells past return
losers (momentum trading strategy) over the period of January 1868 to December 1913.
There is convincing evidence that the momentum strategy is profitable. For instance, the
6/6 strategy that we study in detail here yields an average profit of about 8.52% and
12.7% per annum, for equally and valueweighted portfolios, respectively. Finding
momentum profits in the 19
th
century provide some evidence that the profits found for
the postWorld War II US market are not mainly due to data snooping biases. Detailed
analysis of the 6/6 strategy on beta and size subsamples shows that momentum profit is
not confined to particular beta subsamples. We find no momentum profit for small size
stocks. Except for the small size sample, the momentum profits for beta and large size
samples are not significantly different from the full sample profit.
Investigating the momentum profit for each calendar month shows that the negative
January momentum profit found on the USA market by Jegadeesh and Titman (1993),
(2001) is not in the 19
th
century BSE. January records the third highest momentum profit
relative to the other months of the year. The momentum profit is not robust across our
sample period, as a tenyear subperiod analysis indicates that the profit is not strong in
the first twenty years of our sample. The profit is marginally significant or not significant
in the first and the second ten years respectively. Unsurprisingly, the marginal profit
recorded in the first ten years is mainly due to the middle and the large size groups of
stocks, since the average momentum profit for the small size group is insignificant.
To investigate the source of momentum profit, we resort to the Jegadeesh and Titman
(2001) approach to examine the returns of the momentum portfolios in the postholding
107
period. We study postholding period returns to differentiate between the assertion by
Conrad and Kaul (1998) and behavioral explanations presented by Barberis, Shleifer and
Vishny (1998), Daniel et al. (1998) and Hong and Stein (1999). Conrad and Kaul (1998)
assert that the momentum profit is due to the crosssectional differences in expected
returns, not to the time series' pattern in asset returns. Therefore, the momentum strategy
will continue to yield positive returns in all periods after the holding period. The
behavioral explanations suggest that, if momentum profit is completely due to under
reaction to information, the returns of the momentum portfolio will decay to zero.
Conversely, if the profit is due to lagged overreaction, the momentum portfolio will
possibly reverse in sign before decaying down to zero overtime. Our evidence supports
the Jegadeesh and Titman (2001) results, and suggests that the profit of the momentum
portfolio in the 13 to 60 months after the portfolio formation is significantly negative (
0.31 with tstatistics of 6.83). This evidence supports the behavioral explanations and
clearly rejects the Conrad and Kaul (1998) assertion. Further study of the size
subsamples revealed that the negative profit recorded in the months 13 to 60 after
portfolio formation is mainly contributed by the small groups. This suggests that the
support for the behavioral explanations of the momentum profit should be interpreted
with caution, although the result is still not consistent with the Conrad and Kaul (1998)
assertion.
We also test if momentum profit and longrun reversal in the crosssection of stock
returns depend on the state of the market. Cooper et al. (2004) document that the Daniel
et al. (1998) behavioral theory can be extended to predict differences in momentum profit
across states of the market, like UP and DOWN markets, as aggregate overconfidence
108
should be greater following market gains. For each month between 1868 and 1913, we
define the state of the market as UP (DOWN), if the lagged threeyear returns of the
value weighted market index are positive (negative). The 6 month formation and 6 to 12
month holding period strategies are solely profitable in the periods of market gains, when
three year lagged value weighted returns are used to define the market state. Specifically,
from 1868 to 1913, the average monthly momentum profit for UP market states is
significantly positive (0.88% and 7.42 tstatistic, 0.65% and 7.62 tstatistics for both 6
and 12 months holding periods respectively). In contrast, the DOWN markets in the same
period record insignificant positive momentum profits of 0.15% (tstatistic of 0.37 and
0.62 for both 6 months and 12 months holding periods). We also find that the momentum
profit in the UP market states is reversed in the long term.
109
CHAPTER 5
5 THE COMBINED EFFECT OF DIVIDEND YIELD, SIZE,
TOTAL RISK AND MOMENTUM (18681913 EVIDENCE)
8
In the previous chapters, stock returns do not show a positive relationship with beta
(CAPM) in the 19
th
century BSE data. However, the winners (losers) over the past 3 to
12 months predict the subsequent 3 to 12 month winners (losers) for the 19
th
century BSE
data (momentum). Size seems to have a relationship with average returns in our data.
However, the relationship is completely driven by extremely small stocks, accounting for
about 0.35% of the total market capitalization. In this chapter, we use the 19
th
century
BSE data to study the predictive pattern of stock returns based on size, momentum, total
risk and dividend yield. The next section explains the motivation for repeating size and
momentum and choosing total risk and dividend yield. By studying the predictive power
of these characteristics on stock returns in a new database, we provide outofsample
evidence that may address the datamining critique. The results may also provide
evidence on the combined effect of different characteristics on the crosssection of stock
returns.
Introduction and Literature Review 5.1
The Capital Asset Pricing Model (CAPM) of Sharpe (1964), Lintner (1965) and Mossin
(1966) provides a particularly appealing way to look at the crosssection of expected
returns. The model implies that expected returns are a linear function of the stock‟s beta
coefficient, i.e. the regression slope of the stock‟s return on the market return. It
8
Different versions of this chapter have been presented at conferences: the 14
th
Conference for Swiss
Society for Financial Market Research (SGF), 8
th
April, 2011, Zurich Switzerland and (Blind Reviewed).
Eastern Finance Association 2011 Annual Meetings, 15
th
April, 2011, Savannah Georgia, USA (Blind
Reviewed)
110
measures the security‟s systematic risk, which is the only part of its total risk that should
be rewarded in the market. Although initially the model was corroborated empirically
(e.g. Fama and MacBeth (1973)), its prediction is not borne out in more recent empirical
research, as summarized in Fama and French (1992). Not only is there no longer a
positive relation between beta and average stock returns, the returns are also found to be
related to other firm or stock characteristics, such as size (market capitalization) and its
booktomarket ratio. Later research adds even more characteristics that are associated
with average returns. For instance, Fama and French (2008) dissect the return patterns
based on momentum, asset growth, profitability, net stock issues and accruals, as well as
size and booktomarket.
These patterns beg explanation. One possibility is that the characteristics are proxies for
exposure to common risk factors, which then leads to the question of which factors
should be studied and how to measure them. Another explanation is that they reflect
irrational investor behavior that a rational investor could exploit. This then raises the
issue of to what extent these patterns are tradable, as the patterns are often stronger for
smaller and less liquid stocks and exploitation involves high portfolio turnover. For
instance, Hanna and Ready (2005) demonstrate the vulnerability of some patterns to
turnoverinduced transaction costs. Finally, the patterns could be illusory or simply found
thanks to collective data dredging (see e.g. Schwert (2003)), in which case the patterns
should not be discernable in new datasets.
In this chapter, we shed some light on this issue by studying crosssectional patterns in
Belgian stock returns for the period of 18681914. This unique and highquality data set
provides a genuine outofsample testing environment. In chapter 1, we tested the
111
predictive power of the beta coefficient for average stock returns. We find that, similar to
the post1969 U.S. results of Fama and French (1992), the relation in preWorld War I
Belgian data is flat. We pay particular attention to the computation of beta, as 19
th
century stock markets were less liquid than their modern counterparts, but to no avail.
Also, within the largest group of stocks, representing on average about 96% of stock
market capitalization, beta and size are of no importance in explaining average returns.
We now turn our attention to the test of the presence of some repeated and other
characteristic effects, namely size, momentum, total risk and dividend yield. We repeat
size in order to investigate whether the relationship between stock returns and the other
firm characteristics will not be confined to the extremely small stocks. In addition, Fama
and French (1992) show that booktomarket and size are the most important stock
characteristics related in the crosssection to average returns. As we do not have
accounting information, we cannot include the booktomarket ratio, but we do include
the size (market capitalization) of the firm in our analysis. Given the evidence that the
size effect seems to have attenuated since the publication of Banz (1981) (see Schwert
(2003)), additional robustness analysis, using a new and independent dataset may shed
light on the explanation of the size effect. Horowitz et al. (2000) list three potential
explanations: (a) data mining; (b) the increased popularity of passive investments, which
would have driven up prices of large companies; (c) the awareness of investors after
publication of the research results has eliminated the profit opportunities. As index funds
did not exist in the 19
th
century, finding a size effect would favour the awareness
explanation, whereas not finding a size effect is more consistent with the data mining
argument.
112
Since Jegadeesh and Titman (1993) published their article on momentum, or short to
medium term past returns, this type of stock prediction has received a prominent place in
empirical asset pricing. Although it is still not clear why momentum is positively related
to future average returns, it is complementary to the size and booktomarket effects
(Fama and French (1996)). Also, Jegadeesh and Titman (2001) have updated their
previous results and still consistently observe the momentum effect in the years 1965 to
1998, making it less likely that it is due to data mining. Fama and French (2008) use a
different measure of momentum in their dissection of the anomaly test and find
momentum to be positively related to the average return. We follow Fama and French
(2008) to repeat momentum in the analysis. Including this characteristic in our tests
yields additional outofsample evidence.
Furthermore, there has recently been a debate on the usefulness of idiosyncratic or total
volatility to predict average returns in the crosssection. Theoretically, Merton (1987)
argues that when investors do not or cannot hold the entire market portfolio due to
various exogenous reasons, idiosyncratic risk and total risk should be priced. On this
note, Malkiel and Xu (2006) find that portfolios of stocks with high idiosyncratic
volatility have higher returns than portfolios with low volatility stocks. In contrast, Ang
et al. (2006) find a negative relation between average returns and idiosyncratic volatility.
In addition, Blitz and Van Vliet (2007) based on Sharpe Ratio and alpha from the Fama
French three factor model document a negative relationship between excess returns and
total volatility on global markets. However, Bali and Cakici (2008) show that the
relationship between total risk and excess return is induced by methodology and the
predominance of small illiquid stocks in the sample. Screening stocks using liquidity and
113
price filters destroys the relation. In addition, in the 19
th
century, stocks were expensive
relative to the average daily wage (see Scholliers (1997)). The above literature measures
idiosyncratic risk as the standard deviation of the errors in the Fama and French (1993)
threefactor model. However, we cannot compute the value factor in the FamaFrench
model as the accounting data has not been digitalized in our data. In addition,
idiosyncratic risk estimated from the standard deviation of the error term in Fama
French‟s threefactor model may introduce errorinvariable complications in the cross
sectional regression.
On this note, we considered total risk to minimize the error inherent in the estimation of
risk. Investors might not be able hold the market portfolio due to the high prices of stocks
relative to the daily wage. Therefore, total risk could also be priced to compensate
rational investors for their inability to hold the market portfolio. Total risk, which is the
main arbitrage risk, may also prevent arbitrageurs from exploiting mispricing
opportunities on the market. Therefore, total risk will have a positive relationship with
mispricing (arbitrage limit theory). Given this debate and the household income situation
in the 19
th
century, we decide to include total risk in our list of characteristics. We follow
Blitz and Van Vliet (2007) to measure total risk as the standard deviation of the past two
to five years excess returns.
Lastly, we would have investigated the relationship between booktomarket value and
the crosssection of stock returns in the 19
th
century BSE, following Fama and French
(1992). However, as we do not have data on the booktomarket value of equity, we
follow Grossman and Shore (2006) and use dividend yield as the best available proxy for
value. In addition, the investigation on the interaction between dividend yield (as a
114
measure of value) and momentum has recently been documented in the literature. For
example, Asness (1997) documents a negative correlation between momentum and
dividend yield. This implies investment in high value stocks, and to some extent entails
investment in poorly performing stocks and vice versa. Similarly, Gwilym, Clare, Seaton
and Thomas (2009) use independent doublesort quintile portfolios to document that the
momentum strategy yields a significant return among the lower value quintile, and the
value strategy is less effective within the highest momentum quintile. They advise value
investors to stay away from high momentum firms until they exhibit at least some
relative strength compared to the general market. On historical data, Grossman and Shore
(2006) also record that a negative relationship exists between dividend yield and
momentum. In view of this, we also investigate the relationship between momentum and
dividend yield in the 19
th
century BSE. Moreover, as it has been shown in the literature
that small firms find it necessary or desirable to pay dividends in the earlier periods, we
also investigate the relationship between size and dividend yield in the 19
th
century BSE.
In addition, we investigate the effect of the relationship between dividend yield and total
risk on the average returns.
The rest of the chapter is organized as follows: The next section describes the
measurement of the characteristics. In section 5.3, descriptive statistics of the
characteristic sort portfolios are studied. Section 5.4 describes the average returns of the
singlesort and independent doublesort portfolios. We use the FM crosssectional
regression method to confirm the sorting result in section 5.5. We conclude the chapter
with section 5.6.
115
Measures of Characteristics 5.2
We follow two paths to establish the importance of the characteristics: portfolio sorting
(mainly singlesort and independent doublesort) and FM crosssectional regressions. To
study the pervasiveness of any pattern, we divide our sample into three groups based on a
characteristic and investigate the effect of dividend yield separately in each group.
As in chapter 3, our sample starts in 1868 and Figure 3.1 illustrates our motivation. As
part of our analysis relies on portfolio sorts (sometimes with double sorting), a minimum
of stocks in the crosssection is needed. Thanks to a change in legislation in 1867, it
became much easier to set up a company, and the number of listed firms increased
accordingly. Portfolios are constructed in January of each year based on information
available by the end of December of the previous year. More specifically, to be included
in the analysis, the stock has to comply with the following requirements:
1. In order to obtain some accuracy in the estimation of total risk, a minimum of 24
out of the 60 months‟ return observations are required.
2. Six months‟ return data in the previous year is necessary to compute our
momentum measure (discussed below).
We use excess return, which is computed as the difference between the realised return
and the riskfree rate used in chapter three.
In order to study the relationship between dividend yield, size, total risk and momentum,
we form yearly portfolios. Unlike Asness (1997), we pay more attention to zerodividend
stocks, as they account for, onaverage, more than a quarter of the stocks in the cross
section each year. Figure 5.1 shows the percentage of stocks in the crosssection that did
not pay dividends each year.
116
Figure 5.5.1: Percentage of Zerodividend paying stocks and their Relative Market Capital: Jan. 1868 Dec.1913
It is clear from the figure that about 45% of the stocks in the crosssection did not pay
dividends in 1880 and or in 1903. On average, about 28% of stocks did not pay
dividends across the period of our study. This tends to be smaller stocks compared with
those that pay dividends. The relative market capitalization of these firms does not
exceed 20% across the period of our study. The result is not surprising, as Gwilym et al.
(2009) and Grossman and Shore (2006) record that small stocks do not pay dividends on
the contemporary and historical UK markets respectively. We treat the zerodividend
stocks as a group, rather than incorporating them into the lowest dividend group of
stocks. We do this because both Keim (1985) and Morgan and Thomas (1998) had a
comparable return between the zerodividend firms and the highest dividend yield firms.
In effect, a “Ushaped” relationship between dividend yield and return. Each year, in
order for a stock to be part of this analysis in the following year, it must have at least 6
months of return data to facilitate the computation of the momentum characteristic.
As before, each year, we measure the size (market capitalization) as price multiplied by
shares outstanding in December of the year before the portfolio formation. Momentum is
0
5
10
15
20
25
30
35
40
45
50
1860 1870 1880 1890 1900 1910 1920
P
e
r
c
e
n
t
a
g
e
year
Relative Market Capitalization
% ZeroDividend Paying stocks
117
measured as the compound return from June to November of the previous year. We are
motivated to use the sixmonth compounded return, as the 6/6 strategy emerges as the
most profitable strategy in chapter 4. Fama and French (2008) rely on the most profitable
strategy in Jegadeesh and Titman (1993) to compute the 12month compound return as a
momentum measure. Total risk is the standard deviation of the past 24 to 60 months
excess returns. Finally, each year, we measure the dividend yield as the sum of all
dividends paid in the past 12 months, divided by the price in December of the year. As
Annaert et al. (2004) observe seasonality in dividend yield, summing dividend over a
full year removes any seasonal patterns in dividend payments, and the current price level
is used to incorporate the most recent information in the stock prices.
Each year, we separately sort stocks into three groups based on size, momentum or total
risk. We use the FM breakpoint method in all stock groupings. The lowest size group is
called „Micro‟ cap stocks. The second group is called „Small‟ cap stocks and the largest
size group is called „Big‟ cap stocks. For total risk and momentum, we place the lowest
total risk (momentum) in group 1 and the highest in group 3. In the same year, we
separately sort stocks based on dividend yield. Zerodividend paying stocks are placed in
one group. The rest is then sorted into three groups based on their dividend yield. The
stocks that pay the lowest dividends are in group 1 and the highest dividend paying
stocks are in group 3. We follow Lakonishok, Shleifer and Vishny (1994) and Fama and
French (1996) to form 12 (for example, Micro/D/P= 0, Micro/1, Micro/2, Micro/3,
Small/D/P=0, Small/1, Small/2, Small/3, Big/D/P=0, Big/1, Big/2, Big/3) sets of
portfolios from the intersections of dividend yield sorts and other characteristic sorts.
These are the independent doublesort portfolios. We repeat the portfolio formations for
118
dividend sorts and the size, momentum or idiosyncratic risk sort. Sorting on two
characteristics will test the marginal effect of each characteristic on average excess
returns.
Descriptive Summary Statistics of the Characteristics 5.3
In this section, we report the summary statistics of the twelve sets of portfolios formed on
sizedividend yield, momentumdividend yield and idiosyncratic riskdividend yield. We
also report the univariate sort on the various characteristics.
5.3.1 SizeDividend yield double sorts
Table 5.1 reports the summary statistics of the characteristics for the 12 portfolios formed
from the intersections of the size sort and dividend yield sort stocks. We also report
separate univariate sort portfolios for size and dividend yield. Between 1868 and 1914,
our sample contains 84 to 518 different firms, which results in an average of 240 stocks
traded every year (see Panel A and Figure 3.1). On average, there are seven or more
stocks in every independent doublesorted portfolio. The number of „Micro‟ stocks that
do not pay dividends in the last 12 months exceeds the number of total stocks that pay
dividends. It is not surprising to see fewer numbers of „Big‟ stocks that do not pay
dividends (seven stocks on the average).
In panel B, the average relative market capitalization of each portfolio is reported. The
first column indicates that the „Big‟ portfolio accounts for the largest percentage of
money invested in the stock market, representing, on average, 81% of total stock market
capitalization. On the contrary, the „Micro‟ portfolio accounts for less than 4% of stock
market capitalization. The remaining stocks represent only 15% of market capitalization.
For the doublesorted portfolios, it is surprising to see more money invested in „Micro‟
119
Table 5.1: Summary statistics for SizeDividend doublesorts
Market D/P= 0 1 2 3 Market D/P= 0 1 2 3
Market 72 56 56 56 8.96 43.96 29.67 17.42
Micro 80 46 8 10 16 3.67 1.52 0.47 0.65 1.02
Small 79 19 19 20 22 15.22 3.10 3.86 3.97 4.28
Big 80 7 30 26 18 81.11 7.00 29.50 26.00 17.67
Market D/P= 0 1 2 3 Market D/P= 0 1 2 3
Market 0 3.31 5.00 8.58 12.65 5.90 5.10 6.44
Micro 5.31 0 3.15 5.05 10.34 11.54 14.58 8.86 6.72 7.59
Small 5.05 0 3.25 5.00 7.62 6.96 9.86 6.65 5.27 6.50
Big 4.50 0 3.37 4.99 7.01 5.06 7.58 4.84 4.39 5.24
Market D/P= 0 1 2 3 Market D/P= 0 1 2 3
Market 0.49 10.30 5.09 2.63 226 1006 893 802
Micro 0.63 2.08 7.75 4.83 2.00 244 134 364 409 364
Small 5.86 5.84 10.22 4.90 2.48 606 311 644 716 725
Big 7.33 7.49 10.66 5.39 3.48 1259 668 1376 1226 1330
Market D/P= 0 1 2 3
Market 2.55 15.69 10.84 6.19
Micro 0.94 0.71 1.29 1.30 1.16
Small 3.92 3.59 4.23 3.96 3.87
Big 20.40 12.29 26.01 19.83 14.25
Panel A: Average Number of stocks Panel B: Relative Market Cap
Panel C: Dividend Yield (%) Panel D : Annual time series average of Total risk(%)
Panel E:Annual time series Momentum(%) Panel F: Annual Average of Price (Belgian Franc)
Panel G:Annual time series average Size (*10^6)
are then used to create a set of 12 portofolios (Micro/D/P= 0, Micro/1, Micro/2, Micro/3, Small/D/P= 0, Small/1, Small/2, Small/3, Big/D/P= 0, Big/1,
Big/2 and Big/3.We compute the annual average number of stocks in each portfolio. The annual average relative market capital of the stocks in each
portfolio is reported. Dividend Yield is measured as the sum of all dividends paid in the year before year t divided by the current price.Total risk is
the standard deviation of the past 24 to 60 months excess returns. Momentum is measured as the compound returns for six months before the
In this table we show summary statistics of some characteristic identified to capture the crosssection of stock returns. At the beginning of January
for each year t, we sort stocks into three groups (Micros, Small and Big) based on their size (price time's shares outstanding) at the end of the
previous year. We seperately sort the stocks in the same year based on their dividend yield in the previous year. No dividend paying stocks are
assigned one group. The rest of the stocks is splitted into three groups in order of magnitude of their dividend yield. The lowest dividend paying
stocks are place in 1 and the highest dividend paying stocks are place in 3. The intersections of the stocks in the sizesort and dividendsort
t .We also report the annual averages for univariate sorted characteristic on the market. The sample period is January 1868 December 1913.
120
zerodividend paying stocks as compared to the various dividendpaying stocks. This
might be due to the high number of zero dividend paying stocks in the micro size group
(see Panel A). Unsurprisingly, less money is invested in zerodividend paying „Big‟
stocks. Among the dividend paying „Big‟ stocks, the average market capital for the
highest dividend paying stocks is lower. For the „Small‟ group of stocks, the zero
dividend paying stocks have slightly lower market capital than the dividend paying
stocks.
Panel C shows the average dividend yield of the univariate sort on size. It also shows the
dividend yield for the 12 sizedividend yield doublesorted portfolios. The first column
shows very small reduction in dividends as size increases. Among the doublesorted
dividend paying stocks, there is a positive relationship between the lowest dividend yield
stocks and size. On the contrary, the relationship is less negative for the middle and the
highest dividend paying stocks.
In panel D, there is a strong correspondence between total risk and size. As the first
column shows, total risk monotonically reduces as size increases. This relationship
persists in all doublesorted sizedividend yield portfolios. It is worthy to note the high
total risk recorded for zerodividend paying stocks. For the univariate sort on dividend
yield, the relationship between total risk and dividend yield is „Ushaped‟ among
dividend paying stocks. The „Ushaped‟ relationship persists in all size groups.
Panel E of Table 5.1 shows the average annual past performance (momentum) of the
univariate size and dividend yield sort portfolios. For univariate size sorts, the first
column shows that size increases with momentum. The positive relationship persists in
121
all columns of sizedividend yield doublesorted portfolios. The zerodividend paying
stocks performed poorly in the past six months, compared to the dividend paying stocks.
The poor performance may be due to „Micro‟ zerodividend stocks, as they record an
average momentum of 2.08%. It is important to note the strong negative relationship
between momentum and dividend yield among dividend paying stocks. This confirms
the Asness (1997) results and is consistent with the phenomena that value stocks are
stocks that performed lower in the past 6 to 12 months, and vice versa.
Panel F reports the annual average price of the size and dividend yield sorted portfolios.
Unsurprisingly, there is a positive relationship between size and price in both single and
double sort portfolios. It is clear that zerodividend paying stocks have the lowest price.
This persists for all size groups. This confirms the results of Grossman and Shore (2006),
who found that stocks that do not pay dividends tend to have lower market capitalization
and price.
Finally, panel G reports the average size (in millions) of the stocks in each portfolio
across the period of the study. It is obvious to observe the zerodividend paying stocks
recording the lowest size in the 19
th
century BSE. Size seems to show a negative
relationship with the dividend yield (single sort among dividend paying stocks). Detailed
analysis of the doublesort portfolios reveals that the “Small” and the “Big” stocks
contribute to the negative relationship.
5.3.2 MomentumDividend yield double sorts
Table 5.2 reports the summary statistics of the characteristics for the 12 portfolios formed
from the intersections of the momentum sort and dividend yield sort stocks. We also
report univariate sort portfolios on momentum, and again on the dividend yield. As in the
122
previous section, between 1868 and 1914, our sample contains 84 to 518 different firms,
which result in an average of 240 stocks traded every year (see Panel A and Figure 3.1).
On average, there are twelve or more stocks in every doublesorted portfolio. The
number of stocks that performed poorly in the last six months (displaying low
momentum) and do not pay dividends in the past 12 months exceed the dividendpaying
stocks in each portfolio. For the middle momentum group, zerodividend stocks are less
than the numbers that pay dividends.
In panel B, the average relative market capitalization of each portfolio is reported. The
first column indicates that the middlemomentum portfolio accounts for the largest part
of money invested in the stock market, representing, on average, 45% of total stock
market capitalization. On the contrary, the lowest momentum portfolio accounts for less
than 22% of stock market capitalization. The remaining high momentum stocks represent
only 33% of the market capitalization. For the doublesorted portfolios, it is not
surprising to see less market capital for zerodividend paying stocks in all momentum
categories. For singlesort dividend paying stocks, the amount of money invested in the
various portfolios decreases as dividend yield increases. This is also true for all
categories of doublesort portfolios. Among the dividend paying portfolios, the middle
momentum portfolios show a higher amount of investment than the other portfolios. On
the contrary, the middle momentum records the lowest amount among the zerodividend
paying portfolios.
Panel C shows the average dividend yield of the univariate sort on momentum. It also
shows the dividend yield for the 12 momentumdividend yield doublesorted portfolios.
The first column shows a slight increase in dividend yield as momentum increases to the
123
Table 5.2: Summary statistics for MomentumDividend yield double sorts
Market D/P= 0 1 2 3 Market D/P= 0 1 2 3
Market 72 56 56 56 8.96 43.96 29.67 17.42
1 80 37 12 12 19 21.62 3.55 7.11 5.97 4.99
2 79 13 20 25 21 45.32 2.35 21.11 14.63 7.23
3 80 22 23 19 16 33.05 3.06 15.73 9.07 5.19
Market D/P= 0 1 2 3 Market D/P= 0 1 2 3
Market 0 3.31 5.00 8.58 11.74 5.31 4.55 5.75
1 4.78 0 3.29 4.99 8.55 9.29 12.96 6.20 5.71 6.98
2 5.45 0 3.33 5.03 7.32 5.95 10.48 5.16 4.50 5.73
3 4.62 0 3.33 4.98 8.56 8.30 13.29 6.56 5.78 6.85
Market D/P= 0 1 2 3 Market D/P= 0 1 2 3
Market 0.49 10.30 5.09 2.63 226 1006 893 802
1 16.11 22.63 9.29 9.25 12.63 468 174 829 830 591
2 2.86 2.77 3.21 2.97 2.60 839 308 1064 942 874
3 26.91 37.08 26.76 18.08 21.20 808 261 1061 911 946
Market D/P= 0 1 2 3
Market 2.55 15.69 10.84 6.19
1 5.49 1.94 11.48 8.75 5.25
2 11.55 3.70 19.74 12.63 7.10
3 8.35 2.86 13.97 9.58 5.84
In this table we show summary statistics of some characteristic identified to capture the crosssection of stock returns. At the beginning of January
for each year t, we sort stocks into three groups (1, 2 and 3) based on their momentum ( compound returns in prior sixmonth ) at the end of the
previous year. The lowest momentum stocks are assigned group 1 and the highest assigned group 3. We seperately sort the stocks in the same
year based on their dividend yield in the previous year. No dividend paying stocks are assigned one group (D/P=0). The rest of the stocks is spl
itted into three groups in order of magnitude of their dividend yield. The lowest dividend paying stocks are place in 1 and the highest dividend
paying stocks are place in 3. The intersections of the stocks in the momentumsort and dividendsortare then used to create a set of 12 portofolios
(1/D/P= 0, 1/1, 1/2, 1/3, 2/D/P= 0, 2/1, 2/2, 2/3, 3/D/P= 0, 3/1, 3/2 and 3/3).We compute the annual average number of stocks in each portfolio.
The annual average relative market capital of the stocks in each portfolio is reported. Dividend Yield is measured as the sum of all dividends paid in
the year before year t divided by the current price. Total risk is the standard deviation of the past 24 to 60 months excess returns. Momentum
is measured as the compound returns sixmonths before the t .We also report the annual averages for univariate sorted characteristic
on the market. The sample period is January 1868 December 1913.
Panel A: Average Number of stocks Panel B: Relative Market Cap
Panel C: Dividend Yield (%) Panel D : Annual time series average of Total risk(%)
Panel E:Annual time series Momentum(%) Panel F: Annual Average of Price (Belgian Franc)
Panel G:Annual time series average Size (*10^6)
124
middle portfolio. Dividend yield increases slightly with momentum in the lowest double
sorted dividend paying stocks. On the contrary, the relationship is „cupshaped‟ and „U
shaped‟ for the middle and the highestdividend paying portfolios.
In panel D, the second column shows a very high total risk for the zerodividend paying
portfolios. This is sometimes almost twice the value of the total risk for the dividend
paying portfolios in each momentum category. Surprisingly, the total risk for the middle
momentum portfolios is the lowest. The total risk for high and low momentum portfolios
is almost the same for all dividend yield groups. This is surprising, as one would expect
the stocks that performed well in the past to have lower total risk than stocks that
performed poorly in the past.
Panel E of Table 5.2 shows the average annual past performance (momentum) of the
univariate and doublesort momentum and dividend yield portfolios. As in the previous
section, for the univariate sort, it is important to note the strong negative relationship
between momentum and dividend yield among dividendpaying stocks. The doublesort
portfolios reveal that the strength of the negative relationship is due to the low and the
middle momentum portfolios.
The average price for the univariate sort and double sort portfolios are reported in panel
F. As before, the prices for zerodividend paying portfolios are low, sometimes onethird
of the price of their dividend paying counter parts (univariate sort). This persists in all
doublesort momentum groups. For the univariate sorts, there is a negative relationship
between average price and dividend yield among dividend paying stocks. The double
125
sort portfolios indicate that the negative relationship pulls its strength from the middle
portfolios.
Finally, we show the average size of the single and doublesort portfolios in panel G.
Clearly, the zerodividend paying stocks are the small size stocks. For the univariate sorts
on dividend yield, the zerodividend paying portfolio has an average size that is less than
that of the dividend paying portfolios. This persists in all doublesort portfolios. The
negative relationship between size and dividend yield in dividendpaying portfolios is
obvious for both single and doublesort portfolios.
5.3.3 Total riskDividend yield double sorts
Table 5.3 reports the summary statistics of the characteristics for the 12 portfolios formed
from the intersections of the total risk sort and dividend yield sort stocks. We also report
the univariate sort for total risk and dividend yield. As in the previous section, between
1868 and 1914, our sample contains 84 to 518 different firms, which result in an average
of 240 stocks traded every year (see Panel A and Figure 3.1). On average, there are five
or more stocks in every doublesorted portfolio. Looking at the rows of the doublesorts,
the number of highrisk zerodividend stocks exceeds (sometimes by 8 times) the
dividend paying stocks. This is not surprising, as the small stocks are more likely the
highrisk stocks (see Table 5.1).
In panel B, we report the relative amount of money invested in each doublesort total
riskdividend yield portfolio. For the univariate total risk sorts (first column), there is a
negative relationship between the amount of money invested in each portfolio and its
risk. Perhaps smaller firms are simply less diversified and therefore have a higher total
risk. The doublesort portfolios show that the negative relationship between the amount
126
Table 5.3: Summary statistics of total risk dividend yield double sorts
Market D/P= 0 1 2 3 Market D/P= 0 1 2 3
Market 72 56 56 56 8.96 43.96 29.67 17.42
1 80 5 24 31 20 61.84 1.47 31.85 21.01 7.51
2 79 15 21 18 25 25.31 2.99 7.90 7.00 7.41
3 80 52 10 6 12 12.85 4.50 4.21 1.65 2.49
Market D/P= 0 1 2 3 Market D/P= 0 1 2 3
Market 0 3.31 5.00 8.58 11.74 5.31 4.55 5.75
1 5.55 0 3.52 4.98 7.03 3.13 3.65 2.84 3.05 3.58
2 6.65 0 3.17 5.03 8.00 6.31 6.71 6.31 6.13 6.19
3 2.68 0 2.99 5.04 9.90 14.06 15.35 12.57 11.07 11.46
Market D/P= 0 1 2 3 Market D/P= 0 1 2 3
Market 0.49 10.30 5.09 2.63 226 1006 893 802
1 3.29 1.86 4.97 3.71 2.25 902 450 1096 917 792
2 4.17 2.49 8.13 6.17 3.00 808 336 949 914 885
3 6.35 1.77 22.53 9.92 2.90 406 170 857 659 534
Market D/P= 0 1 2 3
Market 2.55 15.69 10.84 6.19
1 15.54 5.36 25.17 14.07 7.99
2 6.53 3.85 8.16 7.65 5.91
3 3.32 1.88 7.42 4.20 3.48
divided by the current price. Total volatility is the standard deviation of the past 24 to 60 months excess returns. Momentum is measured
as the compound returns for sixmonths before the t . We also report the annual averages for univariate sorted characteristic on the market.
The sample period is January 1868 December 1913
Panel A: Average Number of stocks Panel B: Relative Market Cap
Panel C: DividendYield(%) Panel D: Annual time series average of Total risk(%)
Panel E:Annual time series momentum(%) Panel F: Annual Average of Price (Belgian Franc)
In this table we show summary statistics of some characteristic identified to capture the crosssection of stock returns. At the beginning of January
for each year t, we sort stocks into three groups (1,2 ans 3) based on their total risk (standard deviation of excess returns) at the end of the
previous year. The lowest total risk group are assigned group 1 and highest group 3. We seperately sort the stocks in the same year based on their
dividend yield in the previous year. No dividend paying stocks are assigned one group. The rest of the stocks is splitted into three groups in
order of magnitude of their dividend yield. The lowest dividend paying stocks are place in 1 and the highest dividend paying stocks are place
in 3. The intersections of the stocks in the sizesort and dividendsort are then used to create a set of 12 portofolios (1/D/P= 0, 1/1, 1/2, 1/3, 2
/D/P= 0, 2/1, 2/2, 2/3, 3/D/P= 0, 3/1, 3/2 and 3/3.We compute the annual average number of stocks in each portfolio. The annual average relative
market capital of the stocks in each portfolio is reported. Dividend yield is measured as the sum of all dividends paid in the year before year t
Panel G:Annual time series average size (*10^6)
127
of investments and total risk is completely due to the dividend paying stocks. In fact, the
relationship is positive for zerodividend paying stocks. A negative relationship is clearly
visible between dividend yield and the amount of investments. This relationship is
mainly due to the low total risk stocks.
In panel C, dividend payment does not show any sign of increasing or decreasing with
risk (column one). This is reflected in all doublesorted portfolios. In Panel D, total risk
does not show any sign of decrease or increase in the univariate sort dividend yield
portfolios. An exception is the total risk of the zerodividend portfolio. The doublesort
portfolios show that the high risk recorded by the univariate sort zerodividend portfolio
draws its power from the highest total risk portfolios.
Panel E reports the past performance of the total riskdividend yield doublesort
portfolios. We also report the univariate sort portfolio characteristics. Momentum
increases as the firm‟s specific risk also increases for univariate sort portfolios. The
positive relationship is reflected in the lowest and the middle dividend yield portfolios.
As in the previous sections, the negative relationship between dividend and momentum
are prevalent in both singlesort and doublesort portfolios (among dividend paying
portfolios). The lowest momentum recorded by the zerodividend paying portfolios is
also apparent.
As before, the price for zerodividend portfolios remains smaller (Panel F). There is a
negative relationship between price and total risk (first column). The negative
relationship persists in zerodividend, low and middle dividend yield portfolios. As
128
before, the negative relationship between price and dividend yield are still present in the
univariate sort. The negative relationship is persistent in all doublesort portfolios.
Finally, size shows a negative relationship with firmspecific risk (Panel G, column one).
The negative relationship continues to exist in all doublesorted portfolios. As in the
previous sections, size decreases as dividend yield increases among the dividend paying
stocks. This relationship persists in all dividends paying doublesort portfolios. The
average size for zerodividend paying stocks is lower compared to their dividendpaying
counterparts.
From the previous sections, there is enough evidence to conclude that zerodividend
stocks had low past performance, small market capitalization, low price and high total
risk in the 19
th
century. This is consistent with the notion that dividends play an
important role in communicating to investors in the early capital markets; they can
assume that the small size and highrisk firms that do not pay dividends are distressed
(see Baskin and Miranti (1997), Cheffins (2006) and Cheffins (2008)). Alternatively,
this may be security for smaller, riskier, lowerpriced and younger firms that may have
chosen not to pay a dividend because they were newly established and had more growth
prospects. Often, growth firms in their early lives have low or zero payouts ratios, with
the hope of returning more earnings back to investors at maturity.
To investigate the robustness of the relation among the suggested characteristics, we
compute the crosssectional correlation matrix for each year. In Table 5.4, we report the
timeseries averages of the bivariate correlations. We perform a hypothesis test on the
time series average to test whether it is significantly different from zero. We also include
129
realized annual returns in the matrix. We indeed retrieve the negative correlations
between total risk on the one hand and dividend yield and size on the other hand. Both
are significantly different from zero at the 1% level. Likewise, size and momentum are
significantly positively related.
Table 5.4: Annual Time Series Average of the correlation between the entire characteristic and Average return
This is also visible in panel E of Table 5.1. It is not surprising to see the positive
relationship between size and dividend yield, as the relationship would be influenced by
the high number of zerodividend stocks (see sections 5.3.15.3.3). The dummy variable
is correlated with all the other characteristics except the average returns. Investigating the
correlations with realized returns in the last line of Table 5.4, we notice that all are small
in magnitude. Only two are significantly different from zero at the 5% level: the negative
Dummy DY In(Size) σ
εi
Mom AR
Dummy
DY 0.609
***
In(Size) 0.503 0.187
** **
σ
εi
0.509 0.268 0.570
*** *** ***
Mom 0.139 0.027 0.157 0.015
** ***
AR 0.001 0.005 0.072 0.037 0.071
** **
log of size, σ
ɛi
= total risk, DY = dividend yield, Mom = momentum. We also report
stocks.
a dummy variable which is 1 for zerodividend stocks and 0 for dividend paying
cance at 1 % , ** = signifincance at 5 %. AR=average returns, In(Size) = natural
Note: This table shows the time series averages of the bivariate annual cross
sectional correlation between the various characteristics define to explain average
returns. We test the hypothesis that, the time series average of the bivariate corre
lation is equal to zero.The significance of the test are in parenthesis. *** = signifi
130
correlation with size and the positive correlation with past returns. We now turn to more
formal testing of any predictive relation between stock characteristics and future returns.
Average Excess Returns on Portfolio Sorts 5.4
In this section, we investigate the relationship between stock characteristics and excess
returns using tercile portfolios. As before, we sort the stocks for each year into tercile
groups based on a specific characteristic (size, momentum and total risk) observed the
previous December. We sort stocks separately based on the dividend yield in the
previous year. We assign zerodividend stocks to one group. The rest of the stocks are
divided into terciles. Twelve sets of portfolios are formed from the intersection of stocks
from the dividend yield sort group and the size, momentum or total sort groups. We then
hold these portfolios for the entire year and compute monthly returns. We compute
portfolio returns either using equal weights or value weights. Both weighting schemas
are complementary: with equal weights, the smaller stocks dominate, whereas the large
caps stand out with value weights. In addition, we form portfolios on univariate sort
characteristics and hold the position for one year. The FM breakpoint is used in all
portfolio formations. Repeating the portfolio formations each month yields 552 portfolio
returns in our sample period. We report the average excess returns and their
corresponding NeweyWest autocorrelation, heteroskedastic standarderror adjusted t
statistics for each portfolio. We also report the difference in average excess returns and
their tstatistics for the highest and the lowest terciles characteristic portfolios (hedge
portfolio). For the dividend yield, we show the difference in average excess returns for
the highest dividendpaying portfolios on the one hand, and the lowest and zerodividend
paying portfolio on the other hand. The results are reported in Table 5.3.
131
Table 5.5: Equal and Valueweighted portfolios excess returns (%) of doublesorted characteristics
Market D/P= 0 1 2 3 30 31 Market D/P= 0 1 2 3 30 31
Panel A: Sorti ng on Si ze and Di vi dend Yi el d
Market 0.27 0.14 0.34 0.26 0.01 0.12 1.33 1.43 3.59 2.19 0.05 1.23
Mi cro 0.47 0.53 0.00 0.53 0.17 0.36 0.11 2.73 2.21 0.02 3.15 1.17 1.58 0.50
Smal l 0.22 0.11 0.21 0.27 0.48 0.62 0.26 1.74 0.48 1.48 2.63 2.00 2.19 1.10
Bi g 0.19 0.00 0.14 0.33 0.22 0.35 0.08 1.91 0.02 1.43 3.09 1.69 2.06 0.74
Bi gMi cro 0.28 0.28 0.65 0.07 0.20 0.05 2.20 2.20 2.86 0.34 1.36 0.33
Market 0.01 0.19 0.38 0.42 0.42 0.24 0.04 2.50 4.36 3.03 2.70 1.94
Mi cro 0.16 0.03 0.04 0.53 0.06 0.03 0.01 1.16 0.13 0.25 2.98 0.46 0.17 0.08
Smal l 0.35 0.01 0.26 0.33 0.71 0.70 0.46 2.75 0.06 1.89 3.32 2.43 2.20 1.54
Bi g 0.25 0.09 0.19 0.38 0.33 0.24 0.14 3.15 0.49 2.54 4.18 2.64 1.43 1.37
Bi gMi cro 0.09 0.06 0.14 0.15 0.27 0.83 0.31 0.86 0.93 1.95
EW Mean Monthly Returns(%) tstatistics for EW
VW Mean Monthly Returns(%) tstatistics for VW
dividends in the past one year are assigned one group. The rest of the stocks is split into three groups based on their dividend yield. The stocks in the
intersections of dividend sorts and the Size, Momentum, and Idiosyncratic sorts are grouped to form equally and value weighted portoflios. For the size
and dividend yield sort, we have the following 12 portfolios, Micro/D/P= 0, Micro/1, Micro/2, Micro/3, Small/D/P= 0, Small/1, Small/2, Small/3, Big/D/P= 0,
Big/1, Big/2 and Big/3. Similar portfolio formations are repeated for Momentum and Idiosyncratic risk. Size is measured as the price times the number of
shares outstanding in December of the year before portfolios formations. Momentum is the measured as the compound returns of the stock six month
the portfolios formation year. Total risk is the standard deviation of the past 24 to 60 months excess returns. Dividend yield is the sum of all
dividends paid in the year before the protfolio formation divided by the price at December of the same year. We also report the univariate sort on Divide
nd Yield, Size, Momentum and Total risk. Portfolios are formed annually. We report the NeweyWest Heteroskedastic autocorellation adjsuted
tstatistic for the average excess return for each portfolio. The sample period is Jan. 1868 Dec. 1913.
EW= Equally Weighted, VW= Value Weighted
At the beginning of January for each year t , the 19th centrury BSE stocks are allocated to three groups based on their sorted size (market capitalization),
Momentum and Idiosyncratic Risk in the previous year t1. The BSE stocks are also sorted on dividend yield and grouped into 4. Stocks which do not pay
132
Table 5.5 continued
Market D/P= 0 1 2 3 30 31 Market D/P= 0 1 2 3 30 31
Panel B: Sorti ng on Momentum and Di vi dend Yi el d
Market 0.27 0.14 0.34 0.26 0.01 0.12 1.33 1.43 3.59 2.19 0.05 1.23
1 0.17 0.30 0.02 0.14 0.13 0.43 0.14 1.16 1.39 0.11 0.86 0.72 1.84 0.70
2 0.26 0.18 0.13 0.36 0.42 0.59 0.29 2.86 0.72 1.44 3.67 3.95 2.46 2.93
3 0.60 0.60 0.44 0.62 0.56 0.01 0.12 4.21 2.20 3.35 4.73 4.44 0.06 1.11
31 0.43 0.25 0.42 0.49 0.69 3.05 0.91 2.38 2.78 3.75
Market 0.01 0.19 0.38 0.42 0.42 0.24 0.04 2.50 4.36 3.03 2.70 1.94
1 0.09 0.28 0.21 0.14 0.11 0.39 0.32 0.72 1.53 1.18 0.95 0.42 1.39 1.09
2 0.21 0.22 0.16 0.31 0.36 0.58 0.21 2.86 0.94 2.07 3.31 3.30 2.64 2.03
3 0.56 0.39 0.53 0.60 0.63 0.24 0.10 4.69 1.76 4.13 4.78 4.63 1.25 0.81
31 0.66 0.67 0.74 0.46 0.52 4.90 3.05 3.89 2.64 1.85
Market D/P= 0 1 2 3 30 31 Market D/P= 0 1 2 3 30 31
Panel C: Sorti ng on Total ri sk and Di vi dend Yi el d
Market 0.27 0.14 0.34 0.26 0.01 0.12 1.33 1.43 3.59 2.19 0.05 1.23
1 0.29 0.35 0.17 0.36 0.42 0.77 0.25 4.07 1.30 2.39 4.74 2.91 2.69 1.78
2 0.19 0.13 0.18 0.23 0.20 0.07 0.02 1.58 0.68 1.27 1.69 1.61 0.43 0.15
3 0.39 0.40 0.07 0.90 0.06 0.45 0.13 1.92 1.59 0.38 3.47 0.29 1.94 0.61
31 0.10 0.75 0.10 0.54 0.47 0.56 2.17 0.57 2.29 2.31
Market 0.01 0.19 0.38 0.42 0.42 0.24 0.04 2.50 4.36 3.03 2.70 1.94
1 0.21 0.23 0.16 0.31 0.30 0.53 0.15 3.29 0.81 2.68 3.84 1.67 1.65 0.82
2 0.23 0.11 0.23 0.25 0.29 0.18 0.06 1.86 0.60 1.58 1.78 2.24 1.15 0.53
3 0.08 0.14 0.07 0.82 0.04 0.10 0.03 0.43 0.61 0.36 2.99 0.20 0.48 0.15
31 0.13 0.08 0.23 0.51 0.34 0.86 0.25 1.22 2.05 1.44
EW= Equally Weighted, VW= Value Weighted
EW Mean Monthly Returns(%) tstatistics for EW
VW Mean Monthly Returns(%) tstatistics for VW
EW Mean Monthly Returns(%) tstatistics for EW
VW Mean Monthly Returns(%) tstatistics for VW
133
In panel A, average excess returns and tstatistics for the sorting on size and dividend yield
are shown. When sorting on dividend yield across all stocks, there is no obvious pattern in
equally weighted portfolio returns. There is no significant difference in average returns of the
highest and the lowest dividend yield portfolios (12 basis points with tstatistic of 1.23). The
difference in average returns of the zerodividend yield and highest dividend yield portfolio is
not significant.
We find high returns for zerodividend stocks, which is consistent with Grossman and Shore
(2006) and Gwilym et al. (2009). For valueweighted portfolios, there is a monotonic
increase in average excess returns for univariatesort dividend yield portfolios. The difference
in the average excess returns of the highest and the lowest dividend yield portfolios is 24 basis
points, with a tstatistic of 1.94. The profit improves when the investor goes long on the high
dividend yield and short on the zerodividend yield portfolios. The difference in average
excess return increases to 42 basis points with the tstatistic of 2.70. This shows that
weighting stocks by their market capital in portfolio formations, a strategy that goes long on a
highdividend yield portfolio and short on a lowdividend yield portfolio, results in a
significantly positive profit. The result is consistent with the findings for the UK market by
Dimson, Marsh and Staunton (2002) and Grossman and Shore (2006).
For equally weighted doublesort portfolios, among the size groups, „Small‟ and „Big‟ stocks
show a significantly positive difference in the highest and the zerodividend yield portfolios.
Average returns increase monotonically from zerodividend yield stocks to the highest
dividend yield stocks (except the middle dividend paying portfolio in „Big‟ stocks). The
positive significant difference disappears among the dividend yield stocks. It is obvious that
the size effect exists in the crosssection of the stocks. However, the effect obtains its power
from the „Micro‟ and zerodividend yield stocks. These are the securities for distressed firms,
as they have low prices and negative past performance (see Tables 5.1 to 5.3). For value
134
weighted portfolios, within the „Small‟ size groups, the dividend yield shows a strong positive
relationship with the average excess returns. Surprisingly, the positive relationship between
dividend yield and average returns disappear in the „Big‟ size group. It is not surprising to see
the size effect disappear when stocks are valueweighted to form portfolios. This confirms the
result from Chapter 3. The result adds credence to the view that the initial evidence for the
size effect is due to data mining.
In Panel B, doublesort portfolios are formed based on dividend yield and momentum. In
Table 5.1, dividend yield is negatively related to momentum; however, momentum is
positively related to average returns (Chapter 4). In view of this, one would expect
momentum to be an additional predictor of returns holding the dividend yield constant, and
vice versa. Panel B confirms momentum as a predictor of future returns. An exceptional case
is the insignificant average excess return, spread between the high momentum portfolio and
the lower momentum portfolio, for zerodividend yield portfolios. The spread becomes
significant when stocks are valueweighted. In general, there is a strong momentum effect and
a significant dividend yield effect when stocks are valueweighted.
For the doublesort portfolios, the spread between the average excess returns of the high and
low dividend yield portfolios in the momentum tercile groups is not always significant. The
dividend yield only shows a linear pattern with average excess returns in the middle
momentum group for equally weighted portfolios. The average excess return for the hedge
portfolio is 29 basis points, with a tstatistic of 2.93. This value increases to 59 basis points
when the zerodividend yield portfolio is used to compute the hedge portfolio. The positive
relationship between dividend yield and average excess return in the middle momentum
portfolios is robust to the weighting scheme used to form portfolios. The relationship still
exists for valueweighted portfolios. The notable exception is the negative spread in average
excess returns for the lowest and the highest momentum groups in equally weighted
135
portfolios, which becomes positive when stocks are valueweighted. Surprisingly, the value
weighted portfolios average excess return spread for the largest momentum stocks is not
significant (the average excess return spread of 10 basis points with a tstatistic of 0.81). The
results show that among dividend yield stocks, the relationship between dividend yield and
average excess return in the middle momentum group is strong. However, the relationship
becomes stronger when zerodividend yield stocks are considered. This finding supports those
of Gwilym et al. (2009), who found a strong relationship between dividend yield and average
excess returns among dividendpaying stocks in their quintile momentum portfolios. On zero
dividend yield stocks, Grossman and Shore (2006) have shown that stocks that do not pay
dividends and have performed poorly in the past have very high future returns, while stocks
that do not pay dividends and have a high past return have lower future returns. We record
opposite results on the 19
th
BSE when stocks are equally weighted, and are even stronger
when stocks are valueweighted. This shows that, for stocks that do not pay dividends, past
performance is most likely to determine future returns.
We now turn to total risk. The evidence presented in panel C does not show clear patterns.
There does not seem to be a monotonous relation between total risk and average returns,
neither in the entire market, nor across the dividend yield quartile, regardless of how returns
are weighted. When focusing on the hedge portfolios, the only difference that is statistically
significant is the middle dividendyield portfolio, where high total risk translates into high
average excess returns. The difference is 54 and 51 basis points per month, 2.29 and 2.05
standard deviations from zero for equally weighted and valueweighted portfolios,
respectively. Our result does not support Blitz and Van Vliet (2007), who argue based on
Sharpe Ratio and alpha to confirm a negative relationship between total risk and average
excess returns on international markets. However, the results confirm Bali, Cakici, Yan and
Zhang (2005), which show that the positive relationship between total risk and expected
136
returns documented by Goyal and SantaClara (2003) is driven by small stocks (traded on
NASDAQ, in their case). Looking across dividend yield groups, there is no statistically
significant result in the hedge portfolios, be it equally weighted or value weighted. However,
we repeat that the absence of a monotonous pattern across the total risk portfolios in the entire
market, as well as the lack of any significant results for the high dividend yield portfolios,
indicates that total risk is not a pervasive characteristic of price in the 19
th
century BSE
market. Generally, there is no total effect.
The CrossSectional regressions 5.5
From Table 5.4, it is clear that our characteristics are crosssectionally correlated. In order to
study the marginal effects of the characteristics, we resort to the FM crosssectional
regression method.
Every month, we do a regression of the crosssection of individual stock excess returns on the
characteristics proposed to explain the average excess return. We update characteristics
annually in January, based on information available at the end of December. The fact that
characteristics are recomputed every year introduces some time variations. The full usage of
the available individual stocks ensures the maximum utilization of information about the
crosssectional behavior of individual stocks, which might have been lost when portfolios are
formed. We run the crosssectional regression equation of the form
(13)
where is the excess return with , the return on the individual stocks in the month
t, and is the shortrate used as a proxy for the riskfree rate. DY
j,t1
is the dividend yield
estimated in the previous year.
(price times shares outstanding) is the size of a stock
in December of the year before, and is the total risk measured as the standard deviation
( )
0 1 1 2 1 3 1 4 1 5 1
DY Size Mom Dum
c
¸ ¸ ¸ ¸ o ¸ ¸ ,
÷ ÷ ÷ ÷ ÷
÷ = + + + + + +
, , , , ,
ln ,
jt ft t t j t t j t t j t t j t t j t jt
R R
jt ft
R R ÷
jt
R
ft
R
, 1
Size
j t ÷
, 1 j t c
o
÷
137
of the past 24 to 60 months excess returns.
is the momentum estimated as the
compound gross return from June to November of the year before portfolio formation. Dum
j,t
1
is the dummy variable which takes the value 1 for zerodividend yield stocks and 0 for
dividend paying stocks. From Figure 5.1, the number of zerodividend stocks constitutes
about 28% of the stocks in the crosssection each year. To eliminate the effect of zero
dividend stocks on the dividend yieldexcess return relationship, one has to delete these stocks
or add a dummy variable. The dummy variable takes a value of 1 when a stock does not pay
dividends, and 0 for all other stocks. Deleting the zerodividend stocks has the disadvantage
of reducing the crosssectional information on stocks. The advantage of using the dummy
variable is to allow full usage of the crosssectional information on individual stocks. In
addition, using the dummy variable allows a direct measurement and a test of significance of
the difference in behavior of the excess returns in the zerodividend stocks.
is the
vector of regression coefficients and the regression error. We also estimate the regression
with subsets of the characteristics. The estimation of the crosssectional regression every
month yields 552 time series for regression coefficients. The timeseries average of the slope
coefficients is the estimated premium earned for the different exposures. The averages are
tested for statistical significance using heteroskedastic autocorrelation corrected standard
errors. We use the Newey and West (1987) correction with T
1/4
lags.
9
The average of the coefficients and their corresponding NeweyWest standard error adjusted
tstatistic (parenthesis) is reported in Table 5.6. In Panel A, model 1 confirms the positive
(negative) relationship between momentum (size) and average excess returns. The importance
of size is consistent with the post WWIIUSA evidence presented by Fama and French
(1992). The significance of the size premium is also consistent with our equally weighted
9
We used , where T is 552, the number of months in our sample.
, 1
Mom
j t÷
( )
0 5
, ,
t t
¸ ¸
jt
,
( )
1 4
int T
138
portfolio sorts. This is not surprising, as the regression observations are unweighted, giving
undue importance to the „Micro‟ size stocks. We return to this issue in the next section. Total
risk does not show a significant relationship with average returns (average
Table 5.6: CrossSectional Regression of Excess Returns on Dividend Yield, Size, Total Risk and Momentum
Dividend Total Dividend
Model Intercept Yield Size Risk Momentum Dummy
1 1.88% 0.15% 0.51% 0.71%
(2.74) (4.13) (0.38) (2.21)
2 1.81% 0.51% 0.15% 0.63% 0.69%
(2.66) (0.37) (4.10) (0.47) (2.14)
3 2.58% 1.64% 0.18% 1.13% 0.54% 0.41%
(3.63) (0.82) (4.81) (0.82) (1.63) (2.60)
Subperiods
Jan.1868Dec. 1877 0.08% 6.23% 0.04% 4.35% 0.03% 0.89%
(0.05) (1.83) (0.55) (1.06) (0.03) (2.46)
Jan. 1878Dec. 1887 6.22% 4.80% 0.34% 0.49% 0.83% 0.27%
(2.85) (0.68) (3.08) (0.16) (1.04) (0.61)
Jan. 1888Dec. 1897 1.64% 0.21% 0.18% 1.60% 1.78% 0.16%
(1.31) (0.06) (2.66) (0.61) (3.08) 0.55
Jan. 1898Dec. 1907 3.19% 4.42% 0.25% 0.22% 1.10% 0.58%
(2.66) (1.57) (3.39) (0.11) (2.26) (1.94)
Jan. 1908Dec. 1913 1.68% 2.37% 0.15% 4.65% 0.61% 0.48%
(1.63) (0.75) (2.64) (1.47) (1.35) (1.65)
Panel A
Panel B
is Jan. 1868 Dec. 1913.
This table reports the coefficients of the monthly crosssectional regressions of the excess return on the characte
ristic.The characteristics identified in this regression are dividend yield size, momentum and idiosyncratic risk. We
also consider a dummy variable for dividend paying and zerodividend paying stocks in the last 12 months before
the portfolios formation year. The dummy variable is assigned a value is assigned a value of 1 for zerodividend
paying stocks and 0 for dividend paying stocks. Each year, the characteristics are measured based on the inform
ation available prior to the year. Size is measured as price times number of shares outstanding. Momentum is comp
uted as 6 months compound returns prior to the regression year. Dividend yield is the sum of dividends paid in the
last 12 months dividend by the current price. Total risk is the standard deviation of the past 24 to 60 months excess
returns.The interactions between size, momentum and dividend dummy are considered. NeweyWest autocorrelation
and heteroskedastic adjusted tstatistics are in parentheses.We compute characteristics each year.The sample period
139
coefficient is 0.51 with a tstatistic of only 0.38). This confirms the sorting result, as for the
entire market, the average total risk hedge portfolio is 10 basis points with a tstatistic of 0.56.
Model 1 in Panel A confirms the positive relationship between past returns and future returns
in the equal and value weight sorts. Adding dividend yield to the regressors in model 2 shows
that the dividend yield does not relate to excess returns. The average coefficient is 0.51% with
a tstatistic of 0.37. This is not surprising, as the crosssectional regression gives equal weight
to all stocks, and it confirms the results for equal weight sorts on dividends. The momentum
and size premiums are still significant in model 2. Including the dividend yield and its dummy
in regression model 3 reveals a negative insignificant relationship between dividend and
average excess returns. However, the dummy variable has a negative significant relationship
with average returns. This shows that there is a negative significant difference in average
returns between the zerodividend and dividend paying stocks. Specifically, zerodividend
paying stocks have an average excess return less than the dividend paying stock. The value 
0.41 is significant at a 5% level. The intercepts in models 1 to 3 are significantly different
from zero, and it is far greater than the riskfree rate.
As a further check on the form of these relationships, we run the regression in model 3 on five
(four tenyear and a fiveyear subperiods) nonoverlapping subperiods between 1868 and
1914. The result in Panel B shows that most periods support the overall results in model 3.
The relationship between size and average excess returns does not persist in all subperiods.
Size is not related to excess returns in the first tenyear period. Momentum seems to pull its
power from the last twentyfive years of the study. Dividend yield does not show a significant
relationship in any of the subperiods. The effect of the dummy variable is not always
significant in the subperiods. Table 5.6 confirms the result from the sorting method for
equally weighted portfolio formations.
140
5.5.1 Pervasiveness of the CrossSectional Relationships
From a broad economic point of view, for a characteristic to be accepted and to explain the
crosssection of the returns, its effect should be marketwide. The danger with the cross
sectional regression approach is that illiquid, small stocks drive the results, as the observations
are not weighted. Recent evidence shows that some characteristics are not pervasive. For
example, Bali and Cakici (2008) documents that the relationship between total volatility and
the expected stock returns are not robust, but depend on the (i) weighting scheme used to
compute average portfolio returns, (ii) the frequency of the return data used to estimate total
volatility and, (iii) the breakpoints used to sort stocks into quintiles. Likewise, Horowitz et al.
(2000) indicate that when firms with less than $5 million in value are excluded, the size effect
is considerably reduced and becomes statistically insignificant. Using equally weighted
portfolios and ignoring transaction costs therefore overstates the size effect. This seems to be
consistent with Schwert (2003). He reports that the abnormal performance of the Dimensional
Fund Advisors (DFA) US 910 Small Company Portfolio, which invests in the two lowest
decile of stocks by market value, has been close to zero since 1982. Also, Fama and French
(2008) report that the asset growth anomaly is prevalent in their socalled small and micro
stocks.
We investigate whether our results suffer from this caveat by separately running the Fama
MacBeth analysis for our „Micro‟ and large („Small‟ plus 'Big‟) stocks. Table 5.7 reports the
time series averages of the monthly crosssectional regression coefficients. First, the evidence
for size is not consistent across size groups (model 1). It appears that the microcap group,
which represents, on average, only 3.67% of market capitalization (Panel B), drives the
negative relationship between size and average return.
Unsurprisingly, total risk does not have a relationship with average excess returns in the large
stocks (Panel A). In model 2, introducing the dividend yield in the regressions further reduces
141
the importance of total risk. It is still negative for the large stocks, but is statistically
insignificant. These conclusions do not change when the dividend yield dummy is included in
the regression. As an expectation, dividend yield has a consistent significant positive
relationship with average excess returns in the large stocks. The average coefficient of the
dummy variable is no more significant.
Table 5.7: CrossSectional regression of Excess Returns on Dividend Yield, Size, Total Risk and Momentum of Size
subsamples
Panel B confirms that the size effect pulls its strength from the „Micro‟ stocks. The strength of
the dividend yield effect and the momentum effect is concentrated in the large stocks, which
Dividend Total Dividend
Model intercept Yield Size Risk Momentum Dummy
3 0.86% 0.06% 2.46% 2.23%
(1.13) (1.54) (1.42) (6.68)
4 1.15% 5.55% 0.05% 1.41% 2.12%
(1.63) (2.04) (1.36) (0.95) (6.53)
5 1.42% 8.36% 0.03% 1.41% 1.98% 0.09%
(2.16) (1.64) (0.99) (0.88) (6.06) (0.38)
1 9.02% 0.64% 0.09% 0.05%
(4.66) (4.68) (0.04) (0.10)
2 8.96% 0.59% 0.63% 0.00% 0.01%
(4.62) (0.28) (4.59) (0.00) (0.02)
3 10.27% 4.06% 0.69% 0.74% 0.26% 0.57%
(5.06) (1.35) (4.84) (0.36) (0.51) (2.16)
paying stocks. Each year, the characteristics are measured based on the information available prior to the year.
Size is measured as price times number of shares outstanding. Momentum is computed as 6 months compound retur
ns prior to the regression year. Dividend yield is the sum of dividends paid in the last 12 months dividend by the
current price. Total risk is the standard deviation of the past 24 to 60 months excess returns.The interactions between
size, momentum and dividend dummy are considered. NeweyWest autocorrelation and heteroskedastic adjusted
tstatistics are in parentheses.We compute characteristic each year.The sample period is Jan.1868 Dec.1913.
Panel A (Small and Big)
Panel B (Micro)
This table reports the coefficients of the monthly crosssectional regressions of the excess return on the characte
ristic without the Micro size stocks. We seperately perform similar analysis on Micro stock.The characteristics iden
tified in this regression are dividend yield size, momentum and idiosyncratic risk.We also consider a dummy vari
able for dividend paying and zerodividend paying stocks in the last 12 months before the portfolios formation year.
The dummy variable is assigned a value is assigned a value of 1 for zerodividend paying stocks and 0 for dividend
142
accounts for about 96% of market capital. The dummy variable contributes its effect in the
„Micro‟ stocks. Table 5.7 confirms the results from the sorting method, as the valueweighted
dividend sort portfolios seem to have the positive relationship with average excess returns.
The result is not surprising, as Elton and Gruber (1983) obtained similar results on the USA
market between the years 1927 and 1976, using NYSE data. Finally, the results for
momentum are quite consistent across size groups. In all models, we find a positive
relationship between past returns and average realized returns. The premium for momentum
significantly ranges from 1.98% to 2.23% (compared to 0.54% and 0.71% from the full
sample regressions).
143
Conclusion 5.6
Since beta fails to explain the crosssection of stock returns in the 19
th
century BSE, we
document the crosssectional relationship between average excess returns and size, total risk,
momentum and dividend yield. We investigate these relationships with completely out of
sample data in the 19
th
and first few years of the 20th century from the Brussels Stock
Exchange. We also investigate the pervasiveness of the crosssectional relationships across
different size groups. We use sorting and crosssectional regressions in the analyses.
Unsurprisingly, we find a significantly negative relationship between size and expected
returns. However, further investigation reveals that the negative significant relationship
between average return and size is completely driven by our „Micro‟ stocks, accounting for
about 3.67% of the market capitalization. There is no consistent pattern to be found for total
risk. The momentum effect, on the other hand, does not exist in zerodividend yield stocks but
strong in dividend paying stocks for equally weighted sorts (average regression coefficient
with tstatistics not less than two standard errors from zero). Momentum can explain average
excess returns of stocks, which account for about 96% of the market capital („Small‟ and
„Big‟ stocks groups).
Dividend yield is negatively related to momentum and each of them is positively related to
excess returns. Further investigation reveals that the relationship between excess return and
dividend yield does not exist among „Micro‟ group, which are mostly zerodividend yield
stocks. The relationship is significant in „Small‟ and „Big‟ stocks.
144
6 CONCLUSION
“The behavior of the aggregate U.K stock markets before World War I is similar in many ways to that
of the modern US markets. The relative size of the market and the relative numbers of large and small
stocks are also similar. However, the cross section of stocks looks quite different in the two samples”
Grossman and Shore (2006)
In this doctoral thesis, we answered the following research questions:
 Is beta (systematic risk) stable, unbiased and robust to outliers in the 19
th
century BSE?
 Does the CAPM provide a good description for expected returns in the 19
th
century BSE?
 Does size affect the crosssection of stock returns on the 19
th
century BSE market?
 Does the momentum effect exist in the 19
th
century? If it exists, what is its source?
 Does total risk predict returns on the 19
th
century BSE?
 Does dividend yield predict returns on the 19
th
century BSE?
 What is the marginal effect of the above characteristics on the crosssection of stock returns?
This doctoral dissertation answered the above questions, thereby contributing to the existing
literature on the crosssectional predictability of stock returns. More specifically, we
examined the robustness of the crosssectional predictability of stock returns. The study used
the 19
th
and the first few years of the 20
th
century Brussels stock exchange data. To this end,
we conducted four empirical studies to answer the above research questions. These studies
covered the assessment of betas (the main inputs in the CAPM), testing of the CAPM and the
effects of size, the presence and source of momentum and the combined effects of size,
momentum, dividend yield and total risk on the crosssection of stock returns.
We found that market model betas for individual stocks were poor predictors of future betas.
Predictability was improved by adjusting betas with Blume and Vasicek‟s autoregressive
methods. Grouping stocks to form portfolios also improved beta stability. On the 19
th
century
145
BSE, nonsynchronous trading effects were not prevalent. This may have been due to the
measurement interval of returns, as monthly returns were used to compute betas. We also
found that the iterative reweighted least square method, which accounts for outliers in the
estimation of beta, produced betas that were not significantly different from the market model
betas, in terms of predictive accuracy. By studying the behavior of betas in the 19
th
century,
we produced sufficient evidence of how betas should be adjusted for instability and bias when
testing the CAPM. Overall, our results also suggested that betas on the 19
th
century BSE were
biased and not stable, as in the postWorld War II period.
In the second study, we examined the validity of the CAPM for the 19
th
century BSE. We also
investigated whether size effects determined crosssectional variation in stock returns. First,
sorting and crosssectional regression methods were used to investigate the relationship
between asset beta and the crosssection of stock returns. We found no relationship between
beta and average excess return for the various estimates of beta. Our results indicated that the
CAPM is not a valid model for capturing crosssectional variations in stock returns on the 19
th
century BSE. The results also confirmed Fama and French (1992) finding that beta has a flat
relationship with the crosssection of average returns in the US market. Second, we adopted
sorting and crosssectional regression methods to test the crosssectional relationship between
size and average excess return. We found a strong relationship between size and average
excess return on the 19
th
century BSE. To this end, size could have been used to capture the
crosssectional variation of stock returns on the 19
th
century BSE. This would have confirmed
results from Banz (1981), Reinganum (1981), (1983), Chan et al. (1985) and Chan and Chen
(1988) and Fama and French (1992). However, the relationship was not crosssectionally
robust. Detailed investigation confirmed that the size effect drew its power from a small group
of stocks accounting for about 0.35% of market capital. The size effect disappeared when
these stocks were omitted. In effect, we could not rely on size as a crosssectional predictor of
146
returns on the 19
th
century BSE, as it was confined to a group of stocks representing a small
fraction of market wealth. Our results implied that size should not be considered a systematic
proxy for risk, which corroborated Horowitz et al. (2000) findings for the US market between
1963 and 1981. The results also confirmed the findings of Fama and French (2008), who used
US data from 1963 to 2004 to document that the size effect owes much of its power to micro
caps and that it is marginal for small and big caps. In our data, the size effect did not exist for
small and large stocks.
The third study investigated the relationship between the shortterm past performance of
stocks (momentum) and their future shortterm performance. We found that momentum
existed in stocks, which constituted, on average, more than 90% of market wealth. Our
finding that large stocks had momentum showed that momentum could be used as a firm
attribute when predicting returns in the crosssection of the 19
th
century BSE. The presence of
momentum on the 19
th
century BSE also provided evidence that the momentum profit found
in postWWII US and other markets was not due to data snooping bias. We further
investigated the source of momentum profit in the 19
th
century and found that profit reversed
two to five years after portfolio formation. This evidence supported Jegadeesh and Titman
(2001) results, and it is contrary to that of Conrad and Kaul (1998). Therefore, the cross
section of expected returns cannot explain the momentum profit. However, investigating the
momentum profit in size subsamples showed that postholding period reversal was mainly
due to “Micro” size stocks. Therefore, the Jegadeesh and Titman (2001) behavioral
explanation of momentum profit should be interpreted with caution. Extensions of behavioral
theory have postulated that investors‟ aggregate overconfidence is high following market
gains. This compelled us to test whether momentum profit in the crosssection of stocks
depended on the state of the market. We found that momentum depended on the state of the
market when a threeyear lagged return was used to define the state of the market.
147
Dependence of momentum on market state for the 19
th
century BSE added to the behavioral
explanation of momentum profit. This result was not surprising, as it corroborated Cooper et
al. (2004) and Chabot et al. (2009) results on the contemporary US and Victorian Era UK
markets, respectively.
In the last study, we examined the combined effects of size, momentum, total risk and
dividend yield on average returns. We included size to investigate whether the other effects
were confined to small and illiquid groups of stocks. Given its importance, momentum was
also included, to investigate the marginality of its effect in the presence of other
characteristics. In addition, we did not compute momentum as we did in the third study (i.e.
did not use the portfoliooverlapping method). Here, we measured momentum as the
compound return of individual stocks over sixmonth periods. We used six months compound
returns, because the six formation and six month holding period strategy has been the most
profitable in previous studies. As stated earlier, following Basu (1983), Rosenberg, Reid and
Lanstein (1985) and Fama and French (1992), we would have added priceearnings ratio or
booktomarket ratio as a measure of value or growth. However, data for the computation of
priceearnings and booktomarket ratios were not available for the 19
th
century BSE. As a
result, dividend and price, which were readily available, were used to compute dividend yield,
which serves as a proxy for value. We considered total risk, as investors might not have been
able to hold the market portfolio, due to transaction cost constraints. Therefore, total risk
could also be priced to compensate rational investors for their inability to hold the market
portfolio. In this case, total risk was useful in explaining the crosssection of stock returns.
We found that the sizeeffect that existed in the 19
th
century was driven by our socalled
“Micro” stocks, of which more than 50% had a zerodividend yield, negative momentum, low
price and very high total risk. Low price, coupled with low momentum, high total risk and no
dividend payment, may point to firms that were distressed. Total risk did not show any
148
consistent relationship with average return. This confirmed Bali and Cakici (2008) finding
that (i) the interval of return measurement used to estimate risk (ii) the portfolio breakpoint
method (iii) the weighting scheme use to compute portfolio returns and (iv) using liquidity
and price filters to screen stocks determines the existence and significance of the cross
sectional relationship between risk and expected returns.
Among dividend paying stocks, momentum was negatively related to dividend yield.
However, each was positively related to excess stock return for our large stocks. These
results were not different from what Asness (1997) and Gwilym et al. (2009) documented
recently in the US and UK markets, respectively.
Based on the predictability of stock market returns, we found several similarities between
contemporary stock markets and the 19
th
century Brussels Stocks Exchange. Specifically,
CAPM was not valid for the 19
th
century BSE, as it is with current markets. At first sight,
there seem to be evidence for the size effect, but the effect disappears when stocks in the
lowest size decile are eliminated on the 19
th
century BSE. The findings on size effect confirm
Fama and French (2008) results, who conjectured that the effect draws its power from micro
stocks on the recent US market. The momentum effect existed on the 19
th
century BSE, as on
contemporary markets. Total risk did not show any consistent relationship with average
excess returns on the 19
th
century BSE. Similar results are found for contemporary markets,
where positive, negative and no relationships have been found between total risk and average
excess return (see Ang et al. (2006), Bali and Cakici (2008), Ang, Hodrick, Xing and Zhang
(2009), Fu (2009)). Dividend yield showed a positive relationship with average returns on the
historical BSE. A similarly positive relationship also existed in the current UK and US
markets (as in Asness (1997) and Gwilym et al. (2009)). The positive relationship between
dividend yield and expected returns shows the presence of value effect on the 19
th
century
BSE as the dividend yield is used to represent value.
149
Although any inference across historical periods and systems must be interpreted with care,
the similarities that we found with the 19
th
century BSE mostly supported the conclusions of
current research on the crosssectional predictability of stock returns. Specifically, not finding
the relationship between expected return and beta in our data, we confirm the doubt placed on
the CAPM in the contemporary markets. In this research, size effect is not market wide but
rather found in small size stocks. This indicates that the effect is not due to market
inefficiency as confirmed in the contemporary markets. Therefore, we rule out size as a
predictor of the crosssection of expected return of securities. Total risk should not be
considered as a predictor of the crosssection of stock returns and limit to arbitrage in our
data. This is because the total risk does not show a consistent relationship with average
returns. Furthermore, returns from the dividend yield effect are not positively related to total
risk.
Our findings on momentum effect, dividend yield effect and the interactive effect of these two
on the 19
th
century BSE revealed the robustness of these characteristics as predictors of the
crosssection of stock returns across time. In view of this, we see momentum and value as the
main predictors of the crosssection of stock returns. Even though the environment in which
the historical market operated has changed so much in relation to the contemporary market,
value and momentum turn out to be the common characteristics that can predict returns.
Back to Chapter 1, we differentiate between characteristics that are consistent predictors of
returns and those that are due to data snooping or statistical artifacts. Specifically, in this
dissertation, we reveal that characteristic such as size, total risk, momentum and dividend
yield are not time series and crosssectionally robust in predicting stock returns.
150
References
Altman, E. I., B. Jacquillat, and M Levasseur, 1974, Comparative analysis of risk measures: France and
the united states, Journal of Finance 14951511.
Amihud, Y., 2002, Illiquidity and stocks returns; crosssection and time series effect, Journal of
Financial Markets 5, 3156.
Amihud, Yakov, and Haim Mendelson, 1986, Asset pricing and the bidask spread, Journal of Financial
Economics 17, 223249.
Ang, Andrew, Robert J. Hodrick, Yuhang Xing, and Xiaoyan Zhang, 2006, The crosssection of volatility
and expected returns, Journal of Finance 61, 259299.
Ang, Andrew, Robert J. Hodrick, Yuhang Xing, and Xiaoyan Zhang, 2009, High idiosyncratic volatility
and low returns: International and further u.S. Evidence, Journal of Financial Economics 91,
123.
Ang, Andrew, Jun Liu, and Krista Schwarz, 2008, Using individual stocks or portfolios in tests of factor
models, working Paper, Columbia University.
Annaert, J., Frans Buelens, and M. De Ceuster, 2004, Equity return estimation for 19th century
belgium, working paper ,University of Antwerp.
Asness, Clifford S., 1997, The interaction of value and momentum strategies, Financial Analysts
Journal 2936.
Bairoch, P., 1982, International industrialisation levels from 1750 to 1980, Journal of European
Economic History 269312.
Bali, Turan G., and Nusret Cakici, 2008, Idiosyncratic volatility and the cross section of expected
returns, Journal of Financial and Quantitative Analysis 43, 2958.
Bali, Turan G., Nusret Cakici, Xuemin (Sterling) Yan, and Zhe Zhang, 2005, Does idiosyncratic risk
really matter?, Journal of Finance 60, 905929.
Banz, Rolf W., 1981, The relation between return and market value of common stocks, Journal of
Financial Economics 9, 318.
Barberis, N., A. Shleifer, and R. Vishny, 1998, A model of investor sentiment, Journal of Financial
Economics 49, 307343.
Bartholdy, J., and A. Riding, 1994, Thin trading and the estimation of betas: The efficacy of alternative
techniques, Journal of Financial Research 241254.
Baskin, J.B., and P. J. Miranti, 1997, A history of corporate finance, Cambridge University Press.
Basu, Sanjoy, 1977, Investment performance of common stocks in relation to their priceearnings
ratios: A test of the efficient markets hypothesis, Journal of Finance 32, 291311.
151
Basu, Sanjoy, 1983, The relationship between earnings yield, market value, and return for nyse
common stocks: Further evidence, Journal of Financial Economics 12, 129156.
Beer, Francisca Marie, 1997, Estimation of risk on the brussels stock exchange: Methodological issues
and empirical results, Global Finance Journal 8, 8394.
Bekaert, Geert, Campbell R. Harvey, and Christian T. Lundblad, 2007, Liquidity and expected returns:
Lessons from emerging markets, Review of Financial Studies 20, 17831831.
Black, F. , M. C. Jensen, and Myron Scholes, 1972, "The capital asset pricing model: Some empirical
tests in jensens(ed.), Studies in the theory of Capital Markets New York: Praeger.
Blitz, David C., and Pim Van Vliet, 2007, The volatility effect, Journal of Portfolio Management 34,
102113.
Blume, Marshall E., 1971, On the assessment of risk, Journal of Finance 26, 110.
Blume, Marshall E., 1975, Betas and their regression tendencies, Journal of Finance 30, 785795.
Bulkley, George, and Vivekanand Nawosah, 2009, Can the crosssectional variation in expected
returns explain momentum?, Journal of Financial and Quantitative Analysis 44, 777794.
Chabot, Benjamin, Eric Ghysels, and Ravi Jagannathan, 2009, Momentum cycles and limits to
arbitrage evidence from victorian england and postdepression us stock markets, (NBER).
Chan, K. C., and Naifu Chen, 1988, An unconditional assetpricing test and the role of frim size as an
instrumental variable for risk, Journal of Finance 43, 309325.
Chan, K. C., Naifu Chen, and D. A. Hsieh, 1985, An exploratory investigation of the firm size effect,
Journal of Financial Economics 14, 451471.
Chan, L. K. C., and J. Lakonishok, 1993, Are the reports of beta's death premature?, Journal of
Portfolio Management 19, 5162.
Chan, Louis K. C., and Josef Lakonishok, 1992, Robust measurement of beta risk, Journal of Financial
and Quantitative Analysis 27, 265282.
Chatterjee, S., and W. Jacques, 1994, An outlier resistant approach to risk estimation Financial
Analysts Journal 6975.
Cheffins, B. R., 2006, Dividend as a substitute for corporate law: The separation of ownership and
control, Washington and Lee Law Review 63, 12731338.
Cheffins, B.R., 2008, Corporate ownership and control: British business transformed, Oxford
Univerisity press.
Cohen, Kalman J., Gabriel A. Hawawini, Steven F. Maier, Robert A. Schwartz, and David K. Whitcomb,
1983, Estimating and adjusting for the intervallingeffect bias in beta, Management Science
29, 135148.
Collins, D. W., J. Ledolter, and J. Rayburn, 1987, Some further evidence on the stochastic properties
of systematic risk, Journal of Business 60, 425448.
152
Conrad, Jennifer, and Gautam Kaul, 1998, An anatomy of trading strategies, Review of Financial
Studies 11, 489519.
Cooper, Michael J., Huseyin Gulen, and Michael J. Schill, 2008, Asset growth and the crosssection of
stock returns, Journal of Finance 63, 16091651.
Cooper, Michael J., Roberto C. Gutierrez, Jr., and Allaudeen Hameed, 2004, Market states and
momentum, Journal of Finance 59, 13451365.
Daniel, K., D. Hirshleifer, and Avanidhar Subrahmanyam, 1998, Investor psychology and security
market under and overreactions, Journal of Finance 53, 18391886.
Dimson, E., and P. Marsh, 1999, Murphy's law and market anomalies, Journal of Portfolio
Management 25.
Dimson, E., and P. Marsh, R.,, 1983, The stability of uk risk measures the problem of thin trading,
Journal of Finance 754784.
Dimson, Elroy, 1979, Risk measurement when shares are subject to infrequent trading, Journal of
Financial Economics 7, 197226.
Dimson, Elroy, Paul Marsh, and Mike Staunton, 2002. Triumph of the optimists. 101 years of global
investment returns (Princeton University Press).
Doukas, J. A., Chansog (Francis) Kim, and Pantzailis Christos, 2010, Arbitrage risk and stock mispricing,
Journal of Financial and Quantitative Analysis 45, 907934.
Eisenbeiss, Maik, Göran Kauermann, and Willi Semmler, 2007, Estimating betacoefficients of german
stock data: A nonparametric approach, European Journal of Finance 13, 503522.
Elton, E. J., and M. J. Gruber, 1983, A simple examination of the emipirical relationship between
dividend yields and deviation from the capm, Journal of Banking and Finance 7, 135146.
Faff, R. W., and H John, 1992, Time stationarity of systematic: Some australian evidence, Journal of
Business Finance and Accounting 19, 253270.
Fama, E. F., and K. R. French, 1992, The crosssection of expected stock returns, Journal of Finance
47, 427465.
Fama, E. F., and K. R. French, 1993, Common risk factors in the return on stocks and bonds, Journal of
Financial Economics 33, 356.
Fama, Eugene F., and Kenneth R. French, 1996, Multifactor explanations of asset pricing anomalies,
Journal of Finance 51, 5584.
Fama, Eugene F., and Kenneth R. French, 2008, Dissecting anomalies, Journal of Finance 63, 1653
1678.
Fama, Eugene F., and James MacBeth, 1973, Risk, return, and equilibrium: Empirical tests, Journal of
Political Economy 81, 607636.
153
Fowler, D. J., and C. H. Rorke, 1983, Risk measurement when shares are subjected to infrequent
trading, Journal of Financial Economics 280283.
Fu, Fangjian, 2009, Idiosyncratic risk and the crosssection of expected stock returns, Journal of
Financial Economics 91, 2437.
Gervais, Simon, and Terrance Odean, 2001, Learning to be overconfident, Review of Financial Studies
14, 127.
Ghysels, E., 1998, On stable factor structures in the pricing of risk: Do timevarying betas help or
hurt?, Journal of Finance 53, 549573.
Gibbons, Michael R., Stephen A. Ross, and Jay Shanken, 1989, A test of the efficiency of a given
portfolio, Econometrica 57, 11211152.
Goyal, Amit, and Pedro SantaClara, 2003, Idiosyncratic risk matters!, Journal of Finance 58, 975
1007.
GregoryAllen, Russell, C. Michael Impson, and Imre Karafiath, 1994, An empirical investigation of
beta stability: Portfolios vs. Individual securities, Journal of Business Finance and Accounting
21, 909916.
Grossman, Richard S., and Stephen H. Shore, 2006, The cross section of stock returns before world
war i, Journal of Financial and Quantitative Analysis 41, 271294.
Gwilym, Owain AP, Andre Clare, James Seaton, and Stephen Thomas, 2009, Dividends and
momentum, The Journal of investing 4247.
Hanna, J. Douglas, and Mark J. Ready, 2005, Profitable predictability in the cross section of stock
returns, Journal of Financial Economics 78, 463505.
Harvey, David, Stephen Leybourne, and Paul Newbold, 1997, Testing the equality of prediction mean
squared errors, International Journal of Forecasting 13, 281291.
Haugen, R. A., and N. L. Baker, 1996, Commonality in the determinants of expected stock returns,
Journal of Financial Economics 41, 401439.
Hawawini, G., and P. Michel, 1979, An assessment of risk in the thinner market, belgium case, Journal
of Economics and Business
Hong, Harrison, Terence Lim, and Jeremy C. Stein, 2000, Bad news travels slowly: Size, analyst
coverage, and the profitability of momentum strategies, Journal of Finance 55, 265295.
Hong, Harrison, and Jeremy C. Stein, 1999, A unified theory of underreaction, momentum trading,
and overreaction in asset markets, Journal of Finance 54, 21432184.
Horowitz, Joel L., Tim Loughran, and N.E. Savin, 2000, The disappearing size effect, Research in
Economics 54, 83100.
Huber, P. J., 2004, Robust statistics, Hoboken, New Jersey, Wiley.
154
Ibbotson, R. G., P. D. Kaplan, and J. D. Peterson, 1997, Estimates of smallstock betas are much too
low, Journal of Portfolio Management 23, 104111.
Jagannathan, R., and Z. Wang, 1996, The conditional capm and the crosssection of expected returns,
Journal of Finance 51, 353.
Jegadeesh, N., 1990, Evidence of predictable behavior of security returns, Journal of Finance 45, 881
898.
Jegadeesh, Narasimhan, and Sheridan Titman, 1993, Returns to buying winners and selling losers:
Implications for stock market efficiency, Journal of Finance 48, 6591.
Jegadeesh, Narasimhan, and Sheridan Titman, 2001, Profitability of momentum strategies: An
evaluation of alternative explanations, Journal of Finance 56, 699720.
Jegadeesh, Narasimhan, and Sheridan Titman, 2002, Crosssectional and timeseries determinants of
momentum returns, Review of Financial Studies 15, 143157.
Keim, D. B., 1985, Dividend yields and stock returns: Implications of abnormal january returns,
Journal of Financial Economics 14, 473489.
Klemkosky, R. C., and J. Martin, D., 1975, The adjustment of beta forecasts, Journal of Finance 1123
1128.
Kothari, S. P., Jay Shanken, and Richard G. Sloan, 1995, Another look at the crosssection of expected
stock returns, Journal of Finance 50, 185224.
Lakonishok, J., and Alan C. Shapiro, 1986, Systematic risk, total risk and size as determinants of stocks
market returns, Journal of Banking and Finance 10, 115132.
Lakonishok, J., A. Shleifer, and R. W. Vishny, 1994, Contrarian investment, extrapolation, and risk,
Journal of Finance 49, 15411578.
Lee, Charles M.C., and Bhaskaran Swaminathan, 2000, Price momentum and trading volume, Journal
of Finance 55, 20172069.
Lesmond, David A., Michael J. Schill, and Chunsheng Zhou, 2004, The illusory nature of momentum
profits, Journal of Financial Economics 71, 349380.
Levy, R., A., 1971, On the short term stationarity of beta coefficient, Financial Analysts Journal 5562.
Lintner, John, 1965, Security prices, risk and maximal gains from diversification, Journal of Finance
20, 587615.
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, 1337.
Lo, Andrew W., and A. Craig MacKinlay, 1990, Datasnooping biases in tests of financial asset pricing
models, Review of Financial Studies 3, 431468.
Luoma, M., T. Martikainen, J. Perttunen, and S. Pynnönen, 1994, Different beta estimation
techniques in infrequently traded and inefficient stock markets, Omega 22, 471476.
155
Malkiel, B. G., and Yexiao Xu, 2006, Idiosyncratic risk and security returns,, University of Texas at
Dallas.
Martin, R. Douglas, and Timothy T. Simin, 2003, Outlierresistant estimates of beta, Financial Analysts
Journal 59, 5669.
McLean, R. David, 2010, Idiosyncratic risk, long term reversal and momentum, Journal of
Quantitative and Financial Analysis 45, 883906.
Merton, Robert C., 1987, A simple model of capital market equilibrium with incomplete information,
Journal of Finance 42, 483510.
Morgan, G., and S. H. Thomas, 1998, Taxes,dividend yields and returns in the uk equity market,
Journal of Banking and Finance 405423.
Moskowitz, Tobias J., and Mark Grinblatt, 1999, Do industries explain momentum?, Journal of
Finance 54, 12491290.
Mossin, Jan, 1966, Equilibrium in a capital asset market, Econometrica 34, 768783.
Newey, Withney K., and Kenneth D. West, 1987, A simple, positive semidefinite, heteroskedasticity
and autocorrelation consistent covariance matrix, Econometrica 55, 703708.
Neymarck, A., 1911. Les finances contemporaines. L'epargne française et les valeurs mobilières 1872
1910 (Felix Alcan, Paris).
Nijman, Theo, Laurens Swinkels, and Marno Verbeek, 2004, Do countries or industries explain
momentum in europe?, Journal of Empirical Finance 11, 461481.
Reinganum, M. R., 1981, A new empirical perspective on the capm, Journal of Financial and
Quantitative Analysis 439462.
Reinganum, M. R., 1983, Portfolio strategies based on market capitalization, Journal of Portfolio
Management 9, 2936.
Reinganum, Marc R., 1981, Misspecification of capital asset pricing: Empirical anomalies based on
earnings' yields and market values, Journal of Financial Economics 9, 1946.
Ritter, J. R., and Navin Chopra, 1989, Portfolio rebalancing and the turnoftheyear effect, Journal of
Finance 44, 149166.
Rosenberg, Barr, Kenneth Reid, and Ronald Lanstein, 1985, Persuasive evidence of market
inefficiency, Journal of Portfolio Management 11, 917.
Rouwenhorst, K. Geert, 1998, International momentum strategies, Journal of Finance 53, 267284.
Rouwenhorst, K. Geert, 1999, Local return factors and turnover in emerging stock markets, Journal of
Finance 54, 14391464.
Scholes, Myron, and Joseph Williams, 1977, Estimating betas from nonsynchronous data, Journal of
Financial Economics 5, 309327.
156
Scholliers, P., 1997, Industrial wage differentials in nineteenthcentury belgium, in H. Van der Wee,
and J. Blomme, eds.: The economic development of belgium since 1870 (Edward Elgar,
Cheltenham).
Schwert, G. William, 2003, Anomalies and market efficiency, in G. M. Constantinides, Milton Harris,
and Rene Stulz, eds.: Handbook of the economics of finance (NorthHolland, Amsterdam).
Shanken, Jay, 1992, On the estimation of betapricing models, Review of Financial Studies 5, 133.
Sharpe, W. F., and G. M. Cooper, 1972, Risk return class of new york stock exchange, Financial
Analysts Journal 28, 4652.
Sharpe, William F., 1964, Capital asset prices: A theory of market equilibrium under conditions of risk,
Journal of Finance.
Subrahmanyam, Avanidhar, 2010, The crosssection of expected stock returns: What have we learnt
from the past twentyfive years of research?, European Financial Management 16, 2742.
Van der Wee, Herman, 1996, The industrial revolution in belgium, in M. Teich, and Roy Porter, eds.:
The industrial revolution in national context. Europe and the USA (Cambridge University
Press, Cambridge).
Van Nieuwerburgh, Stijn, Frans Buelens, and Ludo Cuyvers, 2006, Stock market development and
economic growth in belgium, Explorations in Economic History 43, 1338.
Vasicek, Oldrich, 1973, A note on using crosssectional information in bayesian estimation of security
betas, Journal of Finance 28, 12331239.
157
NEDERLANDSTALIGE SAMENVATTING
In dit proefschrift bestuderen we crosssectionele patronen in aandelenrendementen, gebruik
makende van een volledig onafhankelijke databank gebaseerd op gegevens van de Brusselse
Beurs (BSE: Brussels Stock Exchange) tijdens de 19
de
eeuw en de eerste jaren van de 20
ste
eeuw. Deze dataset laat ons toe de invloed te bestuderen op aandelenrendementen van sterk
variërende omstandigheden in de economische en institutionele omgeving. De tijdsperiode die
deze dataset bestijht, vermijdt mogelijke kritiek op datamining. Met deze data:
 Testen we de validiteit van het CAPM.
 Daarenboven testen we of andere kenmerken een verklaring kunnen bieden voor de
crosssectionele variatie in aandelenrendementen.
Op het moment van het schrijven van dit proefschrift was de boekhoudkundige en
transactiedata van de BSE nog niet gedigitaliseerd voor de 19
de
eeuw. Daarom was het niet
mogelijk de verklaringskracht van boekhoudkundige gerelateerde kenmerken voor
aandelenrendementen te onderzoeken. Daarentegen onderzoeken we of grootte
(marktkapitalisatie), momentum (voorbije korte termijn rendementen), totaal risico
(bedrijfsspecifiek risico) en dividendrendement (als indicator voor de waarde van de activa,
net zoals de booktomarket ratio) de aandelenrendementen crosssectioneel kunnen
voorspellen.
Er werden vier empirische studies uitgevoerd aan de hand van de BSE data uit de 19
de
eeuw
en het begin van de 20
ste
eeuw. In studie 1 behandelen we de beoordeling van de bèta. In
studie 2 testen we de validiteit van het CAPM en het grootteeffect. De derde studie test het
momentumeffect. De gecombineerde impact van grootte, momentum, totaal risico en
dividendrendement op de crosssectionele aandelenrendementen wordt behandeld in studie 5.
158
In de eerste studie focussen we op de beoordeling van de bèta. Bèta is de belangrijkste
inputvariabele van het CAPM. Het is een geschatte variabele die mogelijk gemeten wordt met
een bepaalde statistische fout, waardoor de testresultaten van het CAPM vertekend kunnen
zijn. De instabiliteit, vertekening en nietrobuustheid naar outliers toe van bèta zijn dan ook
belangrijke onderzoekstopics geworden sinds de ontwikkeling van het CAPM. In deze studie
onderzoeken we de relatieve prestatie van de verschillende methodes om de bèta te schatten,
gebaseerd op hun vermogen om de erop volgende bèta te voorspellen. Meer specifiek
vergelijken we de bèta‟s van het marktmodel met bèta‟s die geschat werden door de auto
regressieve technieken van Blume (1971) en Vasicek (1973).
We tonen aan dat de individuele bèta‟s van het aandelenmarktmodel niet stabiel zijn. De
voorspelbaarheid van deze bèta‟s kan worden verbeterd door het vormen van portefenilles
met ten minste tien of meer aandelen. Opvallend genoeg zijn er geen significante verschillen
in de verklarende nauwkeurigheid tussen de Blume en Vasicek aangepaste bèta‟s. Wanneer
er gebruik gemaakt wordt van de Dimsonmethode om de bèta‟s te schatten, blijken een klein
aantal aandelen rendementen te hebben die een voorsprong (lead) of achterstand (lag) habben
op de marktrendementen. Om rekening te kunnen houden met outliers werden iterative
reweighted least square (IRLS) technichen gebruikt om de bèta‟s te schatten. De bèta‟s van
de IRLS zijn klein qua grootte in vergelijking met de bèta‟s van het marktmodel, maar hebben
wel dezelfde verklarende nauwkeurigheid als het marktmodel.
In de derde studie werd er gebruik gemaakt van de sorteermethode en de Fama en MacBeth
(1973) (FM) crosssectionele regressiemethode om te kunnen onderzoeken of het CAPM
geldig is op data van vòòr de Eerste Wereldoorlog. Er werd eveneens getest of het grootte
effect (de neiging van kleine aandelen om hogere rendementen te hebben dan grote aandelen)
reeds in de 19
de
eeuw aanwezig was op de BSE.
159
De sorteermethode en de FM crosssectionele regressiemethode leveren geen empirische
bewijskracht voor het CAPM. We tonen eveneens aan dat de relatie tussen de bèta en de
rendementen varieert doorheen de tijd. De nulhypothese die de gelijkheid van de geschatte
hellingen veronderstelt van de crosssectionele regressie en de abnormale marktrendementen
wordt niet verworpen in de periode 18681893. Daarentegen wordt door gebruik te maken van
de op grootte gesorteerde portfolio‟s (met gelijk gewicht) in de crosssectionele regressie, een
verband vastgesteld tussen bèta en abnormaal rendement met bèta als enige verklarende
variabele (het CAPM is geldig). De voorwaardelijke dubbele sorteermethode die toegepast
wordt door Fama en French (1992) scheidt echter het effect van bèta en grootte op de
verwachte rendementen. Deze methode maakt de gemiddelde crosssectionele helling van de
bèta niet significant, zowel wanneer ze met als zonder de grootte in de regressie wordt
geplaatst.
We vinden dat grootte negatief gerelateerd is aan abnormale rendementen. Er bestaat een
significant negatief verband met grootte (grootteeffect), maar de bèta is niet gerelateerd aan
abnormale rendementen wanneer beide variabelen worden opgenomen in de crosssectionele
regressies. Een gedetailleerde analyse van de data onthult dat het grootteeffect voornamelijk
wordt veroorzaakt door de kleinere aandelen, die meetellen voor slechts 0.35% van de totale
marktkapitalisatie. Wanneer deze kleine aandelen buiten beschouwing worden gelaten wordt
er geen relatie gevonden tussen abnormale rendementen enerzijds en bèta‟s of grootte
anderzijds. Zowel de sorteermethode als de crosssectionele regressie onthult dat het grootte
effect verdwijnt wanneer de aandelen op waarde worden gewogen bij het vormen van de
portefenilles. Samenvattend wil dit zeggen dat het CAPM niet geldig is voor de BSE tijdens
de 19
de
eeuw. Door het schatten van de bèta‟s met het marktmodel en de Dimson en Vasicek
methodes zal het model niet tot stand gebracht worden. Het grootteeffect bestaat, maar is
160
voornamelijk toe te schrijven aan een kleine groep van aandelen die slechts een klein
percentage van de totale marktkapitalisatie vertegenwoordigen.
In de vierde studie onderzoeken we of een momentumstrategie een abnormale winst kan
realiseren op de 19
de
eeuwse BSE. De momentumstrategie houdt in dat aandelen worden
gekocht (verkocht) die sterk (zwak) gepresteerd hebben in de laatste 3 tot 12 maanden. Er
bestaat overtuigend bewijs dat deze strategie winstgevend is op de 19
de
eeuwse BSE. Het
vinden van een momentumeffect in deze periode bevestigt de bewering dat momentum
winsten gevonden op markten in de periode na de Tweede Wereldoorlog, niet enkel toe te
schrijven zijn aan vertekeningen door datasnooping. Een gedetailleerde analyse onthult dat
het momentumeffect niet bestaat in de groep van kleine aandelen. Een bijkomende analyse
met betrekking tot de momentumwinsten in elke kalendermaand toont aan dat de winst
positief was voor elk van deze maanden. Een omkening in januari (January reversal effect)
dat gevonden wordt op naoorlogse Amerikaanse markten kan niet gevonden worden op de
19
de
eeuwse BSE. In feite laat de maand januari zelfs de vierde hoogste momentumwinsten
optekenen in vergelijking met de andere maanden van het jaar. We vinden dat de momentum
winsten niet sterk zijn in de eerste twintig jaar die we bestudeerden.
Teneinde de oorzaak na te gaan van de momentumwinsten hebben we gebruik gemaakt van
de Jegadeesh en Titman (2001)benadering om de rendementen te bestuderen van de
momentumportrfenilles in de postholdingperiode. De momentumrendementen keren om in
het tweede tot vijfde jaar na de holdingperiode. Verder onderzoek toont aan dat deze
omkering voornamelijk veroorzaakt wordt door kleine aandelen. We hebben eveneens getest
of de momentumwinsten en de lange termijn ommekeer in de crosssectie van
aandelenrendementen afhankelijk is van de markttoestand. We vonden dat de 6maanden
formatiestrategie en de 6 tot 12maanden holdingperiodestrategie enkel winstgevend waren
in perioden van marktwinsten.
161
In de vijfde studie onderzoeken we of grootte, momentum, total risico en dividendrendement
de crosssectie van aandelenrendementen kunnen verklaren in de periode 18681913. Grootte
en momentum worden opnieuw in de analyse opgenomen om de standvastigheid en de
gecombineerde impact van deze kenmerken te testen op de crosssectie van abnormale
rendementen. Het sorteren op grootte test eveneens of het effect van de andere
karakteristieken niet enkel van toepassing is op de groep van kleine en illiquide aandelen. Het
totaal risico voor elk aandeel wordt gemeten als de standaardafwijking van de residuen van
het Dimsonmodel voor het schatten van de bèta‟s van de aandelen. In elk jaar is het
dividendrendement van een aandeel gelijk aan de som van alle dividenden betaald tijdens de
laatste 12 maanden gedeeld door de eindejaarsprijs. We onderzochten de standvastigheid van
de relaties over alle dividendrendement, grootte, totaal risico en momentumgroepen heen.
De sorteermethode en FM crosssectionele regressie methodes werden toegepast. We
bevestigen dat grootte een significant negatief verband heeft met abnormale rendementen.
Echter, wanneer deze analyse op groepen van verschillende grootte wordt uitgevoerd, komt
aan het licht dat dit negatieve verband volledig is toe te schrijven aan zeer kleine aandelen die
meetellen voor slechts 3.67% van de totale marktkapitalisatie. Dit bevestigt de bevinding dat
het grootteeffect voornamelijk toe te schrijven is aan de eerste deciel grootteportefenille,
waarvan deze zeer kleine aandelen deel uit maken. We hebben geen consistente verbanden
gevonden voor totaal risico. Momentum toont een consistent positief verband met abnormale
aandelenrendementen, die meetellen voor 96% van de totale marktkapitalisatie. Verder
vonden we een negatief verband tussen dividendrendement en momentum bij dividend
betalende aandelen, maar elk van hen is positief gerelateerd aan gemiddelde abnormale
rendementen. Dividendrendement en momentum zijn positief gerelateerd aan gemiddelde
abnormale rendementen bij onze grote aandelen, die meetellen voor ongeveer 96% van de
totale marktkapitalisatie.
162
Hoewel elke conclusie over historische periodes en verschillende systemen heen met enige
voorzichtigheid geïnterpreteerd moet worden, bevestigen de gelijkenissen met de 19
de
eeuwse
BSE de conclusies van actueel onderzoek naar crosssectionele voorspellingen van
aandelenrendementen. Meer specifiek bevestigen we, door het niet vinden van een verband
tussen verwachte rendementen en de bèta in onze data, de twijfel die rust over het CAPM in
de hedendaagse markten. In dit onderzoek is het grootteeffect niet van toepassing op de
gehele markt, maar eerder op de kleine aandelen. Dit toont aan dat het effect niet te wijten is
aan marktinefficiëntie zoals bevestigd wordt in hedendaagse markten. Daarom sluiten we
grootte uit als een voorspeller van de crosssectie van verwachte rendementen van effecten.
Totaal risico mag niet beschouwd worden als een voorspeller van de crosssectie van
aandelenrendementen en is slechts beperkt tot een arbitrage rol in onze data. Dit is omdat
totaal risico geen consistent verband toont met de gemiddelde rendementen. Bovendien zijn
rendementen van het dividendrendementeffect niet positief gerelateerd aan totaal risico.
Onze bevindingen in verband met het momentumeffect, dividendrendementeffect en het
interactieeffect van beide op de 19
de
eeuwe BSE tonen de robuustheid van deze kenmerken
aan als voorspellers van de crosssectie van aandelenrendementen doorheen de tijd. Afgaande
hierop beschouwen we momentum en waarde als de belangrijkste voorspellers van de cross
sectie van aandelenrendementen. Ook al veranderde de omgeving waarin deze historische
markten actief waren enorm in vergelijk met de huidige marktomstandigheden, waarde en
momentum blijken de gemeenschappelijke kenmerken te zijn die rendementen kunnen
voorspellen.
We maken dus een onderscheid tussen karakteristieken die consistente voorspellers zijn voor
rendementen en deze die toe te schrijven zijn aan datasnooping of statistische artefacten. In
het bijzonder onthullen we in dit proefschrift dat karakteristieken zoals grootte, totaal risico,
163
momentum en dividendrendement niet robuust zijn in tijdzeeksen en crosssecties in de
voorspelling van dividendrendementen.
This research was funded by the Faculty of Applied Economics, from the 2007 to 2011 research grant for the Accounting and Finance Research group.
CrossSectional Predictability of Stock return: PreWorld War I Evidence Lord Mensah, Antwerpen, Belgium, 2011 ISBN: 9789089940445 Printed by: Universitas Antwerpen © Copyright 2011, Lord Mensah All rights reserved. No part of this book may be reproduced or transmitted in any form by any electronic or mechanical means (including photocopying, recording or information storage and retrieval) without permission in writing from the author.
Doctoral Jury
Internal members
Prof. Dr. Jan Annaert (Supervisor) Universiteit Antwerpen, Belgium Department of Accounting and Finance Jan.annaert@ua.ac.abe Prof. Dr. Marc De Ceuster (chair/PhD committee chair) Universiteit Antwerpen, Belgium Department of Accounting and Finance marc.deceuster@ua.ac.be Dr. Frans Buelens (PhD committee member) Universiteit Antwerpen, Belgium Department of Accounting and Finance Fran.buelens@ua.ac.be Prof. Dr. Joseph Plasmans Universiteit Antwerpen, Belgium Department of Economics joseph.plasmans@ua.ac.be
External Members
Prof. Dr. Wim Janssens (PhD committee member) Universiteit Hasselt, Belgium Department of Marketing wim.janssens@uhasselt.be Prof. Dr. Laurens Swinkels Erasmus School of Economics, Rotterdam, Netherlands Vice President, Robeco Investment Solutions, Rotterdam lswinkels@ese.eur.nl Prof. Dr. Patrice Fontaine EUROFIDAI and University of Grenoble 2, France patrice.fontaine@eurofidai.org
thank you very much for the updates you gave me on the data. Frans Buelens. You guided me through the research and thought me how to select the literature to support my arguments. In addition. Secondly. Mark De‟Ceuster. Wim Janssens and Dr. Patrice Fontaine for taking your time to read and discuss my work. Joseph Plasmans. Dr. Laurens Swinkels and Prof. Frans Buelens. Primarily. I would like to express my sincere gratitude to my PhD committee members Prof. Prof. Dr. the success marked in this project depend on the environment I worked in and the people around me. In short. I will say you know the right food combination you will give to your child in order to mature quickly. Dr. I am also thankful to the external doctoral jury members Prof. I may like to express my profound gratitude to everyone who contributed to the success of this project both professionally and personally. Dr. I also appreciate how quick you read my chapters and direct me for changes here and there.ACKNOWLEDGEMENTS In as much as writing a PhD thesis requires independent study and scientific research. In a motherchild scenario. I would like to thank my supervisor Prof. . Your comments and suggestions have been of great value to improve and polish the chapters in this thesis. Jan Annaert for the confidence pose in me. this dissertation would not be what it is now without your suggestions. Dr. encouragement and guidance. As a result. Prof. and giving me the opportunity to be part of the research project. Your comments and suggestions have been of immense value in completing this dissertation. Dr. Dr.
Specifically. I have been blessed with supportive family. Completion of my PhD. I have come to understand that only few people are fortunate to have caring and supporting wife like you. my sincere gratitude goes to the secretariat office for their beaming smiles and friendly services. This allows me to release stress at work for the past four years. I do appreciate your support.Genuinely. lunches and visit to museums in Antwerp. I hope that we will stay in touch for the future. Zita LordMensah and Nhyira LordMensah. I want to thank the department head for the annual outing. the period of writing PhD thesis comes with moments of joy and doubt. However. gives me the opportunity to thank you for the unconditional love and support you have given me. Thank you for making my home a place of endless happiness. Firstly. Lord Mensah . how to handle the stress depends on the people around you after office work. I owe an immense debt to my wife Sefam Ekua LordMensah. since we got married. I have enjoyed every moment at work for the past four years. and I will always remember you. Thank you for your patience and not been on your sight ten to eleven hours every day. I want to thank all my colleagues from the Accounting and Finance department of the University of Antwerp. My final words go to all those who made my education a success by contributing financially. Finally. Second words go to my lovely kids Lakeisha LordMensah. On this note. What you have done for the past four years goes beyond your call to duty. Special thanks to my one and only Uncle Maxford Kwadwo Kwakye who contributed to my senior high school education.
.....................1 INTRODUCTION................. 39 2..............................................................................3.............................................................................................................................. 11 Methodology .2 1........................................ 52 3...3....................................................................................................4 2..................................................... 47 Conclusion...................................................................1 1............................................... 6 1................................................................................................................... 1 The aim of this Dissertation ..........2 1............................. 20 Chapter 2 ..............................1 2......................................................................................4 CHAPTER 2 ................. DATA DESCRIPTION AND SUMMARY ................................... 20 Chapter 3 ........ 37 Blume and Vasicek stability adjustment techniques.................5 2......... 33 Beta Stability ................... 1 Introduction ........................................................ 51 CHAPTER 3 ......... 52 Expected returns of portfolios sorted on betas ................. 52 3 THE TEST OF CAPITAL ASSET PRICING MODEL (CAPM) AND THE SIZE EFFECT IN 19th CENTURY BSE .....................................Table of Contents CHAPTER 1 ............................................1 1.......................................................... 44 Impact of outlying observations on Beta ............................. 25 Introduction and Literature Review ............................................................................................ 25 Beta Coefficient Descriptive Statistics .......................3............................................................................................ 23 1...........................2 Data Description and Methodology ...................................... 25 2 2...........................................................................................................................................................................................2 ..2..........3.......................3 ASSESSMENT OF BETA IN THE 19th CENTURY BSE...........................................1 Introduction and Literature Review ..................... 22 Chapter 5 .................................................................................2.....................3 Summary of the chapters .........6 Beta Bias ................................................................................................... 21 Chapter 4 .......1 1...........................................3 1.............. 14 1.......................................1.............................................................................. 7 The Riskfree rate and the Market Index ...........3............................... 54 3..............................................................................1 2............2 2........................................ 1 1 1...........................
...................................................2 5................................. 125 5.......................................................1 4....... 109 5.............2 5.............................................................2............................................... 80 Momentum Trading Strategies and their Returns ..............2 4.................................3 4......................... 103 Conclusion.........................................................................................1 5......................................................... 143 CONCLUSION.............6 Seasonality and Subperiod Analysis of the Momentum Profit ............ 80 CHAPTER 4 ........ 157 ....................................................................................... 98 The Momentum profit and the Market State ........................................ SIZE................5 Average Excess Returns on Portfolio Sorts ........................................................................................................................................................................................5....... 80 4 4..........................................................................1 3.....................................................................6 6 Conclusion ................................ 59 Expected Returns..........3 Introduction and Literature Review ....... 106 CHAPTER 5 ..................................................................3......3.................................. 94 Post holding Period Momentum Profits ......................................................................... and the Size Effect .......................................................4 4............. 115 Descriptive Summary Statistics of the Characteristics ......4 5..........4 Conclusion.. 144 References ..... 130 The CrossSectional regressions ........................................................................................ 109 5 THE COMBINED EFFECT OF DIVIDEND YIELD....... 91 4......... 66 FamaMacBeth CrossSectional Regressions to Test the Size Effect...................... 118 MomentumDividend yield double sorts .. 136 Pervasiveness of the CrossSectional Relationships ..1 5..... 121 Total riskDividend yield double sorts ................................................................................................................................................................... 109 Measures of Characteristics ..................................1 5.........................2.......3..........3............................................................................................................1 Introduction and Literature Review .... 140 5.........................................................5 4.....................1 3......2 4.......2................................... TOTAL RISK AND MOMENTUM (18681913 EVIDENCE)...3...................................... 118 SizeDividend yield double sorts .......................................... 71 3............3 The CrossSectional Regressions ........... Beta... 79 DOES THE MOMENTUM EFFECT EXIST IN THE 19TH CENTRURY? ........ 90 Momentum Profit within Size and Betabased Subsamples ..... 85 Expected Returns and Average Size of Quintile Portfolios ......................................................................................................... 150 NEDERLANDSTALIGE SAMENVATTING .......3 5.....
....................... 1868.........................................TABLES Table 1............ 68 Table 3........... 34 Table 2................... 50 Table 3................ 69 Table 3... 13 Table 2............4: Subperiod look into estimated slopes and excess market returns ........................................................................7: Average time series slopes and intercept from the FamaMacBeth crosssectional regression: Jan 1868Dec.... 44 Table 2..... 36 Table 2............. 1913 ................................. 1913 ............................7: Dimson Aggregate Coefficient (AC) beta Adjustment ............. 1913 ..... 61 Table 3........................1: Beta coefficient descriptive statistics for the 15 estimated periods ... and Postranking Betas of Decile portfolios formed from preranking betas in Jan........................................3: Weighted average of correlation and Spearman rank order correlation across successive periods ..................................9: Test of equal predictive accuracy between MM and IRLS models......................... 48 Table 2............. 73 Table 3......................................................1: Time Series Mean (%).................................................................................................................5: Beta Estimate and Mean Excess Return for the BSE equally weighted size portfolios................................. 65 Table 3............................................................................................. 40 Table 2..............8: Average Time Series Slopes and Intercepts from the FamaFrench CrossSectional Regressions: Jan..............2: Average time series slopes from the FamaMacBeth CrossSectional Regressions in Jan..................................................Dec....................................................................3: Average Time Series Slopes from FamaFrench CrossSectional Regression in Jan.......................................... 1913 ......... 58 Table 3.......... Standard Deviation (%)...................... 38 Table 2..4: Measurement of regression tendency of estimated beta coefficient for individual stocks.................................. 42 Table 2................................................ 76 ............... 64 Table 3.........................5: Predictive performance of Blume and Vasicek (Bayesian) procedures of estimating beta ........................................................6: Modified DieboldMariano test statistics (pvalue in parentheses)........................ 1868Dec....................................................6: Equally weighted portfolios excess returns without the firstsize decile group ........ 1913 ..............2: Average beta and average coefficient of determination of the size based subsamples ................................................................................................................................ 1868Dec............... 1913 ........................................ 1868Dec.. 45 Table 2............ 1868Dec......................................... Jan..........................8: Comparison of the market model betas and the iterative reweighted least square betas ...............1: Summary Statistics for Riskfree rate and the ValueWeighted Index .........
.. 95 Table 4...................................................1: Profitability of momentum Strategies on BSE (Jan............1868Dec..................................... Total Risk and Momentum ...dividend yield double sorts . 91 Table 4.................................3: Portfolio Returns of the Momentum Strategies with Size and Beta Subsamples .........3: Summary statistics of total risk...........7: The Momentum profit and the Market State .................... 36 Figure 2..... 1868Dec........... 100 Table 4................9: Average Time Series Slopes and Intercepts from the FamaFrench CrossSectional Regressions without the FirstSize Decile: Jan... 126 Table 5........ 55 ....................................1: The graph of the average beta of each period for large stocks and small stocks .................................... 89 Table 4.................................................................................................................................................. 10 Figure 1... 129 Table 5.......................... 131 Table 5............... 12 Figure 2.......................................5: Equal and Valueweighted portfolios excess returns (%) of doublesorted characteristics ...4: Seasonality in momentum profits ................................1913.................................................................................................................. 123 Table 5......................... 141 FIGURES Figure 1................ Total Risk and Momentum of Size subsamples .. 104 Table 5.......................................................... Size..............5: Subperiod Analysis of Momentum Profit ...............2: Average Returns and Average Size of Quintile Momentum Portfolios ......... 97 Table 4............................................................. 92 Table 4............................. 78 Table 4...................2: Total market capitalization of the BSE ..............................6: CrossSectional Regression of Excess Returns on Dividend Yield......... Size.............1: Number of listings on the 19th Century Brussels Stock Exchange ......................... 138 Table 5......................7: CrossSectional regression of Excess Returns on Dividend Yield...........3: Evolution of the value weighted market index in the two subperiods of our study ........................................................4: Annual Time Series Average of the correlation between the entire characteristic and Average return ...................... 1913) ..................2: Plot of average market model betas and IRLS betas for stocks with outlier observation less than 4 ................................................................1: Number of stocks in our selection criteria for the entire period of the preworld war I SCOB data ....................................................... 9 Figure 1.............................6: Long Horizon Momentum Profits ........................................1: Summary statistics for SizeDividend doublesorts ...................Table 3.......2: Summary statistics for MomentumDividend yield double sorts ............... 119 Table 5................ 49 Figure 3.............
..3: Average returns of the momentum profit in all calendar months ............. 95 Figure 4......4: Cumulative Returns for Five years after portfolio formation ............... 101 Figure 5...................................................... 86 Figure 4........................................2: Time line of sample periods ..................1: Percentage of Zerodividend paying stocks and their Relative Market Capital: 18681913...............Figure 3....................................................................................................... 116 ..........................1: Number of common stocks in our sample for the momentum studies.......2: Sixty months moving average of the crosssectional slopes and excess market returns using Dimson beta estimates ...................... 70 Figure 4.................... 86 Figure 4......................................................5........................................................................ 65 Figure 3.........................................................3: Size Portfolio betas......
use a comprehensive set 1 of controls and discern whether the results survive simple variations in methodology .CHAPTER 1 1 INTRODUCTION. However. If the presence of an anomaly truly indicates the inadequacy of the CAPM. DATA DESCRIPTION AND SUMMARY More than fifty variables have been used to predict returns. inefficiencies in the market (profit opportunities) or a data snooping bias. Either the anomalies represent inadequacies in the CAPM. 2010. Avanidhar. because more needs to be done to consider the correlation structure among the variables.1 Introduction The issue about why stock returns differ in the crosssection from one another at a particular time has been a hot topic of financial research for the past decades. 2742. 1. since the 1980s. The model postulates a linear relationship between expected returns and the covariance between the market portfolio returns and the returns of an asset (beta). then factors other than beta can predict stock returns. While many have accepted the factors other than beta to predict returns. 1 . The overall picture remains murky. documentation on deviations from the model (anomalies) has been an extremely active area of research. This implies that anomalies would continue to exist before and after their discovery. others believe that they were discovered out of luck and are due to data snooping bias (Lo and MacKinlay 1 Subrahmanyam. European Financial Management 16. Lintner (1965) and Mossin (1966) seems to provide an adequate description of the crosssection of stock returns until the 1980s. This implies that the asset beta with respect to the market portfolio is sufficient to determine its expected returns. The crosssection of expected stock returns: What have we learnt from the past twentyfive years of research?. The capital asset pricing model (CAPM) proposed by Sharpe (1964).
Many studies have tried to differentiate between the two possibilities. or attenuate after they have been documented and analyzed in the academic literature. 2 . or by making predictions using periods that are new to asset pricing research. but most of them concentrate on the postWorld War I data. Haugen and Baker (1996) find some degree of commonality in the characteristics that are most important in determining comparative expected returns among different stocks. Because of this. As a result. testing the asset pricing models on a new data set provides an obvious way to distinguish between the data snooping hypothesis and the persistence of the characteristics identified to predict returns. but they would have no statistical significance outside the sample from which it was discovered. This dissertation fills these gaps in the literature by introducing an independent data set and a different set of characteristics to test the crosssectional predictability of stock returns. usually limited to the USA. The situation remains murky. Schwert (2003) documents that anomalies often seem to disappear. as more needs to be done to distinguish between the data snooping hypothesis and the persistence of the characteristics other than beta. The problem can be addressed by using data from markets that have not been searched exhaustively.(1990)). Indeed. any anomaly found might appear to be valid within the data set. The possible ways that data snooping bias can occur are (i) when researchers continuously test the properties of a data set or the outcomes of other studies on data set (ii) form predictive models based on the characteristics of the previous results and (iii) test the power of their models on the same data set. This is to test whether the anomaly exists in a new and independent sample. reverse. On the international front.
The negative relationship 3 . Models that include market frictions (illiquidity). due to various reasons such as transaction costs. Ang. Jensen and Scholes (1972). In effect. Xing and Zhang (2006) find a negative relationship between idiosyncratic risk and expected returns. However. accruals. return reversal. These are: Theoretical motivation based on riskreturn model variants. dividend yield. Informal Wall Street wisdom (such as value and size investing) Predictors originated from the behavioral biases of investors. asset growth. Investor‟s inability to hold market portfolios will force them to care about the total risk. Subrahmanyam (2010) categorizes the origins of the predictive characteristics based on four principles. etc. Fama and French (1992) find that their data do not appear to support the CAPM. On the contrary. not simply the market risk.Several characteristics have been studied in the literature to predict stock returns. there is a positive relationship between idiosyncratic risk and average returns. Malkiel and Xu (2006) document that the CAPM may not hold when some investors are not able to hold the market portfolio. Sharpe and Cooper (1972) and Fama and MacBeth (1973) find support for the CAPM in the 70s. Black. The traditional CAPM argues that only market risk should be incorporated into asset prices and command compensation. profitability. illiquidity. total risk (idiosyncratic risk). On the contrary. idiosyncratic risk would be priced in the market. as implied by the CAPM. Hodrick. price earnings ratio. Among them are size (market capitalization: measured as price times shares outstanding). Hence. book to market value ratio. momentum. net issues. The theoretically motivated risk/return models conduct a test to see whether a higher return has been associated with higher risk (measured as beta).
He partially based his assertion on the notion that recommending stocks based on price/earnings ratio is common on the Wall Street. Basu (1977) documents the negative relationship between price/earnings ratio and abnormal returns. Similarly. Banz (1981) documents that stocks with low market capitalization outperforms those with high market capitalization. for high idiosyncratic risk stocks.between idiosyncratic risk and expected returns remains debatable. they show the CAPM is not supported in their data. such stocks are more likely to trade at a price far from their fundamental values. Doukas. The Wall Street wisdom characteristics are just found by chance or motivated by informally appealing to the knowledge of finance professionals. which is consistent with the limited arbitrage argument. Idiosyncratic risk may also limit arbitrageurs from exploiting mispricing opportunities on the market. as a specific example. For example. They document the importance of size and book/market value in the crosssection of expected stock returns. hence. In addition. since no theoretical framework has been established to determine the source. Fama and French (1993) build on the role of size and value to postulate that a three factor model based on factors formed on size. it is a challenge to execute arbitrage activity that is free from idiosyncratic risk. That is. The arbitrage theory posits that mispricing can persist whenever the cost of arbitrage exceeds the benefit. In addition. Kim and Christos (2010) find a positive relationship between mispricing and idiosyncratic risk. McLean (2010) finds a strong reversal mispricing in high idiosyncratic risk firms. The literature on the informal Wall Street characteristics is given a magnificent boost by Fama and French (1992). book/market value characteristics and the market returns can explain expected returns. 4 . The premise of these characteristics is not based on any prior theoretical reasoning.
Jegadeesh and Titman (1993) document the prediction of three to twelve months of past returns. This 5 . As information that is more public arrives. Hirshleifer and Subrahmanyam (1998) argues that if an investor places more weight on his private information signal. However. the prices move closer to their fundamental values. The overreaction correction pattern is consistent with a long run negative autocorrelation in stock returns (longterm reversal in returns). it causes stock price overreaction. On the other hand. Jegadeesh and Titman (2001) do not only address the datamining critique in explaining momentum effect but also document that models of investor behavioral bias offer a good explanation of the momentum effect. the source of the momentum effect is subject to debate. Conrad and Kaul (1998) and Bulkley and Nawosah (2009) argue that the effect is mainly due to the crosssectional variation in expected returns. they indicate the support for the behavioral model should be tempered with caution. Their argument is because momentum profit is due to delay in overreactions that are eventually reversed.On the return predictions based on past performance. However. Gutierrez and Hameed (2004) and Chabot. The characteristics derived from behavioral biases or cognitive challenges are based on informal arguments about investor overreaction and underreaction to information. The premise of the behavioral notion is that the conventional financial theory ignores how investors take decisions. Others believe momentum effect can be explained by the state of the market (Cooper. Daniel. Ghysels and Jagannathan (2009)). he tends to see a new public information signal as a confirmation of his private signal. if the investor begins with unbiased beliefs about his private information signal. On the contrary. They find that stocks or portfolios that have performed well in the past 3 to 12 months will continue to do so in the next 3 to 12 months (momentum).
absolute return to the dollar trading ratio.1 The aim of this Dissertation In this dissertation. From emerging markets. defined as the absolute value of stock returns divided by the dollar volume. Amihud (2002) proposes price impact measures. Gulen and Schill (2008) show that growth in book assets are crosssectionally related to future returns. the biased selfattribution implies the shortrun momentum and longterm reversal. 1. In a reaction to the possible problem associated with providing a consistent liquidity measure for all markets. etc. With the notion of investor reactions to value. we study the robustness of the crosssectional patterns in stock returns. Amihud and Mendelson (1986) find a significant premium for the bidask spread measure. He finds a positive relationship between his illiquidity measure and average return. using the completely independent database of the Brussels Stock Exchange (BSE from here onwards) in the 19th century and the first few years of the 20th century.1. market friction is a predictor of stock returns. and the implication is that investors underreact to information in the time series of balance sheets. market capitalization. Some of these include the bidask spread. Harvey and Lundblad (2007) use the proportion of zero returns as a measure of illiquidity to establish a positive relationship with expected returns. Hence. Investors require compensation for market friction (illiquidity). relationship between price changes and order flows. Bekaert. thereby causing momentum in security prices. Cooper. the proportion of zero returns. The most difficult issue in relating market illiquidity to expected returns is the measure of illiquidity. There is a pack of measures in the literature related to illiquidity.suggests that the public can also trigger further overreaction to the preceding private signal. Apart 6 . In effect. share turnover.
from the quality of the data set, it allows us to study the influence of strong varying conditions in the economic and institutional environment on stock returns. The period of the data set avoids the data mining critique. With this data, We will test the validity of the CAPM. In addition, we test whether other characteristics can explain the crosssectional variation in stock returns. As of the time of writing this dissertation, accounting and transaction data has not been digitalized for the 19th century BSE, so we will not be able to investigate the predictability of accountingrelated characteristics on stock returns. However, we investigate whether size, momentum (short run past returns), total risk (firm specific risk) and dividend yield (an asset value' indicator like the booktomarket value ratio) can crosssectionally predict returns. With the importance of illiquidity, in the absence of volume and transaction data, we follow Bekaert et al. (2007) to measure illiquidity as a proportion of zero returns in the last 30 months. However, our unreported results indicate that zeroreturn illiquidity measure is strongly correlated with market capitalization (also a proxy for liquidity) and stock price level.
1.2 Data Description and Methodology
It is worthy to reiterate that empirical research on the stock market return predictability has a long tradition. The large body of the financial literature in this area uses data from the Center for Research in Security Prices (CRSP). The CRSP data consist of comprehensive and accurate historical returns for all stocks listed on the NYSE, the
7
Amex and the NASDAQ stock markets in the USA. Researchers make use of the long time series (NYSE from 1926 to date, Amex and NASDAQ from 1962 and 1973, respectively to date) and high quality financial data to carry out research about general equilibrium asset pricing models and predictable patterns in returns, among others. However, the obsessed and continuous search for predictable patterns in a single data set will likely reveal an interesting (spurious) pattern (Lo and MacKinlay (1990)). Since one of the most studied quantities on the market to date is the stock return, the tests on financial asset pricing models seem especially at risk. Recent anomalies detected in empirical studies call for researchers to suggest a modification to standard economic theories about asset prices. However, the CRSP data are mostly used in asset pricing research due to the nonavailability of reliable and independent data sources. There is a considerable danger of inducing data mining/snooping bias when a single data set is used repeatedly. This may attenuate the reliability of statistical analysis on the data. Data from financial markets of other countries are normally used to investigate asset return predictability. However, common characteristics have been identified that can predict returns on contemporary markets. To reduce the doubts regarding inferences of the test of asset pricing predictability, we analyze a completely new and unique historical data set of stock returns from the BSE during the 19th century and the beginning of the 20th century. The data set was constructed at the University of Antwerp in Belgium (Studiecentrum voor Onderneming en Beurs (SCOB)). The data for our study start from 1832 and ends at 1914, just before the outbreak of World War I. During the World War I period, the BSE was closed and this can be regarded as a natural breaking point of the long time series of the stock
8
returns. The BSE was considered one of the biggest markets in the world at that time, because Belgium was one of the first nations on the European continent to become industrialized (see Van der Wee (1996) and Neymarck (1911)). On the industrial output per head ladder, Belgium stood second after Britain in 1860, and third in 1913, after the UK and the USA (see Bairoch (1982)). During this period, highly developed banking system coupled with liberal stock market regulations attracted a great deal of domestic and foreign capital in Belgium. In confirmation, Van Nieuwerburgh, Buelens and Cuyvers (2006) document that the development of the financial sector, accompanied with the stock marketbased financing of firms, played an important role in the economic growth of 19th century Belgium.
Figure 1.1: Number of listings on the 19 Century Brussels Stock Exchange
th
Source: Annaert, Buelens and De Ceuster (2004)
Figure 1 indicates the number of listings on the 19th century BSE. The figure depicts the rising popularity of the BSE. The bold line shows the number of common stocks, bonds
9
and preferred stocks. Figure 1. From 1868 onwards. Notably. The importance of the common stocks declines to about 65% just before the World War I.2: Total market capitalization of the BSE Source: Annaert. The number of common stock listings on the BSE is quite important. Annaert. the faint line shows the common stock listings of foreign companies. Buelens and De Ceuster (2004) It is obvious from Figure 2 that increasing number of common stock listings increases the market capitalization of the BSE. Buelens and De Ceuster (2004) shows that in the 19th century stock market capitalization was 10 . About one fifth of the common stock listings from the late 1870s are for foreign companies. even if we limit our attention to Belgian common stocks. It is clear from the figure that the BSE almost exclusively lists common stocks until the mid1850s. gradually increasing to about 600 just before the World War I. at least 100 Belgian common stocks are listed. the financial liberalization from the early 1970s onwards reflects in the growth of the total market capitalization. The evolution of common stock listings is represented with the dashed line. To illustrate the international attraction of the BSE.
d‟Anvers.2. The rate is extracted from the newspapers in the period of our study. The total returns of the common stocks on the BSE were computed by considering dividends.2 For the market index. stocks splits. All available information begins in January and ends in December of each year. The short rate is based on the commercial rate from the official quotation list of the Antwerp Stock Exchange. mergers. We concentrate on the valueweighted index as it 2 Journal du Commerce. financials. in this research. based on the geographical location of the major production facilities and the company‟s country of residence. and the analysis of the anomalies identified in the existing literature on asset pricing. Moniteur des Intérêts Matériels and Het Handelsblad. The BSE provides monthly data on stock prices. (2004). mining and extraction. However. utilities and industrials. L‟Avenir. 11 . we restrict ourselves to the analysis of data from the securities of Belgium owned companies. 1. 57% in 1880 and steadily growing to 80% by the end of the period of our study. The stocks on the BSE are well diversified across industries such as transportation. This enables the computation of the market capitalization and total returns for individual common stocks on the exchange.27% of gross national product in 1846. Annaert et al. The data permit us to test the robustness of the predictive power of characteristics on common stock returns. (2004) classified all companies listed on the BSE into five categories. delisting and other capital operations.1 The Riskfree rate and the Market Index We use the annualized short rate converted to a monthly rate as a proxy for the riskfree rate. dividends and the number of shares outstanding for more than 1000 different companies officially quoted on the exchange. we employ the all share monthly value weighted index constructed by Annaert et al.
Evidence from the Figure depicts a market 12 .1913 1000 800 600 400 200 0 1860 1870 1880 1890 1900 1910 1920 It is clear from Figure 1.3 that. 1868Dec.1832Dec.1867 500 450 400 350 300 250 200 150 100 50 0 1830 1835 1840 1845 1850 1855 1860 1865 1870 Valueweighted index evolution in the period of deregulation and expansion: Jan. The index represents the total returns of the stock market investable assets.mirrors the return evolutions of the investable assets' universe.1867). 100 Belgian Franc invested in the index would have grown to about 500 Belgian Franc at the end of the period of strict regulation and industrial revolution (Jan. Figure 1.3: Evolution of the value weighted market index in the two subperiods of our study Valueweighted index evolution in the period of industrial revolution: Jan. The valueweighted index is constructed by considering the market capital of the individual stocks in the index portfolios. 1832Dec.
12% 0.30* ValueWeighted Index 0.60% 15.26% 0.78% 0.19 2. 1832Dec. standard deviation.28* 0.1: Summary Statistics for Riskfree rate and the ValueWeighted Index Mean Riskfree ValueWeighted Index Riskfree ValueWeighted Index Riskfree 0. 1867 0.1 shows the time series summary statistics of the riskfree rate and the valueweighted index.60% 1st Order Autocorr.86% 0. as well as the average value weighted return for the entire period.1913). is almost the same in the two subperiods. Table 1.68% 0.24% Standard Deviation Skewness Kurtosis Min Median Jan. 100 Belgian Franc invested in the beginning of the period would have grown to about 1000 Belgian Franc at the end of the period (Jan. We calculate the time series mean.26 0. As indicated in Annaert et al.92 20.28 8. In the period of the deregulation and expansion period.68% 0.99 4. On the other hand. 1913 0. We show the summary statistics in the entire period and both subperiods.42% 0.34% 11.1868Dec. 1913 0.13 0.That is the period of the industrial revolution (18321867) and the period of deregulation and expansion (18681913).78% 0.15 18.17% 0.16% 0.29% 0.28% 0. These were the periods of the industrial revolution and high regulation (18321867) and the period of deregulation and expansion (18681913).25* Note: In this Table.12% 0. the standard deviation of the valueweighted index is high in the first period as compared to the second period. maximum and the first order autocorrelation of the valueweighted index series of the riskfree rate and the valueweighted market index.27% Jan. minimum.01% 0. The high standard deviation and 13 .25% 2. skewness.08% 0. The average riskfree rate. median.drop in 1948. the period from 1832 to 1913 is divided into two subperiods based on the environment in which the BSE was operating.The Table also displays the descriptive statistics for two subperiods. (2004).53% 15.61 20. the year of industrial revolution across Europe.55 9. we compute the descriptive statistics of the riskfree rate and the valueweighted market index.07% 3.59 3. 1868Dec.54 0.21% Jan. * = significantly different from zero at 5% (two sided).40% 0.99% 0. 0.08% 0. Table 1.41% 0.02 20.40% 1.23% Max 0. 1832Dec.
(2004)). and groups them to form portfolios. Table 2). This might be too low compared to the annual risk premium of 4. it is close to the values recorded on the UK market in the period 18701913 (see Grossman and Shore (2006) page 281.92% per annum. The sorting method is subdivided into the univariate.2. basically.16% for the overall period and the two subperiods. These are the sorting and the regression method. independent and conditional double sorts. However. the low risk premium recorded on the 19th century BSE is not surprising as Grossman and Shore (2006) record far too low annual risk premium on the UK market between 1870 and 1913 ( risk premium of 0. The univariate sort method ranks stocks on beta. The method tests the significance of the difference in average returns between the high and low ranked portfolios. A univariate sort has the disadvantage of not being able to disentangle the effect of two (or more) 14 . a year of revolution across Europe (although not in Belgium) and (ii) strong restrictions on joint stock holdings that may affect the volatility of prices (Annaert et al. The standard deviation recorded on the 19th century BSE is low compared to the high values recorded on the USA market.53% and 1. However.93% for value weighted and equally weighted respectively). or their characteristic. The average returns of the portfolios should be a monotonic function of the characteristic. The average difference between the value weighted index and the riskfree rate is between 0. there are two methods for testing the crosssectional predictability of stock returns. as the period is characterized with (i) a large market drop in 1848.kurtosis in the first period are not surprising. 1. This implies a risk premium of 1.2 Methodology As shown in the diagram below.61% recorded on the USA markets.13% and 0.
The conditional double sorting method keeps one characteristic constant and tests the effect of the other characteristic on the crosssection of expected returns. The advantage of the 15 . Univariate sort Sorting Independent Double Sort Conditional Double Sort Time Series CrossSectional Predictability of Stock Returns Regression CrossSectional Fama and French (1992) adopted the conditional double sort to disentangle the effects of beta and size on average returns. Portfolios are formed from the stocks that exist at the intersections of the groups. one characteristic is expected to predict stock returns better. stocks are separately ranked on each characteristic and grouped. In such a case. the conditional double sort is applied. In the conditional double sort method. holding the other constant and vice versa.characteristics. stocks are first ranked on a single characteristic and split into groups. To distinguish between the effects of two characteristics on the crosssection of stock returns. The stocks in each group are then sorted on the second characteristic. Empirically. The independent double sort method simultaneously tests the effect of two characteristics on the crosssection of stock returns. Asness (1997) used the independent double sort method to test the interactive effect of value and momentum on the crosssection of stock returns.
(1) where rjt is the time series excess returns of asset j (usually portfolio of assets). It also ignores the noisy nature of the stock or portfolio returns. As stated before. The method answers the question of whether the highranked characteristic portfolio outperforms the lowranked characteristic portfolio. usually comprised of the market excess returns and portfolio returns. the method test is based on whether the difference in the average returns of the high and low. However. The method is based on time series regression. The sign and the significance of the difference in average returns determine whether the relationship is negative or positive.ranked portfolios are significantly different from zero. it is weak to use the information on the high and low ranked portfolios to represent the crosssectional effect of a characteristic on expected returns. (1972) initiated the time series method. The two types of regressions used to test the crosssectional predictability of stock returns are the time series and the crosssectional regression method. αj. Ross and Shanken (1989). it is difficult to disentangle the marginal effect of several characteristics because the sorting method cannot sort on more than three characteristics. which was later developed by Gibbons. The estimated intercepts 16 . for which the intercepts are tested to determine whether they are significantly different from zero. bj and ɛjt are the regression coefficients and the regression error.sorting methods is that they do not impose any functional relationship between characteristics and expected returns. Black et al. In addition. Ft is the factor portfolio. formed from the difference in the high.and lowranked characteristic portfolios. The regression method addresses some of the shortcomings of the sorting methods. The method runs the regression of each asset base on the equation rjt j b j Ft jt .
jt1 can be any characteristic determined prior to t. the method compares the statistic to the F. (1989) suggest computing the statistic 1 T N 1 ' 1 1 SR2 ~ F ( N. for each time t. is vector of intercepts from the regression equation above. TN1)). If the model can explain the crosssectional variation in returns. ɛjt is the 17 .distribution with degrees of freedom N and TN1 (F (N. That is. the twopass crosssectional regression method can take on factors that are not returns (see Fama and MacBeth (1973)). which requires regressors that are also returns. Gibbons et al. v 2 C (2) where rjt is the crosssectional excess returns of all assets at the time t. The first pass regression (time series) estimates betas. The intercepts represent pricing errors. Unlike the time series method. jt 1 jt . N where T is the length of the time series. Γv.are grouped together to form a vector. several characteristics can be combined as independent variables or regressors. the crosssectional regression method runs a regression of the form rjt ot 1t jt 1 vt v . In the second pass regression. Under the null hypothesis that the intercept equals zero. ∑ is the covariance matrix of the residuals from the regression above and SR is the Sharpe ratio of the factor. and βjt1 are the timevarying beta estimates. T N 1) . which then serve as input for the second pass crosssectional regression. the intercepts should not be jointly significantly different from zero. The method often involves two regression stages (hence twopass regression) because of the test of CAPM. N is the number of assets or portfolio of assets.
Any characteristic that can explain the crosssectional variation in stock returns has an average coefficient significantly different from zero. There is a tendency to measure betas with error in the first regression. the expected value of 0t should be zero and the expected value of 1t . we estimate their expected values and test if they are significantly different from zero. 18 . The betas are then used in the crosssectional regression for the subsequent period. . the second crosssectional regression estimate is Tconsistent. Using estimated betas in the second pass regressions therefore induces an errorinvariable problem. If the CAPM holds.error term. Betas can be estimated with a rolling window or the full window regressions. In the full window regression method. We now form the ttest t v v v (3) 1 T vt and v is the standard error normally adjusted for the Newey T t 1 West heteroskedastic and autocorrelation consistency (HAC). The NeweyWest standard where v q w ' error is estimated as v 0 1 w w where 1 q w1 w E vt vt vt w vt and q is the number of lags. Estimating betas over the full sample period mitigates the measurement error. biasing the slope coefficient in small samples. Ct is a vector of regression coefficients. In this case. which is the risk premium on the market should be significantly positive. Using the time series of the regression coefficients. portfolios are formed by sorting stocks on betas estimated in the period prior to t. as the time series sample size T increases. The rolling window regressions estimate betas with information in the period prior to t. 0t .
we will use the double sorting methods to disentangle the marginal effect of each characteristic on average returns. since it uses the full information in the sample period and the crosssection. the full sample betas can be assigned to individual stocks in the portfolios. so each time. Shanken (1992) suggests an adjustment factor 1 . A second approach to minimize the errorinvariable problem induced by beta or any other estimated variable is to multiply the denominator in equation 3 by a correction rm 0 2 factor. The main assumption underlying this approach is that stocks in a portfolio have the same exposure to longterm systematic risk. On the other hand. The sorting method provides the advantage of not placing any linear restrictions on the return/characteristic relationship. This approach will yield N and T consistent estimates. Where r2m is the 2 rm standard deviation of the market returns in excess of the riskfree rate. Full sample betas are estimated using these time series of portfolio returns. Based on the assumption that betas are constant throughout the sample period. the full sample betas can be used in the crosssectional regression (see Ibbotson. Fama and French (1992) utilized the fullsample beta estimate in the test of the CAPM. For instance. Kaplan and Peterson (1997)). Finally. asset beta varies with time.Portfolios are then rebalanced by repeating the process of beta estimation and portfolio formation for each time t in the period studied. To test the effect 19 . portfolio returns are computed. rm is the average excess market returns and 0 is the average intercept series from the crosssectional regressions. we will use all the possible sorting methods depending on the task. In this dissertation.
1 Chapter 2 In the second chapter. thereby biasing 20 . If N is even. minus the bottomranked portfolio returns). stocks are allocated to int( N / n) portfolios. The fourth chapter tests the momentum effect. We cover the assessment of beta in chapter 2. Beta serves as the main input of the CAPM. Momentum. we test the validity of the CAPM and the size effect. In chapter 3.3 Summary of the chapters In this doctoral research. This will serve as a confirmation of the result in the sorting method. we focus on the assessment of beta. That is. It is an estimated variable that can be measured with error.3. which are of a great deal of interest because of the formation of the hedge portfolio (topranked portfolio returns. If N is odd. It also ensures that no stock is lost in the portfolio formation process. where int N / n is the nearest integer less or equal to N / n . we conduct four empirical studies with the BSE data from the 19th and the beginning of the 20th century. The last chapter concludes the dissertation. if N is the number of stocks in year t and n is the number of portfolios required. For all portfolio formations we adopt the Fama and MacBeth (1973) breakpoint method. int N / n 1/ 2 N n int N / n stocks will be allocated to the first and the last portfolio. The combined effects of Size. 1. The method sometimes allocates more stocks to the extreme portfolios. 1. we will adopt the crosssectional regression method. The middle portfolios have int N / n stocks each. Total Risk and Dividend yield on the crosssection of stock returns is studied in chapter 5.of several characteristics on returns. one stock will be added to the last portfolio.
We show that the individual stock market model betas are not stable. Using Dimson‟s method of estimating betas. we find no empirical validity for the CAPM. based on their ability to predict the subsequent beta. we determine how to estimate beta for the study of the CAPM.the test results of the CAPM. If anything. small numbers of stocks prove to have returns that lead or lag the market returns. We also show that the relationship between beta and returns varies with time. Specifically. The predictability of the market model betas can be improved by forming portfolios with at least ten or more stocks.3. the null hypothesis of the equality of the estimated slopes 21 . 1. we compare the market model betas to the betas estimated with the Blume (1971) and Vasicek (1973) autoregressive techniques.2 Chapter 3 In the third chapter. The chapter studies the relative performance of different methods of estimating beta. The betas from the IRLS are small in magnitude compared to the market model betas. To account for outliers. Most strikingly. Beta instability. With the sorting and the FM crosssectional regression methods. Based on the study in chapter two. we use the sorting and the Fama and MacBeth (1973) (from here on FM) crosssectional regression methods to investigate whether the CAPM is valid before World War I. but they have the same predictive accuracy as the market model. bias and nonrobustness to outliers have become the center stage of research since the development of the CAPM. there is no significant difference in the predictive accuracy of Blume and Vasicek‟s adjusted betas. We will also test if the size effect (the propensity of small size to have higher returns than large size stocks) exists in the 19th century BSE. we use iterative reweighted least square techniques (IRLS) to estimate betas.
from the crosssectional regression and the market excess returns is not rejected in the period 1868 through to 1893. On the contrary, using sizesorted portfolios (equally weighted) in the crosssectional regression establishes the betaexcess return relationship for beta as the lonely regressor (the CAPM is valid). However, the conditional double sorting method adopted by Fama and French (1992) separates the effect of beta and size on expected returns. The method makes the average crosssectional slope of the beta insignificant, whether placed alone or together with size in the regression. We find size to be negatively related to excess returns. Size is negatively significantly related to excess returns (size effect), but beta does not relate to excess returns when placed together in the cross sectional regressions. Detailed analysis of the data reveals that the size effect is mainly due to small size stocks, which accounts for about 0.35% of the total market size. Eliminating these small stocks destroys the relationship between excess returns and beta or size. Both sorting and crosssectional regression reveals that the size effect disappears when stocks are value weighted to form portfolios. In summary, the CAPM did not work in the 19th century BSE. Estimating betas with the market model, the Dimson and Vasicek method will not establish the model. The size effect exists, but it is mainly due to a small group of stocks that represent a very small portion of the total market capitalization.
1.3.3 Chapter 4
In chapter 4, we investigate whether the momentum strategy can generate abnormal profit in the 19th century BSE. The momentum strategy buys stocks that have performed well in the past 312 months, and sells stocks that have performed poorly in the past 312 months. There is convincing evidence that the strategy is profitable on the 19th century
22
BSE. Finding momentum in this era confirms the assertion that the momentum profit found on the postWorld War II markets is not mainly due to datasnooping biases. Detailed analysis reveals that the momentum effect does not exist in the small size group of stocks. Additional investigation into the momentum profit in each calendar month shows that the profit was positive for all months. The January reversal effect, found in the postWorld War II USA markets, cannot be found in the 19th century BSE. In fact, January records the fourth highest momentum profit relative to the other months of the year. We find that momentum profit is not strong in the first twenty years of our study period. In order to investigate the source of the momentum profit, we use the approach of Jegadeesh and Titman (2001) to study the returns of the momentum portfolios in the postholding period. The momentum returns reverse in the second to fifth years after the holding period. Further study reveals the reversal is mainly due to small size stocks. We also test whether the momentum profit and the long run reversal in the crosssection of stock returns depends on the state of the market. The 6month formation and the 6 to 12month holding period strategies are profitable solely in periods of market gains.
1.3.4 Chapter 5
In this chapter, we investigate whether size, momentum, total risk and dividend yield can explain the crosssection of stock returns between the years 1868 and 1913. We repeat size and momentum characteristics in the analysis to test the pervasiveness and the combined effect of these characteristics on the crosssection of excess returns. The size sort also tests whether the effect of the other characteristics are not confined to the small, illiquid group of stocks. Total risk for each stock is measured as the standard deviation of
23
the past 24 to 60 months excess returns. Each year, dividend yield for a stock is the sum of all dividends paid within the last 12 months divided by the current month price. We investigate whether the relationships are pervasive across all dividend yield, size, total risk and momentum groups. Sorting and FM crosssectional regression methods are adopted in the analysis. We confirm that size has a negative significant relationship with excess returns. However, repeating the analysis on different size groups reveals that the negative relationship is completely driven by microsize stocks accounting for less than 3.67% of the market capital. This confirms the finding in Chapter 3 that the size effect is mainly due to the first decile size portfolio, which forms part of the micro size group. We did not find a consistent relationship for total risk. Momentum shows a consistent positive relationship with excess returns in stocks, which accounts for more than 96% of the market capital. We find a negative relationship between dividend yield and momentum among dividendpaying stocks, but each of them is positively related to average excess returns. Dividend yield and momentum are positively related to the average excess returns in our large stocks, which account for about 96% of the market capital.
24
Beta. plays a key role in CAPM applications. and it is assumed stable in empirical applications. the instability. Studying the various techniques of estimating beta. we may determine relatively unbiased and stable beta for the test of CAPM. The notion of its estimation is of fundamental importance to the testing of CAPM. one of the parameters of a time series' regression. However.CHAPTER 2 2 ASSESSMENT OF BETA IN THE 19th CENTURY BSE3 Beta risk estimation and the testing of asset pricing models have a long tradition in financial literature. which posits a positive relationship between the beta and expected returns. bias and robustness evaluations have become the center stage of research in finance since the development of the CAPM by Sharpe (1964). The MM is a statistical model that relates the return of any given stock to the return of the market index. (4) 3 This chapter was presented as a paper at the First World Finance Conference on the 27 th May 2010 at Viana do Castelo.1 Introduction and Literature Review Beta stability. The objective of this chapter is to introduce the new data set from the 19th century BSE to test the performance of the alternative techniques of estimating beta. Portugal (blind reviewed) 25 . Therefore. for returns of a stock j in period t we write. R jt j j Rmt jt . 2. Beta serves as the main input of the CAPM. Beta is usually estimated with the standard market model (MM). Lintner (1965) and Mossin (1966). bias and nonrobustness (to outliers) of the beta have raised concerns in the literature.
The objective of Blume‟s autoregressive model is based on the tendency of betas from successive periods to revert to the mean. Var (x) is the variance of x. That is. He finds portfolio betas having stronger correlation across successive periods than individual security betas. The jt is the error of the crosssectional regression of betas in successive periods.…. In effect. The autoregressive method adopted by Blume is the crosssectional regression of betas in period t on the betas in period t1. This indicates that portfolio betas are more stable than individual stock betas. The MM assumes that j (beta) is constant over the estimation period. 26 . R jt and Rmt are the period t returns on the stock j and the market index return. but the adjustment also depends on the uncertainty (standard error) 4 E x is the expected value of x . but they revert towards their crosssectional mean. j and 2j are the parameters to be estimated from the MM peculiar to stock j . The Vasicek autoregressive method not only adjusts betas toward their crosssectional mean. the method adjusts historical betas towards their crosssectional mean.where Var jt 2 j 4.N (5) N is the number of stocks in the crosssection. Blume (1971) and Vasicek (1973) show that betas are not stable. substantial evidence in the financial literature has established that security betas are not stable over time. respectively. However. They propose autoregressive models to capture this variation in beta. for j=1. Blume (1971) and (1975) pioneered the research on beta stability based on the crosssectional correlation between beta estimates for successive periods. a and b are the intercept and the slope of the crosssectional regression of betas in the period t on betas in the period t1. jt a b jt 1 jt . The parameters j .
They show that there is a strong connection between beta instability and the market index used to estimate beta. and the more the uncertainty surrounding either estimate of beta. the literature has shown that nonsynchronous trading biases betas estimated with the MM. the lower the weight placed on it. Faff and John (1992) use Australian data to confirm the evidence in beta nonstationarity. When stocks are ranked based on their market capital (size). That is. The Bayesian correction method places weights on the crosssectional average beta and the asset‟s beta estimate. jt 1 . (6) where jt is the mean of the posterior distribution of beta for stock j . Therefore. and var jt 1 is the variance of the crosssection of betas. 2. Impson and Karafiath (1994) use daily data to document that portfolio betas are not more stable than individual security betas on the US market. . Vasicek (1973) applied a Bayesian correction method to capture the differences in standard errors. GregoryAllen. Betas estimated with high standard errors have a greater tendency to deviate from the crosssectional mean of the betas.N . beta is biased 27 . jt 1 is the crosssectional mean of betas in period t 1 . which serves as 2 the beta forecast. They show that. On the contrary. high uncertainties call for greater adjustment. is the standard error from the estimation of the market model regression coefficient. Also. beta is estimated with the model 2 jt var jt 1 2 jt 1 2 var jt 1 var jt 1 jt 1 for j 1. He used crosssectional information from the previous period betas.about beta. The weights sum up to one. accounting for the variance around beta estimation will not make portfolio betas more stable than individual stock betas.
Dimson (1979) adopted the aggregate coefficient method. which computes the lead and the lag betas in a multiple regression of the stock return on a number of lead/lag market returns. Scholes and Williams biased correction model betas. Scholes and Williams (1977). The literature also highlights the impact of outliers (extreme observations) on the beta estimates. On the other hand. This implies that the returns of small stocks are more capable of exhibiting autocorrelation than large stocks. They recommend the inclusion of the lagged market returns in the estimation of beta when securities are traded infrequently. The error term jt follows the i l assumptions of the classical linear regression model. i l l l (7) Dimson‟s beta estimate is then given by dim i . The literature has found that the bias in beta estimates is due to autocorrelation in returns caused by infrequent trading.downward for small size stocks and upward for large size stocks. Outlier observations in returns. Bartholdy and Riding (1994) found that the MM betas are less biased. On the other hand. depending on their location in the stock return/ 28 . Ibbotson et al. Dimson‟s technique involves estimating a multiple regression of the form R jt j j Rmt i jt . Dimson (1979) and Fowler and Rorke (1983) developed methods to correct the bias. more efficient and consistent on the New Zealand Market than the Dimson. (1997) found that beta estimates for small stocks are severely biased downwards. Scholes and Williams (1977) compute the lead and lag beta coefficients by univariate regressions. The sum of the estimated beta coefficients from the multiple regressions is Dimson‟s estimate of beta.
The literature proposes various techniques of minimizing the impact of outliers on beta estimates. Martin and Simin (2003) confirmed that outlier resistance beta is a better predictor of future beta than the market model beta. Nevertheless. Under idealized conditions. The weighted least square estimation technique is one of the methods used to minimize the impact of outliers on beta estimates. the MM does not consider the effect of outlying observations on beta. can have an extreme influence on beta when the MM is used to estimate beta. these conditions often fail in empirical settings. One explanation of the failure of the condition is the occurrence of outlier observations in returns. Based on the assumption that outliers have a lower probability of occurring in the future. less weight is placed on them (see Martin and Simin (2003)). They also reveal that small stocks betas are most vulnerable to outliers.market return plane. An alternative technique would be to remove the extreme 29 . The idealized conditions are that the paired returns of the stock and the market conform to the linear model relationship with zero mean error not correlated with the market returns. Their detailed study compares the betas estimated from the MM and the outlier resistance methods. Stocks that exhibit this behavior are mostly small stocks. the MM beta produces the best estimate because of the ordinary least squares method. dividend cuttings. Chan and Lakonishok (1992) said the robust method (outlier detecting method) of estimating beta is good for stocks susceptible to stock splits. They indicate that the MM cannot detect a large number of outlying returns in their data. Unfortunately. Identifying the outlying observations and estimating betas with the outlying resistance method yield beta estimates that are lower than the MM betas. Chatterjee and Jacques (1994) document the effect of outlying “observations” on the beta parameter. and initial public offerings (IPO).
The least squares approach adopted by the MM minimizes the sum square residuals with respect to the model parameters j and j : min R jt j j Rmt . 30 . j The squaring of the residuals magnifies the effect of the outliers on the estimated parameters. defined as: 2 2 s jt 1 W s jt T 0 for for s jt T s jt T . (8) with weight function W and j the standard deviation of the return residual. To reduce the influence of outliers. We apply Huber or the bisquare weight function. This method estimates beta by iteratively minimizing a weighted function that depends on standardized return residuals: R jt j j Rmt min W j . expected return is likely to shift toward the outliers while the covariance matrix will be inflated toward the outliers. the statistics literature emphasizes the use of iterative reweighted leastsquares (IRLS) method. 2 j . j j .observations and perform the regression with fewer observations. We estimate the regression parameters from the least squares regressions and use these parameters as initial input for iteration. A weight function is applied to the standardized residuals. For instance.
with a vector of ones with the same size as Rmt (market index returns). For the iterations. Therefore. the initial input is the coefficients from the MM. He indicated that timevarying beta model might be mispecified. 31 .with the standardized residual s jt R jt j j Rmt and T . W 0 is an initial standardized weight diagonal matrix and b1 the estimated beta parameter. The new set of parameters serves as input for the next iteration. a tuning constant. A standardized residual observation that exceeds the tuning constant is assigned zero weight (outliers). static beta CAPM should replace with time varying beta conditional CAPM (see Jagannathan and Wang (1996)). The residual from the MM is standardized and the weight function defined above is used to transform them. we did not pursue the timevarying beta because. The literature also suggests that estimated betas vary with time and macroeconomic conditions. Ghysels (1998) used monthly NYSE data to document that pricing errors with static beta models are smaller than those with timevarying beta models. In this dissertation. Small values of T introduce more resistance to outliers. but at the detriment of efficiency when returns are outlier free.685. The procedure above is repeated until the parameter of interest (beta) converges. Estimate the parameters by the weighted least squares: b1 'W 0 'W 0 R j where 1 Rmt . Huber j (2004) sets the value of T to 4. He chose the value of T so the method will have less influence on beta when there is no extreme observation and it still provides protection against outliers.
The subsequent sections of the chapter are organized as follows: Section 2. They found that various beta estimation techniques behave uniquely in different markets. Estimating betas with the Dimson method reveals a very small number of stocks showing a lead and lag relationship with the market returns.2 provides descriptive statistics of beta when estimated with the least squares regressions of the MM. The study also shows no significant difference between the Blume and Vasicek adjusted betas in terms of their predictive accuracy. individual and portfolio betas from the various estimation techniques will be compared based on their predictive accuracy. This chapter reveals that for individual stocks. The iterative reweighted least squares method produces betas that have the same predictive power as the MM betas. the literature has not traced the stability.3 examine the Blume correlation technique of detecting the stability of beta.To the best of my knowledge. but are lower in their magnitude on average. Section 2. The 19th century data provide a very good platform to determine whether betas before 1914 exhibit a similar pattern as betas after 1914. Predictability can be improved when a portfolio of 10 or more stocks components is formed. Martikainen. Luoma. The autoregressive technique of Blume and Vasicek is used to adjust betas in 32 . In this chapter. bias and robustness of beta using data before 1926 in any country. The root mean square error (RMSE) criterion adopted by Blume (1971) and Klemkosky and Martin (1975) is used to determine the predictive accuracy of the betas estimated from the various techniques. the market model betas are weak in predicting their future. Perttunen and Pynnönen (1994) simulated returns data to generate artificial markets.
Stocks with complete returns data for five years are considered in each subperiod..15. The beta coefficient of each stock in each period is estimated by simply regressing the monthly returns of the stock in each period on the corresponding monthly valueweighted index of the market by using the MM above. In Section 2. Leybourne and Newbold (1997) is used to test for equal predictive accuracy of the models. The modified Diebold and Mariano test proposed by Harvey.. respectively. The number of stocks with full returns data in each period ranges from 21 to 424 as shown in Column 2.1 presents the number of stocks in each period under study. The equally weighted average beta and the valueweighted average beta are displayed in Columns 3 and 4. We compute the valueweighted mean beta by considering individual stock relative to market capitalization at the beginning of the period.5.2 Beta Coefficient Descriptive Statistics The data sample is divided into 15 fiveyear nonoverlapping periods ti for i 1.. Table 2. The final section presents the conclusion.3. we compare outlier resistant betas to the benchmark MM betas. starting from January 1837 to December 1911. Crosssectional statistics of the betas in each period are computed and the result is displayed. For the proxies of the market portfolio. From the high values of equally weighted mean beta and low values of valueweighted mean beta in the periods 33 .1. The predictive accuracy of the adjustment techniques based on the RMSE criterion is tested in the same section. 2. the statistical characteristics of the betas estimated in each period and displays the 15 periods studied.. we determine how the stocks in each period lead or lag the market index using the Dimson‟s model. we consider the valueweighted market index. In Section 2. It can be seen in Table 2.4.1 that the number of stocks that have five years of data does not exceed 100 before 1877.Subsection 2.
36 1.52 1.96 0.10 0.83 1.32 1.15 1.01 0.33 0.66 2.96 1.21 1.98 19 2.05 Standard Deviation 0.98 0.69 0.09 1.01 1.83 0.94 1.29 1.91 0.10 0.23 8 5.28 0.55 15 7.27 0.44 1.99 1.02 1.73 1.67 1.05 0.33 0.70 12 2.75 0.66 0.50 1.10 0.35 0. The valueweighted mean is the average of the betas weighted by their market capital at the beginning of the period.18 Fractiles 0.1: Beta coefficient descriptive statistics for the 15 estimated periods Percentage of BETAS <Zero 0 12 0 3 2 4 8 10 12 18 14 12 8 11 9 Minimun 0.83 1.03 2.02 0.03 1.74 0.01 1.74 3.58 1.05 1.25 1.79 0.03 0.14 2.87 2.02 0.03 1.56 1.09 1.25 0.04 0.76 1.77 0.35 0.55 1.92 25 1.06 0.85 0.11 1.50 1.15 1.34 0.09 0.75 1.19 18 3.16 Average beta(VW) 1.67 0.32 0.43 0.24 1.15 0. The last column reports the average coefficient of determination R .70 1.74 1.29 2.76 0.61 1.55 0.70 0.84 0.33 0.90 4.63 2.93 0.38 2.18 0.04 1.48 0.27 0. 34 .24 1.14 0.Table 2.30 0.19 0.03 1.01 0.54 1.40 0.50 0.73 0.04 0.35 2.33 12 Note: This table displays crosssectional descriptive statistics of betas estimated by market model (1) of returns on their market counterparts over the 15 fiveyear subperiods between January 1832 and December 1911.14 0.33 2.38 1.24 1.71 0.57 1.80 17 7.74 27 3.71 1.01 0.24 1.08 0.05 0.39 0.98 1.90 2.92 10 7.95 0.81 22 7.71 3.74 1.03 0.20 0.13 1.11 0. Equally weighted (EW) and valueweighted (VW) mean betas are also displayed in the 2 table.55 0.35 0.59 3.72 0.86 0.64 0. The maximum and minimum beta of each period is also recorded.99 0. The number of stocks in the various periods ranges from 21 to 424 for stocks with full return data within a period.47 14 8. The table also reports the percentage of beta sample in a period that is less than zero.36 11 3.37 1.34 Number of Period 1/183712/1841 1/184212/1846 1/184712/1851 1/185212/1856 1/185712/1861 1/186212/1866 1/186712/1871 1/187212/1877 1/187712/1881 1/188212/1886 1/188712/1891 1/189212/1896 1/189712/1901 1/190212/1906 1/190712/1911 Stocks 21 33 32 33 54 72 76 82 113 154 161 196 252 374 424 Average beta(EW) 1.94 6 7.93 1.07 1.45 Mean 2 Maximum R (%) 2.90 29 10.76 0.71 1.42 0.
The possible explanation for these extreme values of the beta and low coefficients of determination might be linked to the infrequent trading effect and the influence of extreme observations (outliers) in the returns series of the stocks. In order to capture the effect of size (market capitalization) on the average beta and R 2 values. the stocks in each period are subdivided into three mutually exclusive sizebased subsamples.after 1873. However. It is also important to note the very high and low betas in the 19th century BSE compared to the post1926 market betas recorded in the literature. Jacquillat and Levasseur (1974). Altman. Dimson and Marsh (1983) recorded a large number of negative betas in their weekly returns interval estimation of beta on the French and the American markets. It is worthy to note the percentage of negative betas in nearly all the subperiods. For example. The composition of the subsamples is as follows: the sample of stocks in each period is sorted in descending order based on their market capitalization (size) at the beginning of the period. The values in the last column show the average coefficient of determination R 2 in percentages. The removal of stocks that did not trade fully within the fiveyear estimation periods does not introduce survivorship bias since there is no significant difference in the beta coefficient descriptive statistics when stocks with at least 24 observations are added to the sample (not reported). Surprisingly. which is a measure of explanatory power of the MM. 35 . we can conclude that small stocks have high betas in these periods. we computed the same crosssectional statistics for betas in each period. the literature reports negative betas on different markets after World War I.
2 12.1: The graph of the average beta of each period for large stocks and small stocks Note: The average beta estimates for large and small stocks are plotted against their corresponding periods of estimation.73 0. 36 .2 10. M and S subscripts on betas indicate large.2 19.40 1.1 36. The first 30 percent of stocks in the periods are classified as large stocks.4 14.94 1.67 0.1 27.9 5.04 0.22 1.3 4.0 1.3 3.5 43.07 1.2 9.0 12.0 8.5 17.18 30.0 6.40 0.1 10. 2 are Figure 2.0 10.86 1.59 1.19 0.27 1.3 5.01 0.51 1.39 1.03 1.61 0. respectively.8 10. stocks in each period were sorted in descending order based on their market capital.00 0.70 0.6 17. medium and small stocks.30 0.5 22.2 27.2: Average beta and average coefficient of determination of the size based subsamples Period 1/183712/1841 1/184212/1846 1/184712/1851 1/185212/1856 1/185712/1861 1/186212/1866 1/186712/1871 1/187212/1877 1/187712/1881 1/188212/1886 1/188712/1891 1/189212/1896 1/189712/1901 1/190212/1906 1/190712/1911 Large Medium Small mean R2(%) mean(βiM) mean R2(%) mean(βiS) meanR2(%) mean(βiL ) 1.5 20.2 9.95 1.06 1.17 32.69 0.15 1.7 9.6 15.06 1.84 1.4 17.2 20.7 5.66 1.05 1.6 9.5 21.13 23.90 1.08 0.4 16.29 1.16 1.1 11.76 0.4 18. The L.10 1.5 7.1 10.29 1.5 Note: In this table.77 1.02 1.49 0. Crosssectional statistics of the betas of the subsamples were computed and their equally weighted averages of betas and coefficient of determination R displayed. the next 40 percent as medium stocks and last 30 percent as small stocks.6 14. The diamond points indicate the averages of large stocks and the square points for small stocks.73 1.4 18.1 1. This depicts the extreme values of beta estimate recorded for the small stocks relative to the benchmark beta of one.4 40.16 1.Table 2.79 0.98 1.2 35.0 9.6 7.
2 reports the average beta and the R 2 values for each subsample in each period. we adopt the Blume (1971) and Altman et al. found that small stock returns due to infrequent trading show a high degree of autocorrelation and that they are capable of recording high betas. Figure 2. Generally. 2. the high values of R 2 recorded for large stocks are striking because on the UK market. betas of small stocks (infrequently traded stocks) are lower when estimated with the market model. (1974) correlation method of investigating the stability of beta estimates. we confirm their result in the periods after 1867. (1997). It is clear from Figure 2.The first 30 percent of stocks are classified as the large stocks portfolio. On the contrary. Dimson (1979) recorded a similar pattern with 15 years of monthly data.1 that small stocks record comparatively low betas in the periods before 1867 and high betas thereafter. where the small stock average beta lies slightly below the average beta of large stocks. Scholes and Williams (1977) and Beer (1997).3 Beta Stability Here. Ibbotson et al. Table 2. The only anomaly is the period between 1882 and 1886. based on NYSE data between 1926 and 1994. The result in the periods before 1867 corroborates the assertion of Dimson (1979). and the last 30 percent as the small stocks portfolio. This might be attributed to the valueweighted index used in the computation. next 40 percent as medium stocks portfolio. On the 19th century BSE. since the index will have more explanatory power for largesized stocks. Blume (1975) shows that the beta coefficient 37 . They said that if trading frequency is highly correlated with the market capitalization of the stock.1 depicts the crosssectional average beta of each period for large stocks and small stocks.
it has data for two complete consecutive periods. For a stock to be included in this analysis. Betas in period ti are used to rank stocks existing in periods ti and ti 1 in ascending order.4. Table 2. In period ti equally weighted portfolios of s 1. of Stocks per Portfolio 1 2 4 7 10 20 Correlation 0. Assuming equal amounts are invested in each stock.between two successive periods is stationary if5 E ti E ti1 .62 0.. 38 .10.2.. then the portfolio beta will be the mean of the betas of stocks included in the portfolio.92 0. We computed the weighted average correlations by considering the number of portfolios in each tenyear period. corr ti . Var ti Var ti1 .95 Spearman's rank Correlation 0. y) is the correlation between x and y .72 0. ti1 1 where ti are the betas in period ti .7.91 0.20). The weighted average correlation takes into account the number of stocks or portfolios in each adjacent tenyear period.3: Weighted average of correlation and Spearman rank order correlation across successive periods No.64 0. the betas of all portfolios in period ti 1 are also computed. Portfolios are formed with their constituency as follows: the first portfolio is the first s stocks for s = (1.77 0. This process of portfolio formation is repeated until the number of stocks left is less than s.96 Note: This table shows the weighted average correlation and Spearman‟s rank correlation of betas of individual stocks and portfolios in successive periods across the 15 periods studied. The next portfolio consists of stocks with the next smallest beta estimates. The second portfolio contains the following s stocks and so on until available stocks is less than s .72 0..76 0. Table 2.3.56 0. stocks are formed as follows: the first portfolio consists of stocks with s smallest beta estimates.3 reports the weighted average correlation across all the subperiods studied.54 0. For each s.2. 5 corr ( x. We compute the correlation and Spearman rank order correlation of the betas between each two adjacent periods. Betas in a period (estimation period) are used to rank the betas in that period and the next adjacent period (prediction period) in ascending order.
Ledolter and Rayburn (1987). where a and b are the constant term and slope coefficients.44 and 0. we use the Blume (1971) autoregressive adjustment model to improve the stability of beta estimates for both individual stocks and portfolios. GregoryAllen et al. based on 52week estimation periods.4 presents regression tendencies implied between adjacent periods. monthly returns interval measurement and sixtymonth estimation periods to obtain an average correlation of 0. Furthermore. Faff and John (1992). on average.56 and 0.91 for 1. Table 2.3. Levy ( 1971) records 0.The weighted average correlation across the successive periods ranges from 0. The values of the coefficients in the periods between 39 . Blume finds 0.and 10sized portfolios using 84month estimation periods respectively. less information about their future beta than the portfolio beta. There is a tendency for a high beta estimate to overstate its true value and vice versa. respectively.or 10sized portfolios respectively.93) mean correlation (rank correlation) for 1.67) and 0.91(0.54 for individual stocks to 0. Dimson and Marsh (1983) used valueweighted market index.62(0. The correlations in Table 2.3 show that the betas from the 19th century BSE are stable compared to the betas in post1926 US and UK markets. Eisenbeiss. Therefore.1 Blume and Vasicek stability adjustment techniques Individual stock betas estimated from the MM are noted as unstable in the previous section (also in Blume (1971). (1994).95 for portfolios of 20 stocks.and 10sized portfolios on the same exchange. Blume (1971) found a similar result on the US market. Due to the limited number of stocks. we were not able form up to 50sized portfolios on the BSE. Kauermann and Semmler (2007)). These values indicate that the beta of individual stocks have. Collins. On the UK market between 1955 and 1979. 2.82 mean correlation for 1.
99 7. The last two Columns show the tstatistics of the test of a hypothesis of the slope coefficient equal to zero or one.80 7.21 41.70 0.1+η jt R 2 (%) H0:b =0 0 3 20 26 18 3 5 20 39 31 27 27 28 15 20 0.49 18.13 0.55 0.99 11.70 0. the coefficients change over time. The tstatistics of the slope coefficients show that the null hypothesis of the slope coefficient equal to zero is rejected in the adjacent periods after 1872.22 3.40 0.39 0.42 7.39 0.60 0.58 0.66 7.81 8.49 0.94 7.12/1846 and 1/1837.53 0.67 b 0.58 3.48 0.43 0. As can be seen in Table 2.67 0.32 1.75 19. The R2 values also show that the betas in the periods after 1872 have more explanations for their prior betas than those before 1872.4.74 0.11 7. A result not reported shows that increasing the length of the estimation period (such as seven years in Blume (1971)) 40 .1837 and 1867 are striking.4: Measurement of regression tendency of estimated beta coefficient for individual stocks Regression Tendency Implied Between Periods 1/1842.03 10. In addition.33 0. the null hypothesis of the slope coefficient equal to one is rejected in all the adjacent periods (except the first two periods).38 0.44 0. but there are extreme coefficients outside the interval between zero and one in the first two periods.78 0.98 19.27 0. The extreme values may be attributed to the number of stocks in a period as we record less than 50 stocks in our first two fiveyear periods.00 0.67 18.65 0.12/1841 1/18471/18521/18571/18621/18671/18721/18771/18821/18871/18921/18971/19021/190712/1851 12/1856 12/1861 12/1866 12/1871 12/1876 12/1881 12/1886 12/1891 12/1896 12/1901 12/1906 12/1912 and and and and and and and and and and and and and All periods 1/18421/18471/18521/18571/18621/18671/18721/18771/18821/18871/18921/18971/190212/1846 12/1851 12/1856 12/1861 12/1866 12/1871 12/1876 12/1881 12/1886 12/1891 12/1896 12/1902 12/1906 a 1.31 14.46 β jt =a+bβ jt.63 0. The tstatistic of the null hypothesis of the slope coefficient (b ) equal to zero or one is also reported in the last two columns.75 H0:b =1 1. Table 2. R 2 (%) is the percentage of the variance of betas in the period t explained by betas in the period t 1.86 0.96 2.04 5.00 0.77 0.41 5.54 0.34 4.38 1.94 3.07 0.69 0.46 Note: In this table beta of stock existing in a period are regressed on the betas of the same stocks in a prior adjacent period.60 5.91 10.43 10. It is not consistent with the Blume assertion that all the coefficients lie between zero and one.
suppose we want to forecast the beta for any stock or portfolio in the period 18421846.5 displays the average RMSE of the adjusted betas and MM betas across the 15 periods studied.. 4.improves the R 2 values and the tstatistics but at the expense of losing more stocks. N .5 also shows that the predictive performance improves as the number of stocks in a portfolio increases for both adjusted and the MM betas. We compute its beta in 18371841.. For individual stocks. We also introduce the Vasicek adjustment model (6) to adjust betas in successive adjacent periods. 10 and 20. With the regression tendencies. since fewer stocks have complete returns' data for longer periods. The adjustment method is repeated for all adjacent periods on equally weighted portfolios of size 2.. It is clear from the table that the average RMSE of the adjusted betas is lower than that of the market model betas. Table 2. In order to compare the predictive performance of the MM betas and the autoregressiveadjusted betas. βt is then computed from the equation and used as the forecast. the Bayesian adjustment technique proposed by Vasicek is superior to Blume‟s adjustment as reflected in the small average RMSE. Blume (1971) and 6 The root mean square error was calculated by 41 jti jti1 2 N for j 1. we compute the RMSE of betas estimated with MM in adjacent periods. Table 2. The adjustment process is repeated for stocks in the subsequent 13 adjacent periods using their respective coefficients.4. The RMSE tests the performance of the autoregressive methods based on variation and unbiasedness of their beta forecast. We test the predictive performance of the various adjusted betas by using the root mean square error6 (RMSE) criterion. The forecast of the beta is obtained by substituting it for βt1 in equation (5) with the coefficients in the first row of Table 2.. 7.
57 0. we consider the root mean square error function.40 0.19 Bayesian 0.5: Predictive performance of Blume and Vasicek (Bayesian) procedures of estimating beta No.57 0.35 0. In our case.31 0. Harvey et al.27 0. Table 2.29 0. suppose we want to compare the forecasts of Blume (BL) and Vasicek (VA) models. The conclusion is that on the 19th century BSE.08 0.76 0.23 Note: Average RMSE across the 15 periods studied were used to compare the predictive performance of the various adjusted betas and the market model (MM) betas in successive periods. the predictive accuracy of betas estimated by the MM can be improved by adjusting betas using either the Blume or Bayesian adjustment methods and a portfolio with a sizeable number of stocks. The average RMSE across the successive adjacent periods for the various equally weighted portfolio formations are displayed.Klemkosky and Martin (1975) recorded similar patterns of the predictive performance on the NYSE market.45 0. BL and VA are the forecasting errors from the Blume and Vasicek models.27 0.48 0. respectively. The reliability of the conclusion above can be confirmed by performing an additional test on the root mean squared error values. For portfolios of size 7 or more.25 Average RMSE Blume 0. f BL root mean square error of the Blume adjusted betas.23 0.38 0.of Stock per Portfolio 1 2 4 7 10 20 MM 1. one cannot see much difference between the Blume adjustment method and the Bayesian approach. Therefore. The possible method is to test whether the differences in the values of the RMSE‟s are statistically significant.44 0. The test is based on the loss differential 42 . Their adjusted betas mean square errors were smaller than their MM betas. (1997) presented a modified Diebold and Mariano test statistic that will be used for this purpose.
function d j f BL f VA for j 1. a period of strict regulation and a period of deregulation and expansion.6 reports the modified DieboldMariano test statistics between the various models under study. We reject the null hypothesis of equal predictive accuracy when the test statistic is greater than the critical value at level.1 . Blume and Vasicek models in all periods are pooled together to form three series of length H. The Modified DieboldMariano (MDM) statistic is: H 1 2h H 1h h 1 2 d MDM ~ t .. The null hypothesis of expected equal j j predictive performance is H0: E d j 0 and the alternative hypothesis of the Blume model predicting worse than the Vasicek model is H a : E d j 0. Then we perform the test on the three series. H ... The values in the first row of the table reveal that we can confidently reject the null hypothesis of onestep ahead equal predictive accuracy of the 43 . H 1 (0. i cov d j .1) a H j 1 1 student‟s t distribution with H 1 degrees of freedom and v is the significant level usually set at 5%. H 1 0.. d j i . . Table 2. In order to apply this test. h is the horizon of forecast and t . 1 2 H 0 2 i i 1 where d H 1 d j . The test compares the DieboldMariano test statistics to critical values from the student‟s t distribution. betas estimated with the Market. Betas estimated from the various models are considered across the entire period studied. The period studied is divided into two subperiods based on the environment in which the BSE operated.
11 (0.19 (0. β VA = Vasicek's adjusted betas.87 1.6 show that there is no significant difference between the Vasicek betas and the Blume betas in terms of their onestep ahead forecast. 2. β VA Note : This table reports the modified DieboldM ariano test statistics for one step ahead equal forecast accuracy between the market model.18 (0. we might expect that some stocks may not trade every month for economic reasons or because of 44 . During the strictly regulated period. we cannot reject the null hypothesis of equal predictive accuracy between the two models.00) (0.78 (0.Blume and MM betas. Table 2. Between the Vasicek adjusted betas and the MM betas. we find that Blumeadjusted beta significantly outperforms the market model beta. we can reject the null hypothesis of equal predictive accuracy for the deregulated and expansion period.44 (0.08) 0. The hypothesis test H 0 : E(d j ) = 0 and Ha= E(d j ) >0 The significance level of the rejection becomes weak in the strictly regulated period.99 (0.22) Overall period 1/183712/1911 2. β BL Strict Regulatory Deregulation and period expansion period 1/183712/1871 1/187212/1911 2. β BL = Blume's adjusted betas.42) 4. For instance. The values from the bottom row of Table 2.4 Beta Bias Considering the period of the study and the trading frequency of the market.00) 0. The equality in the predictive performance of the Vasicek and the MM betas is strongly rejected in the entire period.6: Modified DieboldMariano test statistics (pvalue in parentheses) Models β MM vs.02) 1. in the overall period of our sample the null hypothesis can be rejected at the 5 percent level. Blume and the Vasicek's adjusted betas.00) 0. β MM =market model betas.75 (0.01) 3.56 (0.29) β MM vs. Therefore. β VA β BL vs.
19 (55) 0. These stocks may systematically lead or lag the market movement.05 (3) 0.03 (4) 0.07 (13) 0.02 (7) 0.19 (42) 0.03 (13) 0.05 (6) 1.15 (4) 0.97 0.05 (5) 0.22 (10) 1.74 0.17 (59) 0.04 (4) 0.11 (9) 0.06 (3) 0.13 (56) 0.76 (73) 0.regulatory conditions.19 (10) 0.96 1.12 (5) β0 1.00 (4) 0.12 (10) 1.00 (5) 0.90 (42) 0.07 (0) β2 0.04 (4) 0.07 (6) 0.07 (9) 0.03 (6) 0.02 (5) Note: This table reports Dimson's aggregate coefficient adjusted beta in each five year period. The numbers in parenthesis are the percentage of stocks in a period that reject the null hypothesis of the coefficient been zero at 5% significant level.00 (3) 0.73 (67) 0.07 (6) 0.02 (60) 0.05 (5) β1 0.00 (0) 0.01 (5) 0.47 0.18 (47) 0.05 (7) 0.37 Mean Lag.16 (0) 0.7: Dimson Aggregate Coefficient (AC) beta Adjustment Number of stocks 21 AC Beta 1.67 0.05 (5) 0. producing biased betas when beta is estimated with the MM.12 (0) 0.17 (7) 0.99 (78) 0.02 (0) β3 0.09 (7) 0.02 (3) 0.02 (5) 0.07 (13) 0.09 0.04 (7) 0.69 0.04 (3) 0.04 (11) 0.07 (10) 0.80 0.03 (3) 0.06 (58) 0.21 (4) 0.03 (10) β2 0.37 (12) 0.07 (53) 0.97 (38) 0.01 (5) 0.15 (3) 0.24 (11) 1.09 (1) 0.09 (8) 0.01 (6) 0.10 (0) 0.58 (52) 0.11 1.08 (8) 0.16 (3) 0.56 0.16 (3) 0.12 (7) 0.07 (6) 0. 45 .42 (95) β1 0.03 (3) 0.10 (18) 1.01 (0) 0.12 (15) 0.04 (3) 0.07 (5) 1/184212/1846 1/184712/1851 1/185212/1856 1/185712/1861 1/186212/1866 1/186712/1871 33 32 33 54 72 76 1.15 (11) 1/188212/1886 154 0.10 (13) 1.09 (5) 1.00 (2) 0.09 (11) 0.12 (6) 1.00 (6) 0.01 (6) 0.08 (5) 0.07 (6) 0.49 0.03 (6) 0.05 (6) 0.08 (3) 1/187212/1877 1/187712/1881 82 112 1.03 (6) 0.Match and the Lead beta estimates Period 1/183712/1841 β3 0.42 0.00 (2) 0.10 (4) 0.03 (6) 0.76 (72) 0.12 (6) 0.68 1.04 (8) 1.11 (3) 0.04 (5) 0.03 (2) 0.07 (6) 0.17 1.06 (8) 0.02 (7) 0.08 (6) 0. In order to expose the presence of possible lead or lag effects.10 (5) 0.18 (4) 0. Table 2.08 (5) 0.12 (7) 0.18 (1) 1/188712/1891 1/189212/1896 1/189712/1901 1/190212/1906 1/190712/1911 Overall periods 160 196 252 374 424 2075 1. we test the significance of the coefficient of the returns on the lagged or lead market index.23 1.02 (7) 0.
We use only three months lag and lead because beta bias has been documented as not prevalent in monthly returns data (see Cohen. that is i 3.. which incorporate the lagged and lead market indexes. For example.7 reports the crosssectional average of the lag and lead betas in each period. This indicates that there is no severe timing problem in the 19th century data. we can interpret this as evidence of the market model (MM) producing statistically reliable beta estimates in relation to the other models. As most of the lead and lagged coefficients are significantly equal to zero. We test the hypothesis H 0 : i 0 against the alternative H 1 : i 0 for each stock..The Dimson bias adjustment equation with maximum lag or lead of three months is considered. the estimates of the parameters i indicate the lagged.. matched and lead beta coefficients. These results can be compared to the results from the postWorld War I markets. Schwartz and Whitcomb (1983)). but their numbers does not exceed the coefficients corresponding to the match. Maier. 46 . The result also follows Cohen et al. Evidence from this table indicates that beta coefficients i for i 0 are not significantly different from zero for the majority of the stocks.3 . Unsurprisingly. Hawawini.. there are some stocks with lead and lag coefficients that are statistically significant. The numbers in parentheses are the percentage of stocks that reject the null hypothesis. This shows that the explanatory power of the model for i 0 is approximately zero for most of the stocks. Table 2. (1983) hypothesis that there is a strong relationship between beta estimates and the length of the interval over which returns are measured. For each stock. Hawawini and Michel (1979) found a similar pattern of results on the Belgium stock exchange by using weekly interval returns data between 1963 and 1976.
reducing the impact of outliers in the estimation of the beta can significantly change the value of beta. The weights given to each return pair observation depends on the distance between the observation and the fitted line (Martin and Simin (2003)). and the bias disappears when the difference of the interval is lengthened (monthly). in the first period out of the 21 stocks. stocks with identified outliers less than or equal to 4 are grouped into Category A and those with identified outliers greater than 4 are grouped in category B. three fall in Category A with an average market model (MM) beta estimate of 1. which minimizes a weighted sum of squares of residuals. we compare the crosssectional average betas of the MM and IRLS for each category.2.1 for the fiveyear periods studied might be due to the influence of outliers or unexpected movement by the stock or the market returns. The number of stocks in each fiveyear period is grouped into two. As explained in Section 2. Similarly. On the contrary.They established that beta bias mostly shows up in the short length interval (daily) of returns.5 Impact of outlying observations on Beta The extreme (maximum/minimum) beta estimates recorded by some stocks in Table 2.8 reports how the presence of outliers affects the beta value. In such cases. 2. We apply the IRLS (outlier resistant). The literature shows that outliers have a tendency to reduce or increase the magnitude of the beta when it is estimated with the MM (Chatterjee and Jacques (1994)). (1997) reports that lagged coefficients should be considered when estimating beta. Ibbotson et al.29 and 47 . For example. Table 2. on the New Zealand market. Bartholdy and Riding (1994) used monthly data to establish that betas estimated from MM are less biased. We compute the average beta of each category. In each period.
61 A B 54 (28) 1.16 0.16 0.46.32 0.12 0.72 0.06 0.60 0. In Category A.15 0.24 A B 9 (23) 0.36 0.00 0.36 0. In each period.72 0.38 1/184212/1846 1/184712/1851 1/185212/1856 1/185712/1861 1/186212/1866 1/186712/1871 1/187212/1877 1/187712/1881 33 32 33 54 72 76 82 112 A B 10 (23) 0.28 0.72 1.62 1.8: Comparison of the market model betas and the iterative reweighted least square betas Total Number of Number of detected Number of Average stocks outlier observations Stocks MM Period 1/183712/1841 Average IRLS 0.06 0.80 1.64 A B 346 (78) 1.07 0. Table 2.81 0.21 A B 41 (13) 1.26 0.45 0.10 0. We also show the number of stocks that falls in each category.77 0.14 0.40 Note: In this Table. the rest of the periods have more stocks in Category A than Category B. except for periods 1 to 3.28 0.IRLS beta of 0.13 0. the difference between the average MM beta and the average IRLS beta is 0.51 A B 48 24 0.26 0.46 A B 23 (10) 0.83 1.20 0.60 0.41 0.08 0.12 0.49 0. 48 .47 0.46 0.47 1.34 1/190212/1906 1/190712/1911 374 424 A B 290 (84) 1.07 0.56 0.42 0.57 A B 120 (40) 1.10 0.33 0.83 1.23 1.19 0.52 0.29 0.62 1.88 0.63 1/189212/1896 1/189712/1901 196 252 A B 146 (50) 1. We compute the cross sectional average of the betas estimated with the market model and the IRLS model.57 A B 42 (34) 0. Looking across periods.71 1/188212/1886 1/188712/1891 154 160 A B 110 (44) 1.67 0.97 0.19 βMMβIRLS 0.20 0.07 0.42 0.75 0.03 0.23 0.50 A B 198 (54) 1.05 0.21 0.26 0.07 0.61 A B 78 (34) 1.10 0.40 1.39 0.29 1.60 1.05 0.79 0.06 0.19 0.07 0.91 1.83.19 21 A B 3 (18) 1. we estimate betas in each category by using the M arket model and iterative reweighted least square (IRLS) model.80 0. stocks with outlier observations less or equal to 4 are grouped in Category A and those with outlying observations greater than 4 are grouped in category B.
2: Plot of average market model betas and IRLS betas for stocks with outlier observation less than 4 Notes: This figure depicts the average market model and iterative reweighted least square betas for stocks with at most 4 detected outliers across the 15 periods of study. This implies that the more the outlier observations in the return series. As in Subsection 2. A pooled sample of betas within the period studied is considered. 49 . we apply the modified DieboldMariano test to compare onestep ahead predictive accuracy of the MM and IRLS estimated betas. The modified Diebold.In each period.37 with pvalue of 0. The modified Diebold Mariano test statistics proposed by Harvey et al. It confirms the result by Chatterjee and Jacques (1994) that the weighted least squares estimation reduces the MM betas by a certain percentage. The MM beta is always greater than the IRLS beta in Category A (across periods in Figure 2.09 for onestep forecasts in the period of our study.1.2). the difference between the MM betas and the IRLS betas for Category B is greater than the difference in Category A (last column). Figure 2. (1997) are employed to test the null hypothesis of equal predictive accuracy against the alternative of IRLS betas forecasting better than the MM betas. the higher the market model overestimates beta.Mariano's test statistic between the two models is 1.3.
the IRLS method can help to curb the influence of outliers on estimated betas.29 1.43 (0. Pvalues are in parenthesis.9. β IRLS Strict Regulatory Deregulation and period expansion period 1/183712/1871 1/187212/1911 2. The hypothesis test H 0 : E(d j ) = 0 and Ha= E(d j ) >0. β MM =market model betas. β IRLS = iterative resistive least squares method betas.01) (0. 50 . we conclude that on the 19th century BSE.08) Overall period 1/183712/1911 1.37 (0. This shows that we cannot reject the null hypothesis of equal predictive accuracy at the 5 percent significance level in the overall period. From Table 2. but it does not significantly outperform the standard MM in terms of their ability to predict onestep ahead in the period of deregulation and expansion. the null hypothesis of equal predictive accuracy is not rejected. From the deregulation and expansion period.Table 2. The null hypothesis can be rejected only at the 10 percent level.09) Note : This table reports the modified DieboldM ariano test statistics for one step ahead equal forecast accuracy between the market model and the iterative resistive least squares model.9: Test of equal predictive accuracy between MM and IRLS models Models β MM vs.
correcting nonsynchronous trading effect reveals that returns of few stocks have a significant relationship with the lead and lag market returns. which is the primary implication of the CAPM. Specifically. the market model beta is weak in its ability to predict the future beta. Applying the Dimson method. The predictability can be improved by grouping 10 or more stocks to form a portfolio or adjusting betas with the Vasicek and Blume autoregressive techniques. There is no significant difference in the predictive accuracy of the betas estimated with the IRLS method and the market model in the deregulation and expansion period.6 Conclusion This chapter evaluates the relative performance of different methods of estimating beta based on their ability to predict subsequent beta on the 19th century BSE.2. 51 . In the next chapter. the study reveals that for individual stocks. The analysis of the different beta techniques reveals that beta estimated with the market model is not stable. we study the ability of beta to explain returns in the crosssection of stocks. The study also shows no significant difference between the Blume and Vasicek adjusted betas in terms of their predictive accuracy.
The effect implies the propensity for stocks with low market capitalization to outperform those with high 52 . using the Brussels Stock Exchange (BSE) data. and the size effect exists. the literature has questioned the validity of the model and suggest other characteristics than beta to explain expected returns. Lintner (1965). Despite the dominance of the CAPM in empirical finance. We use preWorld War I Belgium data. and Mossin (1966). The purpose of this chapter is to investigate if CAPM is valid before World War I. In addition. In addition. Banz (1981) finds a size effect in stock returns. The use of the preWorld War I (18681914) data will provide very good grounds for an outofsample test of the CAPM and the size effect.1 Introduction and Literature Review This section investigates the crosssectional relationship between stock returns and beta. However. we will investigate whether a company‟s market capitalization (size) is related to its average excess returns (size effect). a number of researchers have found evidence against the model. it also examines whether other parameters such as beta square and idiosyncratic risk can explain expected returns. 3.CHAPTER 3 3 THE TEST OF CAPITAL ASSET PRICING MODEL (CAPM) AND THE SIZE EFFECT IN 19th CENTURY BSE The CAPM posits a positive relationship between the systematic risk (beta) and the expected return of an asset. It investigates whether there is a positive linear relationship between expected returns and beta. The empirical study that supports the CAPM model in the 1970s is from Fama and MacBeth (1973). Since the development of the CAPM in the 1960s by Sharpe (1964). This will minimize the data mining critique if the CAPM is valid.
In this case. In addition. Lakonishok and Shapiro (1986) and Ritter and Chopra (1989) do not detect any significant relationship between beta and expected returns. they conclude that size and the booktomarket value ratio can explain the variation in expected returns when placed together in a crosssectional regression.2% of market capitalization. The anomaly is initially discovered outside the period of our study. this chapter introduces the 53 . Grossman and Shore (2006) use preWorld War I UK data to present evidence against the size effect. To distinguish between data snooping and the persistence of size effect. Kothari. which account for about 0. Instead. To add to the existing literature on asset pricing. With the debate on the validity of the CAPM still ongoing.market capitalization. we investigate the effect on another dataset. Fama and French (1992) find no association between betas and average returns. Schwert (2003) documents that the size effect disappears after 1981 in the US market using monthly data for the period 1962 to 2002. Shanken and Sloan (1995) use annual portfolio returns and equally weighted market index to document evidence in support of the CAPM. In addition. even when beta is the only explanatory variable in their crosssectional regressions. Dimson and Marsh (1999) test the presence of the size effect by using FTSE all share monthly returns data from the period 1955 to 1998 and document that the effect has disappeared after 1979. The above literature on beta and size focuses on postWorld War I return data and sometimes mainly on data from the US. the effect should not be found in other periods. Another view is that the characteristic may have been discovered out of luck through data snooping bias (see Lo and MacKinlay (1990)). but the size effect disappears when these stocks are eliminated. They find a size effect among extremely small stocks. In contrast.
3. we resort to the sorting and the FM crosssectional regression methods discussed in Chapter one.35% of the total market capitalization. To this end.3. Section 3.19th century BSE data to test the validity of CAPM and the presence of the size effect. we use FM crosssectional regression analysis to confirm the above sorting results.2. In the previous chapter. we rank stocks based on beta and group them to form portfolios.1. As the aim of this chapter is to test the validity of the CAPM. In Section 3. one needs to form portfolios in order to improve on the precisions of individual betas. the question answered by this method is whether highbeta stocks outperform lowbeta stocks. As stated before. the method needed to estimate its input is worth considering. and Vasicek betas V . The FM crosssectional regressions are used to test the relationship between beta and expected returns (CAPM) in Subsection 3.3 shows that a betasorted portfolio should contain at least seven stocks in order to have a reliably stable portfolio beta estimate. Table 2. Figure 3. The remainder of the chapter is organized as follows: In Section 3. In testing the CAPM.2. we show the expected returns of portfolios sorted on MM betas MM . Dimson betas dim . We also find a size effect on the 19th century BSE.1 54 . We find no relationship between beta and expected returns.2 Expected returns of portfolios sorted on betas In the sorting method.4. we investigate the effect of size and beta on excess returns by using the sorting method. Detailed investigation reveals that the size effect in our data is confined to small stocks. but it disappears when stocks are value weighted to form portfolios. In Section 3. which represent on average 0.4 concludes the chapter.
Based on these reasons. Figure 3. Evidence from this figure shows that until 1868 decile portfolios will not have the minimum of seven stocks. as shown in Chapter 2 (Table 2. this chapter and the subsequent ones will focus on the data between 1868 and 1914. Furthermore. individual betas before 1868 do not predict well their subsequent fiveyear beta.4). which is reflected in the number of stocks listed on the BSE. it must have a minimum of 24 months of observations out of the 60 months required to estimate beta before the 55 .shows the number of stocks included in our portfolio formation every year. The changes in legislation in 1867 ease the establishment of a company.1: Number of stocks in our selection criteria for the entire period of the preworld war I SCOB data For a stock to be included in the portfolio formation. In addition. Van Nieuwerburgh et al. (2006) indicate the importance of the longterm relationship between the development of the BSE and economic growth in Belgium after legal liberalization.
since the market index used in this analysis is heavily weighted towards large stocks.7) that some stocks systematically lead or lag behind the market movement. Including stocks with at least 24 months of returns does not change the descriptive statistics of the prior betas (from here on preranking betas). This enables us to capture more stocks in the crosssection. and Rmt is the market portfolio for period t . It is the slope coefficient from the regression equation R jt R ft MM Rmt R ft jt . converted to a monthly rate. Possible explanation for the significant lead (lagged) relationship is because large (small) firm prices adjust quickly (slowly) to market wide information. (2004) as a proxy for the market portfolio. which may produce biased betas. We recall from Chapter 2 (Table 2. (9) where R jt is the return on a portfolio or stock for period t . We 56 . we study three beta estimates in this chapter. We use the valueweighted market portfolio constructed by Annaert et al. we do not restrict our analysis to stocks with a complete fiveyear returns data as in chapter two. The model is used to compute the posterior Vasicek betas V using the crosssectional information on betas estimated by equation (9). In this chapter. For comparison purposes. small stock returns have the tendency to lead or lag behind in relation to the market wide returns. R ft is the riskfree rate for period t .portfolio formation year. which is the traditional beta. The annualized money market rate. We compute the second beta estimate using the Vasicek model introduced in Chapter 2. when we estimate beta by the MM. The first is the MM MM beta. is used as a proxy for the riskfree rate. Thus.
which captures the correlation between the current period returns of a stock and current and lagged market returns.1 captures the correlation between stock‟s current period returns and the lagged market returns. For our monthly data.0 captures the contemporaneous covariation between the returns of a stock (portfolio) and the market returns.0 j . (10) where j .6) reveals that infrequent trading effect is not a severe problem. documents that the infrequent trading effect is not severe when monthly returns are used to estimate betas. we compute betas (preranking) for all stocks using the past 24 to 60 months of returns data. Dimson (1979) with UK data. New estimates of preranking betas are calculated in December each year. Stocks are assigned to decile portfolios using the FM breakpoint method. That is. we use only onemonth lag because Chapter 2 (Section 2.adjust for the lag effect by using the Dimson model in Chapter 2 to obtain a third beta estimate. and the portfolio formation is repeated. We sort stocks into decile portfolios based on the preranking betas (univariate sort). we run the regression R jt R ft j j . Dimson onemonth lagged beta is estimated as dim j . j . Portfolio 1 contains stocks with the lowest betas. Monthly portfolio 57 . 1 . In addition. while portfolio 10 contains stocks with the highest betas. beginning in January 1868. The postranking valueweighted and equally weighted return for each month is calculated for each portfolio.1 Rmt 1 R ft 1 jt . Specifically. We estimate postranking portfolio betas for the sample period (18681914) by using valueweighted and equally weighted portfolio returns.0 Rmt R ft j . We account for the possible timevariation in betas by rebalancing stocks in each year.
62 0.81 0. Portfolio 1 contains the lowest betas and Portfolio 10 contains the highest betas.99 0.79 2.71 0.40 0.45 0.95 0.02 0.16 3.41 0.52 1.68 1. Vasicek's adjustment (β V ) model and the Dimson's model with one month lag (β dim ).56 0. The FamaMacBeth breakpoint technique is used to assign stocks to decile portfolios.23 2.17 3. From Panel A.09 1.08 1.02 0.14 0.63 0.26 6.37 3.23 0.67 2.27 3.26 2.91 0.32 0.68 0.93 0. standard deviation.64 0.01 1.55 0.36 2.22 1. beta does not exhibit any relationship with average returns. Standard Deviation (%).36 1.30 3.12 1.formation for each year yields 552 monthly returns for each decile portfolio.06 4. Mean (%) is the time series average of the portfolio excess returns for the entire period.30 3.04 0. Table 3.71 0.17 5.32 3.06 1.30 3.36 2.41 0.45 0.32 0.45 0.20 3.34 3.71 1.25 3.35 1.14 4.86 0.05 5.94 0.21 0.88 0.48 2.66 2.58 0.92 0.09 1.19 0.25 3.25 3.12 0. Table 3.95 0.45 1.36 0.06 4.45 3.82 2.17 3.80 0.26 0.10 0.05 5.12 2.93 0.90 1.47 0.79 0.57 0.70 1.14 1.81 0.03 3.32 6.35 3.82 1.17 0.09 4. The result does not change when we consider 58 .We compute time series Standard Deviation(%) of the postranking excess returns. This process is followed for all three beta estimates MM .07 1.33 0.18 0.97 0.80 0.90 1.66 1.19 2.49 0.10 0.26 2.20 5.17 1.23 4.92 0.34 0.34 1.43 3. The average returns do not show any pattern as beta progressively increases from low to high beta portfolios.22 0. 1913 Low 1 Market Model (Equally Weighted) Mean (%) Standard Deviation(%) Beta Market Model (Value Weighted) Mean (%) Standard Deviation(%) Beta Dimson Betas(Equally Weighted) Mean (%) Standard Deviation(%) Beta Dimson Betas(Value Weighted) Mean (%) Standard Deviation(%) Beta Vasicek Betas(Equally Weighted) Mean (%) Standard Deviation(%) Beta Vasicek Betas(Value Weighted) Mean (%) Standard Deviation(%) Beta 2 3 4 5 6 7 8 9 High10 Panel A 0.68 0.48 0.39 0.33 0.32 1.36 2.14 0.37 0.18 1.58 0. stocks are sorted based on preranking betas.07 1.13 2.35 0.01 4.08 0.62 0.64 0.61 1.30 3.48 0.09 1.06 1.33 3.38 2.59 0. and the postranking betas of the ten portfolios.60 0.21 Panel B 0.45 0.75 0.61 0. 1868Dec.33 0.20 3. V and dim .18 3.1: Time Series Mean (%).26 5. Betas are estimated by using the long time series portfolio excess returns and the corresponding excess market returns.12 1.00 1. and Postranking Betas of Decile portfolios formed from preranking betas in Jan.13 2.22 1.58 1.56 1.68 1.13 1.26 2.50 0.19 0.71 0. when both preranking and postranking betas are estimated with the market model.39 At the beginning of each year.23 3.32 0.24 0.49 0. The preranking betas are estimated with market model (β MM ).13 0.45 2.20 2.1 reports the average excess return (time series).18 1.92 0.23 4.28 2.82 0.67 0.53 0.39 1.98 Panel C 0.
The univariate beta sorting results confirm Fama and French (1992) findings. jt is the beta estimated from the fullsample portfolio returns. second. we use FM crosssectional regression to test the robustness of the above sorting result. The slope coefficient from each regression is treated as the reward per unit of the beta risk in that month (risk premium). The most striking of all is that the postranking betas almost surely follow the ordering of the preranking betas (except the first. and Mossin‟s version of the CAPM. Estimating betas with the Dimson and Vasicek methods in Panels B and C does not establish the relationship between beta and expected returns. The method also considers the noisy nature of portfolio or stock returns by running monthly crosssectional regressions of betasorted portfolio returns on betas. In this subsection. The standard deviation of the monthly time series of slopes is used to perform a 59 .the valueweighted portfolio excess returns. The FM approach also provides a straightforward procedure to test whether the reward for bearing beta risk (risk premium) is equal to the excess market returns (the return of the market less the risk free rate) as implied by Sharpe. Lintner. and the sixth decile portfolios). R jt R ft 0t 1t jt t (11) where 0t and 1t are the regression intercept and slope for month t respectively.2. the standard approach to test the validity of the CAPM is sorting and FM (1973) crosssectional regression. The time series average of the monthly coefficient is the average reward for bearing the beta risk.1 The CrossSectional Regressions As stated in chapter one. They use Dimson adjusted betas to establish a flat relationship between beta and average return. We can also compare our result to the evidence of Reinganum (1981) who finds no relationship between beta and average return in the period 19641979. 3. That is.
Fama and French (1992) also rely on full window portfolio betas to mitigate the errorinvariable problem. and Dimson betas) and form portfolios each year. We repeat the process each year to account for time variations in betas. Each month. The postranking betas are estimated by using the postranking long time series' returns of the decile portfolios. This will produce 552 monthly returns of decile portfolios (postranking returns). We adopt the method by Fama and French (1992) to estimate full window portfolio betas. We form equally weighted and valueweighted portfolios from the betasorted group of stocks each month. The postranking beta serves as the input for equation (11) to perform the crosssectional regressions. As in the previous section. the result in Chapter 2 (Section 2. whether the beta risk is priced on average. whether the average slope is statistically significant from zero. V . The only difference is that we replicate Ibbotson et al.4) shows that we can rely on the portfolio betas.ttest. we regress the postranking excess returns of the decile portfolios on their corresponding beta (postranking) estimates. it is common to rely on large sample size statistics to draw inferences. To mitigate a possible errorinvariable problem. in other words. (1997) method and use the portfolio betas for the crosssectional regression instead of assigning the portfolio beta to individual stocks in the portfolio each year. We repeat the process for the various estimates of betas MM . Vasicek betas. This curbs the argument that the test can be incorrect if the size of the sample is not large enough for the asymptotic results to provide a good approximation. 60 . dim . we sort stocks based on their estimated preranking betas (Market model betas. Moreover. Portfolio 1 contains the lowest beta stocks while portfolio 10 contains the highest beta stocks.
12% (0.28% (1. we obtained 552 crosssectional regressions for each estimate of beta.29) 1.02% (0. β MM =Market Model beta. Newey West adjusted tstatistics are in parentheses.19) Panel C: Individual Stocks 0.26% (2.24% (1. 1913 ttest Intercept 0.92) 0.10) 0. It also shows the hypothesis test of mean slope (risk premium) equal to the average excess market returns as implied by the SharpeLintner CAPM.49) βMM 0.16 0.Table 3.06% (0.24% (2. Eventually.10% (0.94 0.41) 0.81) 2.48) 2.11% (0.87) 0.17) βdim βV H0 :Slope=(R m R f ) Panel A: Equally Weighted Portfolio 1.34) 0.63 0.30% (3.2: Average time series slopes from the FamaMacBeth CrossSectional Regressions in Jan. The significance of the average slope is tested by using heteroskedastic and autocorrelation consistent standard errors (Newey 61 . After performing the monthly crosssectional regressions. β V =Vasicek beta and β dim =Dimson's beta with one month lag. the time series mean of the slope coefficients is tested for statistical significance.16 0.16) 1.36% (2.30% (2.93 0.45) 1.02% (0.02% (0.37) 0.70) 0.55 This table reports average time series slopes and intercepts from monthly crosssectional regression of postranking portfolio excess returns on postranking beta estimates. 1868Dec.14) 0.01% (0.42% (1.47 0.57) Panel B: Value Weighted Portfolios 0.41 0.05 0.06% (0.41) 1.29% (2.
The average slope using the Dimson beta is 0. regardless of how beta is estimated. and their corresponding tstatistics in parentheses.and West (1987) correction with default lag of int ( T 1 4 ). The average slope of all the beta estimates in the crosssectional regression is not significantly different from zero. it may be possible that the result is influenced by small stocks. In Panel A.14. we use valueweighted portfolios for the estimation of postranking betas and in the crosssectional regression. the mean estimated slope for the market model beta is negative. slopes. in Panel B. the hypothesis cannot be rejected at the 5% level. The values in the last column show the tstatistics from the hypothesis test of average slope (risk premium) equals the average excess market return as implied by the CAPM. Panel A indicates that the market model postranking beta estimated with equally weighted portfolio returns does not provide a significant relationship with returns. We also adopt the Shanken‟s correction factor discussed in Chapter one for the computation of tstatistics. Specifically. The estimated mean slope with the Vasicek beta is also not significant.17 standard errors from zero. The negative slope is quite surprising as it goes against the notion of positive risk premium (CAPM). and it is only 0. the hypothesis of 62 . where T is 552). since equally weighted portfolios give undue weight to small stocks. in Panel A. Table 3. Therefore. This is to eliminate the possible errorinvariable biased induced by the estimated betas. As shown by the sorting method. Most strikingly.2 reports the average intercepts. Estimating preranking and postranking betas with the Vasicek and Dimson methods does not establish the betareturn relationship. Fama and French (1992) had a negative slope for beta when placed together with size in the crosssectional regression. However.02% with a tstatistic of 0.
A detailed look at Panel A shows that the Dimson beta is weak in explaining average returns (average slope of 0. Ang. For example. This confirms Fama and French (1992) result. In Panel C. This serves as a robustness check of the results in Table 3.3 reports the average crosssectional regression slopes for both equally weighted and valueweighted portfolio betas assigned to individual stocks.17). As a result. The results do not support the CAPM for the three beta estimates. Table 3. we apply the Fama and French (1992) approach of estimating full window portfolio beta and assigning the portfolio beta to the individual constituent stocks of the portfolio in the crosssectional regression. others believe that portfolios may conceal important information contained in the individual stock betas. portfolio betas are used for crosssectional regression.2. 63 . Liu and Schwarz (2008) show that the slope coefficient (risk premium) of crosssectional regression can be estimated more precisely using individual stocks instead of portfolios because creating portfolios reduces the crosssectional variation in betas. Using valueweighted portfolios (Panel B) to estimate postranking betas does not establish the beta return relationship. Although.equality between the average slope and the average excess market return is rejected at the 5% level for the market model and the Dimson betas.02% but with a tstatistic of only 0. This is a predictive test since the preranking betas are estimated over a period prior to the period over which the crosssectional regression is performed. they assert that beta is flat in relationship with average returns for post1960s USA data. The market model beta and the Vasicek beta estimate still maintains the negative nonsignificant relationship with average returns. we follow the traditional FM (1973) rolling window approach by using individual preranking betas in the crosssectional regression.
Mean slope and their corresponding ttstatistic is reported in parenthesis. We also report the tstatistic for the test of hypothesis of the mean slope equal to the average excess market returns. in all cases the hypothesis that the mean slope is equal to the mean excess market return is not rejected.3: Average Time Series Slopes from FamaFrench CrossSectional Regression in Jan.45) βMM 0.17) 0.03% (0.3. 64 .81) 0.19) 0. Portfolios are rebalanced annually. we assign the postranking portfolio beta to the individual stocks in the portfolio.18) 1.56) 0. and it is close to the average riskfree rate as postulated by CAPM.41% (1. Table 3.12% (0.West adjusted tstatistics are parenthesis. Newey. eq=equally weighted vw=value weighted.42% (1.02% (0.24% (2. Panel A) calls for a detailed look into its time series' behavior with the excess market returns. The positive average slope of the Dimson beta crosssectional regressions (Table 3.54) 0.89 0.89 0.17) βdim βV H0 :Slope=(R m R f ) Panel A: FamaFrench approach (eq) 1.97 In this table.40 0. the average intercept is marginally significant.02% (0. βV =Vasicek beta and βdim = Dimson beta with one lag.04) 0.44) 1.54) Panel B: FamaFrench approach (vw) 0.29% (2.Surprisingly. 1913 ttest Intercept 0.12% (0. To investigate the evolution of the slope coefficient and the excess market return through time. 1868Dec.29% (2.03 0.42) 0.39 0. In addition.24% (1.02% (0. β MM =Market Model beta.
1893 0. 1868Dec.03% 2. Avg.32% 0. 1913).02% 0.4: Subperiod look into estimated slopes and excess market returns ttest subperiods Intercept Slope H0 :Slope=Avg. Newey West tstatistic in parenthesis.01 0.015 slope excess market returns 0.75) (0.15% 0. Dimson's beta estimated from equally weighted portfolios is used in the crosssectional regressions for the two subperiods.09) Jan. The graph shows that the relationship between beta and expected returns varies with time.01 slopes/excess market returns 0. 1913 0.15) (0.Figure 3. R m R f =0. 1885) and (1907. =Average. Table 3.005 0.015 1870 1875 1880 1885 1890 year 1895 1900 1905 1910 1915 Surprisingly. Figure 3. there seems to be a close correlation between the slopes and the excess market returns for much of the period except between the years (1880. we report subperiod average slope and intercept from the Fama 65 .005 0 0. 1894Dec.04%) (0. In Table 3.(R m R f ) Jan.04 (Avg.20%) (3.2 presents a fiveyear moving average of the estimated slopes and excess market returns.09 (Avg.21) In this table. R m R f =0.4.2: Sixty months moving average of the crosssectional slopes and excess market returns using Dimson beta estimates 0.
04) in the second subperiod as the difference in magnitude confirms (0. the average excess market return (0.3 Expected Returns. which uses the excess market returns. size. a 66 . and Chan and Chen (1988) first documented the size effect in modern data. Fama and French (1993) build a threefactor model. the propensity for large stocks to have consequent lower returns than small stocks.17%). For the first subperiod. Reinganum (1981). Early works of Banz (1981).02%). and the Size Effect This section examines the wellknown size effect on the 19th century BSE. The null hypothesis of the equal average cannot be rejected. Beta. The finance literature uses the threefactor model as a benchmark model to measure longrun abnormal returns and many other factors. On the contrary. The last column shows the tstatistics for the test of equality of the average slope and average excess market return. Subsequently. (1983). This shows that researchers and practitioners have accepted size as an important characteristic to explain the crosssectional behavior of longrun stock returns. Chan and Lakonishok (1993) document similar results with post1920 Amex and NYSE data and caution researchers and practitioners not to rush into discarding beta. Chen and Hsieh (1985). Chan. In contrast.03% average slope and 0. 3. Fama and French (1992) present evidence that size and booktomarket combine to capture the crosssectional variation in average stock returns in the period 19631990.20% average excess market returns). the null hypothesis that the average slope is equal to the average excess market returns is rejected (tstatistic of 2. That is. and booktomarket factors.04%) is very close to the average slope (0.and French (1992) crosssectional regressions using the Dimson beta. The average slope is significantly less than the average excess market return (a difference of about 0.
As in the previous sections. The market capitalization is measured as the price of stock multiplied by shares outstanding. we sort (univariate sort) stocks based on their size (or market capitalization) at December of the prior year and then split them into decile portfolios. Monthly portfolio returns are calculated as the valueweighted and equally weighted averages of the individual stock returns within each of the ten portfolios. We compute the relative percentage size of a portfolio as the time series average of the crosssectional sum of the market size of the stocks in the portfolio divided by the sum of the size of stocks in our sample. T is the number of years. FM breakpoint method is employed to group the stocks into decile portfolios. With historical data. The relative percentage of markets size is computed as 67 . Schwert (2003) uses US monthly returns data between the year 1962 to 2002 to document that the size effect disappears after 1981. Again. It conjectures the magnitude of size effect is not robust when the transaction costs and very small stocks (the removal of stocks with market capitalization less than $5 million) are taken into account. Grossman and Shore (2006) do not find any presence of the size effect on UK data between the years 1870 to 1913. We present similar evidence on the 19th century BSE covering the same period. Loughran and Savin (2000) presents evidence against the size effect in the USA market. if nt is the number of stocks in a portfolio for the month t . That is. Portfolios are rebalanced each year to capture changes in their constituent stock market capital over time.recent paper by Horowitz. This would imply size is not a systematic risk factor. and the largest size stocks are put in decile ten. Each year. N t is the number of stocks in the crosssection of our sample for the month t . the smallest size stocks are put in decile one.
89 1.45 1.45 2.12 0.02 0.97 2.16 1.12%) which is almost three times the next largest excess return (0. Two beta estimates of the size portfolios are calculated using equation (9) and (10) in the previous section.We test the hypothesis with and without the 1st decile portfolio.14 tstatsitics (3.26 2 0.13 1.54 1. 1868.38 2.12 1.43 0.86 3 1.15 0.43 1. However. portfolios ten contains the largest size stocks.79 In this table.61 1.16 2.43 10 53.84 1.67 8 10.95 0.91 9 15.56 0.73 6 5.37 1.16 1.29 0. Relative market size is reported in column 2.35 1.43 1.43 1.68 mean of hedge portfolio (%) 0.5: Beta Estimate and Mean Excess Return for the BSE equally weighted size portfolios.65 3. stocks are ranked each year based on their size at the end of the prior year.04 0.74 1. We also retport the standard deviation of the portfolio return series.26 1.01 1.21 1.56 6.61 1. Table 3.06 0. They are then grouped deciles for portfolio formation.16 1.94 0.88 0.09 1.45 3.63) Fstatistics with the first decile 4.16 1.17 0.57 5 3.Dec.36 1.42 4.46 1.10 0.16 1.10 1.01 1.1 T % Market Size = iNt1 T t 1 j 1 nt Sizeit jt Size 100 .13 2.06 0.93 5. The negative relation between size and returns is concentrated in the first decile 68 .20 0.37 4 2.52 1. this is not the case when size portfolios are value weighted in our sample.17 1. EW=equallyweighted and VW=value weighted.16 0.12 0.16 1.12 1.67 3.28 3.60 0.30 1.52 3.38 0.04 Fstatistics without the first decile 0.14 1.35 1.43 3.79 0.5 reports the beta estimates and the average excess returns of the tensize portfolios during the period 1868 to 1914.16 5. Portfolio one contains the smallest size stocks.10 1.71 2.16 7 7.56 0. The Fstatistic for the test of hypothesis of equal mean of the porftolio returns is also reported.41 1.74) (0.46 0.77 2. Portfolios are rebalancec each year.19 1.38% from portfolio 6).84 0. It is well known in empirical finance that small stocks have both a higher beta and average return than large stocks.Average excess returns of the decile portflios are reported in column 3 and 4. We also report the Dimson and market model betas for the decile portfolios. Table 3. Column 3 reveals that equally weighted portfolio 1 has an extreme average excess return (1. 1913 R p R f EW VW Standard Deviation Size Portfolio % Market Size EW(%) VW(%) β MM β dim β MM β dim EW(%) VW(%) 1 0.14 0.10 1.17 2. Jan.35 1.
In addition.74) shows that the size effect exists in our data.10 0. the value increases to 0.28 0. In addition.05 0.95% and tstatistic 3.17 0. 1890 and 1905.6: Equally weighted portfolios excess returns without the firstsize decile group Size Portfolio 1 2 3 4 5 6 7 8 9 10 mean of hedge portfolio tstatsitic R p R f (EW%) 0.11 0.35%) of these stocks shows that they are likely so illiquid that it would have been difficult to profit from buying them. Table 3. Looking at the Fstatistics.14 (0. the extreme excess returns of 1.14 0.39 0.63) for the valueweighted hedge portfolio. As the data source has been well checked for outliers. The average excess return of the equally weighted hedge portfolio (mean excess return of 0.12 0.19 0.12% recorded for the first size decile is not due to data error. Surprisingly. the relative market capitalization (0. the null hypothesis of equal average returns is rejected at the 1% significant level when the first decile is included in the test. Excluding the first decile portfolio fails to reject the hypothesis.21 0.80) 69 . Not able to study the events in 1880. the stocks in the lowest size portfolio exhibit very high returns in these years.14% (tstatistic of 0. the effect disappears when stocks are value weighted in the portfolios.portfolio as the average excess return sharply drops from portfolio 1 to portfolio 2.
3: Size Portfolio betas 2 1. (2000). For robustness.Recently. we eliminate the stocks in the firstsize decile each year and perform the size sorting analysis.35%) before the size portfolio formation every year.8 1.2 Beta 1 0. This corroborates Horowitz et al. As mentioned earlier.8 0.4 plots the MM MM and the Dimson dim betas with the onemonth lag for each equally weighted size portfolio. Fama and French (2008) used USA data from 1963 to 2004 to document that the size effect owes much of its power to micro caps and that it is marginal for small and big caps. the size effect disappears when we eliminate the first decile portfolio (portfolio with relative market size of about 0.4 0.6 0. Clearly. As shown in Table 3.6 1. the difference between the MM and dim progressively gets smaller as stock size gets larger.4 1.2 0 1 2 3 4 5 6 7 Market Model Betas Dimson Betas 8 9 10 Size Portfolio 70 . who find no size effect in the period from 1963 through to 1981 when they eliminate firms with less than $5 million in market value on the USA market. Figure 3. Grossman and Shore (2006) find similar results for the UK market in the same period of our study. Figure 3.6.
This is to curb the possible underestimation of smallstock beta. the previous section reveals that size is crosssectionally correlated with beta. they expect the betas of size portfolios to be strongly correlated crosssectionally with size.3. the correlation will increase the standard errors of the estimates. 3. t is the full period estimate of beta for portfolio. However. when both characteristics are included in a regression. Chan and Chen (1988) argue that as size serves as a proxy for betas. Ibbotson et al. Fama 71 . This might be due to nonsynchronous trading as chapter two reveals that some stocks show lead or lag relationship with the market returns.1 FamaMacBeth CrossSectional Regressions to Test the Size Effect In order to support the above evidence on size effect. respectively. we resort to the FM crosssectional regression method adopted by Ibbotson et al. (1997) find similar results on the USA market between the years 1926 and 1994. We regress the crosssection of excess returns for a given month on the beta estimate (full window beta estimate) and natural logarithm of size by using an extension of equation (11): R jt R ft 0t 1t jt 2t ln Size jt 1 t (12) where 0t .79 with a pvalue of 0.4). In our sample. They recommend the inclusion of lagged information of market returns in the estimation of beta. and this will make the outcomes murky for interpretation.This shows that small stock betas are underestimated when estimated with the market model. We also recommend the use of the Dimson beta with the onemonth lag when estimating betas for small stocks in our sample. In addition.0065) between size and portfolio beta (Figure 3. (1997). There is a clear negative correlation (0. 1t and 2t are the regression intercept and slopes for month t .
This is to confirm the effect of the correlation between size and beta on the betareturn relationship. As in the sorting method. The procedure generates fifteen sizebeta portfolios for each beta estimate. However. For all portfolio formations. we regress portfolio excess returns on beta and the natural log of size by 72 . The conditional doublesorting portfolio formation is repeated at the end of each year. We achieve this by conditional double characteristics sorting. We show that when equally weighted portfolios are built on size alone. it breaks the effect of size and beta even on equally weighted portfolio excess returns.and French (1992) show that when portfolios are formed on size alone. The size effect disappears when stocks are valueweighted in portfolios. To separate the effect. The correlation between size and beta makes the test on size portfolios unable to disentangle the effect of size and betas on average returns. there is support for CAPM. Specifically. Each month. we use the FM breakpoint technique. we first sort stocks based on size and then sort within each size group on preranking beta. Postranking betas are estimated with postranking returns over the entire period from 1868 through to 1914. We find a strong relation between size and average excess return but no relation between beta and average return for equally weighted portfolios. there is evidence of a positive relationship between average return and beta (CAPM). allowing the variations in beta that are unrelated to size. Each size group is then sorted into five groups based on their preranking MM or dim beta estimates. we sort stocks into three size groups each year. we form decile size portfolios. The equally and valueweighted return for each portfolio is computed for each month of the following year. The size effect does not exist when we eliminate the firstsize decile portfolio in the analysis each year.
36% (0.7: Average time series slopes and intercept from the FamaMacBeth crosssectional regression: Jan 1868Dec.31% (0.89) 0.86) Panel B1: Sizeβ mm Portfolios 0.42% (0.05) 0.06) 0.12) 0.63% (2.62% (1.20% (0.46) VALUE WEIGHTED Intercept β MM β dim ln (Size) Panel A2: Size Portfolios 0.79) 0.72) (0. we sort stocks into ten portfolios based on their size at the end of the prior year.12% (2.43) 0.64) 0.86% (3.17% 0.45) 0. tstatistics are in parenthesis.10) 0.34% (0. In all portfolio formations we use the FM break point. Post ranking betas are used in the crosssectional regression.62% (1.using equation (12).69% (2.09% (1.35) 0. Size is determined at the end of the year before the portfolio formation year.74) 0.00% (1.72) 1.61) Panel C1: Sizeβ dim Portfolios 0.64% (2.10% (0.62) (3.56) 0.97) 2.08) 0.56) 0. The full period postranking betas are used in the crosssectional regressions.09% (2.32) 0.47) 0.03% (0. Table 3.23% 0.03% (0.88% (3.02% (0.20) 0.35) (0.02% (0.45% (1.71) 2.76) 0.03% (0.91) 1.85% (2.03% (0.12% (1. 73 .56) Panel C2: Sizeβ dim Portfolios 0.69% (2.09% (2. Estimate postranking betas by using the full period postranking excess returns.17) 0.08% (1.67) 1.09) Panel B2: Sizeβ mm Portfolios 0.18) 0.12% (0.05% (3. This will yield 15 sizebeta equally and value weighted portfolios.09% (0.18) (0.55) 2.15) 0.29% 0.09) 0.63) 1.27% 0.10% (0.07) 0.91) 1. Equally and value weighted portfolio returns are computed each month in the year.09% (2.12% (2.03% (0.81) 0.06% (0.08% (3.01% (0.10% (2.50) 0.06% 1.87) 0. 1913 EQUALLY WEIGHTED Intercept β MM β dim ln (Size) Panel A1: Size Portfolios 1.83) (1.31) Each year.13% (2.72) 0.63) 0.50) 0.51) (0.01% (0.09% (2. The joint effect of size and beta is seperated by first forming three size portfolios and splitting each size group into five beta groups.05) 0.01% (2.05% (1.22% (0.86) 2.40) 0.26% 0.
The average 74 . Both MM and dim are positively related to excess return when placed alone in the crosssectional regression. size is sometimes positively but insignificantly related to excess returns when placed simultaneously with beta in the regressions. where size is significant when placed together with market model betas in the regression. from Panel A1. both the market model and the Dimson beta with the onemonth lag can predict returns at the expense of size. This is contrary to Ibbotson et al. The values in Panel A1 show that the CAPM is valid for equally weighted univariate sizesorted portfolios. However. The time series standard deviations of the slopes and the intercepts are used to test whether the average is significantly different from zero. Size is negatively related to excess returns.Table 3. The Shanken‟s adjustment factor is also adopted in the computation of the tstatistics. the average intercept is significantly negative when MM and the Dimson‟s betas are used in the crosssectional regressions.7 reports the time series averages of the slopes and intercept of the regression. When equally weighted portfolios are formed on size alone. (1997) result. The negative intercept becomes significantly positive when size is placed alone as a regressor in the crosssectional regression. We use Newey and West (1987) heteroskedastic autocorrelation corrected standard errors for the computation of the tstatistics (reported in parentheses). Interestingly. When size and any of the beta estimates are placed simultaneously as independent variables. only the beta estimate is significantly related to excess returns. The implication of the CAPM requires that the average crosssectional regression intercept will not be significantly different from zero.
but beta will not. Both betas are no more significantly related to returns. This suggests that the result from the equally weighted portfolio is due to the influence of small stocks since it assigns equal weights to all stocks in portfolio formations and in the crosssectional regressions. and C2). Assigning full period postranking betas to stocks does not mean a stock‟s 75 . It is worthy to note the significance of the intercepts. Panels A2.5. whether placed alone or with size in the regressions. size effect does not exist when stocks are value weighted in portfolios. These show that beta and size cannot combine to capture the crosssectional variation in stock returns. There is a statistically significant relationship in size to excess returns. the postranking betas estimated with the full period postranking returns will be assigned to each stock in the portfolio. At the end of each year. Panels B1 and C1 show the crosssectional regression slope and intercept for conditional doublesorted sizeβMM and sizeβdim portfolios respectively. Most interestingly. when the valueweighted portfolios are used in the analysis (be it univariate size sorting or conditional double sizebeta sorting). beta or size is not significantly related to the average excess return (See Table 3. We repeat the above analysis by adopting Fama and French (1992) method. Therefore. This is in support of Fama and French (1992) evidence that the conditional doublesort portfolio (sizebeta sort) allows variations in beta that are unrelated to size and would break the combined of size and beta on expected returns. This confirms the sorting result in Table 3.intercept is not significantly different from zero when the MM and Dimson‟s betas are placed together with size in the crosssectional regression. size will be related to average returns. whether placed alone or with any of the beta estimates.7. B2.
58) (0.15) 0.65% (2.37) (3.16% (3. we sort stocks into ten portfolios based on their size at the end of the prior year.48) Panel C1: Sizeβ dim Portfolios 0.26) (0.15) 0.54% 0.16% (3.17% 0.17% (2. 76 .54) 2.63% (3.42) 0.31% 0.70% 0.69% (3.12% (2.58) (3.15) 0.62) 0.63% (3.86) 0.16% (3.29) 2.20) 0.00% (2.07% (0.35) 0.78% (3.63% (3.16% (0.05% 1.54) (0.26% 0.08% (1.55) 0.16% (3. 1868Dec.43% (2.04% (1.91) Panel B1: Sizeβ mm Portfolios 0. Portfolios are rebalanced each year.47) 2.29) 2.44) (0.43) 2.55) Each year.04% (3.58) 0.07% (3.16% (3. The method uses the information available for individual stocks in the cross section.85% (3. Estimate postranking betas by using the full period postranking excess returns.57) 0. We assign postranking betas to the constituent stocks in the portfolio.15% (2. In all portfolio formations we use the FM break point. Equally and value weighted portfolio returns are computed each month in the year.63% (3.05% (0.46) 0.10% (1.15) 0.77% (3.59) 0. This will yield 15 sizebeta equally and value weighted portfolios.86% (2.16% (3.63% (3. Table 3.57) (2.76) 2.88) 1.50) 0.27) Panel B2: Sizeβ mm Portfolios 0.48% (2.44) 0.16% (3.15) 0.29) 1.06% (0.16% (0.03) 2.21% 0.16% (3.05% 0.29) 2.58% (2.44) VALUE WEIGHTED Intercept β mm β dim Panel A2: Size Portfolios 0.16% (3.29) 2.10% 0. as stocks can move across portfolios with yearly rebalancing. tstatistics are in parenthesis.53) (0. The joint effect of size and beta is seperated by first forming three size portfolios and splitting each size group into five beta groups.46) 0.79) ln (Size) 0.51% (1.39) 2.63% (3.22) 0.8: Average Time Series Slopes and Intercepts from the FamaFrench CrossSectional Regressions: Jan.beta is constant.38% (1.29) 2.34) 0.52) 0.15) 0.27) (0.87% (3.37% (1.07) 2.03% (2. 1913 EQUALLY WEIGHTED Intercept β mm β dim ln (Size) Panel A1: Size Portfolios 1.28) Panel C2: Sizeβ dim Portfolios 0.
size and beta have no relationship with return as shown in panels A2. conditional doublesorting returns based on size and betas break the hold up between size and beta. serves as a proxy for size when placed alone in the regression. double sorting stocks to form portfolios will not establish the relationship between betas. From Panels B1 and C1. which is correlated with size. They lose their relationship when placed together with size in the regression. When valueweighted portfolio returns are used in the analysis. For robustness. The result is similar when valueweighted postranking returns are used to estimate postranking betas (See Panels A2. but exclude stocks in the firstsize decile each year. when the fullperiod equally weighted portfolio returns used to estimate postranking betas are formed on size alone. In Panels B1 and C1. This shows that any size effect present in our data is driven by a small group of stocks with an average relative market size of about 0. 77 .35%. It can be seen that beta has no relationship with excess return when it is placed alone or together with size.8 reports the average slopes and intercepts of the crosssectional regressions using equally weighted and valueweighted portfolios to estimate the postranking betas. and C2. size. we repeat Fama and French (1992) crosssectional analysis. both MM and dim have a strong relation with returns when placed alone in the regression. From Panel A1. The values in parentheses are the Newey West adjusted tstatistics for the test of a hypothesis of the average slope or intercept significantly different from zero. and returns. As in the sorting method. B2.Table 3.9 does not show the significant relationship between betas and expected returns when placed alone or combined with size for singlesorted size portfolios in panel A1. This indicates that beta. B2 and C2). Table 3.
00% (0.06) 0.67) 0.10% (2. 1868Dec.09% (2.00% (0.22) 0.67) 0.28% 0.27) 0.28% (0.74) Panel B2: Sizeβ mm Portfolios 0.01% (0. This will yield 15 sizebeta equally and value weighted portfolios.09% (0.70) (0.28% 0.06) 0.20% (0.73) (0.22% (1.20% (0.07) 0.10% (0.41) Panel C1: Sizeβ dim Portfolios 0.55) 0.37) ln (Size) 0.10% (0.92) (1.27) 0.27% 0.86) 0.22) 1.20% (0. 78 .60) 0.82) (0.55) 0.21% (0.02% (0.06) 0.24) 0.52% (0.20% (0.65) 0.07) 1.10) 0.02% (0.17% 0.20% (0.67) 0.09% (0.20% (0.06) 0.30) 0.51% (1.10) 0.37% (1.27) 0.77) 1.22% (1.41% (0.06% (1.76) 0. We assign postranking betas to the constituent stocks in the portfolio.49) Each year.39% (1.00% (0.05) Panel B1: Sizeβ mm Portfolios 0.91) 0.28) 0.57) (0.11% (0.50) 0.10% (0.Table 3.21% 0.08% (2.62) Panel C2: Sizeβ dim Portfolios 0. we sort stocks into ten portfolios based on their size at the end of the prior year.1913 EQUALLY WEIGHTED Intercept β mm β dim ln (Size) Panel A1: Size Portfolios 0.27) 0.54% (1. In all portfolio formations we use the FM break point.25) VALUE WEIGHTED Intercept β mm β dim Panel A2: Size Portfolios 0.00% (0.61) 0. Equally and value weighted portfolio returns are computed each month in the year.00% (0.28% 0.55% (0.08% (2.98) 0.04% (0.89) 0.88% (0.06% (1.00% (0.09) 0. tstatistics are in parenthesis.68) 0.65% (1.79) 0.37% (1.72) 0.31% (0.08% (1.00% (0.9: Average Time Series Slopes and Intercepts from the FamaFrench CrossSectional Regressions without the FirstSize Decile: Jan. The joint effect of size and beta is seperated by first forming three size portfolios and splitting each size group into five beta groups. Portfolios are rebalanced each year. Estimate postranking betas by using the full period postranking excess returns.83) 0.28) 0.85) (1.27) 0.34% (0.
beta. Estimating beta with the market model. we find the relationship between excess returns and size or beta. when placed at the same time as regressors in the crosssectional regression. Dimson model. We find that conditional doublesorting portfolios by size and then by beta breaks the effect between size and beta on excess returns.4 Conclusion We used sorting and the crosssectional regression method to investigate whether the CAPM model is valid in the period before World War I. and size. when we use equally weighted size portfolios in the crosssectional regressions.35% of the total market size.3. Both the sorting and the crosssectional regression methods reveal that the size effect disappears when the valueweighted portfolios are used in the regression. 79 . Further investigation reveals that the size effect in our data is mainly due to small stocks with relative market size of about 0. However. We find no support for the CAPM in the 19th century BSE. We recommend researchers estimate betas with the Dimson method with a onemonth lag since smallstock betas are underestimated when estimated with the market model with this data. but beta does not relate to excess returns. and Vasicek model does not establish the crosssectional relationship between expected returns and the beta. There is a negatively significant relationship of size to excess returns (size effect). the average slope of the crosssectional regression of returns on betas becomes insignificant when placed alone in the regression or combined with size. As a result. Eliminating these small stocks destroys the relationship between excess returns. This is due to a strong correlation existing between size and beta.
the source of the momentum profit poses a 80 . Rouwenhorst (1998) documents the momentum effect in 12 European equity markets.). It produces significant profit in the following 6 to 12 months. there has been extensive literature published confirming the robustness of the momentum effect. The strategy buys stocks that have performed well in the past 6 to 12 months and sells stocks that have performed poorly in the same period. However.CHAPTER 4 4 DOES THE MOMENTUM EFFECT EXIST IN THE 19TH CENTRURY? Fama and French (2008) document that the momentum or relative strength strategy produces an average return that is robust and significant across different size categories of stocks. Though researchers and practitioners have subscribed to the view that the momentum strategy yields a significant profit. Europe and Emerging Markets (see Jegadeesh and Titman (1993). In this chapter. This will serve as an outofsample and robustness test of the postWWII momentum profits documented. Subsequent to this. Their result confirms the pervasive. lagpredictable patterns of stock returns documented by Jegadeesh (1990). 4. Nijman. positively significant. we use our preWorld War I BSE data to test the presence and source of momentum effect between the years 1868 and 1914. Rouwenhorst (1999). For example. most studies testing the returns momentum effect in 6 to 12month horizons use postWorld War II (WWII) data from the USA. Swinkels and Verbeek (2004) and etc.1 Introduction and Literature Review Jegadeesh and Titman (1993) document that portfolios or stocks that have performed well in the past 3 to 12 months often continue to deliver high returns in the subsequent 3 to 12 months.
Jegadeesh and Titman (2002) attribute the results found by Conrad and Kaul (1998) to a small sample bias in their empirical test and bootstrap experiments. While some argue that the momentum profit is due to crosssectional variations in expected returns. They find that the profit is not necessarily smaller in the subsamples. Conrad and Kaul (1998) and Bulkley and Nawosah (2009) argue that the momentum profit is due to the crosssectional variation in the expected returns. but not to autocorrelation in stock returns.strong challenge in the literature. Jegadeesh and Titman (1993) compute momentum profit within size and betabased subsamples with lower dispersion in expected returns. On the contrary. the same returns observation in the formation and the holding periods. Jegadeesh and Titman (2002) adopt sampling without replacement to mitigate the bias selection of a particular observation and realized that the crosssectional variations in expected returns contribute very little of the momentum profits. To examine the possibility that the profit can be explained by the crosssectional variation in expected returns. Based on this result. data mining bias. the state of the market and the behavioral biases of investors. Autocorrelation is the correlation between two returns observations of the same returns series at different times (time series pattern in stock returns). The crosssectional variation of expected returns explaining momentum implies that the returns of the momentum portfolio will be positive on average in any postranking period. To counter the 81 . they conclude that the momentum profit is not due to the crosssectional variation in expected returns. others believe the profits may be due to data mining. The literature attributes the source of the profit to crosssectional variation in expected returns. However. The bias is due to the likelihood of drawing with replacement.
Likewise. as aggregate overconfidence should be greater following market gains. In addition. which are easier and less expensive to trade. They find momentum in the first 12 months after portfolio formation. Lesmond. However. Cooper et al. (2004) apply the behavioral theory of Daniel et al. Despite the existence of momentum profits on different markets and in different time periods. they find that the momentum effect still exists in their extended data. which is consistent with the behavioral theories. and that the profit found in Jegadeesh and Titman (1993) is therefore not likely to be a data mining bias. Chabot et al. On the relationship between the state of the market and momentum. This is because the strategy tends to pick stocks that 82 .data mining explanation. momentum profit is just an illusion. especially for large size stocks. Jegadeesh and Titman (2001) extend the data from the period 19651989 to 1998 and perform an outofsample test. They find that momentum profits depend on the state of the market. it will allow them to assess whether investors have changed their investment strategies and if the profit of the momentum strategy does not exist anymore. Extending the data enables them to assess whether the result in the period 19651989 is merely due to chance. Schill and Zhou (2004) argue that when transaction costs are considered. (1998) to predict the differences in momentum profit across different market states. (2009) use 19th century data (Victoria Era) from the UK to document the dependence of momentum profit on the state of the market. the cumulative return in months 13 to 60 after portfolio formation is negative (return reversal). when a three year lag in average return is used to define market states. like a bull versus a bear market. However. It is imperative that executing the momentum strategy involves frequent trading.
value. Since the primary motive of this chapter is to investigate whether momentum profit exists in the 19th BSE. 83 . Frequent trading in these high cost stocks will prevent gainful strategy execution. The literature also shows that momentum is strong in industry portfolios than individual stocks. The only difference is that Jegadeesh and Titman (1993) used decile portfolios in their analysis. In contrast. future research can be conducted on this data to determine whether the momentum profit is due to individual stock or industry effect. while we use quintile portfolios in order to have a sufficient number of stocks in our portfolios. The result is not influenced by forming the portfolios just after the formation period or by skipping one month after the portfolio formation period. In this chapter. We find investors can earn significantly positive returns over 3 to 12month holding periods when they adopt the momentum strategy. we investigate whether a 3 to 12months momentum effect existed in the 19th century and early part of the 20th century BSE using the methodology similar to Jegadeesh and Titman (1993). the crosssectional dispersion in expected returns and possible microstructure effects.have a high trading cost. Nijman et al. Moskowitz and Grinblatt (1999) document a strong and persistent industry momentum effect that cannot be explained by size. individual stock momentum. For example. Finding a momentum effect in the 19th century casts additional doubt on the datamining explanation for momentum. (2004) documents that the momentum profits on the European markets are mainly due to individual stock effects.
When we replicate the postholding period event analysis documented by Jegadeesh and Titman (2001). the profit does not remain positive after 12month holding periods. There is evidence of a sharp rise in profit in the first year. (2004) to predict the cyclicality of the momentum profit across different market states. (2001) who find negative momentum profit in January (a reversal) each calendar year. We can conclude that in the 19th century.A detailed look into an extensively researched momentum strategy (6 months formation period and 6 months holding period) on the 19th century BSE reveals that the strategy tends to pick small size stocks. we find a strong shortterm momentum profit in the first 12 months and long run reversal. except the years between 1878 and 1887. In Subsection 4.1. two years after portfolio formation. We also investigate the momentum profit across market states. However. we will focus on a sixmonth formation/ sixmonth holding period strategy.2. The rest of the chapter is organized as follows: we explain and compute the returns of the various combinations of momentum strategies in Section 4. The momentum strategy yields a positive return in all months throughout the year. Subperiod analysis also shows that momentum profit is robust in all tenyear subperiods. This strategy has 84 .2. Unlike Jegadeesh and Titman (1993). as in the US. We follow Cooper et al. There is no return reversal in the first month after the portfolio formation. as claimed by Conrad and Kaul (1998). further tests also suggest that momentum exists in beta and large size subsamples. but the profit declines and reverses to become negative in the second to fifth year after the portfolio formation. we do not record a January reversal in the 19th century BSE. We find that momentum profit depends on the state of the market when three years lagged value market weighted returns are used to define market states. as it may be consistent with many behavioral explanations.
1. we form quintile portfolios instead of the decile portfolios formed by Jegadeesh and Titman (1993) and Bulkley and Nawosah (2009). we investigate the source of the momentum profit by studying the characteristics of the momentum profit after the 12month holding period. The momentum strategy and its profitability are constructed in this section as follows: At every given month t .4. 4. Section 4.5. we rank stocks in ascending order based on their previous P months (formation period) compound returns. consisting of monthly prices.6 concludes our findings. We compute average returns (equally weighted and value weighted). In section 4. this chapter uses the 19th century BSE data. In order to obtain sufficiently diversified portfolios.2. five equally weighted and value weighted portfolios are formed. total returns (returns adjusted for stock splits and dividends). and outstanding shares listed in Brussels between the years 1868 and 1913.2 Momentum Trading Strategies and their Returns As in the previous chapters.been extensively researched in the literature to represent the other strategies.2. We study the momentum profit across different market states in section 4. Based on these rankings. As can be seen from Figure 4. 85 . the number of stocks in our sample varies from 71 in 1868 to 513 in 1913. The number of stocks increases sharply from 1890 to 1913. We document the returns of the strategy in size and beta based subsamples in Subsection 4. and the average market capitalization of the quintile momentum portfolios. Section 4.3 investigates seasonality in the momentum profits across calendar months and tenyear subperiods. The weights are determined by the market capitalization of each stock in the portfolio at the end of the portfolio formation.
This yields 16 strategies each for equally weighted and valueweighted portfolio formation. where P and Q equal three. or we skip one month after the formation period. We update our portfolio formation every month. the strategy that selects stocks 86 .2: Time line of sample periods Formation Period Holding Period tP t t+Q The bottom quintile portfolio is called “Loser. The strategy holds the position for Q months (holding period).” The momentum strategy is to buy the winner portfolio and sell the loser portfolio in each month t .Figure 4. Either the holding period is immediately after the formation period. six. We calculate holding period returns from the same month in which the stocks were ranked in order to form portfolios.1: Number of common stocks in our sample for the momentum studies 600 500 Number of Stocks 400 300 200 100 0 1860 1870 1880 1890 year 1900 1910 1920 Figure 4. Specifically. Figure 4. Various combinations of P and Q are considered." and the top quintile portfolio is called “Winner. nine and twelve months.2 shows the various periods we considered in the strategy.
we construct 6/6 strategy as follows: at the end of each month we sort stocks in our sample based on their past six month returns (month 6 to month 1) and group the stocks to form five portfolios based on the ranks.1 presents the monthly average returns of the equally weighted and valueweighted portfolios of the various strategies. This will be referred to as P/Q/1 strategy. In Panel B. t7 to t2. We form equally weighted and valueweighted quintile portfolios from the returns of stocks that coexist in the same ranking. we also calculate returns for skipping one month between the formation period and the holding period. In order to improve the power of our test. and it will continue over the period from January to June.based on the past Pmonths return and holds it for Qmonths is known as a P/Q strategy. the winning portfolio in month t will contain the top quintile of stocks ranked over the previous t5 to t. t6 to t1. and it will continue until t11 to t6. in all the portfolio formations. Table 4. we report both the equally weighted and valueweighted average returns for strategies with holding periods starting immediately after portfolio formation. We resort to the FamaMacBeth method for the portfolio break points. over the period from January 1868 to December 1913 on the BSE. the previous May to October. Portfolios are held for six months (month 0 to month 5) following the ranking months. For a comparative purpose. we report equally weighted and valueweighted average portfolio returns for strategies with holding periods of one month after the portfolio formation. the return in December of the winning portfolio will contain the top quintile of stocks ranked over the previous June to November. for the 6/6 strategy. To illustrate this on the calendar months. Specifically. That is. 87 . we replicate the overlapping portfolio method adopted by Jegadeesh and Titman (2001). In Panel A.
The most profitable zerocost strategy selects stocks based on their previous 6 month returns and holds the portfolio for 6 months. yields almost the same profit as the 6/6 equally weighted strategy (0. the average return of the equally and valueweighted zerocost (WL) portfolios for all strategies are positive and significant.70% per month). 88 . This suggests that momentum on the 19th century BSE is mainly due to large size (market capital) stocks.71% and 0.69 % per month) when there is a time lag between the portfolio formation period and the holding period. the average returns of the equally and value weighted zerocost portfolio for all strategies are positive and significant. The tstatistics based on NeweyWest heteroskedasticity and autocorrelationadjusted standard errors are shown in parentheses. the average return of the value weighted zerocost portfolio is higher than the equally weighted zerocost average return. For all strategies. except the equally weighted average return of the 3/3 strategy. It yields 0. the profit for the latter strategy declines (0.75% per month for the equally weighted zerocost portfolio.For each cell in Table 4. the loser and the zero cost (winner minus loser) portfolios of the various strategies. we report the average returns of the winner. The 9/3 strategy. with an equally weighted portfolio formation.06% each for both 6/6 and 6/6/1 strategies. which is marginally significant. For the valueweighted portfolio formations. In Panel B. This is contrary to the results found by Jegadeesh and Titman (1993). which is marginally significant. it yields 1. However. when portfolios are formed immediately and one month after the formation period respectively.1. In Panel A. except the equally weighted 12/12/1 strategy. where the most profitable strategy was a 12/3 strategy.
16) (0.31) 12 EQ VW 0.57) (5.58) (7.14% (2.29) (0.75% (4.57% 0.32) (1.25) (6.40) (7.78% 0.48% 0.77) (7.47% 0.33% 0.46) 0.49) (0.48) 0.98% (7.88) 0.88% 0.67% (2.63) (7.92% 0.68% 1.23% (3.07% 0.95) (6.55% 0.80% (7.63) 0.20% (3.49) (1.60) 9 EQ VW 0.33) 0.04% (3.56) 0.93% 0.29) (0. Different values of P and Q are shown in the second row and first column respectively.06% (4.37) (8. portfolios are formed immediately after the portfolio formation and in Panel B they are formed one month after the portfolio formation.87% (4.44% 0.25) (6.72% (6.78) (6.66) (6.37% 0.42% 0.84) (7.17) (2. In Panel A.99% (7.67% 0.91% 0.23) 1.51% 0.53% 0.42) (4.23% 0.97% (6.39% 0.11% (4.47% 0.48% 0.58% 0. For the value weighted portfolios.04% 0.04% 0.87% 0.25) (7.90% 0.96% (4.10) (8.81% 0.52% 0.57) (0.15% 0.49% 0.68) (5.55% 0.89) (4.02) 0.40% 0.12) (7.51) 0.17) (6.09) (7.36) (7.08% 0.04% 0.83% (6.86% 0.38% 0.76% 0.74% 1.72) (1.00% 0.92) (0.22) (1.80% 0.23% 0.33% 0.81% 0.19% (2.85% (7.26) 0.91% 0.12) (0.69) (0.05) (2.06% (2.87) (5.04) 0.46) (5. Equally weighted and value weighted quintile portfolios are formed.67) 0. The zerocost (WinnerLoser) portfolio is the portfolio formed by going long on the winner and short on loser portfolios.81% (5.86% 0.43) (7.31) (3.12) (1.71) (5.83% 0.79% (4.02% 0.12) 6 EQ VW 0.68) (0.29% 0.04) (0.44) (7.81% (6.44% 0.17% (2.89% 0.35% 0.76% (4.52% 0.42) 0.75) (5.41) 0.94% (7. Each month t.57) 0.72) (6.79) 1.15) 0.56) (8.12% 0.93) 0.06% (2.25) 0.69) (5.22) (7.MacBeth break point method in the portfolio formation. The top quintile is called "Winner" portfolio and the bottom quintile is called "Loser" portfolio.71) (5.72) 12 Winner Loser WL 0. stocks a ranked based on Pmonths compound returns.04% 0.40% 0.61) 0.74% (5.97) 89 . stocks are weighed by their market capital.67% (4.22) (8.35% 0.04% (2.46) (7.72) (8.03% 0.31) (7.07% (1.95) 9 EQ VW 0.67) (0.43) 0.08% 0.63) (1.78% (7.32) (8.80% (7.99) (7.51) 0.08% 0.97% 0.05) (7.76) (0.91% 0.42% 0.48% 0.59) 0.60% (2.46) 0.88) (6.81) (2.Table 4.35) (5.23) 0.44) (6.76) (0.62) 0.84% 0.37) (4.37% 0.75% 1.73% (4.52) (0.50% 0.44) 0.32) 0.44) 0.06% (2.16) (3.05) (7.414 0.58) (7.10) (0.33% 0.12% (2.20) (7.1868Dec.87% (7.75% (2.04) (0.44) (2.71% 0.18% 0.29) 9 Winner Loser WL 1.06% (5.02% (1.12% (1.80% (4.72) 3 EQ VW 0.97) 0.46% 0.07% (2.41) (7.94% 0.78) (8.40% 0.81% (7.16) (7. one month before the holding period.65) 0.90% (6.83) (1.28) 0.50) (6.44% 0.69) 0.88% 0.53% 0.49) (1.06% (4.31) 0.72) (6.44) 0.135 0.40% 0.16) (9.12) 0.38 1.64% 1.17) (0.57% 0.19) (6.13) (1.95) (4.12% (2.90% (6.79% (5. We adopt the Fama.35) (2.08% (2.33% 0.25% (3.88% 0.39% 0.92) 0.37% 0.38% 0.99% 0.25) (0.82) (6.05% (4.19) 0.73) (4.07) (1.75% (6.44% 0.62% 0.73) (8.10) (7.06% 0.43) (7.81) (7.44) (6.03% 0.20) (0.74% (5.51% 0.92% 0.29% 0.64% 0.80% 0.23) (2.66) (5.09) (5.40) (7.82% (5.52% (2.84% 0.02) 0.77% (6.58% 0.69% 1. Panel A: Holding Period Formation Period Portfolio Winner Loser WL Panel B: Holding Period 3 EQ VW 0.63) 0. Qmonths return are calculated based on the ranking.05% (1.98) (1.49) (0.02) 0.60% 1.92) 0.72% (5.14) 12 EQ VW 0.25) (7.26) 0.52% 0.11% 0.86) 3 6 EQ VW 0. EW= Equally Weighted and VW= Value Weighted.06) (6.62% (2.64) 0. 1913) This table reports the returns from momentum strategy based on BSE returns data.79% (6.71% 1.77) 0.99% (6.89) (7.86) 0.98% (5.87% 0.30% 0.60% 0.88) (5.29) 0.03% 0.70% 1.35) (0.50) 0.26) 0.70% 0.92) (7.90% 0.18) 0.48) 0.87% (5.89% 0.35% 0.15% 0.33) (5.95% 0.36% 0.24% 0.56) (2.07% (2.91) (1.53% 0.83% 0.54) (1.35) 0.1: Profitability of momentum Strategies on BSE (Jan. The monthly average of the three portfolios are reported with their NeweyWest heteroskedastic and autocorrelation adjusted standard error tstatistics.32) (2.82% (6.71) 0.50) 0.89) 0.85) (0.13) (0.31% 0.99) (7.98% (6.27) 0.98% 0.47% (1.94% 0.28) (6. W=Winner and L=Loser.41) 0.60% (3.08) (6.42) 0.33% 0.04% (1.53) 6 Winner Loser WL 1.84% 0.23) 0.
All equally weighted strategies yield profits around 0.50%, regardless of the holding period and of the onemonth lag between the formation period and holding period. It is out of the ordinary to note that the positive average returns of the equally weighted loser portfolios are usually significant (except the 6/3, 6/6, 6/9, 9/3, 9/6, 6/3/1, 6/6/1, 6/9/1, 9/3/1 and 9/6/1 strategies, which are marginally significant). This shows that the positive returns of the zero cost portfolios are mostly due to the positive significant returns of the winner portfolios and positive significant returns of the loser momentum portfolios. In effect, 19th century investors who would find it difficult to go short on loser portfolios could have profited from the momentum effect by going long on the winner portfolios. In contrast, the average return of the value weighted loser portfolio for all the strategies is not significant, may even be negative sometimes. The strategy that has been extensively researched in the literature is the 6/6 strategy. After confirming that momentum strategies of all combinations would yield significantly positive returns on the 19th century BSE, the rest of the chapter will focus on the 6/6 strategy (this strategy will be analyzed, and the result will be used to represent the other strategies that comprise formation and holding periods ranging from three to twelve months).
4.2.1 Expected Returns and Average Size of Quintile Portfolios
Table 4.2 shows the portfolios‟ average returns (equally weighted and value weighted) and average market capitalization (price multiplied by the shares outstanding) of the quintile portfolios. As before, we construct portfolios by adopting the overlapping method. For equally weighted portfolios, the average returns increase from the losing
90
portfolio to the winning portfolio, but there is a marginal drop in average returns from the loser portfolio to the second portfolio.
Table 4.2: Average Returns and Average Size of Quintile Momentum Portfolios
This table reports the average returns and average market capital of the quintile portfolios of 6 /6 months momentum strategy. The market capital is the price times the number of shares outstanding. Average market capital is the average of the holding period market capitals of the stocks in each portfolio. Quintile portfolios are equally weighted and value weighted. The sample period is January1868 to December 1913. Market capital is in millions. W=Winner , L=loser, EQ=Average returns of the equally weighted portfolios and VW is the Average returns of the values weighted quintile portfolios. Average Market Capitalization Portfolio EQ VW (Million Belgium Franc) Loser 0.33% 0.07% 3.80 2 0.26% 0.11% 9.23 3 0.51% 0.43% 11.76 4 0.76% 0.70% 11.28 Winner 1.04% 0.98% 7.50 WL 0.71% 1.06% 
In contrast, the average return of the valueweighted portfolios monotonically increases from the losing portfolio to the winning portfolio. The values in the last column of the Table indicate that past losing and winning portfolios pick stocks that have low market capital on the average. This result is consistent with the findings of Jegadeesh and Titman (1993) and Chabot et al. (2009).
4.2.2 Momentum Profit within Size and Betabased Subsamples
In order to investigate whether the profit from the momentum strategy is not confined to any particular group of stocks, we follow Jegadeesh and Titman (1993) to examine the profitability of the strategy on subsamples. We group stocks into subsamples based on the beta and size. Size may serve as a proxy for liquidity and beta for volatility. The subsample analysis of the momentum profit also provides evidence about the source of the momentum profit. It tests whether the momentum profit is due to the crosssectional
91
variations in expected returns or predictable patterns in the time series of stock or portfolio returns. If the profit is due to the crosssectional variations in expected returns, the profit will be reduced in the subsamples, as the dispersion of the crosssection of expected returns is lower in the subsamples than in the full sample. If the momentum profit is due to predictable patterns in individual stock returns, then the subsamples will yield positive profits, as will the entire sample.
Table 4.3: Portfolio Returns of the Momentum Strategies with Size and Beta Subsamples
This table reports the average monthly returns of portfolios formed based on size and beta subsamples. Portfolios are formed based on six month returns and held for six months. The overlapping method is employed in the portfolio formation. Loser/winner portfolio is the equally weighted portfolio of stocks in the lowest/top quintile when stocks are ranked over the previous six months returns. We adopt Fama Macbeth method for all portfolio breakpoints. "Micro" contains the smallest stocks, "Small" contains the next smallest and "Big" contains the largest stocks. β 1, β2 and β3 the lowest to the highest beta subsamples. Size is determine at the beginning of each formation period. We estimate betas for stock returns prior to the portfolio formation period. Dimson's method one month lag is applied to stocks with at least 24 month returns within the five years prior to portfolio formation to estimate betas. The last two rows show the difference in mean and their Newey West tstatistics between full sample profit and each subsample profit.The sample period is January 1968 to December 1913. W=Winner and L=Loser Average Monthly Returns Portfolio All Micro Small Big β1 β2 β3 Loser 0.33% 1.01% 0.04% 0.09% 0.31% 0.42% 0.53%
(1.52) (3.19) (0.19) (0.69) (1.72) (2.14) (1.80)
2 3 4
Winner WL
0.26%
(2.03)
0.47%
(2.17)
0.35%
(3.01)
0.41%
(3.96)
0.32%
(3.89)
0.41%
(3.37)
0.39%
(1.89)
0.51%
(4.58)
0.58%
(3.61)
0.60%
(5.22)
0.54%
(5.48)
0.48%
(8.63)
0.64%
(5.56)
0.58%
(3.16)
0.76%
(6.85)
0.92%
(5.92)
0.75%
(6.34)
0.70%
(7.61)
0.74%
(8.05)
0.80%
(7.23)
0.77%
(4.36)
1.04%
(7.50)
1.30%
(7.12)
0.98%
(7.07)
0.82%
(7.02)
0.98%
(7.93)
1.06%
(7.69)
1.06%
(5.69)
0.71%
(4.84)
0.29%
(1.17)
1.01%
(7.55)
0.73%
(6.26)
0.67%
(5.09)
0.64%
(4.16)
0.53%
(2.84)
Difference

0.42%
(2.86)
0.30%
(3.35)
0.02%
(0.18)
0.04%
(0.32)
0.07%
(0.68)
0.18%
(1.77)
Table 4.3 presents the average return of momentum portfolios, momentum profits and the difference in the mean of the full sample profit and each subsample profit. Portfolios are
92
Not finding momentum in our micro size stocks support Hong. The momentum strategy is applied on beta and size subsamples. the beta subsamples show no significant difference in profit from the full sample. The nonexistence of momentum in the “Micro” is mainly due to the extreme average return recorded by its loser portfolio (average return of 1. As in our sample.01% with a tstatistic of 3. Like Chabot et al. The values in the last two rows show the average of the difference in profit between the full sample momentum profits and the subsample momentum profits. (2009) also find the largest momentum profit for the middle size group. it is positive but not significant. Clearly.01% with tstatistics of 7. (2009). who find momentum profit in all the small size subsamples. We estimate beta by using the Dimson method with a onemonth lag. The result also deviates from Fama and French (2008) who find significantly positive momentum profit for small stocks. (2009) we do not find a strong relationship between size and momentum profit. the momentum strategy is profitable in all subsamples except the “Micro” size subsamples.formed based on the 6/6 strategy. This indicates that the profit does not exist for the “Micro” size subsample of stocks. the result does not support Jegadeesh and Titman (1993) and Chabot et al. it is worthy to note the high profit (the profit of 1. The analysis includes stocks with at least 24month returns within five years before the formation month.19). Lim and Stein (2000) results. They used NYSE and AMEX data between 1980 and 1996 to document that momentum does not exist in their smallest size decile stocks. Consistent with the previous chapter.55) for the “Small” size subsample. The tstatistics for the test of a hypothesis that the difference is equal to zero is reported in parentheses. The “Big” size subsample also shows no significant 93 . However. Chabot et al. On the contrary. Size is determined at the end of each formation period.
94 . The expected return of the past winner portfolio is significantly positive. and the average return of the past loser portfolio is not statistically significant. It is important to note that the profit of the “Big” size subsample is mainly due to the buy side of the transaction rather than the sell side.4 reports the momentum profits in all calendar months on the Brussels stock exchange between January 1868 and December 1913. they infer that most of the negative profits recorded in the month of January by Jegadeesh and Titman (1993) are due to small and low priced stocks. This implies that even when investors are not allowed to short stocks. as the profit is not reduced significantly in these groups. The difference in momentum profits between the full sample and the subsamples (except the “Micro” and “Small” size subsamples) indicates that the crosssectional difference in the expected returns of the stocks may not determine the momentum profits. This will test the possible seasonal effect of the momentum profits in the 19th century. 4. as Jegadeesh and Titman (1993) document positive momentum profits in all calendar months except January.3 Seasonality and Subperiod Analysis of the Momentum Profit We are motivated to look into the seasonal patterns in momentum profits. they can still earn a momentum profit by buying winners. which are likely to be difficult to trade. If anything. Therefore. However. Table 4. Jegadeesh and Titman (2001) also confirm the January reversal (negative) returns in later data. the predictable patterns in individual stock returns may contribute to the momentum profits in the beta subsamples.difference in the mean with the full sample. the reversal becomes marginal when they exclude stocks with a price lower than $5 per share and stocks in the smallest decile in their sample.
May June July Aug.20% 0. related tstatistic and the percentage of positive profits is also reported for each size subsample.44) 87 (5.56% (2. "Small" contains the next smallest stocks and "Big" contains the largest size stocks.39% (1.81) 85 0.52% (1. FebDec Fstats All 0.08) 89 0.97) 80 0.95% (11. Feb.00% Jan.31) 76 0.84) 63 0.3: Average returns of the momentum profit in all calendar months 1.88% (2.49% 0.18) 83 0.69) 65 (3.93% 0.72) 74 (2. Dec.42) 85 0.65 Big 0.50% 0.73) 72 (2.00% Momentum Profits 0.65% 0.30% (4.15) 65 0.96) 54 0. The average return.41) 85 0.66% (2. Oct.76% 0.08) 65 0.34% (0. Mar.56% (2.95) 78 (2.17% (0.10) 72 0.36) 61 0.42) 52 (1.91) 85 1.37) 87 (4. Again.68) 85 0.96) 59 0.93% (4.96% (3.45) 85 1.90% (4.32) 83 0.21% (0.The last column report the Fstatistics for the test of hypothesis of equal average profits in the various calendar months.14) 83 0.34% (0.29% (5.85) 59 0.18) 63 0.80% 0.76% 0. Newey West standard error adjusted tstatistics are in parenthesis.89) 83 0.07% (3.23) 83 0. May June July Aug.89% (6.70% 0. Feb.96% (3.03) 50 0. The equally weighted portfolio returns formed from the past return top rank quintile stocks is called Winner.68) 76 0.04) 72 0.Table 4.75% (4.87% 0.02) 76 (4. their related tstatistics and the percentage of momentum profits that are positive in calendar month. Sep.46) 76 (1.34% (3.78) 80 0.59% (3.52% 0.41 (4.37% (0.96) 67 (8.70% 0.80) 67 (2. The equally weighted portfolio is formed from the lowest past return decile and is called Loser.11) 59 0.25) 78 0. Size is determined at the end of each portfolio formation period.16% 0.38% (0. Apr.66% (2.63% 0.00% 0.56% (3. this table reports the average momentum profit.97% 0. T his figure reports the average returns of the momentum profit by calender month.96% (5.75% (12."Micro" contains the smallest stocks. Mar.94% (5.39) 78 (1.95) 74 (2. we form momentum portfolios based on the 6/6 strategy.72% 0. We also report the 95 . It also shows the percentage of positive profits in all the calendar months.69% (3.81 Figure 4.40) 89 1.61) 78 (3.62% (3.11 Small 1.26) 85 0.89% (3.51) 80 0.92% (3.46) 67 0.84) 74 Micro 0.73) 80 (1.47% (1. Dec.87% (3.94% 0.4: Seasonality in momentum profits For all months in the period 18681913.35% (0. Nov. Nov.42) 61 0. Oct. Months Jan.28) 76 0. Sep. The momentum portfolios are formed based on past 6 months returns and held for 6 months.40% 0.60% 0.49) 80 0. The zerocost portfolio is the winning portfolio minus the losing portfolio return in each month. Apr. The zero cost portfolio is the Winner minus Loser portfolio.20% 1.
As shown in Figure 4. The same reason applies to the BSE. with tstatistic of 4. Specifically. Size is determined at the end of each portfolio formation period.5 96 .average return. corresponding tstatistics and the percentage of positive momentum profits in each calendar month for size subsamples. (2009) record similar results in almost the same period (Victorian Era) in the UK. The absence of the January effect on 19th century BSE is not surprising. “Micro” is the lowest tercile group for size.02 and 76% positive profits) after February and March. This shows that there is no significant evidence of difference in average momentum profit for the calendar months. “Small” is the middle tercile group for size.3. January records the third highest momentum profit (0. The winner portfolio returns exceed the loser portfolio returns in all months.87% average return. the profit decreases gradually from January to August but picks up again from September to December. We compute the Fstatistics in the last column under the null hypothesis that the average returns on the zero cost portfolios are equal in all calendar months. The results from the 19th century BSE do not support the negative momentum profit found on the USA market by Jegadeesh and Titman (1993) and Jegadeesh and Titman (2001). as Chabot et al. The Fstatistics in the last column indicate that the null hypothesis of equal profit in all calendar months cannot be rejected for the entire sample and the size subsamples. This is supported by a significant percentage of positive momentum profits in all calendar months. and “Big” is the highest tercile group for size. Comparing the profits in all calendar months for the various size subsamples. the “Micro” size group records the lowest profits in all months. Jegadeesh and Titman (1993) also recorded the lowest profit in August. They argue that investors. were not taxed on capital gains and that the tax year does not end in December. as of that time. Table 4.
58) 0.27) 0.Dec. The zero cost portfolio is the winner minus loser portfolios.14% (0.09) 1.87% (9. Sample Month 18681877 18781887 18881897 18981908 19081913 All 0. Table 4.55) 0. although still positive.16% (0.Dec.62% (6.13% 0.26) 0.55% 0.85% (11.55) 0.68) 0.43% (1. Feb.73) 0.36) 0.00) Micro Jan.19% 1.57) 0.06% (0. Feb.28% (3.Dec.88) 0.17% (8.21% (9.Dec.95) 1.09) 0.54% (2.43) 0.53% (4.75% (2.08) Small Jan.68% (2.70% (6.69% (2.17% (0.23% (6.72) 1.46) 2.67% (1.13) 0. 0.36% (1.34% (0.87% (2. 1913.31% (1. 1.16) 0.83% (7.18% (0.87) All 0.Dec. Feb.51% (3.08) 0.03) All Jan.30% (1.94) 0.30% (0.02) 0.18) 1.96% (4.60) 1.29) 1.05% (6.96% (1.58) 0.32) 1.45) 97 .91) (0.99) 0.21) 1.84) 0.42% (3. Newey west standard error adjusted tstatistics is reported in parenthesis. 0.96) 0.22% (3."Micro" is the subsample which contains the smallest stocks and "Big" contains the largest size stocks.86) All 0.52) 0.60% (3.76% (7.73) 1.76% (11.25% (7.25) 0.5: Subperiod Analysis of Momentum Profit This table reports the average monthly returns of the zerocost portfolios in ten year subperiods and the last five years before World War I.05% 1.29) 0.73) All 0.70% (11.33% (3.48) 0.17) 1. 1968.The average return of the zerocost portfolio formed using sizebased subsamples of stocks within subperiods is also reported.10) Big Jan.90% (4.75) 1.50% (0. The sample period is Jan.76% (2.86% (3.42% (2. 0.18% (0.26% (1.65) 0. This table shows that the momentum strategy yields positive significant profit in all subperiods except in the period 1878 to 1887.60) (4.59% (1.86% (9.81) 1.18) 0.66) 0.54) 0.17% (6. The zerocost portfolios are formed based on six month past returns and held sixm months.66) 0.09% (11.56) 1.11% (0.82) 0. We sorts stocks in ascending order based on past six months returns and equally weighted portfolios are formed from lowest quintile stocks group.33) 0. This group is called the sell portfolio and equally weighted portfolios formed from the top quintile group of stocks is called Winner portfolio.29% (0.88) 0.89) (4.17% (0.57) 0.72% (6. Feb.52% (3.48) (5.documents the zerocost portfolio for the 6/6 month strategy in tenyear subperiods and the last five years before World War I.30% (2.
It is important to note the significant average returns in January and outside January for “Micro” size subsamples in 1888 and 1897 period. In addition. there seems to be strong evidence of past losers exceeding past winners in the first month after portfolio formation (shortterm reversal). This may counter the assertion that momentum profit does not exist in small stocks for the entire period (the assertion is not robust over time). the time series‟ abnormal holding period return (momentum) is due to investor delay in overreaction to information. past winners continue to exceed the past losers for two to twelve months after portfolio formation (intermediate term continuation). there is a significant positive profit for January and outside January in all the periods for the “Small” and the “Big” size subsamples. and small stocks will contribute to the greater part of the momentum profit across the entire sample period. Except for the periods 18681877. and past losers will reverse to exceed past winners over three to five years. Jegadeesh and Titman (2001) and Chabot et al. it produces significantly positive profits in all subperiods except in the periods 1878 and 1887. 4. This shows that there is a momentum profit for small stocks in this tenyear subperiod. (2009)). we may find momentum profit for all size subsamples. In effect. 18781887 and 18881897. The behavioral explanation of momentum indicates that the time series' variation in individual stock or portfolio returns contributes to momentum profit. Conrad and Kaul (1998). which pushes the 98 .4 Post holding Period Momentum Profits The support for shortterm reversal.When the strategy is applied to the middle and largest size subsamples. intermediate term continuation and long run postholding period reversal of momentum profits has been extensively documented in the literature (Jegadeesh and Titman (1993). That is. If we base our research on this tenyear period.
86% in the fifth year. if the profit is due to crosssectional variation in expected returns.46% in the fifth year. On the other hand. reverse its sign. but not on the variation in the time series of prices for individual stocks in any particular period. in the periods following the holding periods considered in the previous sections.6 presents the average monthly returns for the first five years after portfolio formation. more than double its value in the first year. across all stocks. the profit should continue to increase after the formation period with time. investors force the prices to revert to their fundamental values. the winner portfolio returns drop from 0. From Panel A. However. Table 4. we expect profit to be reduced to zero overtime. Since the predictability under this hypothesis is based on the difference in unconditional drifts across stocks. if momentum profits are completely due to behavioral biases. the profits from the momentum strategy will continue to remain positive in any post formation period. Upon returning to their senses. we examine the returns of the momentum portfolios. The profit in the second year is approximately the same for the winner and the loser 99 . That is. and that every stock has a unique drift.prices of winners (losers) above (below) their fundamental values. the loser portfolio increases from 0. and if possible. In contrast. Empirical evidence presented by Bulkley and Nawosah (2009) support the assertion. Conrad and Kaul (1998) hypothesize that stock prices follow a random path with various drifts.91% in the first year to 0.38% in the first to 0. To differentiate between the behavioral bias hypothesis and the Conrad and Kaul (1998) hypothesis. They show that the difference in unconditional drifts across stocks explains the momentum profits. Conrad and Kaul (1998) argue that the crosssectional variation in expected returns generates momentum profit.
41) 0.43) 1.54) (5.52% (3.64% 0.37) (0.40% 0.64% 0. Portfolios are formed based on past six months returns.6: Long Horizon Momentum Profits This table reports monthly average momemtum profit for zerocost.03) Panel B: Micro 0.59% (4.43) 1.26% (4.43) Winner Portfolio 0. winner and loser portfolios.81% 0. Table 4. Size is the market capitalization of stock and is determined at the end of portfolio formation.70) (3.27% 0.88) 1.06) 0.80% 0.24% (5. one to five years after portfolio formation. Zerocost portfolio is the winner minus loser portfolios.83) 0.15) Winner Portfolio 0.03% 1.02% (0.54% (4.37) 0.90) (4.85) (4.59% (8.11% (5.66% 0.64) (4.86% (5.15% 1. "Micro" is the lowest size group.73% (1.53% (7. From the third year to the fifth year.18) (4.62) (5.73% (8.72) 0.52% 0.94) Zerocost Portfolio 0. The strongly significant average returns of 0.07) (4.26% (3. 1913.53% 0. "Small" is the middle size group and "Big".46% (4.38% (2.97) 0. the zerocost (winner minus loser) portfolio profit becomes significantly negative.99% (5. 1868 to Dec.61) 0.95) (5.91% (7.16) Loser Portfolio 0.31% (2. The sample period is Jan.portfolios.27% (10.92% (4.31% from the second through the fifth year confirms the return reversal.34) (3.25) 0. as the loser portfolio returns exceed the winner portfolio in these periods.35) 0.46% 0. We repeat the portfolio formation for size subsamples.83% (3.95) (4.50) 0. the largest size group.83) 1.41) 1.30% (1.11) 0. Stock are sorted based based on the pased six months returns and equally weighted portfolio is formed from the lower and the upper quintile group of stocks.09) 0.46) Loser Portfolio 1.66) (5.50% (2. The portflio formed from the upper group of stocks is called the "winner portfolio" and the portfolio formed from the lower group is called "loser portfolio".13% (7.46% (3.23% 0.76) 0.79) 100 . Month Months Months Months Months Months Months Portfolios 1 112 1324 2536 3748 4960 1360 Panel A: All Zerocost Portfolio 0.23) (2. Newey West standard error adjusted tstatistics are reported in parenthesis.65) 0.02% 0.90% (10.53% 0.08) (3.29) (6.50% 0.
Table 4.6 Continued
Portfolios
Panel C: Small
Month 1 0.90%
(4.85)
Months Months Months Months Months Months 112 1324 2536 3748 4960 1360 Zerocost Portfolio 0.77% 0.27% 0.02% 0.25% 0.18% 0.09%
(8.02) (2.76) (0.23) (2.46) (1.59) (1.81)
0.82%
(4.64)
0.78%
(6.56)
Winner Portfolio 0.58% 0.48% 0.46%
(4.81) (4.13) (3.37)
0.46%
(4.16)
0.53%
(7.91)
0.08%
(0.36) Panel D: Large
0.01%
(0.06)
Loser Portfolio 0.31% 0.51% 0.71%
(2.09) (3.11) (4.01)
0.63%
(3.54)
0.62%
(7.42)
0.70%
(4.22)
0.61%
(8.27)
Zerocost Portfolio 0.17% 0.10% 0.13%
(2.31) (1.59) (2.04)
0.21%
(2.76)
0.03%
(0.95)
0.82%
(5.03)
0.76%
(7.29)
Winner Portfolio 0.51% 0.50% 0.43%
(4.62) (4.61) (4.30)
0.44%
(4.32)
0.50%
(8.28)
0.12%
(0.83)
0.15%
(1.39)
Loser Portfolio 0.34% 0.40% 0.56%
(3.14) (3.74) (4.73)
0.65%
(5.09)
0.54%
(7.74)
Figure 4.4: Cumulative Returns for Five years after portfolio formation
Figure 4.4: Cumulative Returns for five years after portfolio formation
0.80% 0.70% 0.60% 0.50% 0.40% 0.30% 0.20% 0.10% 0.00% 0.10% 0.20% 0.30% Event Months 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53 55 57 59
It is important to note that there is no shortterm return reversal in the 19th century BSE, as the momentum profit one month after portfolio formation is positively significant (0.50% with tstatistic of 2.66). Figure 4.4 depicts the cumulative momentum profits for
101
Cumulative Returns
five years after portfolio formation. For our 1868 to 1913 periods, the graph shows that the momentum profit increases monotonically from the first month to the seventh month. The profit declines from the seventh month to the twelfth month, when it falls below its first month value. As shown in Table 4.6, the momentum profits become negative around the second year after portfolio formation, but the negative profit is not strong enough to offset the positive profit recorded in the first year. This keeps the cumulative profits above zero until the fourth year, in which the negative profits exceed the positive profits. Therefore, the cumulative profit falls below zero on the fourth year after portfolio formation. In effect, there is a strong return reversal from the second to the fifth year after portfolio formation on the 19th century BSE. This result is consistent with the behavioral explanation found on the USA market by Jegadeesh and Titman (2001). However, similar analysis on size subsamples (Panels B, C and D) shows that the strong longterm return reversal recorded for all stocks is mainly due to the lower long run return reversal in the “Micro” size group of stocks. For the “Micro” size stocks in Panel B, the first month after portfolio formation records an insignificantly negative momentum profit. The momentum profit one year after portfolio formation is significant (0.30% with a tstatistic of 1.94). There is a strong return reversal from the second to the fifth year (0.53, tstatistic of 7.83). From Panels C and D, the first year records a very strong momentum profit, but the profit reverses from the second year to the fifth year, and it is not significant (0.09 and 0.03 with tstatistics of 1.81 and 0.95 respectively). The evidence that the return reversal draws its power from small size stocks cautions us to interpret the behavioral model explanation of momentum with care. In general, the result is still not consistent with the argument by Conrad and Kaul (1998) either.
102
4.5 The Momentum profit and the Market State
In this section, we investigate whether the condition of the state of the market can determine the profitability of the momentum strategy. As indicated in Cooper et al. (2004), the behavioral theory can be extended to predict differences in momentum profit across the state of the market. The confidence of a group of investors is expected to be higher following market gains (Daniel et al. (1998) and Gervais and Odean (2001)). The upwards adjustment of market prices will tend to be attributed unduly to investor skill. This will result in greater aggregate overconfidence, as aggregate investors hold the long positions in the equity market. The high overconfidence in following up markets will trigger strong overreactions, and it will eventually lead to shortrun momentum. We follow Cooper et al. (2004) and Chabot et al. (2009) in defining two market states: for each month t, the UP (DOWN) market is when the lagged threeyear value weighted market return is positive (negative). We compute average momentum profits across market states by taking the momentum profit (winner minus loser quintiles) for each formation month, and taking the average across all formation months that qualify for a particular market state. Our sample period is still January 1868 to December 1913. For each month t, we sort stocks into a quintile portfolio based on the past sixmonth return. We adopt the portfoliooverlapping method in all portfolio formations. We do not skip one month between the formation period and the holding period, since the previous section shows no sign of return reversal in the first month after the formation period. The Fama and MacBeth (1973) breakpoint method is used to form portfolios. The holding period profits are computed for three horizons: t to t +5, t to t +11 and t +12 to t +59.
103
three years. The momentum portfolio is formed based on past six month returns and held for six months. Market is classified as UP (DOWN) if the value weighted market index over the period t 36 to t 1 is positive (negative). Each month.27% (7. To corroborate the 7 The result is not robust when less than 3 years return is used to define the market states.65 Table 4.60%. etc.15% (0. In Panel A.59 3.88% 0. Panel A and B reports the momentum profit following UP and DOWN markets. for the 6 and 12 months holding periods. two years. The profit of the momentum portfolio (winner minus loser quintiles) is cumulated across the holding periods : months t to t +5. 104 . t to t +11 and t +12 to t +59 if t 1 is the end month of the formation period.85) Panel C: Test of Equality of the Mean Momentum profit (UP=DOWN) 4. stocks are sorted in ascending order based on past six months returns. The sample period is January 1868 to December 1913.We report the mean monthly profits. (2009) find similar results. The DOWN market in panel B shows that momentum profit for 6 and 12 month holding periods are not significantly different from zero.7 reports the mean of the momentum profit for both UP and DOWN markets in the period January 1868 to December 1913.15% 0.65% 0.7: The Momentum profit and the Market State This table reports momentum profits for the UP and DOWN market states. N is the number of observation for each state of the market.61) (6. Cooper et al. we report the UP and DOWN market momentum profit for the threeyear definition of market state.) different definitions of the market state.Table 4. Holding Periods Months t to t +5 Months t to t +11 Month t +12 to t +59 Panel A: UP market states Market States Definition 3years N 424 424 376 Mean Momentum profit 0.7 For brevity. when the three year lagged value weighted returns index is used to define the market state.43% (4. we investigate the profit of the 6/6 strategy on three (one year. and are all above 0. (2004) and Chabot et al.42) (7. In order to verify whether the number of months used to define the market state has any influence on the results.07) Market States Definition N Mean Momentum profit tstatistics Panel B: Down market states 3years 128 128 0. the UP market momentum profits are significant.28 4. Equally weighted quintile portfolios are formed. NeweyWest tstatistics are in parenthesis.62) 128 0.37) (0. The First quintile portfolio is called the loser portfolio and the top quintile is called the Winner portfolio.
105 .20% over the holding period of 13 to 60 months. (2004) find similar results.unconditional results from Lee and Swaminathan (2000). We also find significant longrun reversal for DOWN markets. we test the hypothesis of equals in the mean of the momentum profit between the UP and DOWN markets. Chabot et al. The hypothesis is rejected at 5% for the fiveyear holding period horizons considered after portfolio formation. we find that UP market momentum profit significantly reverses over the long run. In Panel C. Cooper et al. (2009) did not find a significant difference in the mean between the UP and DOWN momentum profits. and they assert that long run reversals are not solely due to corrections of prior momentum. The average momentum profits are significantly below 0. (2004). Jegadeesh and Titman (2001) and Cooper et al.
January records the third highest momentum profit relative to the other months of the year. the momentum profits for beta and large size samples are not significantly different from the full sample profit. the marginal profit recorded in the first ten years is mainly due to the middle and the large size groups of stocks. as a tenyear subperiod analysis indicates that the profit is not strong in the first twenty years of our sample. The momentum profit is not robust across our sample period.6 Conclusion We investigate the trading strategy that buys past return winners and sells past return losers (momentum trading strategy) over the period of January 1868 to December 1913.4. we resort to the Jegadeesh and Titman (2001) approach to examine the returns of the momentum portfolios in the postholding 106 . The profit is marginally significant or not significant in the first and the second ten years respectively. To investigate the source of momentum profit.7% per annum. Unsurprisingly. There is convincing evidence that the momentum strategy is profitable. Detailed analysis of the 6/6 strategy on beta and size subsamples shows that momentum profit is not confined to particular beta subsamples. Investigating the momentum profit for each calendar month shows that the negative January momentum profit found on the USA market by Jegadeesh and Titman (1993).52% and 12. the 6/6 strategy that we study in detail here yields an average profit of about 8. We find no momentum profit for small size stocks. Except for the small size sample. For instance. Finding momentum profits in the 19th century provide some evidence that the profits found for the postWorld War II US market are not mainly due to data snooping biases. since the average momentum profit for the small size group is insignificant. for equally and valueweighted portfolios. (2001) is not in the 19th century BSE. respectively.
as aggregate overconfidence 107 . Further study of the size subsamples revealed that the negative profit recorded in the months 13 to 60 after portfolio formation is mainly contributed by the small groups. Daniel et al. We also test if momentum profit and longrun reversal in the crosssection of stock returns depend on the state of the market. Cooper et al. the returns of the momentum portfolio will decay to zero. This evidence supports the behavioral explanations and clearly rejects the Conrad and Kaul (1998) assertion.83). like UP and DOWN markets.period. Our evidence supports the Jegadeesh and Titman (2001) results. if the profit is due to lagged overreaction. Conversely. The behavioral explanations suggest that. (1998) and Hong and Stein (1999). Conrad and Kaul (1998) assert that the momentum profit is due to the crosssectional differences in expected returns.31 with tstatistics of 6. Shleifer and Vishny (1998). not to the time series' pattern in asset returns. (2004) document that the Daniel et al. (1998) behavioral theory can be extended to predict differences in momentum profit across states of the market. and suggests that the profit of the momentum portfolio in the 13 to 60 months after the portfolio formation is significantly negative (0. although the result is still not consistent with the Conrad and Kaul (1998) assertion. if momentum profit is completely due to under reaction to information. Therefore. We study postholding period returns to differentiate between the assertion by Conrad and Kaul (1998) and behavioral explanations presented by Barberis. the momentum strategy will continue to yield positive returns in all periods after the holding period. the momentum portfolio will possibly reverse in sign before decaying down to zero overtime. This suggests that the support for the behavioral explanations of the momentum profit should be interpreted with caution.
the DOWN markets in the same period record insignificant positive momentum profits of 0.88% and 7.15% (tstatistic of 0.42 tstatistic. from 1868 to 1913.65% and 7. We also find that the momentum profit in the UP market states is reversed in the long term.should be greater following market gains. we define the state of the market as UP (DOWN). 0. the average monthly momentum profit for UP market states is significantly positive (0. The 6 month formation and 6 to 12 month holding period strategies are solely profitable in the periods of market gains. For each month between 1868 and 1913.62 tstatistics for both 6 and 12 months holding periods respectively).37 and 0. In contrast. Specifically. 108 .62 for both 6 months and 12 months holding periods). when three year lagged value weighted returns are used to define the market state. if the lagged threeyear returns of the value weighted market index are positive (negative).
The results may also provide evidence on the combined effect of different characteristics on the crosssection of stock returns. 2011.Georgia. total risk and dividend yield. USA (Blind Reviewed) 109 .1 Introduction and Literature Review The Capital Asset Pricing Model (CAPM) of Sharpe (1964). 5. the winners (losers) over the past 3 to 12 months predict the subsequent 3 to 12 month winners (losers) for the 19th century BSE data (momentum). 15th April. However. accounting for about 0. TOTAL RISK AND MOMENTUM (18681913 EVIDENCE)8 In the previous chapters. stock returns do not show a positive relationship with beta (CAPM) in the 19th century BSE data. Savannah. momentum. the relationship is completely driven by extremely small stocks. Size seems to have a relationship with average returns in our data. Eastern Finance Association 2011 Annual Meetings. Lintner (1965) and Mossin (1966) provides a particularly appealing way to look at the crosssection of expected returns. The next section explains the motivation for repeating size and momentum and choosing total risk and dividend yield. The model implies that expected returns are a linear function of the stock‟s beta coefficient. However. In this chapter.CHAPTER 5 5 THE COMBINED EFFECT OF DIVIDEND YIELD.e. 2011.35% of the total market capitalization. SIZE. 8th April. It 8 Different versions of this chapter have been presented at conferences: the 14 th Conference for Swiss Society for Financial Market Research (SGF). we provide outofsample evidence that may address the datamining critique. Zurich Switzerland and (Blind Reviewed). By studying the predictive power of these characteristics on stock returns in a new database. we use the 19th century BSE data to study the predictive pattern of stock returns based on size. i. the regression slope of the stock‟s return on the market return.
Another explanation is that they reflect irrational investor behavior that a rational investor could exploit. we tested the 110 . net stock issues and accruals. Hanna and Ready (2005) demonstrate the vulnerability of some patterns to turnoverinduced transaction costs. Not only is there no longer a positive relation between beta and average stock returns. which is the only part of its total risk that should be rewarded in the market. In chapter 1. as summarized in Fama and French (1992). Although initially the model was corroborated empirically (e. Fama and MacBeth (1973)). For instance. as the patterns are often stronger for smaller and less liquid stocks and exploitation involves high portfolio turnover. such as size (market capitalization) and its booktomarket ratio. its prediction is not borne out in more recent empirical research. Later research adds even more characteristics that are associated with average returns. This then raises the issue of to what extent these patterns are tradable.measures the security‟s systematic risk.g. In this chapter. Fama and French (2008) dissect the return patterns based on momentum. One possibility is that the characteristics are proxies for exposure to common risk factors. we shed some light on this issue by studying crosssectional patterns in Belgian stock returns for the period of 18681914. These patterns beg explanation. the patterns could be illusory or simply found thanks to collective data dredging (see e. Schwert (2003)). For instance. asset growth. as well as size and booktomarket. This unique and highquality data set provides a genuine outofsample testing environment.g. the returns are also found to be related to other firm or stock characteristics. Finally. profitability. which then leads to the question of which factors should be studied and how to measure them. in which case the patterns should not be discernable in new datasets.
momentum. Horowitz et al. Given the evidence that the size effect seems to have attenuated since the publication of Banz (1981) (see Schwert (2003)). Fama and French (1992) show that booktomarket and size are the most important stock characteristics related in the crosssection to average returns. as 19th century stock markets were less liquid than their modern counterparts. but to no avail. In addition. We pay particular attention to the computation of beta. (2000) list three potential explanations: (a) data mining. We repeat size in order to investigate whether the relationship between stock returns and the other firm characteristics will not be confined to the extremely small stocks. beta and size are of no importance in explaining average returns. namely size. finding a size effect would favour the awareness explanation. results of Fama and French (1992). representing on average about 96% of stock market capitalization. the relation in preWorld War I Belgian data is flat.S. whereas not finding a size effect is more consistent with the data mining argument. which would have driven up prices of large companies. we cannot include the booktomarket ratio. Also. total risk and dividend yield. As index funds did not exist in the 19th century. We find that. 111 . but we do include the size (market capitalization) of the firm in our analysis. (c) the awareness of investors after publication of the research results has eliminated the profit opportunities. We now turn our attention to the test of the presence of some repeated and other characteristic effects. (b) the increased popularity of passive investments. As we do not have accounting information. similar to the post1969 U. additional robustness analysis. within the largest group of stocks. using a new and independent dataset may shed light on the explanation of the size effect.predictive power of the beta coefficient for average stock returns.
there has recently been a debate on the usefulness of idiosyncratic or total volatility to predict average returns in the crosssection. Jegadeesh and Titman (2001) have updated their previous results and still consistently observe the momentum effect in the years 1965 to 1998. Malkiel and Xu (2006) find that portfolios of stocks with high idiosyncratic volatility have higher returns than portfolios with low volatility stocks. Although it is still not clear why momentum is positively related to future average returns. Ang et al. Including this characteristic in our tests yields additional outofsample evidence. it is complementary to the size and booktomarket effects (Fama and French (1996)). making it less likely that it is due to data mining. idiosyncratic risk and total risk should be priced. In contrast. Merton (1987) argues that when investors do not or cannot hold the entire market portfolio due to various exogenous reasons. Also. (2006) find a negative relation between average returns and idiosyncratic volatility. Blitz and Van Vliet (2007) based on Sharpe Ratio and alpha from the FamaFrench three factor model document a negative relationship between excess returns and total volatility on global markets. We follow Fama and French (2008) to repeat momentum in the analysis. In addition. Fama and French (2008) use a different measure of momentum in their dissection of the anomaly test and find momentum to be positively related to the average return.Since Jegadeesh and Titman (1993) published their article on momentum. However. Theoretically. or short to medium term past returns. Furthermore. Bali and Cakici (2008) show that the relationship between total risk and excess return is induced by methodology and the predominance of small illiquid stocks in the sample. On this note. Screening stocks using liquidity and 112 . this type of stock prediction has received a prominent place in empirical asset pricing.
following Fama and French (1992). Given this debate and the household income situation in the 19th century. Investors might not be able hold the market portfolio due to the high prices of stocks relative to the daily wage. In addition. idiosyncratic risk estimated from the standard deviation of the error term in FamaFrench‟s threefactor model may introduce errorinvariable complications in the crosssectional regression. However. may also prevent arbitrageurs from exploiting mispricing opportunities on the market. the investigation on the interaction between dividend yield (as a 113 . In addition. In addition. which is the main arbitrage risk. we follow Grossman and Shore (2006) and use dividend yield as the best available proxy for value. The above literature measures idiosyncratic risk as the standard deviation of the errors in the Fama and French (1993) threefactor model. stocks were expensive relative to the average daily wage (see Scholliers (1997)). total risk will have a positive relationship with mispricing (arbitrage limit theory). we decide to include total risk in our list of characteristics. Therefore. total risk could also be priced to compensate rational investors for their inability to hold the market portfolio. However. Total risk. as we do not have data on the booktomarket value of equity. Therefore. we cannot compute the value factor in the FamaFrench model as the accounting data has not been digitalized in our data. we considered total risk to minimize the error inherent in the estimation of risk. in the 19th century.price filters destroys the relation. Lastly. we would have investigated the relationship between booktomarket value and the crosssection of stock returns in the 19th century BSE. On this note. We follow Blitz and Van Vliet (2007) to measure total risk as the standard deviation of the past two to five years excess returns.
5. and the value strategy is less effective within the highest momentum quintile. Clare. We conclude the chapter with section 5. For example. Asness (1997) documents a negative correlation between momentum and dividend yield. Section 5. In addition. descriptive statistics of the characteristic sort portfolios are studied.6. as it has been shown in the literature that small firms find it necessary or desirable to pay dividends in the earlier periods. They advise value investors to stay away from high momentum firms until they exhibit at least some relative strength compared to the general market. we investigate the effect of the relationship between dividend yield and total risk on the average returns.4 describes the average returns of the singlesort and independent doublesort portfolios. and to some extent entails investment in poorly performing stocks and vice versa. We use the FM crosssectional regression method to confirm the sorting result in section 5. Seaton and Thomas (2009) use independent doublesort quintile portfolios to document that the momentum strategy yields a significant return among the lower value quintile. In view of this. 114 . we also investigate the relationship between momentum and dividend yield in the 19th century BSE. Gwilym. Similarly. The rest of the chapter is organized as follows: The next section describes the measurement of the characteristics. Moreover. we also investigate the relationship between size and dividend yield in the 19th century BSE.measure of value) and momentum has recently been documented in the literature. In section 5.3. Grossman and Shore (2006) also record that a negative relationship exists between dividend yield and momentum. On historical data. This implies investment in high value stocks.
Thanks to a change in legislation in 1867. the stock has to comply with the following requirements: 1. As part of our analysis relies on portfolio sorts (sometimes with double sorting). a minimum of stocks in the crosssection is needed. Portfolios are constructed in January of each year based on information available by the end of December of the previous year. which is computed as the difference between the realised return and the riskfree rate used in chapter three. As in chapter 3. We use excess return.1 illustrates our motivation. Six months‟ return data in the previous year is necessary to compute our momentum measure (discussed below). we form yearly portfolios. we pay more attention to zerodividend stocks. total risk and momentum. as they account for. it became much easier to set up a company. In order to study the relationship between dividend yield.5. to be included in the analysis. and the number of listed firms increased accordingly. size. 115 . More specifically. In order to obtain some accuracy in the estimation of total risk. Unlike Asness (1997). onaverage. Figure 5. more than a quarter of the stocks in the crosssection each year.1 shows the percentage of stocks in the crosssection that did not pay dividends each year. we divide our sample into three groups based on a characteristic and investigate the effect of dividend yield separately in each group. a minimum of 24 out of the 60 months‟ return observations are required. To study the pervasiveness of any pattern.2 Measures of Characteristics We follow two paths to establish the importance of the characteristics: portfolio sorting (mainly singlesort and independent doublesort) and FM crosssectional regressions. our sample starts in 1868 and Figure 3. 2.
1868. (2009) and Grossman and Shore (2006) record that small stocks do not pay dividends on the contemporary and historical UK markets respectively. rather than incorporating them into the lowest dividend group of stocks. a “Ushaped” relationship between dividend yield and return. Each year. The relative market capitalization of these firms does not exceed 20% across the period of our study. each year. On average. The result is not surprising.5. it must have at least 6 months of return data to facilitate the computation of the momentum characteristic. in order for a stock to be part of this analysis in the following year. As before.1913 Relative Market Capitalization 50 45 40 35 30 25 20 15 10 5 0 1860 % ZeroDividend Paying stocks Percentage 1870 1880 1890 year 1900 1910 1920 It is clear from the figure that about 45% of the stocks in the crosssection did not pay dividends in 1880 and or in 1903.1: Percentage of Zerodividend paying stocks and their Relative Market Capital: Jan. We treat the zerodividend stocks as a group. We do this because both Keim (1985) and Morgan and Thomas (1998) had a comparable return between the zerodividend firms and the highest dividend yield firms.Dec. Momentum is 116 . we measure the size (market capitalization) as price multiplied by shares outstanding in December of the year before the portfolio formation. about 28% of stocks did not pay dividends across the period of our study. In effect. This tends to be smaller stocks compared with those that pay dividends.Figure 5. as Gwilym et al.
and the current price level is used to incorporate the most recent information in the stock prices. Small/3. Big/2. momentum or total risk. we separately sort stocks into three groups based on size. Small/D/P=0. Fama and French (2008) rely on the most profitable strategy in Jegadeesh and Titman (1993) to compute the 12month compound return as a momentum measure. The stocks that pay the lowest dividends are in group 1 and the highest dividend paying stocks are in group 3. Big/3) sets of portfolios from the intersections of dividend yield sorts and other characteristic sorts. each year. Each year. summing dividend over a full year removes any seasonal patterns in dividend payments. The lowest size group is called „Micro‟ cap stocks. (2004) observe seasonality in dividend yield. Small/2. Shleifer and Vishny (1994) and Fama and French (1996) to form 12 (for example. The second group is called „Small‟ cap stocks and the largest size group is called „Big‟ cap stocks. We follow Lakonishok. Micro/2. Small/1. Micro/D/P= 0. The rest is then sorted into three groups based on their dividend yield. Zerodividend paying stocks are placed in one group. we place the lowest total risk (momentum) in group 1 and the highest in group 3. Micro/3. In the same year. Finally. Big/D/P=0.measured as the compound return from June to November of the previous year. divided by the price in December of the year. These are the independent doublesort portfolios. Micro/1. Total risk is the standard deviation of the past 24 to 60 months excess returns. As Annaert et al. we separately sort stocks based on dividend yield. For total risk and momentum. Big/1. as the 6/6 strategy emerges as the most profitable strategy in chapter 4. We are motivated to use the sixmonth compounded return. We repeat the portfolio formations for 117 . we measure the dividend yield as the sum of all dividends paid in the past 12 months. We use the FM breakpoint method in all stock groupings.
there are seven or more stocks in every independent doublesorted portfolio. the average relative market capitalization of each portfolio is reported. 5. Between 1868 and 1914. it is surprising to see more money invested in „Micro‟ 118 .dividend sorts and the size. We also report separate univariate sort portfolios for size and dividend yield. 5.3 Descriptive Summary Statistics of the Characteristics In this section. we report the summary statistics of the twelve sets of portfolios formed on sizedividend yield. In panel B. 81% of total stock market capitalization. For the doublesorted portfolios. on average.1 reports the summary statistics of the characteristics for the 12 portfolios formed from the intersections of the size sort and dividend yield sort stocks. On the contrary. The first column indicates that the „Big‟ portfolio accounts for the largest percentage of money invested in the stock market.3. the „Micro‟ portfolio accounts for less than 4% of stock market capitalization. momentum or idiosyncratic risk sort. Sorting on two characteristics will test the marginal effect of each characteristic on average excess returns. momentumdividend yield and idiosyncratic riskdividend yield. The number of „Micro‟ stocks that do not pay dividends in the last 12 months exceeds the number of total stocks that pay dividends. We also report the univariate sort on the various characteristics. which results in an average of 240 stocks traded every year (see Panel A and Figure 3.1 SizeDividend yield double sorts Table 5.1). It is not surprising to see fewer numbers of „Big‟ stocks that do not pay dividends (seven stocks on the average). On average. The remaining stocks represent only 15% of market capitalization. representing. our sample contains 84 to 518 different firms.
90 8.65 3.83 4.96 19.10 7.59 12.10 6.44 7.37 1 10.72 5.29 1 56 8 19 30 1 3. The intersections of the stocks in the sizesort and dividendsort are then used to create a set of 12 portofolios (Micro/D/P= 0.50 5.05 4. Big/D/P= 0.63 2.34 7. Micro/1.33 Market Market Micro Small Big 0. The annual average relative market capital of the stocks in each portfolio is reported.00 2 5.31 3.24 3 802 364 725 1330 Panel A: Average Number of stocks Panel B: Relative Market Cap Panel C: Dividend Yield (% ) Panel D : Annual time series average of Total risk(% ) Panel E:Annual time series Momentum(% ) Panel F: Annual Average of Price (Belgian Franc) Panel G:Annual time series average Size (*10^6) 119 .08 5.67 0.55 0.30 3. Big/1. Small and Big) based on their size (price time's shares outstanding) at the end of the previous year. we sort stocks into three groups (Micros.49 D/P= 0 2.We compute the annual average number of stocks in each portfolio.39 2 893 409 716 1226 3 17.87 14.48 3.65 14.86 7.00 2.67 3 6.65 4. Small/D/P= 0.92 20.86 7. The rest of the stocks is splitted into three groups in order of magnitude of their dividend yield.52 3.96 1.83 3 56 16 22 18 3 8.84 7.00 4.99 2 5. Micro/3.16 3. We seperately sort the stocks in the same year based on their dividend yield in the previous year.67 15.09 4.84 1. The sample period is January 1868 December 1913.54 6.02 4. No dividend paying stocks are assigned one group.97 26. The lowest dividend paying stocks are place in 1 and the highest dividend paying stocks are place in 3.96 5.58 10.31 5.42 1. Big/2 and Big/3. Small/3.22 81.58 9.75 10.49 2.06 Market 3.84 1 1006 364 644 1376 2 29.29 4.23 26.39 2 10.01 3 2.58 D/P= 0 226 134 311 668 1 43. Momentum is measured as the compound returns for six months before the t .96 0.00 5.19 1.00 D/P= 0 12. Small/1.66 1 15.48 3 6.22 10.30 7. Small/2.59 6.94 3.25 3. Micro/2.1: Summary statistics for SizeDividend doublesorts In this table we show summary statistics of some characteristic identified to capture the crosssection of stock returns. At the beginning of January for each year t.27 4.15 3.Total risk is the standard deviation of the past 24 to 60 months excess returns.71 3.25 244 606 1259 11. Dividend Yield is measured as the sum of all dividends paid in the year before year t divided by the current price.62 7.11 Market Market D/P= 0 8.50 Market Market Micro Small Big 0.05 5.We also report the annual averages for univariate sorted characteristic on the market.86 6.86 29.Table 5.63 5.50 1 5. Market Market Micro Small Big 80 79 80 Market Market Micro Small Big 5.90 5.47 3.40 D/P= 0 72 46 19 7 D/P= 0 0 0 0 0 D/P= 0 0.01 2 56 10 20 26 2 5.28 17.69 1.
For the univariate sort on dividend yield. Unsurprisingly. the first column shows that size increases with momentum. Among the dividend paying „Big‟ stocks. total risk monotonically reduces as size increases. the relationship between total risk and dividend yield is „Ushaped‟ among dividend paying stocks. It is worthy to note the high total risk recorded for zerodividend paying stocks. It also shows the dividend yield for the 12 sizedividend yield doublesorted portfolios. This might be due to the high number of zero dividend paying stocks in the micro size group (see Panel A). Panel E of Table 5. The first column shows very small reduction in dividends as size increases. the relationship is less negative for the middle and the highest dividend paying stocks.1 shows the average annual past performance (momentum) of the univariate size and dividend yield sort portfolios. In panel D. there is a positive relationship between the lowest dividend yield stocks and size.zerodividend paying stocks as compared to the various dividendpaying stocks. For univariate size sorts. The „Ushaped‟ relationship persists in all size groups. For the „Small‟ group of stocks. the zerodividend paying stocks have slightly lower market capital than the dividend paying stocks. This relationship persists in all doublesorted sizedividend yield portfolios. As the first column shows. On the contrary. less money is invested in zerodividend paying „Big‟ stocks. the average market capital for the highest dividend paying stocks is lower. Panel C shows the average dividend yield of the univariate sort on size. there is a strong correspondence between total risk and size. The positive relationship persists in 120 . Among the doublesorted dividend paying stocks.
The poor performance may be due to „Micro‟ zerodividend stocks. The zerodividend paying stocks performed poorly in the past six months. This persists for all size groups. panel G reports the average size (in millions) of the stocks in each portfolio across the period of the study.2 MomentumDividend yield double sorts Table 5.3. We also report univariate sort portfolios on momentum. there is a positive relationship between size and price in both single and double sort portfolios. It is important to note the strong negative relationship between momentum and dividend yield among dividend paying stocks. Unsurprisingly. compared to the dividend paying stocks.all columns of sizedividend yield doublesorted portfolios. and vice versa. and again on the dividend yield. who found that stocks that do not pay dividends tend to have lower market capitalization and price. It is clear that zerodividend paying stocks have the lowest price. This confirms the results of Grossman and Shore (2006). This confirms the Asness (1997) results and is consistent with the phenomena that value stocks are stocks that performed lower in the past 6 to 12 months.2 reports the summary statistics of the characteristics for the 12 portfolios formed from the intersections of the momentum sort and dividend yield sort stocks. It is obvious to observe the zerodividend paying stocks recording the lowest size in the 19th century BSE. as they record an average momentum of 2. 5. Finally. Panel F reports the annual average price of the size and dividend yield sorted portfolios. Size seems to show a negative relationship with the dividend yield (single sort among dividend paying stocks).08%. Detailed analysis of the doublesort portfolios reveals that the “Small” and the “Big” stocks contribute to the negative relationship. As in the 121 .
the amount of money invested in the various portfolios decreases as dividend yield increases. On the contrary. zerodividend stocks are less than the numbers that pay dividends. In panel B. it is not surprising to see less market capital for zerodividend paying stocks in all momentum categories. This is also true for all categories of doublesort portfolios. The first column shows a slight increase in dividend yield as momentum increases to the 122 . For singlesort dividend paying stocks. On the contrary.previous section. The first column indicates that the middlemomentum portfolio accounts for the largest part of money invested in the stock market. It also shows the dividend yield for the 12 momentumdividend yield doublesorted portfolios. between 1868 and 1914.1). there are twelve or more stocks in every doublesorted portfolio. the middle momentum records the lowest amount among the zerodividend paying portfolios. the average relative market capitalization of each portfolio is reported. For the middle momentum group. On average. representing. the lowest momentum portfolio accounts for less than 22% of stock market capitalization. Among the dividend paying portfolios. For the doublesorted portfolios. 45% of total stock market capitalization. Panel C shows the average dividend yield of the univariate sort on momentum. The number of stocks that performed poorly in the last six months (displaying low momentum) and do not pay dividends in the past 12 months exceed the dividendpaying stocks in each portfolio. on average. the middle momentum portfolios show a higher amount of investment than the other portfolios. which result in an average of 240 stocks traded every year (see Panel A and Figure 3. The remaining high momentum stocks represent only 33% of the market capitalization. our sample contains 84 to 518 different firms.
29 3.78 2 893 830 942 911 3 17.23 5.20 5.16 6.19 5.11 21.03 4.86 1 56 12 20 23 1 3.08 2 10.31 3.00 4.73 1 5.55 8. The annual average relative market capital of the stocks in each portfolio is reported.69 11.56 3 2.74 13.30 9.We compute the annual average number of stocks in each portfolio. 1/1.48 19.62 Market Market 1 2 3 16.99 5. We seperately sort the stocks in the same year based on their dividend yield in the previous year.63 2.25 7.09 9.33 1 10.29 3.70 2.58 8. The rest of the stocks is splitted into three groups in order of magnitude of their dividend yield.63 2. 2 and 3) based on their momentum ( compound returns in prior six month ) at the end of the previous year. The lowest momentum stocks are assigned group 1 and the highest assigned group 3.55 7. 3/2 and 3/3).97 14.35 3.97 2 56 12 25 19 2 5.71 4.19 3 5. At the beginning of January for each year t.50 5.05 Market Market D/P= 0 8.29 D/P= 0 226 174 308 261 1 43.95 8.84 468 839 808 9.74 12.21 26.06 D/P= 0 11.We also report the annual averages for univariate sorted characteristic on the market.33 3.98 2 5.97 18. 2/2.Table 5.11 15. The intersections of the stocks in the momentumsort and dividendsortare then used to create a set of 12 portofolios (1/D/P= 0.99 7.73 6.62 45.98 5.10 5.49 11.63 9. Market Market 1 2 3 80 79 80 Market Market 1 2 3 4.11 2.20 3 6.2: Summary statistics for MomentumDividend yield double sorts In this table we show summary statistics of some characteristic identified to capture the crosssection of stock returns. 2/3.67 5. The sample period is January 1868 December 1913.86 26.63 12.55 1. The lowest dividend paying stocks are place in 1 and the highest dividend paying stocks are place in 3. 3/1. No dividend paying stocks are assigned one group (D/P=0).91 Market Market 1 2 3 5.49 22.32 8.63 9.75 12.31 6.45 4.29 5. 2/D/P= 0. Total risk is the standard deviation of the past 24 to 60 months excess returns.48 13.85 3 802 591 874 946 Panel A: Average Number of stocks Panel B: Relative Market Cap Panel C: Dividend Yield (% ) Panel D : Annual time series average of Total risk(% ) Panel E:Annual time series Momentum(% ) Panel F: Annual Average of Price (Belgian Franc) Panel G:Annual time series average Size (*10^6) 123 .96 7.84 8. Dividend Yield is measured as the sum of all dividends paid in the year before year t divided by the current price. 1/2.42 4.35 D/P= 0 72 37 13 22 D/P= 0 0 0 0 0 D/P= 0 0.77 37.55 5.60 21. we sort stocks into three groups (1.56 1 1006 829 1064 1061 2 29.55 2.75 6.76 1 15.07 2 4.96 3.25 2.08 D/P= 0 2.30 Market 21.94 3. 3/D/P= 0. Momentum is measured as the compound returns six months before the t .32 33.58 3 56 19 21 16 3 8.78 5. 2/1.96 10. 1/3.
On the contrary.middle portfolio. the total risk for the middle momentum portfolios is the lowest. The doublesort portfolios reveal that the strength of the negative relationship is due to the low and the middle momentum portfolios. This persists in all doublesort momentum groups. The total risk for high and low momentum portfolios is almost the same for all dividend yield groups. the prices for zerodividend paying portfolios are low. This is sometimes almost twice the value of the total risk for the dividend paying portfolios in each momentum category.2 shows the average annual past performance (momentum) of the univariate and doublesort momentum and dividend yield portfolios. Panel E of Table 5. This is surprising. the second column shows a very high total risk for the zerodividend paying portfolios. As in the previous section. Dividend yield increases slightly with momentum in the lowest doublesorted dividend paying stocks. In panel D. it is important to note the strong negative relationship between momentum and dividend yield among dividendpaying stocks. The double 124 . as one would expect the stocks that performed well in the past to have lower total risk than stocks that performed poorly in the past. the relationship is „cupshaped‟ and „Ushaped‟ for the middle and the highestdividend paying portfolios. there is a negative relationship between average price and dividend yield among dividend paying stocks. sometimes onethird of the price of their dividend paying counter parts (univariate sort). for the univariate sort. The average price for the univariate sort and double sort portfolios are reported in panel F. For the univariate sorts. As before. Surprisingly.
As in the previous section. the zerodividend paying stocks are the small size stocks. our sample contains 84 to 518 different firms. The negative relationship between size and dividend yield in dividendpaying portfolios is obvious for both single and doublesort portfolios. For the univariate sorts on dividend yield. For the univariate total risk sorts (first column).sort portfolios indicate that the negative relationship pulls its strength from the middle portfolios. Perhaps smaller firms are simply less diversified and therefore have a higher total risk. which result in an average of 240 stocks traded every year (see Panel A and Figure 3. Finally. the number of highrisk zerodividend stocks exceeds (sometimes by 8 times) the dividend paying stocks. Clearly. Looking at the rows of the doublesorts. as the small stocks are more likely the highrisk stocks (see Table 5. This persists in all doublesort portfolios.1).3 reports the summary statistics of the characteristics for the 12 portfolios formed from the intersections of the total risk sort and dividend yield sort stocks. The doublesort portfolios show that the negative relationship between the amount 125 . On average.3 Total riskDividend yield double sorts Table 5. there is a negative relationship between the amount of money invested in each portfolio and its risk.1). This is not surprising. In panel B. the zerodividend paying portfolio has an average size that is less than that of the dividend paying portfolios. between 1868 and 1914. we report the relative amount of money invested in each doublesort total riskdividend yield portfolio. we show the average size of the single and doublesort portfolios in panel G. 5. We also report the univariate sort for total risk and dividend yield.3. there are five or more stocks in every doublesorted portfolio.
55 Panel G:Annual time series average size (*10^6) Market 1 2 3 15.53 1 15.25 3.48 902 808 406 3.90 3 2.92 2 10.00 1.09 3.17 2. Momentum is measured as the compound returns for six months before the t .42 7. 2/1.55 3.53 3.31 2.84 6.36 3.77 D/P= 0 2.2 ans 3) based on their total risk (standard deviation of excess returns) at the end of the previous year.32 5.99 5.07 2 893 917 914 659 3 17.17 6.65 2.96 31.42 2 56 31 18 6 2 5.65 2 4.19 11.65 4. The lowest dividend paying stocks are place in 1 and the highest dividend paying stocks are place in 3.17 9.49 3 5. Dividend yield is measured as the sum of all dividends paid in the year before year t divided by the current price.35 D/P= 0 226 450 336 170 1 43.50 D/P= 0 11.03 8.91 3. 1/3.99 1 10.05 6. We also report the annual averages for univariate sorted characteristic on the market. 3/1.Table 5.31 14.19 7.31 3.20 3 56 20 25 12 3 8.85 7.04 2 5. The annual average relative market capital of the stocks in each portfolio is reported. 2/2.58 7.29 4.74 3. 1/2.We compute the annual average number of stocks in each portfolio. At the beginning of January for each year t.84 25.06 Market 61.03 5.90 4.dividend yield double sorts In this table we show summary statistics of some characteristic identified to capture the crosssection of stock returns.07 7. No dividend paying stocks are assigned one group.17 8.49 1.85 1.52 3.58 6.13 11.67 21.51 7.99 4.35 Market 1.98 5. The rest of the stocks is splitted into three groups in order of magnitude of their dividend yield.71 15. 3/2 and 3/3.13 22.01 7.47 2.00 2. Total volatility is the standard deviation of the past 24 to 60 months excess returns. 3/D/P= 0.16 7.75 3.57 1 1006 1096 949 857 2 29.69 25.88 126 . The intersections of the stocks in the sizesort and dividendsort are then used to create a set of 12 portofolios (1/D/P= 0.54 6. The lowest total risk group are assigned group 1 and highest group 3.90 3 6.31 12.30 4. We seperately sort the stocks in the same year based on their dividend yield in the previous year. 2 /D/P= 0.49 1 56 24 21 10 1 3.13 6.41 2.00 4.97 8.71 6. 1/1.86 2.85 Market Market D/P= 0 8.96 1.65 6. 2/3.31 12.00 9. we sort stocks into three groups (1.55 6.3: Summary statistics of total risk. The sample period is January 1868 December 1913 Market Market 1 2 3 80 79 80 Market Market 1 2 3 5.21 1 5.46 3 802 792 885 534 Panel A: Average Number of stocks Panel B: Relative Market Cap Panel C: Dividend Yield (% ) Panel D: Annual time series average of Total risk(% ) Panel E:Annual time series momentum(% ) Panel F: Annual Average of Price (Belgian Franc) 3.68 Market Market 1 2 3 D/P= 0 72 5 15 52 D/P= 0 0 0 0 0 D/P= 0 0.63 2.84 14.
A negative relationship is clearly visible between dividend yield and the amount of investments. As before. The negative relationship persists in zerodividend. An exception is the total risk of the zerodividend portfolio. Momentum increases as the firm‟s specific risk also increases for univariate sort portfolios. The doublesort portfolios show that the high risk recorded by the univariate sort zerodividend portfolio draws its power from the highest total risk portfolios. The lowest momentum recorded by the zerodividend paying portfolios is also apparent. In Panel D. As in the previous sections. the price for zerodividend portfolios remains smaller (Panel F). the relationship is positive for zerodividend paying stocks. In fact. In panel C. There is a negative relationship between price and total risk (first column).of investments and total risk is completely due to the dividend paying stocks. The positive relationship is reflected in the lowest and the middle dividend yield portfolios. This relationship is mainly due to the low total risk stocks. dividend payment does not show any sign of increasing or decreasing with risk (column one). We also report the univariate sort portfolio characteristics. the negative relationship between dividend and momentum are prevalent in both singlesort and doublesort portfolios (among dividend paying portfolios). This is reflected in all doublesorted portfolios. total risk does not show any sign of decrease or increase in the univariate sort dividend yield portfolios. Panel E reports the past performance of the total riskdividend yield doublesort portfolios. low and middle dividend yield portfolios. As 127 .
To investigate the robustness of the relation among the suggested characteristics. The negative relationship is persistent in all doublesort portfolios. with the hope of returning more earnings back to investors at maturity. this may be security for smaller. Finally.4. size decreases as dividend yield increases among the dividend paying stocks. The negative relationship continues to exist in all doublesorted portfolios. The average size for zerodividend paying stocks is lower compared to their dividendpaying counterparts. there is enough evidence to conclude that zerodividend stocks had low past performance. column one). growth firms in their early lives have low or zero payouts ratios. they can assume that the small size and highrisk firms that do not pay dividends are distressed (see Baskin and Miranti (1997). Often.before. we compute the crosssectional correlation matrix for each year. size shows a negative relationship with firmspecific risk (Panel G. riskier. small market capitalization. we report the timeseries averages of the bivariate correlations. From the previous sections. We perform a hypothesis test on the time series average to test whether it is significantly different from zero. As in the previous sections. the negative relationship between price and dividend yield are still present in the univariate sort. In Table 5. This relationship persists in all dividends paying doublesort portfolios. lowerpriced and younger firms that may have chosen not to pay a dividend because they were newly established and had more growth prospects. This is consistent with the notion that dividends play an important role in communicating to investors in the early capital markets. low price and high total risk in the 19th century. Cheffins (2006) and Cheffins (2008)). Alternatively. We also include 128 .
037 0. AR=average returns.268 *** 0.503 ** 0. The dummy variable is correlated with all the other characteristics except the average returns. We also report a dummy variable which is 1 for zerodividend stocks and 0 for dividend paying stocks.3.027 0.071 ** Note : This table shows the time series averages of the bivariate annual crosssectional correlation between the various characteristics define to explain average returns. we notice that all are small in magnitude. as the relationship would be influenced by the high number of zerodividend stocks (see sections 5.187 ** 0.The significance of the test are in parenthesis. Both are significantly different from zero at the 1% level. Likewise.1.570 *** 0. Table 5. Mom = momentum. Only two are significantly different from zero at the 5% level: the negative 129 .139 ** 0.015 0.005 0. It is not surprising to see the positive relationship between size and dividend yield.15.609 *** 0. We indeed retrieve the negative correlations between total risk on the one hand and dividend yield and size on the other hand.509 *** 0. In(Size) = natural log of size.3).001 DY In(Size) σεi Mom AR 0. We test the hypothesis that.4.4: Annual Time Series Average of the correlation between the entire characteristic and Average return Dummy Dummy DY In(Size) σεi Mom AR 0. DY = dividend yield.3. This is also visible in panel E of Table 5. size and momentum are significantly positively related. the time series average of the bivariate correlation is equal to zero.realized annual returns in the matrix. Investigating the correlations with realized returns in the last line of Table 5. *** = significance at 1 % .157 *** 0.072 ** 0. ** = signifincance at 5 %. σɛi = total risk.
and the lowest and zerodividend paying portfolio on the other hand.4 Average Excess Returns on Portfolio Sorts In this section. The rest of the stocks are divided into terciles. heteroskedastic standarderror adjusted tstatistics for each portfolio. We report the average excess returns and their corresponding NeweyWest autocorrelation.correlation with size and the positive correlation with past returns. whereas the large caps stand out with value weights.3. the smaller stocks dominate. We now turn to more formal testing of any predictive relation between stock characteristics and future returns. we sort the stocks for each year into tercile groups based on a specific characteristic (size. We assign zerodividend stocks to one group. We then hold these portfolios for the entire year and compute monthly returns. we form portfolios on univariate sort characteristics and hold the position for one year. momentum or total sort groups. For the dividend yield. We sort stocks separately based on the dividend yield in the previous year. Both weighting schemas are complementary: with equal weights. We also report the difference in average excess returns and their tstatistics for the highest and the lowest terciles characteristic portfolios (hedge portfolio). Repeating the portfolio formations each month yields 552 portfolio returns in our sample period. 130 . The FM breakpoint is used in all portfolio formations. In addition. The results are reported in Table 5. we show the difference in average excess returns for the highest dividendpaying portfolios on the one hand. As before. momentum and total risk) observed the previous December. Twelve sets of portfolios are formed from the intersection of stocks from the dividend yield sort group and the size. we investigate the relationship between stock characteristics and excess returns using tercile portfolios. We compute portfolio returns either using equal weights or value weights. 5.
19 0. Micro/D/P= 0. 1913.19 1.37 1.74 Market Micro Small Big Bi gMi cro 0.25 0.89 2.19 0.59 3.15 2. Momentum and Idiosyncratic Risk in the previous year t1.20 0.06 0. Momentum and Total risk.43 0. Big/1. we have the following 12 portfolios.63 3.43 2.20 0. VW= Value Weighted 131 .19 2.02 2.20 1.11 0.48 0. Micro/2. and Idiosyncratic sorts are grouped to form equally and value weighted portoflios.54 0.01 0.35 0.14 0.06 0.10 0. The rest of the stocks is split into three groups based on their dividend yield.11 0.53 0.65 0.22 0.38 0.62 0.98 3.35 0.21 0.47 0.70 0. Portfolios are formed annually.09 0.09 0. Market D/P= 0 1 2 Panel A: Sorting on Size and Dividend Yield EW Mean Monthly Returns(%) 3 30 31 Market D/P= 0 1 2 3 30 31 tstatistics for EW 0.73 1.17 0.64 1. Small/1. Small/2. Micro/1.36 3.00 0. The BSE stocks are also sorted on dividend yield and grouped into 4.24 0.08 2.43 1.17 2.50 0.34 0.33 2. Stocks which do not pay dividends in the past one year are assigned one group.05 1.33 2.26 0.42 0.14 0.27 0.00 1.03 0.33 0.15 0.01 0.94 0.53 0. The sample period is Jan.34 4.36 2.13 0.50 1.32 4.5: Equal and Valueweighted portfolios excess returns (%) of doublesorted characteristics At the beginning of January for each year t . Size.48 1.08 1. We also report the univariate sort on Dividend Yield.27 0.17 2. The stocks in the intersections of dividend sorts and the Size.23 0.71 0. Momentum is the measured as the compound returns of the stock six month the portfolios formation year.20 1.15 0.38 0.86 3.03 0. Momentum.69 1.91 2.14 0. Similar portfolio formations are repeated for Momentum and Idiosyncratic risk.00 0.58 2.24 0.22 0.Dec.01 0.46 0. Size is measured as the price times the number of shares outstanding in December of the year before portfolios formations.33 0.28 0.Table 5.83 0.53 0.16 2.06 0.21 0. Small/3.18 0.09 0. Dividend yield is the sum of all dividends paid in the year before the protfolio formation divided by the price at December of the same year.49 0.25 1.16 0.26 0.27 VW Mean Monthly Returns(%) tstatistics for VW Market Micro Small Big Bi gMi cro EW= Equally Weighted. Big/2 and Big/3.19 0.86 2. Big/D/P= 0.02 1.93 2.42 0. Total risk is the standard deviation of the past 24 to 60 months excess returns.04 0. For the size and dividend yield sort.14 1.06 0.75 3.54 1.03 0.70 0.43 2.28 0.48 0.07 0.05 0. We report the NeweyWest Heteroskedastic autocorellation adjsuted tstatistic for the average excess return for each portfolio.26 0.74 1.36 0. Micro/3.33 0. Small/D/P= 0. 1868.04 0.95 0.01 0.12 0.31 1.46 2. the 19th centrury BSE stocks are allocated to three groups based on their sorted size (market capitalization).
20 0.36 0.61 0.40 0.59 2.31 0.11 Market 1 2 3 31 0.69 4.23 0.51 0.36 0.25 0.18 2.90 1.05 1.72 2.94 1.01 0.28 0.91 1.81 0.13 0.19 0.17 0.03 1.46 0.60 0.75 3.01 0.38 0.39 0.43 0.21 0.21 0.21 0.70 1.08 0.22 0.48 1.74 0.36 1.53 0.44 2.07 0.75 0.02 0.60 0.05 1.07 1.39 2.43 2.42 0.86 1.54 0.10 31 0.67 0.15 0.86 4.18 0.27 0.86 3.35 0.05 D/P= 0 1.12 0.81 0.89 1 3.15 0.94 1.27 0.31 4.16 0.23 0.42 0.18 0.26 0.27 0.19 0.Table 5.82 0.61 0.59 4.32 0.69 0.29 2.47 0.30 0.22 3.50 2.07 0.16 0.09 2.18 0.39 0.07 0.85 3 30 31 2.21 3.17 0.31 0.10 0.34 2.13 0.78 0.59 0.99 2.21 0.13 0.94 1.78 2.39 0.11 0.25 1.29 1.29 0.16 2.13 4.11 0.20 1.12 1.56 0.53 0.57 2.5 continued Market D/P= 0 1 2 Panel B: Sorting on Momentum and Dividend Yield EW Mean Monthly Returns(%) 3 30 31 Market D/P= 0 1 2 3 30 31 tstatistics for EW 0.38 0.15 0.38 2.10 0.68 1.09 0.72 2.53 0.56 0.78 4.53 0.29 0.14 0.04 0.01 0.20 0.44 3.92 0.23 0.03 0.72 3.11 1.25 0.64 1.58 1.95 4.13 3.69 3.95 3.03 3.50 1.31 3.76 3.15 0.49 0.59 0.38 0.19 0.29 0.58 0.35 2.42 0.14 0.91 0.42 0.30 4.14 0.07 4.46 0.26 0.43 1.42 0.39 0.60 0.05 2.36 0.67 4.43 0.58 0.10 0.24 0.23 0.19 0.68 1.33 1.63 0.67 2.65 1.14 0.30 0. VW= Value Weighted 132 .25 0.43 0.44 3.44 0.02 0.74 1.06 0.86 4.84 1.78 2.04 0.42 3.66 0.45 0.34 0.42 0.52 3 VW Mean Monthly Returns(%) tstatistics for VW Market 1 2 3 31 Market D/P= 0 1 2 Panel C: Sorting on Total risk and Dividend Yield EW Mean Monthly Returns(%) Market tstatistics for EW 0.56 0.08 0.43 0.93 1.19 2.63 1.47 2.17 0.94 0.05 0.24 0.03 0.12 0.04 1.26 0.84 2.39 1.73 2.70 2.82 0.23 0.64 2 2.23 Market 1 2 3 31 0.29 4.86 0.30 0.36 0.01 0.62 0.36 3.06 0.77 0.34 0.13 0.61 VW Mean Monthly Returns(%) tstatistics for VW Market 1 2 3 31 EW= Equally Weighted.42 0.90 0.06 1.25 1.14 0.70 1.24 30 0.69 0.14 0.23 0.60 0.33 1.24 0.01 0.
a strategy that goes long on a highdividend yield portfolio and short on a lowdividend yield portfolio. Average returns increase monotonically from zerodividend yield stocks to the highest dividend yield stocks (except the middle dividend paying portfolio in „Big‟ stocks).94. The difference in average returns of the zerodividend yield and highest dividend yield portfolio is not significant.In panel A. We find high returns for zerodividend stocks.23). Marsh and Staunton (2002) and Grossman and Shore (2006). The positive significant difference disappears among the dividend yield stocks. with a tstatistic of 1. For valueweighted portfolios. These are the securities for distressed firms. When sorting on dividend yield across all stocks. which is consistent with Grossman and Shore (2006) and Gwilym et al. This shows that weighting stocks by their market capital in portfolio formations. The result is consistent with the findings for the UK market by Dimson. there is a monotonic increase in average excess returns for univariatesort dividend yield portfolios.3). results in a significantly positive profit. (2009). as they have low prices and negative past performance (see Tables 5. average excess returns and tstatistics for the sorting on size and dividend yield are shown. the effect obtains its power from the „Micro‟ and zerodividend yield stocks. It is obvious that the size effect exists in the crosssection of the stocks. For value133 . „Small‟ and „Big‟ stocks show a significantly positive difference in the highest and the zerodividend yield portfolios.1 to 5. The profit improves when the investor goes long on the high dividend yield and short on the zerodividend yield portfolios. The difference in the average excess returns of the highest and the lowest dividend yield portfolios is 24 basis points.70. There is no significant difference in average returns of the highest and the lowest dividend yield portfolios (12 basis points with tstatistic of 1. However. among the size groups. The difference in average excess return increases to 42 basis points with the tstatistic of 2. For equally weighted doublesort portfolios. there is no obvious pattern in equally weighted portfolio returns.
The average excess return for the hedge portfolio is 29 basis points. The notable exception is the negative spread in average excess returns for the lowest and the highest momentum groups in equally weighted 134 . there is a strong momentum effect and a significant dividend yield effect when stocks are valueweighted. In Table 5. doublesort portfolios are formed based on dividend yield and momentum. The result adds credence to the view that the initial evidence for the size effect is due to data mining. The positive relationship between dividend yield and average excess return in the middle momentum portfolios is robust to the weighting scheme used to form portfolios. The dividend yield only shows a linear pattern with average excess returns in the middle momentum group for equally weighted portfolios. and vice versa. one would expect momentum to be an additional predictor of returns holding the dividend yield constant. however. In general. This value increases to 59 basis points when the zerodividend yield portfolio is used to compute the hedge portfolio. Surprisingly. with a tstatistic of 2. for zerodividend yield portfolios. In Panel B. the positive relationship between dividend yield and average returns disappear in the „Big‟ size group. spread between the high momentum portfolio and the lower momentum portfolio. The relationship still exists for valueweighted portfolios. An exceptional case is the insignificant average excess return. It is not surprising to see the size effect disappear when stocks are valueweighted to form portfolios. dividend yield is negatively related to momentum. within the „Small‟ size groups. This confirms the result from Chapter 3. the spread between the average excess returns of the high and low dividend yield portfolios in the momentum tercile groups is not always significant.93. The spread becomes significant when stocks are valueweighted.weighted portfolios. In view of this.1. For the doublesort portfolios. momentum is positively related to average returns (Chapter 4). Panel B confirms momentum as a predictor of future returns. the dividend yield shows a strong positive relationship with the average excess returns.
respectively. (2009). We record opposite results on the 19th BSE when stocks are equally weighted. Yan and Zhang (2005). The difference is 54 and 51 basis points per month. 2. which show that the positive relationship between total risk and expected 135 . This finding supports those of Gwilym et al.portfolios. where high total risk translates into high average excess returns. Surprisingly. who found a strong relationship between dividend yield and average excess returns among dividendpaying stocks in their quintile momentum portfolios. The results show that among dividend yield stocks. Grossman and Shore (2006) have shown that stocks that do not pay dividends and have performed poorly in the past have very high future returns. Cakici. for stocks that do not pay dividends. who argue based on Sharpe Ratio and alpha to confirm a negative relationship between total risk and average excess returns on international markets. while stocks that do not pay dividends and have a high past return have lower future returns.29 and 2. We now turn to total risk. On zerodividend yield stocks.05 standard deviations from zero for equally weighted and valueweighted portfolios. and are even stronger when stocks are valueweighted. The evidence presented in panel C does not show clear patterns. This shows that.81). When focusing on the hedge portfolios. However. regardless of how returns are weighted. Our result does not support Blitz and Van Vliet (2007). the only difference that is statistically significant is the middle dividendyield portfolio. which becomes positive when stocks are valueweighted. the results confirm Bali. neither in the entire market. nor across the dividend yield quartile. the relationship between dividend yield and average excess return in the middle momentum group is strong. the relationship becomes stronger when zerodividend yield stocks are considered. past performance is most likely to determine future returns. However. the valueweighted portfolios average excess return spread for the largest momentum stocks is not significant (the average excess return spread of 10 basis points with a tstatistic of 0. There does not seem to be a monotonous relation between total risk and average returns.
DYj. it is clear that our characteristics are crosssectionally correlated. However.returns documented by Goyal and SantaClara (2003) is driven by small stocks (traded on NASDAQ. based on information available at the end of December. which might have been lost when portfolios are formed.t1 is the dividend yield estimated in the previous year.t 1 2t lnSizej . the return on the individual stocks in the month t. (13) where Rjt Rft is the excess return with R jt . indicates that total risk is not a pervasive characteristic of price in the 19th century BSE market.t1 (price times shares outstanding) is the size of a stock in December of the year before. Looking across dividend yield groups. Size j . we do a regression of the crosssection of individual stock excess returns on the characteristics proposed to explain the average excess return. The full usage of the available individual stocks ensures the maximum utilization of information about the crosssectional behavior of individual stocks.t 1 4t Momj . in their case). Every month. as well as the lack of any significant results for the high dividend yield portfolios. we repeat that the absence of a monotonous pattern across the total risk portfolios in the entire market.4. be it equally weighted or value weighted. We run the crosssectional regression equation of the form Rjt Rft 0t 1t DYj . We update characteristics annually in January. and j. The fact that characteristics are recomputed every year introduces some time variations.t 1 is the total risk measured as the standard deviation 136 .t 1 5t Dumj .t 1 jt . In order to study the marginal effects of the characteristics. and R ft is the shortrate used as a proxy for the riskfree rate.5 The CrossSectional regressions From Table 5. there is no total effect. Generally. we resort to the FM crosssectional regression method.t 1 3t j . 5. there is no statistically significant result in the hedge portfolios.
Mom j . We also estimate the regression with subsets of the characteristics. The estimation of the crosssectional regression every month yields 552 time series for regression coefficients. The averages are tested for statistical significance using heteroskedastic autocorrelation corrected standard errors. The advantage of using the dummy variable is to allow full usage of the crosssectional information on individual stocks. To eliminate the effect of zerodividend stocks on the dividend yieldexcess return relationship.6.t1 is the dummy variable which takes the value 1 for zerodividend yield stocks and 0 for dividend paying stocks. where T is 552.9 The average of the coefficients and their corresponding NeweyWest standard error adjusted tstatistic (parenthesis) is reported in Table 5. Deleting the zerodividend stocks has the disadvantage of reducing the crosssectional information on stocks. The importance of size is consistent with the post WWIIUSA evidence presented by Fama and French (1992).1.t1 is the momentum estimated as the compound gross return from June to November of the year before portfolio formation. model 1 confirms the positive (negative) relationship between momentum (size) and average excess returns. The significance of the size premium is also consistent with our equally weighted We used int T1 4 . 9 137 . In addition. The timeseries average of the slope coefficients is the estimated premium earned for the different exposures. the number of zerodividend stocks constitutes about 28% of the stocks in the crosssection each year. one has to delete these stocks or add a dummy variable. 0t . The dummy variable takes a value of 1 when a stock does not pay dividends. and 0 for all other stocks. . using the dummy variable allows a direct measurement and a test of significance of the difference in behavior of the excess returns in the zerodividend stocks. Dumj.of the past 24 to 60 months excess returns. We use the Newey and West (1987) correction with T1/4 lags. In Panel A. the number of months in our sample. From Figure 5. 5t is the vector of regression coefficients and jt the regression error.
48% (1.13% (0.The sample period is Jan.64% (0.1868Dec.83% (1.11) 1. giving undue importance to the „Micro‟ size stocks. Model 1 Intercept 1.06) 0.64) 4.05) 6.41% (2.83) 0.13) Total Risk 0.51% (0.85) 4. This is not surprising.49% (0.68% (1. 1898Dec.08) 0.61% (1.42% (1.63% (0. momentum and idiosyncratic risk. 1888Dec.04) 0.54% (1.27% (0. We also consider a dummy variable for dividend paying and zerodividend paying stocks in the last 12 months before the portfolios formation year.74) Dividend Yield Size Panel A 0.03) 0.61) Jan.80% (0. 1913.03% (0.08% (0.Dec.89% (2.15% (4. Momentum is computed as 6 months compound returns prior to the regression year.66) 4.21) Dividend Dummy 2 1.15% (2.10% (2.19% (2. momentum and dividend dummy are considered.82) 0.75) 0.portfolio sorts. Dividend yield is the sum of dividends paid in the last 12 months dividend by the current price.81% (2.61) 1.18% (4.65) 138 . 1897 1. 1887 6. Total risk does not show a significant relationship with average returns (average Table 5.26) 0.47) 0. NeweyWest autocorrelation and heteroskedastic adjusted tstatistics are in parentheses.65% (1.64% (1.22% (0.35% (1. 1908Dec.66) 0.18% (2.The interactions between size.We compute characteristics each year. 1907 3.35) 0.10) 0.23% (1.06) 0. Total Risk and Momentum This table reports the coefficients of the monthly crosssectional regressions of the excess return on the characteristic.21% (0.16% 0.63) 2.37) 0.51% (0.37% (0.46) Jan.69% (2.60) Panel B Subperiods Jan.47) 0.81) 1.58% (3.82) 0. The dummy variable is assigned a value is assigned a value of 1 for zerodividend paying stocks and 0 for dividend paying stocks.63) 1.04% (0. Total risk is the standard deviation of the past 24 to 60 months excess returns.55 Jan.88% (2.71% (2.38) Momentum 0. as the regression observations are unweighted.15% (4. 1877 0. Size is measured as price times number of shares outstanding.16) 0. 1878Dec. 1868.The characteristics identified in this regression are dividend yield size.34% (3.60% (0. We return to this issue in the next section.39) 0.31) 0. the characteristics are measured based on the information available prior to the year.6: CrossSectional Regression of Excess Returns on Dividend Yield.94) Jan.25% (3.78% (3. Each year. Size.66) 1.58% (1.55) 4.57) 0. 1913 1.63) 0.08) 0.22% (2.14) 3 2.68) 0.
The value 0. Momentum seems to pull its power from the last twentyfive years of the study. However.51% with a tstatistic of 0. and it confirms the results for equal weight sorts on dividends. As a further check on the form of these relationships. Size is not related to excess returns in the first tenyear period.coefficient is 0. the average total risk hedge portfolio is 10 basis points with a tstatistic of 0. Including the dividend yield and its dummy in regression model 3 reveals a negative insignificant relationship between dividend and average excess returns. The average coefficient is 0. and it is far greater than the riskfree rate.38). Specifically. This shows that there is a negative significant difference in average returns between the zerodividend and dividend paying stocks. The result in Panel B shows that most periods support the overall results in model 3. Dividend yield does not show a significant relationship in any of the subperiods. Table 5. we run the regression in model 3 on five (four tenyear and a fiveyear subperiods) nonoverlapping subperiods between 1868 and 1914. as for the entire market. Adding dividend yield to the regressors in model 2 shows that the dividend yield does not relate to excess returns. The intercepts in models 1 to 3 are significantly different from zero.41 is significant at a 5% level. as the crosssectional regression gives equal weight to all stocks. The effect of the dummy variable is not always significant in the subperiods. This is not surprising. the dummy variable has a negative significant relationship with average returns.37.56. This confirms the sorting result. 139 . The momentum and size premiums are still significant in model 2. Model 1 in Panel A confirms the positive relationship between past returns and future returns in the equal and value weight sorts. The relationship between size and average excess returns does not persist in all subperiods. zerodividend paying stocks have an average excess return less than the dividend paying stock.51 with a tstatistic of only 0.6 confirms the result from the sorting method for equally weighted portfolio formations.
1 Pervasiveness of the CrossSectional Relationships From a broad economic point of view. Also. but depend on the (i) weighting scheme used to compute average portfolio returns. the evidence for size is not consistent across size groups (model 1). In model 2. Horowitz et al. has been close to zero since 1982. Unsurprisingly. (2000) indicate that when firms with less than $5 million in value are excluded. for a characteristic to be accepted and to explain the crosssection of the returns. drives the negative relationship between size and average return. Bali and Cakici (2008) documents that the relationship between total volatility and the expected stock returns are not robust. on average.5. total risk does not have a relationship with average excess returns in the large stocks (Panel A). (ii) the frequency of the return data used to estimate total volatility and. This seems to be consistent with Schwert (2003).7 reports the time series averages of the monthly crosssectional regression coefficients. only 3. Table 5. as the observations are not weighted. which invests in the two lowest decile of stocks by market value. First. For example. which represents. It appears that the microcap group.5. small stocks drive the results. Recent evidence shows that some characteristics are not pervasive. He reports that the abnormal performance of the Dimensional Fund Advisors (DFA) US 910 Small Company Portfolio. Likewise. (iii) the breakpoints used to sort stocks into quintiles. its effect should be marketwide.67% of market capitalization (Panel B). Fama and French (2008) report that the asset growth anomaly is prevalent in their socalled small and micro stocks. The danger with the crosssectional regression approach is that illiquid. the size effect is considerably reduced and becomes statistically insignificant. We investigate whether our results suffer from this caveat by separately running the FamaMacBeth analysis for our „Micro‟ and large („Small‟ plus 'Big‟) stocks. Using equally weighted portfolios and ignoring transaction costs therefore overstates the size effect. introducing the dividend yield in the regressions further reduces 140 .
95) 2. Total Risk and Momentum of Size subsamples This table reports the coefficients of the monthly crosssectional regressions of the excess return on the characteristic without the Micro size stocks.36) 0.02) 3 10. dividend yield has a consistent significant positive relationship with average excess returns in the large stocks. momentum and idiosyncratic risk.88) 1.09% (0.We also consider a dummy variable for dividend paying and zerodividend paying stocks in the last 12 months before the portfolios formation year.26% (0.13) Dividend Total Yield Size Risk Panel A (Small and Big) 0.41% (0.64% (4.68) 0.63% (4.7: CrossSectional regression of Excess Returns on Dividend Yield.06) 0.69% (4.27% (5.59) 0.00% (0.57% (2. These conclusions do not change when the dividend yield dummy is included in the regression.54) (1.09% (0.41% (0.74% (0.59% (0.The sample period is Jan.86% (1.96% (4. but is statistically insignificant.42) Momentum 2.04) 0.10) 2 8.The interactions between size.23% (6.1868.00) 0. Size is measured as price times number of shares outstanding.02% (4.The characteristics identified in this regression are dividend yield size.42% (2. The average coefficient of the dummy variable is no more significant.62) 0. Model 3 intercept 0.16) 8.15% (1.64) 0.06) 4.03% (0.06% 2. We seperately perform similar analysis on Micro stock.98% (6.1913. The strength of the dividend yield effect and the momentum effect is concentrated in the large stocks. Table 5.51) 0.the importance of total risk.06% (1.16) Panel B confirms that the size effect pulls its strength from the „Micro‟ stocks.28) 0.63) 5.Dec. which 141 . Size.68) Dividend Dummy 4 1.05% (0.35) 0.55% (2. It is still negative for the large stocks. momentum and dividend dummy are considered. As an expectation.36% (1. Dividend yield is the sum of dividends paid in the last 12 months dividend by the current price.84) 0.01% (0.99) 1.05% (1. Total risk is the standard deviation of the past 24 to 60 months excess returns. NeweyWest autocorrelation and heteroskedastic adjusted tstatistics are in parentheses.46% (1.We compute characteristic each year.38) 1 9. The dummy variable is assigned a value is assigned a value of 1 for zerodividend paying stocks and 0 for dividend paying stocks.36) 1.04) 0.66) Panel B (Micro) 0.12% (6. Each year.53) 5 1. the characteristics are measured based on the information available prior to the year. Momentum is computed as 6 months compound returns prior to the regression year.
23% (compared to 0.accounts for about 96% of market capital.98% to 2. using NYSE data. the results for momentum are quite consistent across size groups. Table 5. 142 . In all models. as Elton and Gruber (1983) obtained similar results on the USA market between the years 1927 and 1976. Finally. we find a positive relationship between past returns and average realized returns.7 confirms the results from the sorting method. The premium for momentum significantly ranges from 1.54% and 0. The dummy variable contributes its effect in the „Micro‟ stocks. The result is not surprising.71% from the full sample regressions). as the valueweighted dividend sort portfolios seem to have the positive relationship with average excess returns.
we find a significantly negative relationship between size and expected returns.5. The momentum effect. further investigation reveals that the negative significant relationship between average return and size is completely driven by our „Micro‟ stocks. Unsurprisingly. 143 . We use sorting and crosssectional regressions in the analyses. We investigate these relationships with completely out of sample data in the 19th and first few years of the 20th century from the Brussels Stock Exchange.6 Conclusion Since beta fails to explain the crosssection of stock returns in the 19th century BSE. We also investigate the pervasiveness of the crosssectional relationships across different size groups. total risk. which account for about 96% of the market capital („Small‟ and „Big‟ stocks groups). Dividend yield is negatively related to momentum and each of them is positively related to excess returns. However. we document the crosssectional relationship between average excess returns and size. on the other hand. Momentum can explain average excess returns of stocks.67% of the market capitalization. momentum and dividend yield. There is no consistent pattern to be found for total risk. which are mostly zerodividend yield stocks. does not exist in zerodividend yield stocks but strong in dividend paying stocks for equally weighted sorts (average regression coefficient with tstatistics not less than two standard errors from zero). Further investigation reveals that the relationship between excess return and dividend yield does not exist among „Micro‟ group. The relationship is significant in „Small‟ and „Big‟ stocks. accounting for about 3.
Predictability was improved by adjusting betas with Blume and Vasicek‟s autoregressive methods. dividend yield and total risk on the crosssection of stock returns. we examined the robustness of the crosssectional predictability of stock returns. On the 19th century 144 . we conducted four empirical studies to answer the above research questions. The relative size of the market and the relative numbers of large and small stocks are also similar. The study used the 19th and the first few years of the 20th century Brussels stock exchange data. we answered the following research questions: Is beta (systematic risk) stable. the presence and source of momentum and the combined effects of size. Grouping stocks to form portfolios also improved beta stability. thereby contributing to the existing literature on the crosssectional predictability of stock returns. testing of the CAPM and the effects of size.6 CONCLUSION “The behavior of the aggregate U. We found that market model betas for individual stocks were poor predictors of future betas. These studies covered the assessment of betas (the main inputs in the CAPM). More specifically. what is its source? Does total risk predict returns on the 19th century BSE? Does dividend yield predict returns on the 19th century BSE? What is the marginal effect of the above characteristics on the crosssection of stock returns? This doctoral dissertation answered the above questions. However. To this end. the cross section of stocks looks quite different in the two samples” Grossman and Shore (2006) In this doctoral thesis.K stock markets before World War I is similar in many ways to that of the modern US markets. unbiased and robust to outliers in the 19th century BSE? Does the CAPM provide a good description for expected returns in the 19th century BSE? Does size affect the crosssection of stock returns on the 19th century BSE market? Does the momentum effect exist in the 19th century? If it exists. momentum.
(1985) and Chan and Chen (1988) and Fama and French (1992). we adopted sorting and crosssectional regression methods to test the crosssectional relationship between size and average excess return. sorting and crosssectional regression methods were used to investigate the relationship between asset beta and the crosssection of stock returns. This would have confirmed results from Banz (1981). In the second study. We found no relationship between beta and average excess return for the various estimates of beta. Overall. The size effect disappeared when these stocks were omitted. We found a strong relationship between size and average excess return on the 19th century BSE. we could not rely on size as a crosssectional predictor of 145 . which accounts for outliers in the estimation of beta. nonsynchronous trading effects were not prevalent. However. our results also suggested that betas on the 19th century BSE were biased and not stable. We also investigated whether size effects determined crosssectional variation in stock returns. To this end. as monthly returns were used to compute betas. We also found that the iterative reweighted least square method. Our results indicated that the CAPM is not a valid model for capturing crosssectional variations in stock returns on the 19th century BSE. size could have been used to capture the crosssectional variation of stock returns on the 19th century BSE. produced betas that were not significantly different from the market model betas. Detailed investigation confirmed that the size effect drew its power from a small group of stocks accounting for about 0. we examined the validity of the CAPM for the 19th century BSE. This may have been due to the measurement interval of returns. in terms of predictive accuracy. we produced sufficient evidence of how betas should be adjusted for instability and bias when testing the CAPM. Reinganum (1981). Chan et al. The results also confirmed Fama and French (1992) finding that beta has a flat relationship with the crosssection of average returns in the US market. (1983). Second. First.BSE. the relationship was not crosssectionally robust. By studying the behavior of betas in the 19th century.35% of market capital. as in the postWorld War II period. In effect.
and it is contrary to that of Conrad and Kaul (1998). the crosssection of expected returns cannot explain the momentum profit. which constituted. on average. as it was confined to a group of stocks representing a small fraction of market wealth. We found that momentum existed in stocks.returns on the 19th century BSE. This evidence supported Jegadeesh and Titman (2001) results. The presence of momentum on the 19th century BSE also provided evidence that the momentum profit found in postWWII US and other markets was not due to data snooping bias. 146 . the size effect did not exist for small and large stocks. We found that momentum depended on the state of the market when a threeyear lagged return was used to define the state of the market. more than 90% of market wealth. In our data. Our finding that large stocks had momentum showed that momentum could be used as a firm attribute when predicting returns in the crosssection of the 19th century BSE. (2000) findings for the US market between 1963 and 1981. Extensions of behavioral theory have postulated that investors‟ aggregate overconfidence is high following market gains. which corroborated Horowitz et al. This compelled us to test whether momentum profit in the crosssection of stocks depended on the state of the market. investigating the momentum profit in size subsamples showed that postholding period reversal was mainly due to “Micro” size stocks. Therefore. the Jegadeesh and Titman (2001) behavioral explanation of momentum profit should be interpreted with caution. We further investigated the source of momentum profit in the 19th century and found that profit reversed two to five years after portfolio formation. who used US data from 1963 to 2004 to document that the size effect owes much of its power to micro caps and that it is marginal for small and big caps. Therefore. Our results implied that size should not be considered a systematic proxy for risk. The third study investigated the relationship between the shortterm past performance of stocks (momentum) and their future shortterm performance. The results also confirmed the findings of Fama and French (2008). However.
we measured momentum as the compound return of individual stocks over sixmonth periods. We found that the sizeeffect that existed in the 19th century was driven by our socalled “Micro” stocks. In the last study. which serves as a proxy for value.Dependence of momentum on market state for the 19th century BSE added to the behavioral explanation of momentum profit. total risk was useful in explaining the crosssection of stock returns. As a result. following Basu (1983). (2004) and Chabot et al. However. Given its importance. Here. respectively. This result was not surprising. of which more than 50% had a zerodividend yield. coupled with low momentum. negative momentum. (2009) results on the contemporary US and Victorian Era UK markets. may point to firms that were distressed. total risk could also be priced to compensate rational investors for their inability to hold the market portfolio. As stated earlier. Low price. did not use the portfoliooverlapping method).e. dividend and price. due to transaction cost constraints. because the six formation and six month holding period strategy has been the most profitable in previous studies. We included size to investigate whether the other effects were confined to small and illiquid groups of stocks. Therefore. Total risk did not show any 147 . data for the computation of priceearnings and booktomarket ratios were not available for the 19th century BSE. total risk and dividend yield on average returns. high total risk and no dividend payment. to investigate the marginality of its effect in the presence of other characteristics. Reid and Lanstein (1985) and Fama and French (1992). as it corroborated Cooper et al. momentum. In addition. as investors might not have been able to hold the market portfolio. momentum was also included. which were readily available. We considered total risk. we examined the combined effects of size. were used to compute dividend yield. In this case. we would have added priceearnings ratio or booktomarket ratio as a measure of value or growth. low price and very high total risk. We used six months compound returns. Rosenberg. we did not compute momentum as we did in the third study (i.
Among dividend paying stocks. as it is with current markets. (2009)). (2009) documented recently in the US and UK markets. there seem to be evidence for the size effect. The positive relationship between dividend yield and expected returns shows the presence of value effect on the 19th century BSE as the dividend yield is used to represent value. CAPM was not valid for the 19th century BSE. Xing and Zhang (2009). A similarly positive relationship also existed in the current UK and US markets (as in Asness (1997) and Gwilym et al. who conjectured that the effect draws its power from micro stocks on the recent US market. Fu (2009)). but the effect disappears when stocks in the lowest size decile are eliminated on the 19th century BSE. However. Ang. respectively. These results were not different from what Asness (1997) and Gwilym et al. The momentum effect existed on the 19th century BSE. This confirmed Bali and Cakici (2008) finding that (i) the interval of return measurement used to estimate risk (ii) the portfolio breakpoint method (iii) the weighting scheme use to compute portfolio returns and (iv) using liquidity and price filters to screen stocks determines the existence and significance of the crosssectional relationship between risk and expected returns. where positive. as on contemporary markets. Bali and Cakici (2008). Total risk did not show any consistent relationship with average excess returns on the 19th century BSE. negative and no relationships have been found between total risk and average excess return (see Ang et al.consistent relationship with average return. each was positively related to excess stock return for our large stocks. we found several similarities between contemporary stock markets and the 19th century Brussels Stocks Exchange. (2006). Dividend yield showed a positive relationship with average returns on the historical BSE. Hodrick. Based on the predictability of stock market returns. Similar results are found for contemporary markets. Specifically. At first sight. momentum was negatively related to dividend yield. 148 . The findings on size effect confirm Fama and French (2008) results.
149 .Although any inference across historical periods and systems must be interpreted with care. Total risk should not be considered as a predictor of the crosssection of stock returns and limit to arbitrage in our data. Furthermore. total risk. the similarities that we found with the 19th century BSE mostly supported the conclusions of current research on the crosssectional predictability of stock returns. momentum and dividend yield are not time series and crosssectionally robust in predicting stock returns. This is because the total risk does not show a consistent relationship with average returns. not finding the relationship between expected return and beta in our data. size effect is not market wide but rather found in small size stocks. This indicates that the effect is not due to market inefficiency as confirmed in the contemporary markets. we rule out size as a predictor of the crosssection of expected return of securities. in this dissertation. value and momentum turn out to be the common characteristics that can predict returns. we differentiate between characteristics that are consistent predictors of returns and those that are due to data snooping or statistical artifacts. we see momentum and value as the main predictors of the crosssection of stock returns. Back to Chapter 1. Therefore. dividend yield effect and the interactive effect of these two on the 19th century BSE revealed the robustness of these characteristics as predictors of the crosssection of stock returns across time. In view of this. Even though the environment in which the historical market operated has changed so much in relation to the contemporary market. we confirm the doubt placed on the CAPM in the contemporary markets. Specifically. Specifically. we reveal that characteristic such as size. returns from the dividend yield effect are not positively related to total risk. In this research. Our findings on momentum effect.
Sanjoy. 1974. Journal of Financial Economics 49. 2008. crosssection and time series effect. 1982. Riding.. Clifford S. Baskin. Idiosyncratic volatility and the cross section of expected returns. 259299. P. 223249... Journal of Finance 14951511. Hodrick. Ang. Financial Analysts Journal 2936. B. 905929. 1998. and Krista Schwarz.S. A. Hodrick. Vishny. J. 318. Turan G. E. 1997. 3156. Thin trading and the estimation of betas: The efficacy of alternative techniques. and P. Andrew. International industrialisation levels from 1750 to 1980.. Amihud.. J. 2005. Journal of Finance 61. 2009. Bali. Andrew. Journal of Financial Economics 91. High idiosyncratic volatility and low returns: International and further u. Journal of Financial Economics 17. Robert J.References Altman. Columbia University. Journal of Financial Markets 5. Journal of Finance 60. Barberis.. Bartholdy.. and M Levasseur.University of Antwerp. 291311. 123. The relation between return and market value of common stocks. Banz. Ang.. and Xiaoyan Zhang. Frans Buelens. I. Comparative analysis of risk measures: France and the united states. Cambridge University Press. Bairoch. 1997. A model of investor sentiment. Ang. J. Asset pricing and the bidask spread. and Zhe Zhang. Investment performance of common stocks in relation to their priceearnings ratios: A test of the efficient markets hypothesis. Amihud. and Haim Mendelson. 2002. 1994. Basu. The crosssection of volatility and expected returns. Xuemin (Sterling) Yan. Annaert. 1977. Miranti. Nusret Cakici. Does idiosyncratic risk really matter?. A history of corporate finance. working paper . 2004. 307343. and R. Yakov.. Jacquillat. Using individual stocks or portfolios in tests of factor models. Illiquidity and stocks returns. 2958. Y. 150 . 1981. 2008. Andrew.B.. and Xiaoyan Zhang. 1986. Journal of Financial Research 241254. Equity return estimation for 19th century belgium. 2006. Jun Liu.. Shleifer. working Paper. Evidence. The interaction of value and momentum strategies. Robert J. and M. J. Bali. Journal of Financial Economics 9. Yuhang Xing. Journal of Financial and Quantitative Analysis 43. Yuhang Xing. N. Asness. Journal of Finance 32. Journal of European Economic History 269312. and Nusret Cakici. De Ceuster. Turan G. and A. Rolf W.
Jacques. Momentum cycles and limits to arbitrage evidence from victorian england and postdepression us stock markets. Are the reports of beta's death premature?. . Whitcomb. and Naifu Chen. L. Bulkley. 785795. K. S. F. Maier. Chan. Review of Financial Studies 20. R. Betas and their regression tendencies. Cheffins. Hsieh. M. David C. and Ravi Jagannathan. D. Naifu Chen. The volatility effect. Cheffins.. Marshall E. Campbell R. Global Finance Journal 8. 12731338. Can the crosssectional variation in expected returns explain momentum?. and Christian T. Chan.. and Myron Scholes. 2009. 2008. Gabriel A. Chan. and J. Collins. and D. Chatterjee. Harvey. B. 102113. A. Journal of Financial and Quantitative Analysis 27. An outlier resistant approach to risk estimation Financial Analysts Journal 6975. 8394. Journal of Financial and Quantitative Analysis 44. 2007. and W. An unconditional assetpricing test and the role of frim size as an instrumental variable for risk.. Geert. Journal of Financial Economics 14. Schwartz. 1971. Kalman J. C. Lundblad. Chan. K. and J. 1983. J. Francisca Marie. 129156. 1975. 1985. Rayburn. 151 . Ledolter. C. 2006. Studies in the theory of Capital Markets New York: Praeger. Eric Ghysels. C. Oxford Univerisity press. K. 1987. 1988. Blume.). On the assessment of risk. Estimating and adjusting for the intervallingeffect bias in beta. Blitz. 451471. Estimation of risk on the brussels stock exchange: Methodological issues and empirical results.. and Josef Lakonishok. 2007. market value. 1983. 1994. 1993. 265282. 110. Cohen. George. 1972. Louis K.R. Journal of Business 60. 135148.. (NBER).Basu. An exploratory investigation of the firm size effect. 777794. Some further evidence on the stochastic properties of systematic risk.. 2009. W. C.. "The capital asset pricing model: Some empirical tests in jensens(ed. Steven F. Dividend as a substitute for corporate law: The separation of ownership and control. 17831831. Journal of Financial Economics 12. Jensen. 425448. Journal of Finance 26. Marshall E. Journal of Finance 43. Journal of Portfolio Management 34. 1992.. Lakonishok. Chabot. Corporate ownership and control: British business transformed. The relationship between earnings yield.. Robert A. 1997. Robust measurement of beta risk. B. and David K. Journal of Portfolio Management 19. 309325. Beer. and Vivekanand Nawosah. Benjamin. Sanjoy. 5162.. Washington and Lee Law Review 63. and Pim Van Vliet. Black.. Liquidity and expected returns: Lessons from emerging markets. Blume. Bekaert. Journal of Finance 30.. and return for nyse common stocks: Further evidence. Management Science 29. Hawawini. C.
1992. Market states and momentum.. 152 . 2007. Journal of Finance 53. Fama. French. The stability of uk risk measures the problem of thin trading. Göran Kauermann. and Willi Semmler. Estimating betacoefficients of german stock data: A nonparametric approach. Journal of Finance 63. Fama. R. A. French. Fama. European Journal of Finance 13. 2008. Maik. and James MacBeth. 607636. 1999. 13451365. and Michael J. F. 2010. return. and Mike Staunton. and equilibrium: Empirical tests.. Asset growth and the crosssection of stock returns. 356. 489519. Elroy. Common risk factors in the return on stocks and bonds. E. Journal of Financial Economics 33. 2008. and K. 197226. Journal of Finance 63. Journal of Political Economy 81. Michael J. Multifactor explanations of asset pricing anomalies. and H John. Arbitrage risk and stock mispricing. and P. Michael J. The crosssection of expected stock returns. E. Cooper. Risk. An anatomy of trading strategies. J. 427465. and Kenneth R. Journal of Financial Economics 7. Huseyin Gulen. Journal of Portfolio Management 25.. Dimson. 2004. Investor psychology and security market under. and Pantzailis Christos. Dissecting anomalies. and K. D.. Journal of Business Finance and Accounting 19. Daniel. Journal of Finance 754784. 1998. Journal of Finance 47. Chansog (Francis) Kim.. Review of Financial Studies 11. Journal of Financial and Quantitative Analysis 45. Paul Marsh. E. Dimson. Eugene F. Journal of Banking and Finance 7.. Roberto C.. Eugene F. J.. 2002. Elroy. Marsh. Risk measurement when shares are subject to infrequent trading. Marsh. and M. Eisenbeiss.. 1998.Conrad. and Allaudeen Hameed. R. K. 135146.. 253270.and overreactions.. Journal of Finance 59. French. 1993. Doukas. 1983. J. and P. 1992. Eugene F. 16091651. R. Hirshleifer. Faff. 16531678. and Kenneth R. 5584. Cooper. 101 years of global investment returns (Princeton University Press).. Fama. E.. 503522.. and Avanidhar Subrahmanyam. Elton. 907934. W. Gruber. Gutierrez.. Jennifer. Journal of Finance 51. 18391886. Triumph of the optimists. Time stationarity of systematic: Some australian evidence. Dimson. A simple examination of the emipirical relationship between dividend yields and deviation from the capm.. 1979. Dimson. F. Murphy's law and market anomalies. French. E. 1996. 1973. 1983. Fama. Jr. and Gautam Kaul. Schill. R.
Journal of Finance 53. Amit. and Mark J. Ross. Journal of Financial and Quantitative Analysis 41. Haugen. Profitable predictability in the cross section of stock returns.. J. Harvey. On stable factor structures in the pricing of risk: Do timevarying betas help or hurt?. 2009. and Terrance Odean. Journal of Financial Economics 78. Stein. 1979. and Jeremy C. Risk measurement when shares are subjected to infrequent trading. A unified theory of underreaction. Commonality in the determinants of expected stock returns. 1983. and Jay Shanken. New Jersey. Testing the equality of prediction mean squared errors. Gervais. 401439. 1989. 271294. International Journal of Forecasting 13. Journal of Financial Economics 41. Fangjian. E. Robust statistics. Hawawini. Journal of Financial Economics 280283. Grossman. and Stephen H. Savin. 9751007. 2004. Individual securities. and Paul Newbold. R. 153 . A test of the efficiency of a given portfolio. 1996. 463505. Rorke. and Jeremy C. Hoboken. 549573. Harrison. The cross section of stock returns before world war i. 2001. Ready. A. and Pedro SantaClara. G. J. 21432184. Journal of Economics and Business Hong. analyst coverage. Econometrica 57. 1997. H. Baker. belgium case. 2005. Hanna. Hong. and the profitability of momentum strategies. D. and Stephen Thomas. Terence Lim.. Idiosyncratic risk matters!. 1994. Fu. Shore. Tim Loughran. Review of Financial Studies 14. Douglas. 2437. Gwilym. Ghysels. Joel L. Richard S. Goyal. L. 2000. 83100. P.. 127.. The disappearing size effect. Journal of Financial Economics 91..E. Research in Economics 54. Stephen A. and N. Journal of Finance 55. GregoryAllen. Horowitz. Journal of Finance 58. Michael Impson. Stein. Andre Clare. momentum trading. 11211152. Huber. Michel. 2000. Harrison. Russell. Gibbons.Fowler. Idiosyncratic risk and the crosssection of expected stock returns. J. Journal of Finance 54. C. Wiley. and C. Learning to be overconfident. James Seaton. An empirical investigation of beta stability: Portfolios vs. 265295. Owain AP. The Journal of investing 4247. 909916. and N. 1999. David.. Michael R. Simon. and overreaction in asset markets. 2009. An assessment of risk in the thinner market. 2003. 281291.. Bad news travels slowly: Size.. Journal of Business Finance and Accounting 21. 1998. Dividends and momentum. and Imre Karafiath. 2006. and P. Stephen Leybourne.
The illusory nature of momentum profits. T. 1993. Jegadeesh. 587615. and J. 1994. A. total risk and size as determinants of stocks market returns. and S. 1995. 1996.. risk and maximal gains from diversification. 20172069. 143157. P. 431468. 2004. Narasimhan. R. Price momentum and trading volume.Ibbotson. Andrew W. and J. Journal of Finance 56. Martikainen. Lee. Martin. 15411578. Jegadeesh. and Z. Lintner. D.. 353. 1337. Journal of Banking and Finance 10. 1965. Journal of Financial Economics 14. Contrarian investment. Evidence of predictable behavior of security returns. C. Journal of Finance 45. Systematic risk. and Sheridan Titman. 1990. 1975. Lesmond. Financial Analysts Journal 5562.. Journal of Portfolio Management 23. 1971. P. Returns to buying winners and selling losers: Implications for stock market efficiency. Journal of Finance 55. 2000. G.. Lakonishok. Keim. The conditional capm and the crosssection of expected returns. Journal of Finance 49. Security prices. and Sheridan Titman. Luoma. extrapolation. and A.. Review of Financial Studies 15. 1997. D. The adjustment of beta forecasts. D. Kaplan. Review of Economics and Statistics 47. 1994. Dividend yields and stock returns: Implications of abnormal january returns. Jay Shanken. Kothari. Crosssectional and timeseries determinants of momentum returns. Charles M. Peterson. Lintner. R. Pynnönen. 1985. M. J. and risk.. Journal of Finance 20.. J. David A. John. A. 154 . On the short term stationarity of beta coefficient. and Bhaskaran Swaminathan. Journal of Finance 48.. 185224. Narasimhan. 6591. 2001... Lo. Omega 22.. Perttunen. Jagannathan.C. B. Different beta estimation techniques in infrequently traded and inefficient stock markets. N. 1986. 115132. R. Shleifer. Klemkosky. Vishny. 104111. Craig MacKinlay. and Chunsheng Zhou. R. Journal of Finance 50. 473489. 1990. Estimates of smallstock betas are much too low. 1965. and Sheridan Titman. S. Shapiro. J. Wang. 471476. Jegadeesh. Lakonishok.. W.. Another look at the crosssection of expected stock returns.. Journal of Finance 11231128. Levy. Schill. and Alan C. The valuation of risk assets and the selection of risky investments in stock portfolios and capital budgets. Journal of Financial Economics 71. Sloan. D.. Narasimhan. 349380. John. 2002. and Richard G. Datasnooping biases in tests of financial asset pricing models. Michael J. 881898. and R. Jegadeesh. 699720. Journal of Finance 51. Review of Financial Studies 3. Profitability of momentum strategies: An evaluation of alternative explanations.
. 883906. K. Idiosyncratic risk. Theo. A. 1987. Martin. 267284. Moskowitz.. G. and Yexiao Xu. Outlierresistant estimates of beta.. B. 2004. 461481. 155 . Paris). Kenneth Reid.. Journal of Finance 54. Myron. 5669. Jan. Marc R. 1966. Reinganum. R. 768783. Journal of Finance 44. Journal of Quantitative and Financial Analysis 45. long term reversal and momentum. Do countries or industries explain momentum in europe?. Nijman.. and Timothy T.. Journal of Portfolio Management 11. Douglas. 2010. A simple. 1911. A simple model of capital market equilibrium with incomplete information. Journal of Banking and Finance 405423. 2936. 14391464. 1987. R. Rouwenhorst. Les finances contemporaines. Laurens Swinkels. Journal of Finance 54. Econometrica 34. Journal of Financial and Quantitative Analysis 439462. Misspecification of capital asset pricing: Empirical anomalies based on earnings' yields and market values. 703708.Malkiel. H. 149166. Equilibrium in a capital asset market. R. L'epargne française et les valeurs mobilières 18721910 (Felix Alcan. Do industries explain momentum?. Neymarck. and Kenneth D. and Ronald Lanstein. David. M. R. 1998. Journal of Finance 42. Mossin. Geert. Journal of Financial Economics 5. and S. and Mark Grinblatt. Reinganum. Portfolio strategies based on market capitalization. Estimating betas from nonsynchronous data. Portfolio rebalancing and the turnoftheyear effect. 1985. 12491290. Idiosyncratic risk and security returns. heteroskedasticity and autocorrelation consistent covariance matrix. West. 1989. Merton. Journal of Empirical Finance 11. positive semidefinite. J. Geert. and Joseph Williams. McLean. 2003. R. K. Newey. 1981. Morgan.. Journal of Portfolio Management 9. Reinganum. Robert C. Journal of Financial Economics 9. 1999. and Navin Chopra. G. 917. International momentum strategies.. 1998. Scholes. Rosenberg. Barr. University of Texas at Dallas. Thomas. Persuasive evidence of market inefficiency. 1977.dividend yields and returns in the uk equity market. Rouwenhorst. Tobias J. 309327. and Marno Verbeek. Local return factors and turnover in emerging stock markets. Taxes. Financial Analysts Journal 59. M. 1983. 2006.. 483510. Simin. 1946. 1981. Econometrica 55. Withney K. Journal of Finance 53.. Ritter.. 1999. A new empirical perspective on the capm.
. Journal of Finance 28.. 1996. 2742. Amsterdam). Cooper. F. W. Van der Wee. Blomme.: Handbook of the economics of finance (NorthHolland. G. and Roy Porter. Cambridge). 2010. Constantinides. Frans Buelens. 1972. eds. eds. Explorations in Economic History 43. Oldrich. Stijn. 156 . M. Vasicek. 2003.: The industrial revolution in national context. William. and Rene Stulz. Capital asset prices: A theory of market equilibrium under conditions of risk. M. The crosssection of expected stock returns: What have we learnt from the past twentyfive years of research?. P. Van Nieuwerburgh. Shanken. A note on using crosssectional information in bayesian estimation of security betas.return class of new york stock exchange. in M. eds.. Herman. Avanidhar. and G. 2006.Scholliers. Sharpe. Stock market development and economic growth in belgium. Subrahmanyam. Cheltenham).: The economic development of belgium since 1870 (Edward Elgar. Milton Harris. Europe and the USA (Cambridge University Press. On the estimation of betapricing models. Financial Analysts Journal 28. Review of Financial Studies 5. Schwert. 12331239. Industrial wage differentials in nineteenthcentury belgium. 1338. in G. 1992. 1973. The industrial revolution in belgium. Jay. and J. 4652. and Ludo Cuyvers. 1964. European Financial Management 16. Teich. in H. Journal of Finance. William F. Van der Wee. Risk. Anomalies and market efficiency. Sharpe. 1997. 133.
In studie 2 testen we de validiteit van het CAPM en het grootteeffect. In studie 1 behandelen we de beoordeling van de bèta. net zoals de booktomarket ratio) de aandelenrendementen crosssectioneel kunnen voorspellen. momentum. De tijdsperiode die deze dataset bestijht.NEDERLANDSTALIGE SAMENVATTING In dit proefschrift bestuderen we crosssectionele patronen in aandelenrendementen. Met deze data: Testen we de validiteit van het CAPM. Daarom was het niet mogelijk de verklaringskracht van boekhoudkundige gerelateerde kenmerken voor aandelenrendementen te onderzoeken. Deze dataset laat ons toe de invloed te bestuderen op aandelenrendementen van sterk variërende omstandigheden in de economische en institutionele omgeving. totaal risico en dividendrendement op de crosssectionele aandelenrendementen wordt behandeld in studie 5. Daarentegen onderzoeken we of grootte (marktkapitalisatie). vermijdt mogelijke kritiek op datamining. gebruik makende van een volledig onafhankelijke databank gebaseerd op gegevens van de Brusselse Beurs (BSE: Brussels Stock Exchange) tijdens de 19de eeuw en de eerste jaren van de 20ste eeuw. Op het moment van het schrijven van dit proefschrift was de boekhoudkundige en transactiedata van de BSE nog niet gedigitaliseerd voor de 19de eeuw. Er werden vier empirische studies uitgevoerd aan de hand van de BSE data uit de 19de eeuw en het begin van de 20ste eeuw. 157 . Daarenboven testen we of andere kenmerken een verklaring kunnen bieden voor de crosssectionele variatie in aandelenrendementen. De derde studie test het momentumeffect. momentum (voorbije korte termijn rendementen). totaal risico (bedrijfsspecifiek risico) en dividendrendement (als indicator voor de waarde van de activa. De gecombineerde impact van grootte.
maar hebben wel dezelfde verklarende nauwkeurigheid als het marktmodel. Bèta is de belangrijkste inputvariabele van het CAPM. gebaseerd op hun vermogen om de erop volgende bèta te voorspellen. De instabiliteit. Het is een geschatte variabele die mogelijk gemeten wordt met een bepaalde statistische fout. Opvallend genoeg zijn er geen significante verschillen in de verklarende nauwkeurigheid tussen de Blume en Vasicek aangepaste bèta‟s. 158 . In de derde studie werd er gebruik gemaakt van de sorteermethode en de Fama en MacBeth (1973) (FM) crosssectionele regressiemethode om te kunnen onderzoeken of het CAPM geldig is op data van vòòr de Eerste Wereldoorlog. waardoor de testresultaten van het CAPM vertekend kunnen zijn.In de eerste studie focussen we op de beoordeling van de bèta. De bèta‟s van de IRLS zijn klein qua grootte in vergelijking met de bèta‟s van het marktmodel. Om rekening te kunnen houden met outliers werden iterative reweighted least square (IRLS) technichen gebruikt om de bèta‟s te schatten. Er werd eveneens getest of het grootteeffect (de neiging van kleine aandelen om hogere rendementen te hebben dan grote aandelen) reeds in de 19de eeuw aanwezig was op de BSE. Meer specifiek vergelijken we de bèta‟s van het marktmodel met bèta‟s die geschat werden door de autoregressieve technieken van Blume (1971) en Vasicek (1973). We tonen aan dat de individuele bèta‟s van het aandelenmarktmodel niet stabiel zijn. blijken een klein aantal aandelen rendementen te hebben die een voorsprong (lead) of achterstand (lag) habben op de marktrendementen. Wanneer er gebruik gemaakt wordt van de Dimsonmethode om de bèta‟s te schatten. In deze studie onderzoeken we de relatieve prestatie van de verschillende methodes om de bèta te schatten. vertekening en nietrobuustheid naar outliers toe van bèta zijn dan ook belangrijke onderzoekstopics geworden sinds de ontwikkeling van het CAPM. De voorspelbaarheid van deze bèta‟s kan worden verbeterd door het vormen van portefenilles met ten minste tien of meer aandelen.
35% van de totale marktkapitalisatie. Wanneer deze kleine aandelen buiten beschouwing worden gelaten wordt er geen relatie gevonden tussen abnormale rendementen enerzijds en bèta‟s of grootte anderzijds. Deze methode maakt de gemiddelde crosssectionele helling van de bèta niet significant. Een gedetailleerde analyse van de data onthult dat het grootteeffect voornamelijk wordt veroorzaakt door de kleinere aandelen. maar is 159 . We tonen eveneens aan dat de relatie tussen de bèta en de rendementen varieert doorheen de tijd. De nulhypothese die de gelijkheid van de geschatte hellingen veronderstelt van de crosssectionele regressie en de abnormale marktrendementen wordt niet verworpen in de periode 18681893. zowel wanneer ze met als zonder de grootte in de regressie wordt geplaatst. Zowel de sorteermethode als de crosssectionele regressie onthult dat het grootteeffect verdwijnt wanneer de aandelen op waarde worden gewogen bij het vormen van de portefenilles. Samenvattend wil dit zeggen dat het CAPM niet geldig is voor de BSE tijdens de 19de eeuw. De voorwaardelijke dubbele sorteermethode die toegepast wordt door Fama en French (1992) scheidt echter het effect van bèta en grootte op de verwachte rendementen.De sorteermethode en de FM crosssectionele regressiemethode leveren geen empirische bewijskracht voor het CAPM. Door het schatten van de bèta‟s met het marktmodel en de Dimson en Vasicek methodes zal het model niet tot stand gebracht worden. Er bestaat een significant negatief verband met grootte (grootteeffect). maar de bèta is niet gerelateerd aan abnormale rendementen wanneer beide variabelen worden opgenomen in de crosssectionele regressies. We vinden dat grootte negatief gerelateerd is aan abnormale rendementen. Het grootteeffect bestaat. een verband vastgesteld tussen bèta en abnormaal rendement met bèta als enige verklarende variabele (het CAPM is geldig). Daarentegen wordt door gebruik te maken van de op grootte gesorteerde portfolio‟s (met gelijk gewicht) in de crosssectionele regressie. die meetellen voor slechts 0.
Het vinden van een momentumeffect in deze periode bevestigt de bewering dat momentumwinsten gevonden op markten in de periode na de Tweede Wereldoorlog. Er bestaat overtuigend bewijs dat deze strategie winstgevend is op de 19deeeuwse BSE. De momentumstrategie houdt in dat aandelen worden gekocht (verkocht) die sterk (zwak) gepresteerd hebben in de laatste 3 tot 12 maanden. Teneinde de oorzaak na te gaan van de momentumwinsten hebben we gebruik gemaakt van de Jegadeesh en Titman (2001)benadering om de rendementen te bestuderen van de momentumportrfenilles in de postholdingperiode. Verder onderzoek toont aan dat deze omkering voornamelijk veroorzaakt wordt door kleine aandelen.voornamelijk toe te schrijven aan een kleine groep van aandelen die slechts een klein percentage van de totale marktkapitalisatie vertegenwoordigen. In de vierde studie onderzoeken we of een momentumstrategie een abnormale winst kan realiseren op de 19deeeuwse BSE. Een omkening in januari (January reversal effect) dat gevonden wordt op naoorlogse Amerikaanse markten kan niet gevonden worden op de 19deeeuwse BSE. Een bijkomende analyse met betrekking tot de momentumwinsten in elke kalendermaand toont aan dat de winst positief was voor elk van deze maanden. De momentumrendementen keren om in het tweede tot vijfde jaar na de holdingperiode. We vonden dat de 6maanden formatiestrategie en de 6. We vinden dat de momentumwinsten niet sterk zijn in de eerste twintig jaar die we bestudeerden. We hebben eveneens getest of de momentumwinsten en de lange termijn ommekeer in de crosssectie van aandelenrendementen afhankelijk is van de markttoestand. 160 . Een gedetailleerde analyse onthult dat het momentumeffect niet bestaat in de groep van kleine aandelen. niet enkel toe te schrijven zijn aan vertekeningen door datasnooping. In feite laat de maand januari zelfs de vierde hoogste momentumwinsten optekenen in vergelijking met de andere maanden van het jaar.tot 12maanden holdingperiodestrategie enkel winstgevend waren in perioden van marktwinsten.
totaal risico en momentumgroepen heen. die meetellen voor ongeveer 96% van de totale marktkapitalisatie. De sorteermethode en FM crosssectionele regressie methodes werden toegepast. Echter. Verder vonden we een negatief verband tussen dividendrendement en momentum bij dividendbetalende aandelen. momentum. We hebben geen consistente verbanden gevonden voor totaal risico. waarvan deze zeer kleine aandelen deel uit maken. total risico en dividendrendement de crosssectie van aandelenrendementen kunnen verklaren in de periode 18681913. Dividendrendement en momentum zijn positief gerelateerd aan gemiddelde abnormale rendementen bij onze grote aandelen. komt aan het licht dat dit negatieve verband volledig is toe te schrijven aan zeer kleine aandelen die meetellen voor slechts 3.67% van de totale marktkapitalisatie. grootte. maar elk van hen is positief gerelateerd aan gemiddelde abnormale rendementen. We onderzochten de standvastigheid van de relaties over alle dividendrendement.In de vijfde studie onderzoeken we of grootte. In elk jaar is het dividendrendement van een aandeel gelijk aan de som van alle dividenden betaald tijdens de laatste 12 maanden gedeeld door de eindejaarsprijs. Momentum toont een consistent positief verband met abnormale aandelenrendementen. Dit bevestigt de bevinding dat het grootteeffect voornamelijk toe te schrijven is aan de eerste deciel grootteportefenille. Het sorteren op grootte test eveneens of het effect van de andere karakteristieken niet enkel van toepassing is op de groep van kleine en illiquide aandelen. Grootte en momentum worden opnieuw in de analyse opgenomen om de standvastigheid en de gecombineerde impact van deze kenmerken te testen op de crosssectie van abnormale rendementen. Het totaal risico voor elk aandeel wordt gemeten als de standaardafwijking van de residuen van het Dimsonmodel voor het schatten van de bèta‟s van de aandelen. die meetellen voor 96% van de totale marktkapitalisatie. We bevestigen dat grootte een significant negatief verband heeft met abnormale rendementen. wanneer deze analyse op groepen van verschillende grootte wordt uitgevoerd. 161 .
Hoewel elke conclusie over historische periodes en verschillende systemen heen met enige voorzichtigheid geïnterpreteerd moet worden. 162 . door het niet vinden van een verband tussen verwachte rendementen en de bèta in onze data. Dit is omdat totaal risico geen consistent verband toont met de gemiddelde rendementen. maar eerder op de kleine aandelen. totaal risico. bevestigen de gelijkenissen met de 19deeeuwse BSE de conclusies van actueel onderzoek naar crosssectionele voorspellingen van aandelenrendementen. We maken dus een onderscheid tussen karakteristieken die consistente voorspellers zijn voor rendementen en deze die toe te schrijven zijn aan datasnooping of statistische artefacten. dividendrendementeffect en het interactieeffect van beide op de 19deeeuwe BSE tonen de robuustheid van deze kenmerken aan als voorspellers van de crosssectie van aandelenrendementen doorheen de tijd. Afgaande hierop beschouwen we momentum en waarde als de belangrijkste voorspellers van de crosssectie van aandelenrendementen. Totaal risico mag niet beschouwd worden als een voorspeller van de crosssectie van aandelenrendementen en is slechts beperkt tot een arbitrage rol in onze data. Dit toont aan dat het effect niet te wijten is aan marktinefficiëntie zoals bevestigd wordt in hedendaagse markten. Meer specifiek bevestigen we. In dit onderzoek is het grootteeffect niet van toepassing op de gehele markt. de twijfel die rust over het CAPM in de hedendaagse markten. Bovendien zijn rendementen van het dividendrendementeffect niet positief gerelateerd aan totaal risico. waarde en momentum blijken de gemeenschappelijke kenmerken te zijn die rendementen kunnen voorspellen. Daarom sluiten we grootte uit als een voorspeller van de crosssectie van verwachte rendementen van effecten. In het bijzonder onthullen we in dit proefschrift dat karakteristieken zoals grootte. Onze bevindingen in verband met het momentumeffect. Ook al veranderde de omgeving waarin deze historische markten actief waren enorm in vergelijk met de huidige marktomstandigheden.
163 .momentum en dividendrendement niet robuust zijn in tijdzeeksen en crosssecties in de voorspelling van dividendrendementen.
This action might not be possible to undo. Are you sure you want to continue?