SIZE AND EXPERIENCE EFFECTS ON GLOBAL FUNDS
PERFORMANCE
Luis Ferruz
Departamento de Contabilidad y Finanzas. Universidad de Zaragoza
María Vargas
Departamento de Contabilidad y Finanzas. Universidad de Zaragoza
Luis Vicente
Departamento de Contabilidad y Finanzas. Universidad de Zaragoza
Abstract
This work analyzes the impact of different effects (fund size, fund management company
size and experience of the fund management company) on Spanish global funds
performance. Additionally, it is analyzed the survivorship bias effect on these
performance results.
The conditional performance evaluation lets us to test the use of private information by
fund managers. In this sense, we employ a set of stochastically detrended European
conditional variables in order to avoid a spurious regression bias in the prediction of
returns.
KEYWORDS: FUND SIZE EFFECT; FUND MANAGEMENT COMPANY SIZE;
MANAGERS
EXPERIENCE;
GLOBAL
FUNDS;
STOCHASTICALLY
DETRENDED EUROPEAN VARIABLES.
JEL CODES: G12; G23

The authors would like to express their thanks to the financial institution Ibercaja and to the University
of Zaragoza for the Project 268-96 grant from the Vice-Rectorship of Research, also to the Regional
Government of Aragon for the funding that we receive as a Public and Official Research Group during
2005-2006-2007 from the General Directorate for Research, Innovation and Development.
In addition, we would like to thank the Spanish Ministry of Education and Science (MEC) for the Project
SEJ 2006-04208 grant with European co-financing of FEDER funds (European Commission, Brussels).
Finally, the authors like to thank Dr. Gonzalo Rubio for his helpful comments in a Work Seminar held in
Zaragoza (May 2007), and which will undoubtedly let us to improve our paper.
1
1. Introduction
The interest in testing the market efficiency hypothesis and the implications of its
findings have motivated extensive academic research on mutual fund performance
evaluation since Treynor (1965), Sharpe (1966) and Jensen (1968) first proposed the
conventional measures.
These unconditional measures have been proved to have important limitations for the
appropriate assessment of fund management results.1 This is because the consideration
of constant expectations of return and risk mean unconditional measures estimate
performance incorrectly in dynamic financial contexts with time-varying risk exposures.
Conditional performance evaluation, meanwhile, considers public information available
to investors in the estimation of expected returns and risk, assuming that these
expectations are time-varying because the information changes as a consequence of
time-varying economic conditions. By including such time-varying public information,
conditional models evaluate the valued added by fund managers as a result of the
possession and appropriate use of private information, thereby obtaining a better
assessment of performance.
This public information is configured by a set of predetermined information variables
that represents the state of the economy. Several papers in the literature have provided
evidence that some of these information variables are significant in the prediction of
future market returns, allowing investors to establish their return and risk expectations.
Fama and French (1989) use default spread and term spread variables successfully to
explain expected stock and bond returns.2 In addition to these information variables,
dividend yield is usually considered to explain predictable components of expected
returns (e.g. Fama and French, 1989; Chen, 1991; Ferson and Harvey, 1991; Pesaran
and Timmermann, 1995; Patelis, 1997 and Lamont, 1998).
Conditional models incorporating public economic information should, therefore, obtain
more statistically significant estimations than unconditional measures. Ferson and
Schadt (1996), Ferson and Warther (1996), Chen and Knez (1996), Christopherson et al.
(1998), Christopherson et al. (1999), Silva et al. (2003) and Ferson and Qian (2004) are
some studies that provide evidence of the statistical significance of conditional models.
1
The dynamic strategies followed by fund managers lead to biased unconditional performance
assessments (Jensen, 1972; Grant, 1977; Dybvig and Ross, 1985; Grinblatt and Titman, 1989).
2
Similar results were found by Chen et al. (1986), Keim and Stambaugh (1986), Campbell (1987) and
Schwert (1989). Jensen et al. (1996) find evidence that the ability of these variables to forecast expected
returns is dependent upon the monetary environment. Based on this finding, Black (2000) uses a different
measure of these variables to obtain independent forecasts.
2
Otten and Bams (2004), meanwhile, state that conditional models are economically
relevant to detect dynamic patterns in the management of mutual funds.
Most of the conditional performance models have been applied empirically to the US
and UK mutual fund industries, leaving the rest of the world fund industries largely
unexplored.3 As stated by Hallahan and Faff (2001) and Ayadi and Kryzanowski
(2004), it is necessary to minimise possible data snooping biases that arise from
analysing repeatedly the same markets. Furthermore, studies in unexplored fund
industries would allow academics and practitioners to analyse findings from markets
outside the US with different institutional features, thereby making an original
contribution to the financial literature.
Our research analyses the impact of public information in the evaluation of the
performance of Spanish mutual fund industry, which, though far behind US market
figures, is one of the most important in Europe, ranking 3rd by number of funds and 6th
in terms of total assets.4
Some key aspects of our study make it especially interesting for the financial literature.
First, our research is focused on a survivorship-free sample of Global Funds, a special
category of Spanish mutual funds. Spanish fund regulation imposes strict investment
restrictions on mutual funds in order to protect investors and inform about each fund’s
investment vocation. Thus, the Spanish Securities Exchange Commission (CNMV)
regulates the portfolio assets that may be held through its official classification of
mutual funds. In this restrictive classification, global funds are allowed to invest in any
world market and assets without limit and to change portfolio allocations on a
discretionary basis in view of the information they have.5
Consequently, time variations of global fund returns and risk are not affected by the
investment restrictions imposed by the official mutual fund classification, allowing an
assessment of performance that is free of the classification bias that may have affected
previous results found in the literature on mutual funds.
3
Some exceptions are Sawicki and Ong (2000) for the Australian market, Ayadi and Kryzanowski (2004)
for the Canadian market, Cortez and Silva (2002) and Silva et al. (2003) for the Portuguese market.
Finally, Otten and Bams (2002) and Blake et al. (2002) analyse the European market.
4
Approximately €254,000 million are managed by more than 2,800 Spanish mutual funds, ranking 6 th in
the world by number of funds and 11 th by total assets (Source: European Fund and Asset Management
Association, and Investment Company Institute).
5
Global funds are a category of mutual fund and are not comparable to the recently created Spanish
hedge funds.
3
Our study provides evidence of poor performance estimates – conditional and
unconditional – obtained by Spanish global funds. However, these conditional estimates
are very sensible to spurious regression problems that may affect conditional models.
Our study also measures the survivorship bias in both conditional and unconditional
performance evaluation models, comparing the results found previously in other
empirical research described in the literature. The magnitude and significance of this
bias on performance evaluation is a controversial topic in literature.6
The most representative works applying conditional models (Ferson and Schadt, 1996;
Ferson and Warther, 1996; Christopherson et al., 1998) are affected by this bias. As far
as we know, Ayadi and Kryzanowski (2004) and Leite and Cortez (2006) are the only
studies that analyse the effects of survivorship bias on conditional evaluation models,
and our study therefore makes an original contribution to this largely unexplored topic.
Our work provides evidence that non-surviving funds are slightly worse performers than
surviving ones. Furthermore, this result is valid for every unconditional and conditional
performance model, and it is reported that survivorship bias increases when more
conditional information is included to the models.
A second institutional feature that makes the analysis of the Spanish fund industry
particularly interesting is market concentration, with the 2 largest fund management
companies sharing nearly half of the total assets in the market.7 As a result of the
universal banking model existing in Spain, credit institutions dominate the fund industry
with more than 90% of mutual fund assets managed by banks and saving banks.
In such a concentrated industry, where possible competitive distortions may arise from
the market map, it is necessary to explore the fund-size and management company-size
effects on both conditional and unconditional performance evaluation.
Engstrom (2004) asserts that the previous evidence for a negative fund-size effect on
fund performance is partly explained by a negative management company-size effect.
Chan et al. (2005) provides a direct test between fund performance and economies of
scale, finding that fund performance is negatively affected by fund manager size. Chen
et al. (2003) also find diseconomies of scale in fund management. On the other hand,
6
Grinblatt and Titman (1989), Brown et al. (1992) and Brown and Goetzmann (1995) are examples of
studies providing support for a weak impact of survivorship bias on performance results. In contrast,
Malkiel (1995) and Blake and Timmermann (1998) find evidence for an important impact in performance
estimates.
7
There are 116 fund management companies registered in the Spanish fund market, but the 10 largest
manage more than 75% of the industry’s total fund assets. Nine of these fund management companies are
owned by Spanish banks and saving banks.
4
Gallagher and Martin (2005) conclude that the incidence of fund size and management
company size on fund performance is not statistically significant, rejecting the
hypothesis that performance declines with fund size. Finally, Bauer et al. (2006) provide
recent evidence for a positive relationship between performance and fund size, but no
tests of fund company size are reported in their work.8
Our study shows that large-sized funds are slightly cheaper and worse performed than
their smaller counterparts, even though their performance estimates are poor. On the
other hand, there is no evidence for a positive fund company-size effect on the
performance results obtained by global funds.
In addition to the highly concentrated Spanish fund market, this industry is relatively
young and growing compared with other relevant world fund industries, presenting a
competition map in which experienced fund management companies owned by
traditional Spanish banks and saving banks compete against recently established fund
management companies owned by other financial intermediaries.9 In order to address
this relevant feature of the Spanish fund industry, our study explores the experience
effect on both conditional and unconditional performance evaluation.
Ding and Wermers (2005) find that experienced managers outperform their less
experienced counterparts when large funds are considered. However, experience is
negatively correlated with performance in small funds.
Ferreira et al. (2007) also provide evidence that fund managers with high levels of
experience obtain better performance results than the rest of the market, while
Christoffersen and Sarkissian (2007) and Nowak et al. (2004) also support the
hypothesis of a positive and significant relationship between performance results and
the experience of fund managers.
Our paper finds evidence of a positive but not important experience effect on the
performance results obtained using unconditional and conditional methodologies.
Furthermore, we find that large global funds managed by small fund companies obtain
the best conditional and unconditional performance results regardless of the experience
of the fund company.
8
Van Dijk (2007) makes an exhaustive review of the research on the size effect in stock returns.
The average age of the less experienced management company quartile is 3.39 years. More than 80% of
these recently established management companies are not owned by Spanish banks and saving banks.
In contrast, the average age of the more experienced fund management company quartile is 20.75 years.
Nearly 60% of these experienced firms are owned by traditional Spanish banks and saving banks.
9
5
Finally, the paper also deals with the problem of the significance of the predetermined
information variables to predict future stock excess returns when persistent regressors
are used. This problem has been largely ignored in previous research on conditional
performance. Ferson et al. (2003a) note that many previous empirical studies obtain
spurious regressions as a consequence of the high correlation coefficients exhibited by
some of these information variables. Leite and Cortez (2006) support this result for the
Portuguese market, although they find a limited impact on the conditional performance
estimates.
In our research, we find that the use of stochastically detrended information variables
has a limited impact on the magnitude of conditional performance results, but the sign
and financial interpretation of these performance estimates change sharply when
compared to those obtained using the traditional lagged and demeaned information
variables.
In order to address this spurious regression problem, all the insights of this paper on
conditional performance evaluation have been tested by considering a set of
traditionally demeaned information variables and an additional set of stochastically
detrended information variables to prevent spurious regressions in conditional models.
The paper is organised as follows: Section 2 explains the performance models applied in
the empirical analysis. Section 3 describes the sample of funds and predetermined
information variables considered in the empirical study. Section 4 presents the
performance estimates resulting from the application of unconditional and conditional
models to our sample. Survivorship bias, size and experience effects are also discussed
in this section. Finally, section 5 summarises the main findings of the study.
2. Performance methodology
Both conditional and unconditional models are used to evaluate Spanish global fund
performance. The following models are linear and therefore they are estimated by the
least squares method by using the Newey and West (1987) consistent estimator for
covariance matrix, making standard errors to be robust to autocorrelation and
heteroskedasticity problems.
Unconditional model
In this study we use Jensen’s alpha (1968) as the unconditional measure for
performance evaluation. This measure is the intercept (p) of the Capital Asset Pricing
6
Model (CAPM). In this unconditional model, alpha and beta parameters are assumed to
be invariant over time:
rp ,t 1   p   p rm ,t 1   p ,t 1
(1)
Where rp,t+1 = Rp,t+1 – Rf,t+1 is the portfolio excess return, being Rp,t+1 the rate of return
on the portfolio p between time t and t+1, and Rf,t+1 the return of the risk free asset;
rm,t+1 is the excess return of the market factor during the same period; p is the
systematic risk of the portfolio p and p,t+1 is the error term in period t+1, with the
following properties: E(p,t)=0; Var(p,t)=2p,t and Cov(p,t,Rm,t)=Cov(p,t,j,t)=0.
Our portfolio outperforms (underperforms) the market when a statistically significant
positive (negative) alpha coefficient is estimated in expression (1).
Partial conditional model: The Ferson and Schadt (1996) approach
Ferson and Schadt (1996) use a linear function to assess fund performance, which is a
natural extension of the traditional approaches to the risk of funds. In this model the
portfolio weights are a linear function of the information (however, the alpha parameter
remains invariant). Hence, if the betas of the underlying assets are fixed over time, the
portfolio beta will be also a linear function of this information. This idea is not entirely
true for a managed portfolio but it prevails in the interests of obtaining a simple
regression.
The model uses a vector of instruments representative of the information available in
period t, which is denoted by Zt. It is also established as a hypothesis that the only
information used by the manager is Zt. The portfolio beta p(Zt) is a function of the
information vector Zt. Using a Taylor series and approximating to a linear function, we
obtain:
 p ( Z t )   0 p  B ' p zt
(2)
Where zt = Zt – E(Zt) is a vector of the deviations of Zt from the unconditional mean,
and B’p is a vector with the same dimension that Zt. The B’p elements are the response
coefficients of the conditional beta with respect to the information variables Zt.
Furthermore, 0p can be interpreted as the unconditional mean of the conditional beta:
E(p(Zt)).
Incorporating this portfolio beta into the conditional CAPM framework we obtain the
following generating process for portfolio returns:
rp ,t 1   0 p rm,t 1  B' p zt rm,t 1    p ,t 1
7
(3)
Where E  p,t 1 Z t   0 and E  p,t 1 , rm,t 1 Z t   0 ; Zt represents the lagged information
and p,t+1 is the random disturbance.
The first expression of the previous paragraph arises from the assumption of market
efficiency and the second one indicates that rm,t+1 is orthogonal to the random
disturbance of the model and, therefore, that B’p(zt) are the regression coefficients
conditioned to the information Zt.
If we consider a regression of the portfolio excess return with respect to the market
factor and to the product of the market factor with the lagged information, we get the
following expression:
rp ,t 1   p   0 p rm,t 1  B' p zt rm,t 1    p ,t 1
(4)
The alpha coefficient in equation (4) represents the average difference between the
portfolio p excess return and the excess return of the dynamic strategies that replicate
the exposure to the time-varying risk. When the average return achieved by the manager
of portfolio p is higher than that obtained by the dynamic strategies, then portfolio p
obtains a positive conditional alpha.
Equation (4) could be also interpreted as a multi-factor model, in which the market
excess return is the first factor and the products of the market excess return with each of
the lagged information variables the additional factors. These additional factors may be
interpreted as the returns from dynamic investment strategies, which are constructed
with the aim of replicating the dynamic behaviour of the fund’s market beta over time.
Therefore, a positive alpha would mean that the manager’s average performance is
higher than the average performance obtained by the dynamic strategies.
In this partial conditional model, the covariance between fund’s betas and market
expected returns, given Zt, is captured by the factor ztrm,t+1, so this covariance is
somewhat controlled through the use of conditioning information.
Full conditional model: The Christopherson, Ferson and Glassman (1998) approach
Christopherson et al. (1998) conclude that if betas can be dynamic and change with
market conditions, then alphas might also exhibit a similar behaviour. These authors
criticize the partial conditional model given that it assumes constant alphas thereby
failing to provide much power in predicting superior performance.
In the partial conditional model, conditional alpha will be zero if the manager’s
portfolio weights do not add information about future returns and the unique
8
information about them is contained in public information variables represented by Zt.
Then, Christopherson et al. (1998) consider that if the manager uses more information
than that contained in Zt, the portfolio weights will be conditionally correlated with
future returns, given Zt, and the conditional alpha will be a function of this conditional
covariance.
So, Christopherson et al. (1998) incorporate time-varying alphas to the Ferson and
Schadt (1996) model, with alpha a linear function of Zt:
 p ( z t )   0 p  A' p z t
(5)
Where p is the average conditional alpha and the vector A’p represents the response of
the conditional alpha to the information variables.
If we introduce equation (5) into the partial conditional model we obtain the full
conditional model, which allows us to track the variation in alphas in response to public
information changes:
rp ,t 1   0 p  A' p zt   0 p rm,t 1  B' p zt rm,t 1   u p ,t 1
(6)
3. Data
A sample of Spanish global funds is considered for our empirical work over the period
December 1999-December 2006. Longer evaluation periods are analysed in most of
studies on US and UK fund markets, but the Spanish fund industry is a recent one and a
longer time horizon would have reduced the number of funds included in our sample,
weakening the significance of the conclusions.
Our database comprises all global funds that existed in Spain from December 1999 to
December 2006. The sample is therefore unaffected by survivorship bias, because it also
includes those mutual funds that were terminated during the study period. This sample
was constructed by considering mergers, name changes and changes in funds’
investment policy over the study horizon.
Summary statistics for the monthly returns of an equally-weighted portfolio composed
of the surviving Spanish global funds are shown in panel A of table 1. Similar
information is exhibited for an equally-weighted portfolio of all existing funds over the
sample horizon.10 These monthly returns are gross of management and custodial fees in
order to analyse the performance results without the incidence of the management costs.
10
The number of funds included in this portfolio varies over the study period because of the entry of new
funds and the exit of terminated funds.
9
Table 1. Summary statistics of the equally-weighted portfolios of Spanish global funds
Panel A reports some statistics of monthly returns of both equally-weighted portfolios formed by
surviving funds and all existing funds over the period December 1999-December 2006. Panel B exhibits
the attrition and mortality rates of the sample during each year. The attrition rate is calculated as the
number of exiting funds each year divided by the number of existing funds at the end of the year. The
mortality rate for each year is obtained as one minus the number of surviving funds in December 2006
that also existed at the end of each year divided by the number of existing funds at the end of the year.
PANEL A: Summary statistics of monthly returns
Surviving funds
All funds
Mean
0.169%
0.139%
Median
0.441%
0.442%
Maximum
6.908%
5.465%
Minimum
-6.541%
-6.571%
Standard Deviation
0.0260
0.0246
Skewness
-0.219
-0.400
Kurtosis
2.487
2.556
Jarque-Bera test
1.611
2.965
Number of funds
347
438
PANEL B: Attrition and mortality rates
Attrition rate
Mortality rate
1999
52.1%
2000
0.93%
48.6%
2001
23.71%
37.1%
2002
14.41%
29.7%
2003
9.03%
20.6%
2004
1.87%
15.4%
2005
9.77%
4.9%
2006
3.75%
0.0%
The attrition rates are also exhibited in panel B of table 1 ranging between 0.93% and
23.71% with an average rate of 8.59%. This figure is higher than the rates reported in
other empirical studies on non European fund markets11, although it is very similar to
the 9.37% found by Dahlquist et al. (2000) for Swedish bond funds. We also find that
the mortality rates are very significant although the trend decreases over the study
horizon.
In spite of the global investment objective of this category of Spanish mutual funds,
their normal portfolios are invested mainly in European markets, requiring the use of a
European benchmark to assess their performance results on an appropriate basis.12
Therefore, we use the MSCI Europe TR index in our analysis (Source: Morgan Stanley
Capital International-Barra).
In addition to this benchmark, the 1-month Euribor interest rate is used as a proxy for
the risk-free rate to obtain the excess returns included in the performance models.
11
For example, Elton et al. (1996) and Carhart et al. (2002) for the US equity fund market, and Ayadi and
Kryzanowski (2004) for Canadian fixed-income mutual funds.
12
The average geographical distribution of the normal portfolios is 89.37% in European markets, 5.36%
in US market and 5.27% in other international markets.
10
In order to apply the conditional models, three public information variables are
considered in view of their extensive use in previous empirical studies and their
relevance for stock returns predictability (e.g. Ferson and Schadt, 1996; Christopherson,
Ferson and Glassman, 1998; Cortez and Silva, 2002; Roy and Deb, 2004)
European information variables are used in our work because of the significant tendency
of our sample of Spanish global funds to invest in European markets and the
convergence of European economies as a consequence of the European Monetary
Union.13
First, the European dividend yield variable (DY) is calculated from the dividends paid
by the MSCI Europe benchmark in the prior 12 months divided by the current price of
the MSCI Europe index (Source: Morgan Stanley-Barra). Second, the slope of the
European term structure (TERM) is computed as the annualized yield spread between
10-year European Monetary Union government bonds and the 3-month Euribor interest
rate, both obtained from the Spanish Central Bank. Finally, the 3-month Euribor interest
rate is used as a proxy for the European short-term interest rate (SR).
These three public information variables are one month-lagged to be included in the
conditional models. Moreover, following Ferson and Schadt (1996), these information
variables are demeaned to allow an appropriate interpretation in the conditional
performance models.
Table 2 reflects significant problems in these three predetermined information variables,
questioning their stock return predictability, which could be induced by a spurious
regression bias as a consequence of econometric problems. Thus, significantly high
correlation coefficients are found between the three information variables, as shown in
panel C of table 2.
In addition to this linearity problem, panel B shows very high autocorrelation
coefficients of order 1 for the three variables considered in the analysis, confirming a
possible spurious regression bias to predict stock excess returns.14
13
Hardouvelis, Malliaropulos and Pristley (2006) find that the expected returns of European stock
markets are explained more by European Union market risk and less by local risks since the creation of
the European Monetary Union.
14
The stationarity of the three predetermined information variables was tested using the augmented
Dickey-Fuller (1979) test. The null hypothesis of a unit root can not be rejected in any case for a 5%
MacKinnon critical value and using lags from 1 up to 6 periods.
11
Table 2. Summary statistics of the predetermined information variables
Panel A reports annual statistics of the three 1-month lagged and demeaned information variables:
European dividend yield (DY), European short-term interest rate (SR) and slope of the European term
structure (TERM). These statistics are computed for the period December 1999-December 2006. Panel B
shows the autocorrelation coefficients of order 1, 3, 6 and 12 for DY, SR and TERM. Panel C shows the
correlation matrix of these three predetermined information variables.
PANEL A: Summary statistics
DY
SR
TERM
Mean
0.0000%
0.0000%
0.0000%
Median
0.2886%
-0.1084%
0.0755%
Maximum
0.7509%
1.9916%
1.0055%
Minimum
-1.0263%
-1.0684%
-1.2245%
Standard Deviation
0.0059
0.0095
0.0060
Skewness
-0.2786
0.5805
-0.2619
Kurtosis
1.4755
2.1048
2.2011
Jarque-Bera test
9.3307
7.6117
3.2324
PANEL B: Autocorrelation coefficients
DY
SR
TERM
0.936
0.976
0.989

0.749
0.904
0.941

0.466
0.780
0.822

-0.020
0.536
0.517

DY
SR
TERM
PANEL C: Correlation matrix
DY
SR
1.000
-0.859
1.000
TERM
0.335
-0.673
1.000
Table 3. Summary statistics of the predetermined information variables (stochastically detrended)
Panel A reports some annual statistics of the three 1-month lagged, stochastically detrended and
demeaned information variables: European dividend yield (DY), European short-term interest rate (SR)
and slope of the European term structure (TERM). These statistics are computed for the period December
1999-December 2006. Panel B shows the autocorrelation coefficients of order 1, 3, 6 and 12 for DY, SR
and TERM. Panel C shows the correlation matrix of these three predetermined information variables.
PANEL A: Summary statistics
DY
SR
TERM
Mean
0.0000%
0.0000%
0.0000%
Median
-0.0875%
-0.0255%
-0.0550%
Maximum
0.6526%
1.1553%
1.0633%
Minimum
-0.3744%
-1.1697%
-1.0525%
Standard Deviation
0.0025
0.0056
0.0052
Skewness
0.9502
0.1681
-0.0298
Kurtosis
3.1840
2.4635
2.3815
Jarque-Bera test
12.9113
1.4197
1.3675
PANEL B: Autocorrelation coefficients
DY
SR
TERM
0.931
0.931
0.965

0.689
0.658
0.834

0.271
0.161
0.550

-0.355
-0.408
0.019

DY
SR
TERM
PANEL C: Correlation matrix
DY
SR
1.000
-0.595
1.000
12
TERM
0.410
-0.748
1.000
In order to solve this spurious regression problem induced by persistent lagged
regressors, we follow the simple approach of Ferson, Sarkissian and Simin (2003b) by
subtracting a 12-month moving average from the information variables. These
stochastically detrended information variables should have lower autocorrelation
coefficients, which would not cause a spurious regression bias to estimate future stock
excess returns. These stochastically detrended variables are also demeaned to allow
inclusion in the conditional performance models.
Lower autocorrelation coefficients of order 1 are reported for the three stochastically
detrended variables ranging between 0.93 and 0.96 which are not significant values to
induce spurious regressions as suggested by Ferson et al. (2003b). Finally, the null
hypothesis of a unit root is rejected in every case by using the augmented Dickey-Fuller
(1979) test for a 5% MacKinnon critical value and using lags from 1 up to 6 months.
These findings question the stock return predictability of the demeaned and 1-lagged
information variables described in table 2, supporting therefore the use of the
stochastically detrended information variables in the conditional performance models
(e.g. Ferson et al., 2003b; Ayadi and Kryzanowski, 2004; Leite and Cortez, 2006).
Our analysis will test the impact of the spurious regression problem on excess return
predictability, but it will also analyse the impact of the stochastically detrended
variables on the performance estimates of conditional models.
4. Empirical analysis
An equally-weighted portfolio that includes all surviving and terminated global funds
during the period December 1999-December 2006 has been formed to evaluate the
performance of these funds on an appropriate basis. The results shown in this section
are, therefore, not affected by survivorship bias.
The use of this equally-weighted portfolio instead an individual analysis of each global
fund is justified by the impact that survivorship bias could have on the average
performance of the individual performance estimates of surviving global funds.
However, the impact of survivorship bias on performance evaluation (conditional and
unconditional) will be tested in the last section of this empirical analysis.
The performance estimates of a size-weighted portfolio including all surviving and
terminated global funds instead an equally-weighted one would present a significant
size bias because this size-weighted portfolio would be dominated by the effect of a
13
small number of large global funds.15 This size bias motivates the use of an equallyweighted portfolio in order to assess the Spanish global fund performance appropriately
by using both unconditional and conditional measures.
4.1. Unconditional performance evaluation
The Jensen’s alpha estimate reported in table 4 shows that the European stock market
(MSCI Europe TR) is not outperformed by global fund managers, obtaining a negative
value of -0.0894% per month, which is not statistically different from zero.
This negative and non-significant result is very similar to the evidence provided by most
of the previous empirical studies. Taking into account that this result is obtained before
subtracting management and custodial fees from the gross returns, this unconditional
alpha estimate proves to be a bad performance result.16
The low but significant value of the unconditional beta p suggests that Spanish global
funds do not take high levels of systematic risk as a result of the allocation of
conservative portfolios with respect to the European stock benchmark. This might be
seen as a surprising result taking into account that Spanish global funds are allowed to
diversify their portfolio holdings with fewer restrictions than the rest of Spanish mutual
funds.
However, the adjusted determination coefficient obtained in the unconditional model is
relatively high, supporting the significance of the regression.
In addition to the European stock market, the conditional approach of the following
sections will allow us to determine the significance of other variables in the
performance estimates of these portfolios –European dividend yield, European short
term interest rates and European term structure–.
Table 4. Performance and risk estimates – Unconditional model –
This table shows the unconditional performance estimates ( p) and the systematic risk (p) obtained by
applying the Jensen’s measure on the returns obtained by an equally-weighted portfolio that includes all
surviving and terminated Spanish global funds from December 1999 to December 2006. The adjusted R2
coefficient is also presented in the last column of the table.
Adjusted R2
p
p
***
-0.000894
0.464029
73.69%
Equally-weighted portfolio
Standard errors are consistent with heteroskedasticity and autocorrelation problems
(Newey and West, 1987)
*
10% statistically significant **5% statistically significant ***1% statistically significant
15
The ten largest surviving global funds share approximately 31% of total net assets managed by all
global funds that existed at the end of December 2006, clearly highlighting the level of concentration in
the Spanish global fund market.
16
Management and custodial fees are daily subtracted from the gross returns obtained by Spanish funds.
14
4.2. Conditional performance evaluation
Stock return predictability
Before applying the conditional performance models described in section 2 of this
study, the statistical significance of the predetermined information variables is tested in
order to predict stock excess returns.
First, simple linear regressions of monthly excess returns of MSCI Europe TR on each
1-month lagged and demeaned information variable are applied. The left section of table
5 displays statistical significance for the three predetermined information variables
included in these simple linear regressions. These results provide evidence that higher
European dividend yields, a higher slope of the European term structure and lower
European short term interest rates predict higher European stock excess returns.17
Moreover, we also run the same simple linear regressions but using the excess returns of
the equally-weighted portfolio as the dependent variable. The results obtained in these
models are very similar to those obtained previously for the excess returns of the MSCI
Europe TR, although the adjusted R2 coefficients are slightly higher when European
short term interest rate and the slope of the European term structure are considered as
the explanatory variables in these new simple linear regressions.
In addition to these regressions, a multiple linear regression that jointly includes the
three predetermined information variables is also applied for both excess returns of
MSCI Europe TR and the equally-weighted portfolio, finding that the short term interest
rate is the only information variable with statistical significance for the prediction of
excess returns in the equally-weighted portfolio.
However, the Wald test rejects the hypothesis that the predetermined information
variables are jointly equal to zero for the excess returns of both MSCI Europe TR and
the equally-weighted portfolio, providing evidence that expected excess returns are
time-varying as public information changes over time.
These results should motivate the application of conditional performance models, but
following the approach of Ferson et al. (2003b), the evidence of persistent lagged
regressors reported in table 2 suggests the transformation of these predetermined
information variables in order to avoid any spurious regression bias.
17
Short term interest rate is the information variable with the highest significance level for the estimation
of expected excess returns of the stock market.
15
Table 5. Regressions on the 1-month lagged and demeaned information variables
The table reports the simple and multiple linear regressions of monthly excess returns of both European
stock market and equally-weighted portfolio on the 1-month lagged and demeaned information variables.
The slope coefficients and their statistical significance are reported for every simple and multiple
regressions: European dividend yield (DY), European Short term interest rate (SR) and the slope of the
European term structure (TERM). The figures in brackets represent the adjusted R 2 obtained in every
linear regression. Wald p-value is the probability value of the 2 statistic and it tests for the null
hypothesis that all variables included in the multiple linear regressions have a coefficient equal to zero.
Dependent variable:
Dependent variable:
MSCI Europe TR
Equally-weighted portfolio
2.014*** (5.54%)
1.011** (4.64%)
Slope DY
Slope SR
-1.359*** (6.76%)
-0.799*** (8.25%)
Slope TERM
0.779
0.641
Slope DY
-1.117
Slope SR
-2.685*
Slope TERM
-1.726
Wald p-value
(-0.17%)
(1.20%)
-1.244
-1.859**
(6.24%)
(7.88%)
-0.939
0.0014
0.0083
Standard errors are consistent with heteroskedasticity and autocorrelation problems
(Newey and West, 1987)
*
10% statistically significant **5% statistically significant ***1% statistically significant
Table 6 shows the results of the simple and multiple linear regressions of the excess
returns of both MSCI Europe TR and equally-weighted portfolio on the 1-month lagged,
stochastically detrended and demeaned information variables. As shown in table 3, the
use of these stochastically detrended information variables seems to remove any
spurious regression bias in the statistical significance of the information variables.
However, it considerably reduces the significance of the predetermined information
variables for the prediction of excess returns, throwing up very different results from
those presented in table 5.
In fact, there is no slope coefficient with 10% statistical significance in the new linear
regressions. This finding means that the high statistical significance of the European
predetermined information variables for the prediction of excess returns seems to be a
consequence of spurious regression problems.
In order to test the impact of these spurious regression problems on the conditional
performance estimates, both sets of predetermined information variables – stochastically
detrended and non-stochastically detrended – will be considered in the following
conditional performance analyses.
16
Table 6. Regressions on the 1-month lagged and demeaned information variables
(stochastically detrended)
The table reports the simple and multiple linear regressions of monthly excess returns of both European
stock market and equally-weighted portfolio on the 1-month lagged, stochastically detrended and
demeaned information variables. The slope coefficients and their statistical significance are reported for
every simple and multiple regressions: European dividend yield (DY), European Short term interest rate
(SR) and the slope of the European term structure (TERM). The figures in brackets represent the adjusted
R2 obtained in every linear regression. Wald p-value is the probability value of the 2 statistic and it tests
for the null hypothesis that all variables included in the multiple linear regressions have a coefficient
equal to zero.
Dependent variable:
Dependent variable:
MSCI Europe TR
Equally-weighted portfolios
1.076 (-0.85%)
0.563 (-0.87%)
Slope DY
Slope SR
0.036
Slope TERM
-0.234
Slope DY
1.694
Slope SR
0.212
Slope TERM
-0.400
Wald p-value
(-1.20%)
(-1.13%)
-0.267
(-0.84%)
0.196
(-1.03%)
0.325
(-2.99%)
-0.198
(-3.25%)
-0.027
0.9301
0.9549
Standard errors are consistent with heteroskedasticity and autocorrelation problems
(Newey and West, 1987)
*
10% statistically significant **5% statistically significant ***1% statistically significant
Partial conditional model
The results of the partial conditional model shown in panel A of table 7 provide
evidence that the incorporation of the non stochastically detrended information variables
improves the unconditional performance estimate and its explanatory power, which may
indicate that fund managers use private information. These results are in line with most
empirical studies, which support the hypothesis of a negative correlation between
conditional betas and expected returns of the market.18
In contrast, the unconditional performance estimate is not improved by using the
stochastically detrended information variables. Moreover, the individual statistical
significance of the conditional betas disappears, with the sole exception of the European
short term interest rate. In fact, the Wald p-value reported in panel B of table 7 does not
reject the null hypothesis of conditional betas equal to zero.
This lack of signification proves that the evidence of time-varying betas found in panel
A of table 7 is a consequence of a spurious regression problem, as anticipated in the
previous section of this empirical analysis.
18
Ferson and Schadt (1996) and Ferson and Warther (1996) suggest that this negative relationship may be
caused by the opposite change of the fund portfolio betas with respect the market, and the important cash
flows into funds, which may not be invested immediately, when the expected market returns are high.
17
Although the impact of the use of stochastically detrended variables on the performance
estimates is not important in absolute terms, 0.2217%, the sign and the financial
interpretation of this performance result change in an important way.
In fact, the partial conditional alpha, 0.1122%, obtained from the use of nonstochastically detrended information variables leads us to affirm that global funds
outperform the European market before charging management and custodial fees.
However, the use of stochastically detrended information variables sharply diminishes
this conditional alpha estimate, giving a negative value of 0.1095%.
This result confirms that the consideration of stochastically detrended information
variables has a relevant impact in terms of sign and financial interpretation of the partial
conditional performance estimates.
Table 7. Performance and risk estimates - Partial conditional model Panel A of the table presents the performance and risk estimates of the partial conditional model using 1month lagged and demeaned information variables. Panel B shows the performance and risk estimates of
the partial conditional model using 1-month lagged, stochastically detrended and demeaned information
variables. Both panels compute the partial conditional model for an equally-weighted portfolio that
includes all surviving and terminated funds from December 1999 to December 2006. Each panel exhibits
the performance estimate p, the average conditional beta 0p, and the coefficient estimates for the
conditional beta function: European dividend yield (DY), European Short term interest rate (SR) and the
slope of the European term structure (TERM). The adjusted determination coefficient is also presented
for each model. Wald is the probability value of the 2 statistic and it tests for the null hypothesis that all
variables included in the partial conditional model have a coefficient equal to zero.
Panel A: Partial conditional model (non stochastically detrended information variables)
DY
SR
TERM
Adj-R2 Wald
p
0p
Equally-weighted
0.001122
0.4365*** 22.9516*** 31.4985***
7.2020
79.94% 0.000
portfolio
Panel B: Partial conditional model (stochastically detrended information variables)
DY
SR
TERM
Adj-R2 Wald
p
0p
Equally-weighted
-0.001095 0.4828***
26.6278
19.5978*
-0.7685
74.63% 0.200
portfolio
Standard errors are consistent with heteroskedasticity and autocorrelation problems
(Newey and West, 1987)
*
10% statistically significant **5% statistically significant ***1% statistically significant
Full conditional model
Table 8 shows that the explanatory power of full conditional models is higher than that
obtained using the unconditional methodology, but it is slightly lower than the adjusted
R2 coefficients obtained in the partial conditional framework.
Full conditional betas are very similar to those obtained in partial models, providing
evidence that the statistical significance of the 1-month lagged and demeaned
information variables disappears when they are stochastically detrended. In this sense,
Wald probability values (W2 and W3) also jointly confirm that the conditional beta
parameters lack statistical significance.
18
Table 8. Performance and risk estimates - Full conditional model Panel A of the table presents the performance and risk estimates of the full conditional model using 1month lagged and demeaned information variables. Panel B shows the performance and risk estimates of
the full conditional model using 1-month lagged, stochastically detrended and demeaned information
variables. Both panels compute the full conditional model for an equally-weighted portfolio that includes
all surviving and terminated funds from December 1999 to December 2006. Each panel exhibits the
average conditional performance estimate 0p, the average conditional beta 0p, and the coefficient
estimates for the conditional beta function: European dividend yield (DY), European Short term interest
rate (SR) and the slope of the European term structure (TERM). The adjusted determination coefficient is
also presented for each model. W1, W2 and W3 are the Wald probability values of the 2 statistic that test
for the null hypothesis that the coefficients of the conditional alphas, the conditional betas and both
conditional alphas and betas included in the full conditional model have a coefficient equal to zero,
respectively.
Panel A: Full conditional model (non stochastically detrended information variables)
DY
SR
TERM Adj-R2
W1
W2
W3
0p
0p
Equally-weighted
0.001084 0.4303*** 22.322*** 30.918*** 7.5360 79.64% 0.3988 0.0000 0.0000
portfolio
Panel B: Full conditional model (stochastically detrended information variables)
DY
SR
TERM Adj-R2
W1
W2
W3
0p
0p
Equally-weighted
-0.001069 0.4839*** 24.325
19.988
0.8476 74.12% 0.7027 0.2410 0.5319
portfolio
Standard errors are consistent with heteroskedasticity and autocorrelation problems
(Newey and West, 1987)
*
10% statistically significant **5% statistically significant ***1% statistically significant
In addition to this result, the full conditional alpha estimates are very similar to those
obtained by partial conditional models.
In line with the results reported by Otten and Bams (2004), Wald probability values
(W1) suggest that full conditional alphas are not time-varying. This finding is exhibited
for both full conditional models using non-stochastically and stochastically detrended
information variables.
Meanwhile, the use of stochastically detrended information variables seems to have a
relevant impact on the full conditional performance estimates. The consideration of
these transformed information variables leads to a negative alpha of 0.1069% per
month, which is a slightly better performance result than that obtained by the partial
conditional model but very similar to the unconditional alpha estimate.
On the other hand, the use of 1-month lagged and demeaned information variables
improves the performance estimate as a consequence of the spurious regression
problems previously detected, obtaining a positive alpha of 0.1084%.
These differences between the signs and interpretations of full conditional performance
estimates allow us to conclude that conditional alphas are very sensible to the spurious
regression problems that may affect conditional models when the information variables
are not stochastically detrended.
19
4.3. Size and experience effect
Both size and experience effects on the performance estimates of Spanish global funds
are analysed in this sub-section based on the unconditional and conditional performance
models.
These performance models are applied by introducing different dependent variables
according to the specific effect to be analysed and testing the appropriate hypotheses.
As a consequence of the spurious regression problems detected in the previous sections
of the empirical analysis, the conditional models use the stochastically detrended
information variables in order to obtain rigorous conclusions.
Fund size effect
In Spain’s highly concentrated mutual fund industry, where the average fund size is one
of the lowest in Europe, a few large global funds coexist with much smaller
counterparts.
An obvious question arises from this market map: Do the largest global funds obtain
better performance results than the rest of global funds? In other words, are there
economies of scale in the performance results obtained by Spanish global fund
industry?
In order to address this question on an appropriate basis, total net assets of the funds
(TNA) is taken as a proxy for the fund-size variable. We then proceed to test the null
hypothesis of economies of scale (7) by comparing the performance estimate of an
equally-weighted portfolio with that obtained by a size-weighted portfolio. This
approach has been widely used in international research to analyse the fund size effect
on performance evaluation.

H0: S-W > E-W (Fund size effect exists)
(7)
Where E-W (S-W) represents the performance estimate of the equally-weighted
(size-weighted) portfolio. Both portfolios include all surviving and terminated
global funds during the whole study horizon.
The results shown in table 9 provide evidence of a negative and irrelevant effect of fund
size on performance estimates, rejecting the aforementioned null hypothesis (7).
The conditional and unconditional alpha estimates for the fund size-weighted portfolio
are negative and not statistically significant, which means a bad performance result for
large global funds.
20
Table 9: Unconditional and conditional performance estimates – size and experience effects –
This table reports the performance and risk estimates obtained from the application of the unconditional,
partial conditional and full conditional evaluation models on different weighted portfolios. Comparison
of the alpha estimates of the equally-weighted portfolio and the size-weighted portfolio permits the null
hypothesis (7) to be tested. Comparison of the alpha estimates of the equally-weighted portfolio and the
company size-weighted portfolio permits the null hypothesis (8) to be tested. Comparison of the alpha
estimates of the equally-weighted portfolio and the experience-weighted portfolio permits the null
hypothesis (9) to be tested. The information variables used in the conditional models are 1-month lagged,
stochastically detrended and demeaned.
Panel A: Unconditional model
Adj-R2
p
p
Equally-weighted
-0.000894 0.464029***
73.69%
portfolio
Size-weighted
-0.001492 0.423646***
54.12%
portfolio
Fund size effect -0.000598
Company size-0.000935 0.3576***
62.14%
weighted portfolio
Fund company size
-0.000041
effect
Experience-0.000830 0.456405***
73.21%
weighted portfolio
Experience effect 0.000064
Panel B: Partial conditional model (stochastically detrended information variables)
DY
SR
TERM Adj-R2
Wald
p
0p
Equally-weighted
-0.001095 0.4828*** 26.6278 19.5978* -0.7685 74.63%
0.200
portfolio
Size-weighted
-0.001713 0.4649*** 27.9828 36.3378*** -1.1205 58.62%
0.0257
portfolio
Fund size effect -0.000618
Company size-0.001097 0.3853*** 26.0053 30.2212*** 5.7454 66.05%
0.0025
weighted portfolio
Fund company size
-0.000002
effect
Experience-0.001028 0.4752*** 25.8071 18.9639* -1.4323 74.13%
0.2163
weighted portfolio
Experience effect 0.000067
Panel C: Full conditional model (stochastically detrended information variables)
DY
SR
TERM Adj-R2
W1
W2
W3
0p
0p
Equally-weighted
-0.001069 0.4839***
24.325
19.988
0.8476 74.12% 0.7027 0.2410 0.5319
portfolio
Size-weighted
-0.001685 0.4658***
25.042
35.131** -0.1866 57.50% 0.8474 0.0573 0.0480
portfolio
Fund size effect -0.000616
Company size-0.001075 0.3862*** 23.9465 30.1581*** 6.9169 65.18% 0.8419 0.0101 0.0043
weighted portfolio
Fund company size
-0.000006
effect
Experience-0.001001 0.4762*** 23.3797
19.3184
0.2312 73.66% 0.6814 0.2586 0.5599
weighted portfolio
Experience effect 0.000068
Standard errors are consistent with heteroskedasticity and autocorrelation problems
(Newey and West, 1987)
*
10% statistically significant **5% statistically significant ***1% statistically significant
21
Taking into account the management and custodial fees charged by Spanish global
funds, we conclude that large global funds are slightly cheaper than their small
counterparts and obtain worse performance results than the rest of the market.19
The individual significance tests and the Wald p-values obtained in both the partial and
full conditional models reveal that the beta coefficients of the stochastically detrended
information variables obtain much higher levels of statistical significance in the fund
size-weighted portfolio than in the equally-weighted, especially for European short-term
interest rate. This result indicates that predetermined information has a major impact on
the performance of large funds.
We also find that large global funds allocate more conservative portfolios obtaining
lower p (0p) parameters in the unconditional (conditional) models than those obtained
by the rest of funds. However, the adjusted determination coefficients are lower for the
size-weighted portfolio in every performance model.
Fund company size effect
In addition to the analysis of the impact of fund size on performance results, the
concentration of Spanish fund management companies makes it necessary to test the
economies of scale phenomenon in those global funds that are managed by large fund
companies.
Large companies should have more and better management resources than their small
counterparts. The question arising from this market map is whether fund management
companies with large market shares perform better than their counterparts with a modest
market share.
To answer this question and assess the potential incidence of the highly concentrated
Spanish fund industry on performance results, we compare the performance estimate of
an equally-weighted portfolio with that obtained by a fund company size-weighted
portfolio.20
In order to form these portfolios on an appropriate basis, fund company size is
computed as the total net assets managed by each fund company at the end of each
month.
19
The size-weighted management and custodial fee, 1.3512% per year, is slightly lower than the
arithmetic average of the management and custodial fees charged by the Spanish global funds, 1.3572%
per year.
20
Please, notice the high concentration of the Spanish global fund market where the two largest fund
companies - Santander Gestión and BBVA Gestión - manage 41.93% of the total assets of all global funds
that existed in December 2006.
22
We then construct the hypothesis (8) to test the fund company size effect on the
performance estimates of Spanish global funds:

H0 : CS-W > E-W (Fund company size effect exists)
(8)
Where E-W (CS-W) represents the performance estimate of the equally-weighted
(company size-weighted) portfolio. Both portfolios include all surviving and
terminated global funds during the whole study horizon.
In view of the results shown in table 9, we reject the null hypothesis (8) of a positive
effect of the fund company size on the performance results of Spanish global funds.
This evidence is found in every unconditional and conditional performance model, but
the negative difference is not significant, -0.0041% per month in the unconditional
model and -0.0002% (-0.0006%) per month in partial (full) conditional models.
The negative performance estimates obtained by the company size-weighted portfolio
allow us to conclude that Spanish global funds managed by large companies are not
good performers. However, these negative alphas are not statistically significant.
The analysis of the management and custodial fees charged by these funds provides
evidence that global funds managed by large companies are more expensive than those
managed by small fund companies.21
We also find joint statistical significance (Wald p-values) of the predetermined
information variables of the company size-weighted portfolio, although the European
short term interest rate is the only information variable that obtains individual
significance.
Moreover, global funds managed by large companies seem to allocate more
conservative portfolios, obtaining lower and significant European stock market betas
than those obtained by the rest of funds.
To sum up, the evidence leads us to the conclusion that global funds managed by large
companies are more expensive and allocate more conservative portfolios. Furthermore,
they do not obtain better performance results than those funds managed by companies
with a modest market share.
21
The fund company size-weighted management and custodial fee, 1.3824% per year, is higher than the
arithmetic average of the management and custodial fees charged by the Spanish global funds, 1.3572%
per year.
23
Experience effect
Given that older fund management companies may be expected to display greater
experience in non-local markets than their more recent peers in the Spanish fund
industry, we have examined the effect of experience on performance estimates.
In order to test this effect, the registration date of the fund company provided by the
Spanish market supervisor (CNMV) is used as a proxy for the age, and therefore the
experience, of the fund management company. Those companies with earlier
registration dates are assumed to have more experience than recently registered
companies in the Spanish market.
As a result of their superior abilities, the performance estimates of the funds managed
by the experienced companies should be higher than those obtained by the novice
companies.
This experience effect should be especially identifiable in global funds, because they are
allowed to invest in international markets with fewer restrictions than the rest of
Spanish mutual funds.
Thus, the question to be analysed is: Do older fund management companies perform
better than their less experienced counterparts?
In order to answer this question, we compare the performance estimate of an equallyweighted portfolio with that obtained by a company experience-weighted portfolio. The
fund company experience is computed as the age of each fund company at the end of
each month included in the study horizon.22
The null hypothesis to be tested is the following one:

H0: CE-W > E-W
(Experience economies effect exists)
(9)
Where E-W represents the performance estimate of the equally-weighted
portfolio and CE-W represents the performance result of the company
experience-weighted portfolio
Table 9 provides evidence for a positive and irrelevant effect of the experience on
performance estimates of Spanish global funds. Thus, the performance of the funds
managed by the more experienced companies is slightly higher than that obtained by
funds managed by other less experienced companies.
22
At December 2006, the average age of the ten fund companies with the earliest registration dates is
21.14 years, which means that these companies have actually existed since the very beginning of the
recent Spanish fund industry. At the same time, the average age of the ten most recent fund companies is
1.22 years.
24
However, the unconditional and conditional performance estimates obtained by the
experience-weighted portfolios are negative and not significant, what means a bad
performance result for those experienced fund companies.
These findings prove that global funds managed by more experienced companies are
more expensive23 and perform slightly better than funds managed by companies
recently established in the industry.
The high Wald p-values found in the experience-weighted portfolios may be interpreted
as a non significant relationship between fund performance and the economic context.
4.4. Size and experience effect: Detecting common patterns
In order to detect the common impact of the size and experience on the performance
results, we form eight equally-weighted portfolios based on the median criterion.
Thus, all global funds existing in each month of the study horizon are classified in two
groups according to the median of the fund size, the fund company size and the
company experience, respectively.
The combination of these three median-groups24 allows us to form eight different
clusters of funds for each month according to the three aforementioned characteristics.
We then apply the unconditional and conditional performance models on the equallyweighted portfolios obtained from the eight different clusters of funds. The analysis of
these results allows us to detect common effects on the performance estimates.
The results found in table 10 identify important differences when joint effects are
considered. In this sense, the difference of the unconditional and conditional
performance results between the top and bottom performers is higher than 0.38% per
month, which proves to be a significant common effect.
In this sense, table 10 provides evidence of a significant joint impact of the fund size
and the fund company size on the performance estimates of global funds. Thus, large
global funds managed by small fund companies obtain the best conditional and
unconditional performance results regardless of the experience of the fund company.
Although these alpha estimates are not significant, their positive signs are representative
of a very good performance result for these funds.
23
The company experience-weighted management and custodial fee, 1.3708% per year, is higher than the
arithmetic average of the management and custodial fees charged by the Spanish global funds, 1.3572%
per year.
24
The median-groups are: Large funds vs. Small funds, Large companies vs. Small companies and Highexperienced companies vs. Low-Experienced companies.
25
Table 10: Unconditional and conditional performance estimates – common effects –
This table reports the performance and risk estimates obtained from the application of the unconditional,
partial conditional and full conditional evaluation models on eight different equally-weighted portfolios
based on the median criterion for the three following characteristics: Fund size, fund company size and
fund company experience. The information variables used in the conditional models are 1-month lagged,
stochastically detrended and demeaned.
Panel A: Unconditional model
Fund Company Company
Size
Size
Exper.
Adj-R2
p
p
***
0.001270 0.4517
78.57%
Large Small
Low
0.000822 0.5894***
72.07%
Large Small
High
0.000107 0.4650***
74.78%
Small Small
Low
68.68%
Small Large
High -0.000556 0.3950***
73.51%
Small Small
High -0.000584 0.5220***
46.32%
Large Large
High -0.000695 0.2406***
-0.001894 0.5876***
66.27%
Small Large
Low
-0.002624 0.4788***
62.32%
Large Large
Low
Panel B: Partial conditional model (stochastically detrended information variables)
Fund Company Company


Size Size
Exper.
DY
SR
TERM Adj-R2
p
0p
***
***
0.001170 0.4709
19.6169 23.9558
8.1875 80.21%
Large Small
Low
0.000755 0.6161*** 26.4588* 40.5862*** 23.2249* 74.95%
Large Small
High
0.000011 0.4957*** 22.6265 36.5572*** 13.5405 79.13%
Small Small
Low
9.5879
-5.3671 73.22%
Small Small
High -0.000725 0.5349*** 14.7892
Small Large
High -0.000744 0.4123*** 25.7294 18.8677* 0.2569 69.68%
Large Large
High -0.000831 0.2792*** 15.3366* 30.7248*** -2.3650 58.22%
-0.002474 0.5909*** 60.9017
-0.0045 -25.8795 69.28%
Small Large
Low
-0.002798* 0.4870*** 15.6402
2.9461 -10.6093 61.67%
Large Large
Low
Panel C: Full conditional model (stochastically detrended information variables)
Fund Company Company


Size Size
Exper.
DY
SR
TERM Adj-R2
W1
0p
0p
***
**
0.001191 0.4712
17.2855 23.1907
8.9048 79.86% 0.5951
Large Small
Low
0.000786 0.6163*** 22.7513 38.5430*** 23.8050** 74.55% 0.7779
Large Small
High
0.0000399 0.4961*** 19.4221 34.9669*** 14.2280* 79.00% 0.3774
Small Small
Low
9.3794
-3.4696 72.91% 0.3915
Small Small
High -0.000687 0.5362*** 11.2441
17.8532
0.8088 69.00% 0.7031
Small Large
High -0.000723 0.4124*** 23.3665
Large Large
High -0.000818 0.2791*** 14.1616* 31.4208*** -1.4340 57.34% 0.6525
-0.002442 0.5943*** 59.1654*
2.6984 -22.4779 69.03% 0.3850
Small Large
Low
-0.002782* 0.4881*** 15.0206
6.0050
-8.1826 60.89% 0.6667
Large Large
Low
Standard errors are consistent with heteroskedasticity and autocorrelation problems
(Newey and West, 1987)
*
10% statistically significant **5% statistically significant ***1% statistically significant
Wald
0.0113
0.0115
0.0000
0.6399
0.1922
0.0001
0.1451
0.8601
W2
0.0919
0.0127
0.0001
0.7671
0.2878
0.0001
0.1241
0.8662
Furthermore, the Wald p-values obtained by large funds managed by small fund
companies in both the partial and full conditional models reveal the joint statistical
significance of the predetermined information parameters, which may indicate that
predetermined information has a major impact on the performance of these funds.
The possible explanation for this result may be that small fund companies with fewer
management resources than their large counterparts concentrate their management
efforts on a small number of large funds obtaining good performance results.
26
W3
0.0083
0.0635
0.0001
0.7350
0.3753
0.0000
0.0340
0.3112
In contrast, large fund companies may have to share their management abilities in a
high number of funds, which seems to lead to bad performance results. In this sense,
large fund companies without experience obtain the worst alpha estimates for every
performance model, even with statistical significance for large global funds.
In addition to this result, those small funds managed by small and non experienced
companies also obtain positive but not significant performance results with very
significant Wald p-values for the full conditional model, which rejects the null
hypothesis of both conditional alphas and betas equal to zero.
Furthermore, the conditional performance results are lower than the unconditional
estimates for every cluster of funds, which may indicate the lack of use of private
information by Spanish global funds, which supports the results previously found in our
study.
An additional but interesting result are the relevant differences found in the risk
exposure to the European stock market, ranging from the 0.27 allocated by large funds
managed by large and experienced companies to the 0.61 allocated by large funds
managed by small and experienced companies. However, this risk exposure does not
seem to have a relevant impact on the performance results as shown in table 10.
4.5. Survivorship bias
The performance findings presented so far in this paper are not affected by survivorship
bias, since the weighted portfolios formed in the preceding sections include all funds
that existed in each month for the whole study period, allowing us to consider the
performance results of terminated global funds from December 1999 to December 2006.
The magnitude and significance of survivorship bias on mutual fund performance is a
controversial topic in the literature with some authors arguing that the impact is
marginal and other studies providing evidence for a significant effect on performance
results.25
In this light, the question addressed in this section is: Are surviving funds better
performers than terminated funds? An affirmative answer would confirm the hypothesis
that the worst performers are expelled from the market.
As suggested in the introduction of our work, the analysis of survivorship bias is
motivated by the largely unexplored impact of this bias on conditional performance
25
Relevant examples of these controversial findings are Brown et al. (1992), Malkiel (1995), Wermers
(1997), Blake and Timmermann (1998), Hallahan and Faff (2001) and Carhart et al. (2002).
27
evaluation. Our analysis is based on the conditional performance results obtained by
both stochastically and non-stochastically detrended information variables.26 The effect
of this bias on the unconditional performance estimates is also reported.
The impact of survivorship bias on both unconditional and conditional performance
models is computed by measuring the difference between the performance estimates of
an equally-weighted portfolio made up of surviving funds at the end of the study period
and the results obtained by an equally-weighted portfolio that includes all the funds of
the study sample (surviving and non-surviving funds).
Table 11: Unconditional and conditional performance estimates –Survivorship bias effect–
This table reports the performance and risk estimates obtained from the application of the unconditional,
partial conditional and full conditional evaluation models on both equally-weighted portfolios including
surviving funds and all funds (surviving and terminated). The information variables used in the
conditional models that are reported in panels B and C are 1-month lagged, stochastically detrended and
demeaned.
Panel A: Unconditional model
Adj-R2
p
p
***
Surviving funds -0.000576 0.493599
74.61%
All funds
-0.000894 0.464029***
73.69%
Survivorship
0.000318
bias
Panel B: Partial conditional model (stochastically detrended information variables)
DY
SR
TERM
Adj-R2
Wald
p
0p
***
**
Surviving funds -0.000755 0.5144
25.2439 22.1924
1.5388
75.61%
0.097
All funds
-0.001095 0.4828*** 26.6278 19.5978*
-0.7685 74.63%
0.200
Survivorship
0.000340
bias
Panel C: Full conditional model (stochastically detrended information variables)
DY
SR
TERM
Adj-R2
W1
W2
W3
0p
0p
***
**
Surviving funds -0.000728 0.5151
22.7923 22.3278
2.9877
75.10% 0.7210 0.1683 0.2486
All funds
-0.001069 0.4839*** 24.325
19.988
0.8476
74.12% 0.7027 0.2410 0.5319
Survivorship
0.000341
bias
Standard errors are consistent with heteroskedasticity and autocorrelation problems
(Newey and West, 1987)
*
10% statistically significant **5% statistically significant ***1% statistically significant
The results displayed in table 11 provide evidence of a positive survivorship bias
estimate for the Spanish global funds analysed during the period December 1999December 2006. These monthly estimates are not significant in magnitude and range
26
As far as we know, the only studies to analyse the impact of survivorship bias on conditional
performance results are Ayadi and Kryzanowski (2004) and Leite and Cortez (2006). However, both
studies analyse the impact on the performance estimates obtained by stochastically-detrended information
variables, while our work the first to test this bias on both conditional estimates based on nonstochastically and stochastically detrended variables.
28
from 0.0318% to 0.0341%, which are very similar to the recent findings of Leite and
Cortez (2006) for the Portuguese market.
Survivorship bias is also positive using non-stochastically detrended variables, and it is
slightly higher in magnitude (0.037% for the partial conditional model and 0.0345% for
the full conditional model).
The positive sign and the small magnitude of survivorship bias lead us to conclude that
non-surviving funds are slightly worse performers than surviving funds. However, the
poor performance of these non-surviving funds might not be the only reason they were
expelled from the market. Thus, a relevant number of non-surviving funds disappeared
from database because they switched their global investment vocation to other
international categories, with the result that the Spanish market supervisor (CNMV)
changed their official investment category.
The survivorship bias estimates are higher when more conditional information is
included in the performance models.27 Thus, the bias estimate is highest in the full
conditional performance model and lowest in the unconditional approach. This result is
consistent with that found by Leite and Cortez (2006) for the Portuguese market, but it
contrasts with the findings of Ayadi and Kryzanowski (2004) for the Canadian fund
market.
5. Conclusions
Our study proves that a set of non-stochastically detrended European conditional
variables leads to spurious regression problems in order to predict stock excess returns,
thereby questioning the use of these information variables in conditional models to
evaluate the performance results of Spanish global funds.
The magnitude of these performance estimates is not seriously affected by the use of
stochastically detrended conditional variables, but the sign and financial interpretation
of these results change sharply as a consequence of the correction of the spurious
regressions bias.
The results obtained in our empirical research provide evidence that the incorporation of
non-stochastically detrended information variables improves the unconditional
performance estimates, which may indicate that fund managers use private information.
In contrast, the unconditional performance estimates are not improved by incorporating
27
This statement is not true for non-stochastically detrended variables.
29
stochastically detrended information variables, being the European Short-term interest
rate the only information variable with statistical significance.
Our study also finds a positive but irrelevant effect of the level of experience of the
management company on the performance estimates of Spanish global funds.
Moreover, the null hypotheses of economies of scale for the fund size and the
management company size are rejected.
The specific characteristics of the Spanish fund industry lead us to analyse joint effects
of the size and experience on the conditional and unconditional performance estimates,
revealing a significant impact on performance results. In this sense, large global funds
managed by small companies obtain the best performance estimates, which are positive
and much higher than those negative and significant results obtained by global funds
managed by large companies without experience.
Finally, our research proves that those funds that disappear from the data base are
slightly worse performers than the surviving funds at the end of the study horizon,
providing evidence for a positive, though small, impact of survivorship bias on both
unconditional and conditional performance results.
30
REFERENCES
Ayadi, M.A. and L. Kryzanowsky (2004), “Performance of canadian fixed-income
mutual funds”, Portuguese Finance Network.
Bauer, R.; R. Otten and A.T. Rad (2006), “New Zealand mutual funds: measuring
performance and persistence in performance”, Accounting and Finance, 46, pp. 347363.
Black, A. (2000), “Expected returns and business conditions: A commentary on Fama
and French”, Applied Financial Economics, 2000, 10(4), pp. 389-400.
Blake, D.; B. Lehmann and A. Timmermann (2002), “Performance Clustering and
Incentives in the UK Pension Fund Industry”, Journal of Asset Management, 3 (2), 173194.
Blake, D. and A. Timmermann (1998), “Mutual fund performance: Evidence from the
UK”, European Finance Review, 2, pp. 57-77.
Brown, S.; W. Goetzmann; R. Ibbotson and S. Ross (1992), “Survivorship bias in
performance studies”, Review of Financial Studies, 5, pp. 553-580.
Brown, S. and W. Goetzmann (1995), “Performance Persistence”, Journal of Finance,
50, pp. 679-698.
Campbell, J. Y. (1987), “Stock returns and the term structure”, Journal of Financial
Economics, 18(2), pp. 373-399.
Carhart, M., J. Carpenter, A. W. Lynch, and D. K. Musto (2002), “Mutual Fund
Survivorship”, Review of Financial Studies, 15, 1439-1463.
Chan, H.; R. W. Faff; D.R. Gallagher and A. Looi (2005), “Fund size, fund flow,
transaction costs and performance: size matters”, Working Paper. Monash University,
Clayton and the University of New South Wales, Sydney.
Chen, N. F. (1991), “Financial Investment Opportunities and the Macroeconomy” The
Journal of Finance, 46(2), pp. 529-554.
Chen, J. S., H. G. Hong, M. Huang and J. D. Kubik (2003), "Does Fund Size Erode
Mutual Fund Performance? The Role of Liquidity and Organization" Working Paper
Available at SSRN: http://ssrn.com/abstract=372721
Chen, Z., and P. J. Knez (1996), “Portfolio Performance Measurement: Theory and
Applications”, Review of Financial Studies, 9, 511-555.
Chen, N. F.; R. Roll and S. A. Ross (1986), “Economic Forces and the Stock Market”,
The Journal of Business, 59(3), pp. 383-403.
31
Christoffersen, S. K. and S. Sarkissian (2007), "City Size and Fund Performance", AFA
2007 Chicago Meetings Paper.
Christopherson, J.; W. Ferson and D. Glassman (1998), “Conditioning manager alphas
on economic information: another look at the persistence of performance”, Review of
Financial Studies, 11, pp. 111-142.
Christopherson, J.; W. Ferson and. A. Turner (1999), “Performance Evaluation Using
Conditional Alphas and Betas”, Journal of Portfolio Management, 26 (1), 59-72.
Cortez, M.C. and F. Silva (2002), “Conditioning information on portfolio performance
evaluation: A reexamination of performance persistente in the portuguese mutual fund
Market”, Finance India, 16, pp. 1393-1408.
Dahlquist, M., S. Engtsrom, and P. Soderlind (2000), “Performance and Characteristics
of Swedish Mutual Funds”, Journal of Financial and Quantitative Analysis, 35, 409423.
Dickey, D.A. and W. A. Fuller (1979), “Distribution of the Estimators for
Autoregressive Time Series with a Unit Root”, Journal of the American Statistical
Association, 74, pp. 427–431.
Ding, B. and R. Wermers (2005), “Mutual fund performance and governance structure:
The role of portfolio managers and boards of directors”, EFA 2005 Moscow Meetings
Paper.
Dybvig, P. and R. Stephen (1985), “Differential Information and Performance
Measurement Using a Security Market Line”, Journal of Finance, 40 (2), 383-399.
Elton, E.; M. Gruber and C. Blake (1996), “Survivorship Bias and Mutual Fund
Performance”, Review of Financial Studies, 9 (4), 1097-1120.
Engstrom, S. (2004), “Investment strategies, fund performance and portfolio
characteristics”, Stockholm School of Economics SSE/EFI Working Paper Series Nº
554.
Fama, E. F., and K. French (1989), “Business Conditions and Expected Returns on
Stocks and Bonds”, Journal of Financial Economics, 25, 23-49.
Ferreira, M.A.; A. Freitas and S. Brito (2007), “The determinants of mutual fund
performance: A cross-country study”, EFMA annual meeting. Vienna.
Ferson W. and C. R. Harvey (1991), “The Variation of Economic Risk Premiums”, The
Journal of Political Economy, 99(2), pp. 385-415.
Ferson, W., and M. Qian (2004), “Conditional Performance Evaluation, Revisited”,
Working Paper, Boston College – EUA.
32
Ferson, W.; S. Sarkissian and T. Simin (2003a), “Spurious regressions in financial
economics?”, Journal of Finance, 58, pp.1393-1413.
--------------------------------------------- (2003b): “Is stock return predictability spurious”,
Journal of Investment Management, 1, pp. 1-10.
Ferson, W. and R. Schadt (1996), “Measuring fund strategy and performance in
changing economic conditions”, Journal of Finance, 51, pp. 425-461.
Ferson, W. and V. Warther (1996), “Evaluating fund performance in a dynamic
market”, Financial Analysts Journal, 52, pp. 20-28.
Gallagher, D. R. and K. M. Martin (2005), "Size and Investment Performance: A
Research Note" , Abacus, Vol. 41, No. 1, pp. 55-65.
Grant, D. (1977), “Portfolio Performance and the “Cost” of Timing Decisions”, Journal
of Finance, 32 (3), 837-846.
Grinblatt, M. and S. Titman (1989), “Portfolio performance evaluation: Old issues and
new insights”, Review of Financial Studies, 2, pp. 393-421.
Hallahan, T.A. and R. W. Faff (2001), “Induced persistence or reversals in fund
performance?: The effect of survivor bias”, Applied Financial Economics, 11, pp. 119126.
Hardouvelis, G., M. Dimitrios and R. Priestley (2006), “EMU and European Stock
Market Integration”, Journal of Business, 79(1), 365-392.
Jensen, M.C. (1968), “The performance of mutual funds in the period 1945-1964”,
Journal of Finance, 23, pp. 346-389.
----------------- (1972), “Optimal Utilization of Market Forecasts and the Evaluation of
Investment Performance”. In Szego, G.; Shell, K. (eds.), Mathematical Methods in
Investment and Finance, North-Holland, 310-335.
Jensen, G. R.; J. M. Mercer and R. R. Johnson (1996), “Business conditions, monetary
policy, and expected security returns”, Journal of Financial Economics, 40(2), pp. 213237.
Keim, D.B. and R.F. Stambaugh (1986), “Predicting returns in the stock and bond
markets”, Journal of Financial Economics, 17(2), 357-390.
Lamont, O. (1998), “Earnings and Expected Returns”, The Journal of Finance, 53(5),
1563.
Leite, P.A. and M. C. Cortez (2006), “Conditional performance evaluation: Evidence
for the Portuguese mutual fund Market”, Portuguese Finance Network.
33
Malkiel, B. (1995), “Returns from investing in equity mutual funds 1971 to 1991”,
Journal of Finance, 50, pp. 549-572.
Newey, W. and K. West (1987), “A simple positive semi-definite, heteroskedasticity
and autocorrelation consistent covariance matrix”, Econometrica, 55, pp. 703-708.
Nowak, E., A. Knigge and D. Schmidt (2004), "On the Performance of Private Equity
Investments: Does Market Timing Matter?", EFMA 2004 Basel Meetings Paper.
Otten, R. and D. Bams (2002), “European Mutual Fund Performance”, European
Financial Management, 8, 75-101.
------------------------------- (2004): “How to measure mutual fund performance:
economic versus statistical relevance”, Journal of Accounting and Finance, 44, pp. 203222.
Patelis, A. D. (1997), “Stock Return Predictability and The Role of Monetary Policy”,
The Journal of Finance, 52(5), pp. 1951-1972.
Pesaran, M. H. and A. Timmermann (1995), “Predictability of Stock Returns:
Robustness and Economic Significance”, Journal of Finance, 50 (4), 1201-1228.
Roy, B. and S. S. Deb (2004), “Conditional alpha and performance persistence for
Indian mutual funds: Empirical evidence”, ICFAI Journal of Applied Finance, pp. 3048.
Sawicki, J. and F. Ong (2000), “Evaluating Managed Fund Performance Using
Conditional Measures: Australian Evidence”, Pacific-Basin Finance Journal, 8 (4),
505-528.
Schwert, G. W. (1989), “Why Does Stock Market Volatility Change Over Time?”, The
Journal of Finance, 44(5), pp. 1115-1153.
Sharpe, W. (1966), “Mutual Fund Performance”, Journal of Business, 39 (1), 119-138.
Silva, F., M. D. C. Cortez, and M. J. R. Armada (2003), “Conditioning Information and
European Bond Fund Performance”, European Financial Management, 9, 201-230.
Treynor, J. (1965), “How to Rate Management of Investment Funds”, Harvard Business
Review, 43(1), 63-75.
Van Dijk, M. A. (2007): “Is size dead? A review of the size effect in equity returns”,
Working
Paper
available
at
SSRN
(Social
Science
Research
Network):
http://ssrn.com/abstract=879282.
Wermers, R. (1997), “Momentum Investment Strategies of Mutual Funds, Performance
Persistence, and Survivorship Bias”, Working Paper, University of Colorado at Boulder.
34