Available online at www.sciencedirect.

com

European Journal of Operational Research 189 (2008) 120–131

www.elsevier.com/locate/ejor

Decision Support

Contrarian investment strategy with data envelopment

analysis concept
a,*
Susumu Kadoya , Takashi Kuroko b, Takashi Namatame c

a
Russell Investments Japan Co., Ltd., Product & Research Division,Place Canada 7-3-37 Akasaka, Minato-ku, Tokyo 107-0052, Japan
b
Nikko Asset Management Co., Ltd., 1-1-3 Yurakucho, Chiyoda-ku, Tokyo 100-0006, Japan
c
School of Commerce, Senshu University, 2-1-1 Higashimita, Tama-ku, Kawasaki, Kanagawa 214-8580, Japan

Received 12 May 2005; accepted 7 May 2007

Available online 25 May 2007

Abstract

One of the typical issues in financial literature is that the market tends to be overly pessimistic about value stocks, many
of which are past losers. Therefore, over-reactions might capture by measuring earnings surprise vary with past return lev-
els. In this paper, we propose a new index for an effective investment strategy to capture the return-reversal effect using
both Data Envelopment Analysis (DEA) and Inverted DEA in order to consider the above characteristics of the market.
Our investment strategy using the new index exhibits better performance than the naive return-reversal strategy that only
uses past returns or earnings surprise. In addition, the correlations between our new index and commonly used value indi-
ces are insignificant, and the value indices cannot represent the over-valued (under-valued) situations perfectly. Hence,
considering both proposed and value indices like book-to-price one, we could select value stocks more effectively than
by using only one of these indices.
Ó 2007 Elsevier B.V. All rights reserved.

Keywords: Investment analysis; Contrarian investment strategy; Data envelopment analysis

1. Introduction high BP stocks are high because of price adjustment.

Financial literature has advocated the strategy of
buying stocks with low prices relative to value mea-
sures such as earnings, book values, or dividend
yields. Lakonishok et al. (1994) show that high
Earning to Price (EP) or high Book to Price (BP)
is the result of over-pessimism in the market. More-
over, they insist that the future returns of high EP or
*
Corresponding author. Tel.: +81 48 822 2977.
E-mail address: skadoya@Russell.com (S. Kadoya).
Dreman and Berry (1995) show that positive earn-
ings surprise affects value stocks (i.e. low EP stocks)
more significantly than growth stocks (i.e. high EP
stocks). Further, the past performances of many
value stocks are poor. Matsumura (1998) substitutes
BP for EP and shows similar results through an
analysis of the Japanese market. Therefore, over-
reactions captured by measuring earnings surprise
vary with past return levels. Good investment
opportunities can be expected if both over-reactions
and the amendment of value stock prices in the mar-
ket are observed. The motivation for our study

0377-2217/$ - see front matter Ó 2007 Elsevier B.V. All rights reserved.
doi:10.1016/j.ejor.2007.05.033
 S. Kadoya et al. / European Journal of Operational Research 189 (2008) 120–131
stems from the above stock price pattern observed
in the Japanese market.
In this paper, we propose a new mispricing index
for stock investment using the concept of Data
Envelopment Analysis (DEA) and Inverted DEA
(IDEA) that has been proposed by Yamada et al.
(1994), and attempt to select value stocks whose
future performances are considered to be high.
Then, we show that commonly used value indices
cannot represent over-valued or under-valued situa-
tions of stocks perfectly, and our new index can pro-
vide value-added information in this regard.
DEA is a multi-criteria evaluating method that
can select the most favorable alternatives from
among large sets when there is no parametric
assumption among variables using mathematical
programming technique. Therefore, we can deal
with multi-horizon past returns and several surprise
indices that are harmonically defined in various
ways. DEA presents a composite score—referred
to as efficiency—for each subject. This helps to sim-
plify the complexity of analysis by evaluating the
multi-criteria. In the field of financial investment,
Murthy et al. (1997) applies DEA to empirical study
of performance evaluation. However, to our best
knowledge, apart from performance evaluation
study, few proceedings have been undertaken on
DEA application study in the field.
This paper is organized as follows: In Section 2,
we review preceding studies on investment strategies
and their applications in the field of operational
research. Section 3 presents the idea of our mispric-
ing index and investment strategy. In Section 4, we
outline the formulation of the DEA model and pro-
pose a new mispricing index. Section 5 deals with
data description. Empirical experiments design is
presented in Section 6, and the results are shown
in Section 7. Finally, Section 8 concludes this study.
2. Preceding studies
Recently, academic attention has been focused on
a particular anomaly in finance: the high perfor-
mance of value stocks. Lakonishok et al. (1994)
point out that many past losers (low performance
stocks) are future winners (high performance
stocks). Most of the past losers and winners are high
BP stocks and low BP stocks, respectively. There-
fore, Lakonishok et al. (1994) insist that low BP is
the result of the trade carried out by optimistic trad-
ers. In other words, the return-reversal phenomenon
is observed in the financial market, and the contrar-
121
ian investment strategy is an efficient investment
method. Matsumura (1998) reports similar results
for the Japanese market. Furthermore, Levis and
Liodakis (2001) also present similar results for the
UK market. One possible reason for the effectiveness
of the value strategy is some noise traders’1 cognitive
errors. Some noise traders believe that good stocks
are stocks of excellent companies. Such traders are
too optimistic about growth stocks and too pessimis-
tic about value stocks. On the other hand, informa-
tion traders who can evaluate stock accurately, such
as managers of pension funds, do not play the role of
adjusting the cognitive errors of noise traders
through arbitrage because their clients are more for-
giving of losses on stocks of good companies than
losses on stocks of companies with poor fundamen-
tals. This phenomenon can be explained by the moti-
vation of managers of pension funds in order to
show their prudent investments. The details of pru-
dent investment effects on stock returns are dis-
cussed by Shefrin and Statman (1995).
Fama and French (1992) show that individual
stock size and BP differences can explain return
spreads among US stocks. They also propose a
three-factor pricing model and insist on the exis-
tence of three common factors in the financial mar-
ket. One of the three factors is related to BP.
Therefore, if their hypothesis is true, the value effect
will be consistent with the risk-return framework in
modern finance study. However, Lakonishok et al.
(1994) also argue that the BP ratio may capture var-
ious undesirable factors. For example, a low BP
may describe a company’s many intangible assets
that are not reflected in book values. Therefore,
we should consider various indices other than BP.
The central issue of our study is not to discuss the
relationship between the value factor and return-
reversal phenomenon,2 but to select over-valued
stocks effectively. Therefore, our argument focuses
on portfolio selection techniques.
In this study, we propose a new method to
capture the return–reversal effect using resulting
evidence by Matsumura (1998). Our proposed
method provides an effective contrarian investment
1
A noise trader is defined as an investor who does not possess
sufficient information to evaluate a stock accurately or trade a
stock in accordance with his liquidity demand.
2
Watabe and Kobayashi (2001) discuss this problem. They
conclude that value effect cannot be explained by the return-
reversal phenomenon and that the return-reversal effect cannot be
negated by subtracting the value effect.
122 S. Kadoya et al. / European Journal of Operational Research 189 (2008) 120–131
strategy, and the future performance of this strategy
is better than a naive strategy that only employs
past returns or the magnitude of earnings surprise.
We also show that the correlations between our pro-
posed index and commonly used value indices are
not significant. Therefore, we can overlook over-
valued or under-valued stocks if we focus only on
the BP index for the valuation. Furthermore, our
method can consolidate the BP index and the pro-
posed mispricing index easily and can provide
value-added information, besides the BP index, for
the contrarian investment strategy.
Apart from the by Powers and McMullen (2002),
not many DEA application studies have been con-
ducted in the field of financial investment. They
apply DEA to a portfolio selection problem and
consider high performances, high earnings per share
(EPS), and low risk in terms of beta and volatility as
favorable features for stock selection. Therefore,
EPS and returns are assigned to the outputs, and
beta and standard deviation of returns are treated
as inputs. In their paper, the most efficient stocks
from the viewpoint of DEA are those that remain
efficient over a long period of time. Here, each stock
corresponds to Decition Making Units (DMUs) in
their study. In this case, these DMUs have favorable
and robust features. However, they do not examine
the performances (earning from investment) of the
selected stocks. In investment business issues, the
portfolio selection method should be evaluated
from the viewpoint of performance.
3. Our concept of mispricing index and investment
strategy
Suppose that the market is expecting a 10%
increase in a company’s earnings. In this case, a
30% increase in earnings would be a positive sur-
prise. Therefore, we would expect that the com-
pany’s stock will rise. On the other hand, if the
market is expecting a 40% increase in earnings, there
would be a negative surprise and we would expect
the stock price to fall. We hypothesize that the mar-
ket tends to be overly pessimistic about value stocks
and the performances of such stock are excessively
low. Therefore, we assume that the over-reactions
caused by pessimism are captured by the positive
surprise and historical returns in the market. In addi-
tion, we assume that the stock prices are gradually
adjusted after the news of the earnings surprise.
Considering the above stock price patterns, we
define mispricing index for period t as follows:
1 þ past rettk:t
Idxt ¼ : ð1Þ
Surt
Here,
Actualt  Forecastt
Surt ¼ 1 þ ; ð2Þ
Sizet
where
Past_rettk:t: average monthly stock return over
the period between the end of account per-
iod t and the date as far back as k months
from the end of period t;
Actualt: actual financial data (earnings, sales, etc.)
during the settlement period t;
Forecastt: consensus forecasts of Actualt; and
Sizet: stock market value at the end of period t.
Suppose that Surt is positive, except in the case of
charging extraordinary loss or in the case of extreme
sales deterioration. The numerator of Eq. (1) is also
positive. When the market is pessimistic about a
stock, 1 + Past_rett and Surt of that stock are small
and large, respectively. As a result, Idxt would be
relatively small and positive for optimistic priced
stocks. If this hypothesis is true, we should buy rel-
atively small Idxt stocks and sell relatively large
ones. We can utilize this index for stock evaluation.
Eq. (1) is defined using one surprise index and
one past return. However, we can consider several
combination patterns of financial data and histori-
cal return horizons. In order to make use of this
equation in the context of the real world, we need
to expand it to account for multiple combinations.
4. New mispricing index using the DEA concept
DEA is a multi-criteria evaluating method pro-
posed by Charnes et al. (1978). In this section, we out-
line the method of applying DEA to Eq. (1) and
expand this equation to the multi-criteria index. In
the financial industry, analysts check multiple posi-
tive or negative surprises in financial business results.
Therefore, we should consider several combination
patterns of financial data and historical return hori-
zons in Eq. (1). For example, we would like to com-
pare earnings surprise and sales surprise with both
past 6-month returns and past 1-year returns. Accord-
ing to DEA methodology, the historical returns are
treated as the outputs, while the surprise indices rep-
resent the inputs. Moreover, DEA can deal with mul-
tiple surprise indices and multi-horizon returns
simultaneously. Further, we do not have to take into
 S. Kadoya et al. / European Journal of Operational Research 189 (2008) 120–131 123

consideration the co-linearity between indices. Meric uated as large as possible in the mathematical prob-
and Meric (2001) also indicates this feature. lem solving. Therefore, human judgment is excluded.
In order to expand Eq. (1), we introduce variable If a stock has a relative low efficiency score, that
weights u1, u2, . . . , ul for the historical returns. There- stock is considered as under-valued stock. Our pro-
fore, there are l kinds of return horizons in this case. posed score is conservative because the values of
We also introduce k kinds of valuation weights weights u1, u2, . . . , ul and v1, v2, . . . , vk are set to max-
v1, v2, . . . , vk for multiple surprise indices. Therefore, imize (3). When there are multi-criteria for stock
Eq. (1) can be expanded as follows: value and if even one of these criteria indicates that
, the corresponding stock is relatively over-valued, we
Xl X
k
DEA Idx ¼ uj ð1 þ rmj Þ vi Surmi ; ð3Þ would like to consider the stock as not relatively
j¼1 i¼1 under-valued. In this respect, DEA is a favorable
where m is an individual stock, i represents the system that provides a conservative estimation from
financial data (e.g., sales), and j is the term horizon. this perspective. However, there are two ways in
In addition, rmj represents stock m’s average which we value stocks: whether the stock is over-
monthly returns for the past j terms. Surmi is the sur- valued or whether it is under-valued. Problem (5)
prise index for stock m’s financial data i and is de- is only one aspect and the method is not adequate
fined as follows: to select over-valued stocks.
Therefore, we introduce the inverted version of
Actualmi  Forecastmi
Surmi ¼ 1 þ ; ð4Þ Eq. (3), as follows:
Sizem ,
X k X l
Actualmi: actual financial data i (e.g., earnings, IDEA Idx ¼ vi Surmi uj ð1 þ rmj Þ: ð6Þ
sales, etc.) for stock m; i¼1 j¼1
Forecastmi: consensus forecasts of Actualmi; and This problem is referred to as Inverted DEA
Sizem: stock m’s market value. (IDEA).3 The securities having low IDEA scores
represent over-valued stocks in view of the conser-
DEA determines the optimal weights vj, ui to vative estimation because the output variables and
maximize Eq. (3) for each DMU (in this case, input variables replace each other. If we deal with
DMUs are stocks) under normalizing constraints. only one surprise index and historical return hori-
This is written as the following mathematical pro- zon in Eq. (6), the IDEA score will not be valuable
gramming problem: because it will represent only the inversion formula.
, However, we would like to evaluate stocks by multi-
Xl X
k
max uj ð1 þ rmj Þ vi Surmi ; criteria, which measure the most pessimistic under-
uj ;vi valued stocks as the highest value 1. Therefore,
j¼1 i¼1
, IDEA_Idx provides an evaluation from the view-
X
l X
k
s:t: uj ð1 þ rpj Þ vi Surpi 6 1; ð5Þ point of the inverse of DEA_Idx.
j¼1 i¼1 After all, these DEA and IDEA approaches to
ðp ¼ 1; . . . ; nÞ capture mispricing stocks are adequate for under-
u1 ; . . . ; ul ; v1 ; . . . ; vk P 0: valued stock selection and over-valued stock search,
respectively. Therefore, for the contrarian invest-
Here, n represents the number of stocks. Problem ment strategy, we employ a new mispricing index
(5) is intended to obtain the maximum efficiency that is defined as the difference between the DEA
value for stock m. The constraints in (5) work to and IDEA efficiency scores.
ensure that the efficiency score of any stock never We suggest the new mispricing index4 as follows:
exceeds 1. Therefore, even the most efficient stocks
Miss-pricing Index ¼ DEA Idx  IDEA Idx: ð7Þ
can only have the optimal value of 1. The maxi-
mized object represents the possible high efficiency
3
value of each stock, while simultaneously adhering The concept of IDEA is explained in Appendix 1.
4
to the rules that restrict a stock from having an We can define new index as ratio (DEA_Idx/DEA_Idx).
efficiency that exceeds 1. However, both DEA_Idx and IDEA_Idx are standardized
between 0 and 1 and the difference defined in Eq. (7) is logical.
DEA is a data oriented evaluating technique. In this study, we also test the new index, defined as the ratio of
Then, human bias causes over-valuation in the both indices. Then, the tests yield similar results that have been
market and over-reaction in this study’s aspect is val- omitted from the paper for the reason of space.
124 S. Kadoya et al. / European Journal of Operational Research 189 (2008) 120–131
When we define the DEA efficient frontier as effi-
cient, efficient stocks should be evaluated as having
low efficiency under the IDEA concept. If DEA effi-
cient stocks are assigned high IDEA_Idx scores, we
cannot conclude that the stocks are efficient. We can
avoid an inaccurate evaluation using both frontiers.
On the other hand, the DEA score may be affected
by outliers. However, because the index is con-
structed using both DEA_Idx and IDEA_Idx, we
expect it to be robust.
In following sections, this new index is applied to
empirical tests and its validity is discussed.
5. Data description
We use the Japanese stock data, according to
which the following criteria hold for empirical
study:
(1) The analyzed stocks were confined to those
listed on Tokyo Stock Exchange section one
during the study period—from May 2000 to
December 2004.
(2) The close of the company’s financial year is
March 31, and all business results have to be
declared by the end of May in the subsequent
business term.5
(3) The companies have positive surprise indices.
We use consolidated accounting data published
by Toyo Keizai Inc. However, there are several
companies for which the business results are not
available for all the years. Therefore, the numbers
of available stocks are different for different periods.
The efficiency scores of the analyzed companies are
calculated on the relevant date, that is, the end of
May every year. We calculate the average returns
of the stocks over the past 12, 36, and 60 months
from the relevant dates. The calculated returns are
applied to the numerator of the object function in
(5). The denominator of the function is calculated
using financial forecast data that is available at the
end of April, financial actual data that is announced
at the end of May, and the market values on the rel-
evant day. The same data are used to calculate the
IDEA score.
5
The end of March marks the end of the financial year for
many Japanese companies, and Japanese commercial law stipu-
lates that each company must declare its financial results within
two months from the end date.
Table 1
Applied variables
DEA input (IDEA output) DEA output (IDEA input)
variables variables
Surprise index (sales) Historical return (1-year)
Surprise index (operating income) Historical return (3-year)
Surprise index (ordinary income) Historical return (5-year)
The outputs of the three return variables—the
past 1, 3, and 5 years average returns—are repre-
sented as r1, r2, and r3, respectively. We use three
surprise indices that are calculated using sales, oper-
ating profit, and ordinary profit, respectively. These
indices are represented by Sur1, Sur2, and Sur3,
respectively.
We use the financial data as described above.
The number of objective companies is 942 in fiscal
2000 (from April 2000 to March 2001), 1067 in fis-
cal 2001 (from April 2001 to March 2002), 1121 in
fiscal 2002 (from April 2002 to March 2003), 1139
in fiscal 2003 (from April 2003 to March 2004),
and 1146 in fiscal 2004 (from April 2004 to March
2005). The results of our method will be compared
with that of naive value stock selection using past
returns or surprise indices. The validity is discussed
using future returns that are calculated during the
period from the day following the stock selection
to the next year’s relevant date.6 We applied several
surprise indices because sales may be an important
factor for some companies, and operating profit
may be important for other companies. For exam-
ple, the financial interest account may affect the sur-
prise index value. Table 1 shows the inputs and
outputs for our method.7
6. Methodology
In this section, we examine three tests in order
to verify the advantages of using our method. The
6
However, the last day of the validity-test falls at the end of
December in 2004 because of data availability. Therefore, we
present the future returns in terms of the monthly average for all
the periods from May 2000 to December 2004.
7
Further, we show the statistics of the DEA scores (see Table A
in Appendix 2). The proposed index may be affected by outliers.
The table shows several extreme input weight values in the
DEA_Idx calculation results. However, there are no extreme
values in the IDEA_Idx calculation results. Applying the robust
method using both DEA_Idx and IDEA_Idx, we believe our
results to be reliable.
 S. Kadoya et al. / European Journal of Operational Research 189 (2008) 120–131
first test aims to check the validity of the contrarian
investment strategy using our proposed index. The
second test is performed to verify that our new index
is not significantly related to the BP index and that
our new misprice index provides value-added infor-
mation for naı̈ve value stock portfolios that are con-
structed by using only the BP index. The third test
shows that we can construct an effective contrarian
investment strategy using both the BP index and
our proposed index. The following is the test
methodology.
First, the stocks are sorted in ascending order by
index (7) and are divided into quintile groups. The
average returns of the quintile groups are computed
and the spread return between the top and bottom
groups or between the second and fourth groups
are examined to ascertain the feasibility of the con-
trarian investment. Here, contrarian investment
refers to the strategy in which the most over-valued
stock portfolio (the bottom group by quintile) is
sold and the most under-valued portfolio (the top
group by quintile) is bought. Therefore, if the differ-
ence between the top group and bottom group per-
formances is significant, it can be said that the
contrarian investment strategy is valid.
Next, we conduct two additional tests to discuss
the advantages of our method. One of these tests is
an investigation of the two-stage sorted 25 equal-
weighted asset portfolios’ performances as well as
the Fama and French (1993) size and book-to-mar-
ket sorted portfolios. The first sorting is based on
BP. In this case, all the stocks are sorted into quin-
tiles. The second sorting is a conditional one within
each BP quintile (first sorting) and is based on our
proposed index. Subsequently, all the stocks are
sorted into quintiles within the BP quintile groups.
Therefore, 25 equal-weighted asset portfolios are
formed. These portfolios are used to test whether
the differences between the future returns of the port-
folios are significant. If our index provides value-
added information for the existing index (BP), the
future returns of the sorted layers will be signifi-
cantly different. The second additional test is con-
ducted using DEA and IDEA, considering the
hierarchic category. The details of the methodolo-
gies are shown by Cooper et al. (1999). First, we sort
out all the stocks based on BP in ascending ranking.
Then, we trisect all the securities depending on the
magnitude of the BP. The top portfolio consists of
the lowest BP group stocks and the bottom portfolio
consists of the highest BP group stocks. In order to
exclude the BP effect, we evaluate the stocks that
125
belong to the top portfolio within the top category
using the DEA model. Next, we evaluate the second
group using DEA. Evaluating the stocks that belong
to the second group, we calculate the efficiency
within the second portfolio together with the top
portfolio. Finally, we evaluate the bottom portfolio
stocks across all three-category stocks using DEA.
This method can give low BP stocks a handicap
DEA score for the evaluation. Therefore, low BP
stocks are treated as over-valued stocks right from
the start of the evaluation. When we assign high
BP stocks a handicap score for the evaluation using
IDEA, we should sort out all the stocks based on BP
in descending ranking and trisect all securities
depending on the order. Then, we applied the same
procedure for the evaluation. Hence, using both
the DEA and IDEA scores, we can calculate our
proposed score (Eq. (7)) as a handicapped index.
All the stocks are sorted into quintiles based on this
handicapped index. Subsequently, we test the spread
return between the top and bottom portfolio returns.
If the spread return is more significant than both the
spread returns non-handicapped proposed index
quintiles and the BP quintiles, then the proposed
method can be said to provide additional value for
the investment strategy based on the BP index.
7. Results
7.1. Asymmetry of surprise affection
In this subsection, we show the asymmetric
impacts of earnings and other financial data sur-
prises on the future returns of the stocks.
At first, we construct 25 equal-weighted asset
portfolios by two-stage sorting, as mentioned in Sec-
tion 6. The first sorting is unconditional and based
on the magnitude of the surprise that is measured
on the relevant date. The second sorting is condi-
tional, within each surprise quintile, and is based
on the past return levels. The future returns of the
25 portfolios formed are shown below in Table 2.
The numbers in the upper part of the row are the
returns and those in the lower part (in parentheses)
are t-statistics in the table. The figures represent
the monthly future return averages during the exper-
iment period. Here, we use a surprise measured by
the current profit (Sur3) and the past return levels
measured by r1. Due to space constraints, we refrain
from citing other results that were obtained using
other surprise measures and past return indices.
Table 2 shows that the average returns decline with
126 S. Kadoya et al. / European Journal of Operational Research 189 (2008) 120–131

Table 2 the upper part of the row are the returns and those in
Future returns of cross sorting portfolios by Sur3 and r1 the lower part (in parentheses) are t-statistics.
r1 In this table, we also show the results using other
Top 2 3 4 Bottom quintile portfolios based on other indices (r1, r2, r3,
Sur3 Top 2.5% 2.2% 1.6% 1.2% 0.9% Sur1, Sur2, and Sur3). The ‘‘1–5’’ column in the table
(1.88) (2.02) (1.80) (1.44) (0.85) shows the return differences between the top and
2 2.2% 1.1% 0.9% 0.5% 0.2% bottom portfolios. The ‘‘2–4’’ column also repre-
(2.23) (1.35) (1.12) (0.66) (0.25) sents the spread return between the second and
3 1.4% 1.1% 0.5% 0.3% 1.0%
(1.61) (1.48) (0.58) (0.38) (1.07)
fourth quintiles. These numbers are monthly aver-
4 1.6% 0.9% 0.4% 0.6% 1.4% ages of the differences. The t-values show that the
(1.68) (1.16) (0.63) (0.70) (1.34) differences are statistically meaningful in most cases.
Bottom 2.2% 1.1% 0.9% 0.5% 0.2% However, the t-values of the spread returns based on
(1.26) (1.08) (1.08) (0.49) (0.55) the mispricing index are larger than the others. These
results show that our proposed index is more infor-
both r1 base quintile numbers and Sur3 base quintile mative and helpful for the contrarian investment
numbers, in increasing order. Hence, in order to strategy than other naive indices like past returns.8
select future high return stocks, we should consider
both past return levels and earnings surprise.
7.3. BP effect and mispricing index
7.2. Comparison of quintile portfolio performances
We show the two-stage sorting portfolio perfor-
Quintile groups are formed by using each effi- mances based on BP and the proposed index. The
ciency score on the relevant date for each year. results are shown in Table 4. Here, the numbers in
Therefore, the portfolio rebalances are performed the upper part of the rows are the returns and those
on a year by year basis. The quintile portfolio perfor- in the lower part (in parentheses) are t-statistics.
mances are also measured. The computation period In this table, ‘‘1–5’’ represents the spread returns
is from the day following the relevant date to the between the top and bottom portfolios. The t-statis-
next year’s reference date. The spread returns are tics show that the spread returns between the top
also calculated using monthly portfolio return aver- and bottom portfolios are significant within each
ages. The same procedure is applied to form quintile BP quintile, except the BP top portfolio. The future
portfolios based on surprise indices and past return return of the portfolio that belongs to both the BP
levels. These spreads are also computed. The spreads top quintile and the proposed index top quintile is
between the top and the bottom portfolios and those 2.1%, and the future return of the portfolio that
between the second and fourth portfolios are showed belongs to both the bottom quintiles is 0.6%. There-
in Fig. 1. Here, the spread returns are cumulative fore, the new index that provides value-added infor-
accounting numbers of monthly difference in mation for portfolio selection and the future return
returns. The lines in this figure show the stability of the contrarian investment strategy, using both
of the spread returns based on the proposed index. BP and the new index, should be higher than that
The statistical significances are also tested and the of the value stock selection strategy using only BP.9
results are presented in Table 3. Here, the numbers in Finally, we show the result of the test using DEA
and IDEA, taking into consideration the hierarchic
0.8
0.7
category. The cumulative spread return between the
0.6 top and bottom quintiles is represented in Fig. 2.
0.5 This figure shows an almost constant trend of
0.4
0.3
0.2 8
We conducted 18 tests using various combinations of past
0.1 returns and surprise indices. Essentially, t-statistics improved
0 with an increase in the number of inputs and outputs (see Table B
20000531 20010531 20020531 20030530 20040531
in Appendix 3).
Spread Return (proposing Index quitile,2—4) 9
Spread Return (proposing Index quitile,Top—Bottom) We also calculate rank correlations between our proposed
index and valuation indices, BP, EP, Cash-to-Price (CP) and
Fig. 1. Spread returns between quintiles based on proposed Sales-to-Price (SP). These numbers are enough low and the
index. averages absolute ones are about 0.20.
 S. Kadoya et al. / European Journal of Operational Research 189 (2008) 120–131 127

Table 3
Comparing contrarian investment strategy performances
Top 2 3 4 Bottom 1–5 2–4
Proposed index 1.69% 1.23% 0.81% 0.48% 0.30% 1.39% 0.75%
(1.81) (1.65) (1.14) (0.68) (0.34) (3.90) (4.27)
Sur1 1.46% 0.70% 0.55% 0.64% 1.15% 0.31% 0.05%
(1.65) (0.94) (0.77) (0.87) (1.34) (1.94) (0.44)
Sur2 1.49% 0.88% 0.54% 0.50% 1.10% 0.39% 0.38%
(1.70) (1.17) (0.79) (0.68) (1.20) (2.36) (2.83)
Sur3 1.53% 0.82% 0.62% 0.53% 1.00% 0.53% 0.29%
(1.77) (1.10) (0.87) (0.70) (1.17) (3.31) (1.97)
r1 1.18% 1.15% 1.04% 0.72% 0.42% 0.76% 0.42%
(1.15) (1.52) (1.50) (1.05) (0.49) (1.45) (1.77)
r2 1.67% 1.31% 0.99% 0.52% 0.03% 1.63% 0.79%
(1.67) (1.76) (1.40) (0.70) (0.04) (2.81) (3.76)
r3 1.69% 1.30% 0.98% 0.61% 0.07% 1.76% 0.69%
(1.68) (1.67) (1.35) (0.83) (0.09) (2.49) (2.78)

Table 4
Future returns of two-step sorted portfolios
Proposed index
Top 2 3 4 Bottom 1–5
BP quintile Top 2.1% 1.7% 1.9% 1.4% 2.1% 0.1%
(2.0) (2.0) (2.2) (1.7) (2.0) (0.2)
2 1.8% 1.6% 1.4% 1.0% 0.8% 1.0%
(1.9) (2.0) (1.8) (1.4) (1.1) (2.6)
3 1.6% 1.1% 1.1% 0.7% 0.6% 1.0%
(1.7) (1.6) (1.5) (1.0) (0.7) (2.8)
4 1.0% 0.4% 0.2% 0.3% 0.2% 0.8%
(1.2) (0.7) (0.3) (0.4) (0.2) (2.1)
Bottom 1.0% 0.1% 0.3% 0.4% 0.6% 1.6%
(1.0) (0.1) (0.4) (0.4) (0.6) (2.2)

increase. In Table 5, we compare this spread return Table 5

with those of the BP quintiles and the non-handi- Interaction between the BP index and proposed index
capped proposed index quintiles. Here, the upper Proposed index BP index
row represents the spread returns between the top Handicapped Non-handicapped
and bottom and the numbers in the lower row are 1.78% 1.39% 1.90%
t-statistics. The t-values show that we can acquire 4.42 3.90 4.27
the most significant spread return using the handi-
capped proposed index. Using this index, we can due to the mutual interaction between the BP index
select under-valued or over-valued stocks effectively and the proposed method.

1
0.9
8. Concluding remarks
0.8
0.7
0.6 In this study, we drew a comparison between past
0.5 returns and the surprise indices of business results
0.4 and proposed a new mispricing index. The results
0.3
0.2
revealed that our new index is more helpful in for-
0.1 mulating a contrarian investment strategy than
0 naive indices such as past returns and surprise indi-
20000531 20010531 20020531 20030530 20040531
ces. However, the new index is uncorrelated with
Fig. 2. Spread returns between quintiles based on the handi- other commonly used value indices like BP. There-
capped proposed index. fore, we can select value stocks using both the new
128 S. Kadoya et al. / European Journal of Operational Research 189 (2008) 120–131

index and value indices like BP more efficiently than [IDEA]

using only the value index. In fact, as shown in ,
X
k X
l
Table 5, we demonstrate that the interaction maxuj ;vi vi X mi uj Y mj ;
between the BP index and the proposed method i¼1 j¼1
contributes to the improvement of the spread return ,
X
k X
l
significance. We did not focus on an examination of s:t: vi X pi uj Y pj 6 1 ðp ¼ 1; . . . ; nÞ;
the interaction in this study. The reason for the i¼1 j¼1
effectiveness of the interaction may be the feature
of our proposed method considering multi-criteria: u1 ; . . . ; ul ; v1 ; . . . ; vk P 0:
intangible assets may skew valuation using BP or
the BP index contains some sort of risk factor in Xmi and Ymj is the ith input and the jth output for
the market. By way of future study, the focus should the mth DMU, respectively, in the ordinal DEA
be on revealing the relationship between resources model (i.e. problem [DEA]). Then, these values are
of the BP effect and misprice. positive.The score obtained from IDEA represents
If investors focus on past low returns during the the degree of inefficiency among the DMUs. There-
specific time-horizon for contrarian investment fore, comparing these efficiency scores, we can grasp
strategy, they may be misled in their stock selections. a more detailed idea of the characteristics from the
However, DEA can deal with multi-criteria, and our viewpoint of the efficiency of the DMU than by
proposed index—using both DEA and IDEA— using only the ordinal DEA model.
helps stabilize investment performance. This is The relationship between two frontiers is shown
because we consider that our proposed method pro- in Fig. A (the case of one input and two outputs).
vides a conservative estimation taking into consider- Inverted DEA measures the inefficiency of
ation multi-horizon for the stock values. DMU. If a DMU has a high efficiency score from
To our best knowledge, there have been only a DEA and a low score from IDEA, the DMU is said
few preceding studies in which DEA has been to have complete efficiency (DMU A in Fig. A).
applied to a portfolio selection problem. The speci- Conversely, in the case of a low efficiency score from
fication of the stock return structure is not necessary DEA and a high score from IDEA, the DMU is said
and we do not need to take into account the co-lin- to be very inefficient. A DMU on the end point of
earity of financial variables when we use this both efficient frontiers, like DMU C and D in
method. Therefore, we can expect future progress Fig. A, is considered to be an outlier.
if we use DEA in financial investment problems.
Appendix 2
Appendix 1
See Table A.
In the appendix, we demonstrate the concept of
Inverted DEA (IDEA) proposed by Yamada et al.
(1994) and the relationship between the ordinal Appendix 3
DEA and IDEA.
In IDEA, the inputs and outputs are treated con- See Table B.
versely. In other words, in IDEA, the system creates
inputs from outputs. The formulations of the ordinal
DEA (CCR model) and IDEA are shown as follows:
Output2
[DEA] Input C
A Efficient Frontier of
, Ordinal DEA
X
l X
k
maxuj ;vi uj Y mj vi X mi ;
Efficient Frontier of
j¼1 i¼1
Inverted DEA
,
X
l X
k B
D
s:t: uj Y pj vi X pi 6 1 ðp ¼ 1; . . . ; nÞ; Output1
j¼1 i¼1 Input

u1 ; . . . ; ul ; v1 ; . . . ; vk P 0: Fig. A. Comparison between DEA and IDEA.

 S. Kadoya et al. / European Journal of Operational Research 189 (2008) 120–131
Table A
Statistics of DEA score
Year Freq. Input Output Input weight Output weight DEA_Idx
Sur1 Sur2 Sur3 r1 r2 r3 v1 v2 v3 u1 u2 u3
DEA_Idx
2000 942 Mean 1.000 1.006 1.004 1.003 0.998 1.003 0.247 0.196 0.644 0.182 0.476 0.102 0.766
942 Std 0.219 0.038 0.033 0.038 0.019 0.014 2.460 0.270 0.284 0.317 0.364 0.262 0.041
942 Max 4.627 1.301 1.268 1.227 1.136 1.091 75.635 1.245 1.311 0.979 1.010 1.024 1.000
942 Min 0.013 0.675 0.696 0.914 0.953 0.974 0.000 0.000 0.000 0.000 0.000 0.000 0.558
2001 1067 Mean 1.013 1.000 1.004 1.010 1.010 0.998 0.011 0.552 0.443 0.164 0.095 0.665 0.930
1067 Std 0.130 0.022 0.016 0.026 0.016 0.012 0.204 0.051 0.048 0.188 0.251 0.273 0.020
1067 Max 2.167 1.277 1.231 1.170 1.084 1.081 5.801 1.090 0.999 0.984 1.012 0.938 1.000
1067 Min 0.172 0.838 0.884 0.926 0.972 0.969 0.000 0.000 0.000 0.000 0.000 0.000 0.831
2002 1121 Mean 1.011 1.000 1.003 0.998 1.005 1.001 0.044 0.580 0.382 0.087 0.345 0.417 0.857
1121 Std 0.144 0.031 0.022 0.021 0.014 0.011 0.141 0.109 0.103 0.252 0.412 0.424 0.023
1121 Max 2.881 1.394 1.308 1.196 1.069 1.075 4.563 1.069 1.071 1.005 0.970 1.026 1.000
1121 Min 0.219 0.691 0.757 0.888 0.958 0.968 0.000 0.000 0.000 0.000 0.000 0.000 0.684
2003 1139 Mean 1.007 1.005 1.003 0.996 1.001 1.005 0.023 0.815 0.179 0.347 0.061 0.363 0.780
1139 Std 0.119 0.025 0.019 0.028 0.015 0.011 0.375 0.385 0.382 0.385 0.207 0.384 0.024
1139 Max 2.283 1.196 1.156 1.217 1.064 1.072 8.977 1.069 1.123 0.984 1.009 0.876 1.000
1139 Min 0.111 0.758 0.753 0.879 0.951 0.967 0.000 0.000 0.000 0.000 0.000 0.000 0.692
2004 1146 Mean 1.007 1.002 1.003 1.035 1.010 1.009 0.038 0.741 0.256 0.021 0.337 0.550 0.922
1146 Std 0.099 0.015 0.014 0.030 0.014 0.010 1.283 0.056 0.050 0.112 0.436 0.442 0.016
1146 Max 1.981 1.091 1.098 1.208 1.079 1.072 43.430 1.105 1.074 0.942 0.985 0.998 1.000

129
1146 Min 0.023 0.883 0.773 0.926 0.949 0.980 0.000 0.000 0.000 0.000 0.000 0.000 0.853
 130
r1 r2 r3 Sur1 Sur2 Sur3 v1 v2 v3 u1 u2 u3 IDEA_Idx
IDEA_Idx
2000 942 Mean 1.003 0.998 1.003 1.000 1.006 1.004 0.123 0.336 0.546 0.000 0.051 0.726 0.785
942 Std 0.038 0.019 0.014 0.219 0.038 0.033 0.335 0.477 0.494 0.007 0.191 0.195 0.031
942 Max 1.227 1.136 1.091 4.627 1.301 1.268 1.094 1.049 1.025 0.216 0.803 0.829 1.000
942 Min 0.914 0.953 0.974 0.013 0.675 0.696 0.000 0.000 0.000 0.000 0.000 0.000 0.558
2001 1067 Mean 1.010 1.010 0.998 1.013 1.000 1.004 0.529 0.256 0.220 0.077 0.378 0.438 0.898
1067 Std 0.026 0.016 0.012 0.130 0.022 0.016 0.491 0.430 0.407 0.030 0.030 0.039 0.022

S. Kadoya et al. / European Journal of Operational Research 189 (2008) 120–131

1067 Max 1.170 1.084 1.081 2.167 1.277 1.231 1.080 1.026 1.031 0.461 0.783 0.813 1.000
1067 Min 0.926 0.972 0.969 0.172 0.838 0.884 0.000 0.000 0.000 0.000 0.000 0.000 0.805
2002 1121 Mean 0.998 1.005 1.001 1.011 1.000 1.003 0.324 0.016 0.668 0.003 0.469 0.342 0.818
1121 Std 0.021 0.014 0.011 0.144 0.031 0.022 0.476 0.128 0.472 0.023 0.062 0.064 0.022
1121 Max 1.196 1.069 1.075 2.881 1.394 1.308 1.126 1.044 1.033 0.347 0.717 0.838 1.000
1121 Min 0.888 0.958 0.968 0.219 0.691 0.757 0.000 0.000 0.000 0.000 0.000 0.000 0.704
2003 1139 Mean 0.996 1.001 1.005 1.007 1.005 1.003 0.376 0.143 0.488 0.028 0.329 0.530 0.894
1139 Std 0.028 0.015 0.011 0.119 0.025 0.019 0.262 0.238 0.380 0.022 0.149 0.135 0.022
1139 Max 1.217 1.064 1.072 2.283 1.196 1.156 1.137 1.044 1.020 0.440 0.852 0.909 1.000
1139 Min 0.879 0.951 0.967 0.111 0.758 0.753 0.000 0.000 0.000 0.000 0.000 0.000 0.677
2004 1146 Mean 1.035 1.010 1.009 1.007 1.002 1.003 0.080 0.853 0.059 0.061 0.558 0.316 0.940
1146 Std 0.030 0.014 0.010 0.099 0.015 0.014 0.233 0.315 0.228 0.020 0.199 0.191 0.017
1146 Max 1.208 1.079 1.072 1.981 1.091 1.098 1.080 1.042 1.016 0.505 0.970 0.991 1.000
1146 Min 0.926 0.949 0.980 0.023 0.883 0.773 0.000 0.000 0.000 0.000 0.000 0.000 0.869
 S. Kadoya et al. / European Journal of Operational Research 189 (2008) 120–131 131
t-value (%)
References
0.86
0.86
0.99
0.42
0.30
0.39
0.47
0.17
0.14
0.45
0.52
0.34
0.36
0.35
0.13
0.37
0.26
0.22
0.34
Charnes, A., Cooper, W.W., Rhodes, E., 1978. Measuring
efficiency of decision-making units. European Journal of
Operational Research 2, 429–444.
t-value (%)

Cooper, W.W., Seiford, L.M., Tone, K., 1999. Data envelopment

analysis: A comprehensive text with models, applications,
2.07
3.49
3.12
1.76
2.75
2.69
2.33
3.56
2.67
2.15
2.23
3.00
2.90
3.72
2.57
3.14
3.28
3.12
4.27
references and DEA-solver software. Kluwer Academic
Publishers.
Dreman, D.N., Berry, M.A., 1995. Overreaction, under reaction,
2–4 (%)

and the low-P/E effect. Financial Analysts Journal 51 (4), 21–

0.49
0.88
0.89
0.38
0.57
0.62
0.47
0.72
0.66
0.44
0.47
0.62
0.62
0.71
0.59
0.64
0.60
0.67
0.75
30.
Fama, E.F., French, K.R., 1992. The cross-section of expected
t-value (%)

stock returns. Journal of Finance 47, 427–465.

Fama, E.F., French, K.R., 1993. Common risk factors in the
returns on stocks and bonds. Journal of Financial Economics
2.73
3.01
2.88
2.36
3.54
3.42
2.04
3.60
3.29
3.46
3.34
3.29
3.44
3.47
3.59
3.47
4.03
3.58
3.90
Return spread

33, 3–56.
Lakonishok, J., Shleifer, A., Vinshny, R.W., 1994. Contrarian
1–5 (%)

investment, extrapolation, and risk. Journal of Finance 49,

0.71
0.77
0.70
0.96
1.38
1.32
0.96
1.59
1.59
1.15
1.08
1.32
1.35
1.39
1.71
1.27
1.48
1.58
1.39
1541–1578.
Levis, M., Liodakis, M., 2001. Contrarian strategies and inves-
tors’ expectations. Financial Analysts Journal 57 (5), 43–56.
5 (%)
0.75
0.74
0.85
0.37
0.27
0.34
0.41
0.15
0.12
0.39
0.45
0.30
0.32
0.30
0.11
0.33
0.23
0.20
0.30

Matsumura, H., 1998. Stock market anomalies and the mispric-

ing-correction hypothesis: Some evidence in the Japanese
stock. Security Analyst Journal 37 (2), 16–30 (in Japanese).
4 (%)
0.52
0.32
0.25
0.78
0.60
0.51
0.67
0.52
0.58
0.70
0.61
0.53
0.52
0.46
0.60
0.53
0.55
0.47
0.48

Meric, G., Meric, I., 2001. Risk and return in the world’s major
stock markets. Journal of Investing 10 (1), 63–66.
Quintile portfolio performance

Murthy, B.P.S., Choi, Y.K., Desai, P., 1997. Efficiency of mutual

3 (%)

funds and portfolio performance measurement: A non-para-

0.76
0.73
0.71
0.84
0.83
0.86
0.90
0.86
0.87
0.73
0.83
0.92
0.86
0.90
0.79
0.86
0.87
0.92
0.81

metric approach. European Journal of Operational Research

98, 408–418.
2 (%)

Powers, J., McMullen, P.R., 2002. Using data envelopment

1.01
1.20
1.14
1.17
1.17
1.13
1.15
1.24
1.24
1.14
1.09
1.15
1.14
1.17
1.18
1.17
1.15
1.15
1.23

analysis to select efficient large cap securities. Journal of

Business and Management 7 (2), 31–42.
1 (%)

Shefrin, H., Statman, M., 1995. Making sense of beta, size and
1.47
1.51
1.55
1.34
1.64
1.66
1.38
1.75
1.70
1.54
1.53
1.62
1.66
1.68
1.82
1.60
1.71
1.77
1.69

book-to-market. The Journal of Portfolio Management,

Winter, 26–34.
Watabe, H., Kobayashi, T., 2001. Value anomaly and market
DEA output (IDEA input)

overreaction: Analysis using earning forecasts data. Modern

r3

r3

r3

r3
r3

r3
r3

r3
r3
r3

Finance 9, 41–66 (in Japanese).

Yamada, Y., Matsui, T., Sugiyama, M., 1994. An inefficiency
measurement method for management system. Journal of the
Comparison of portfolio performances for variable combinations

Operations Research Society of Japan 37 (2), 158–168 (in

r2

r2

r2

r2

r2
r2

r2
r2

r2
r2

Japanese).
r1

r1

r1

r1
r1

r1
r1

r1
r1

r1
Sur3
Sur3
Sur3

Sur3
Sur3
Sur3
Sur3
Sur3
Sur3
Sur3
DEA input (IDEA output)

Sur2
Sur2
Sur2

Sur2
Sur2
Sur2

Sur2
Sur2
Sur2
Sur2
Sur1
Sur1
Sur1

Sur1
Sur1
Sur1
Sur1
Sur1
Sur1

Sur1
Table B

Model

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19