Dynamic strategy and performance of german equity and bond mutual funds

Diplomarbeit von Nikola Jeličić

Introduction:

Measuring performance of fund managers is a topic equally interesting to practitioners and researchers. Most common performance measures rely on the assumption of constant risk during the entire evaluation period. The measure of risk is the beta from the Capital Asset Pricing Model (CAPM). In order to better assess a manager’s investment ability, additional factors could be employed to capture the different sources of risk. The manager owes each portion of the achieved return to a certain risk factor. The risks a manager is running can be summed up to form his personal benchmark, which thus reflects the investment style. Still, the exposures to the included risk factors are assumed to be constant.

The dynamics of the capital markets had not been captured by the prevailing performance measures before an approach that controlled for varying economic conditions was suggested. Models that are based on this approach deliver a beta conditional on the market state. The manager’s exposure to the risk of the own benchmark was thus allowed to vary in time. Consequently, the search for indicators of the market states was launched and a model framework which could accommodate the chosen indicators as part of the benchmark had to be chosen. Two model frameworks emerged and a couple of indicators established themselves as standard. This study largely follows the approach of Ferson and Schadt. They introduced a linear model that can be perceived as a conditional version of the CAPM.

The aim of this study is not only to obtain performance measures which result from the conditional models. Since the variation in the exposure to market risk is accounted for, one who employs conditional models gains insight into fund manager’s trading. If the trading is reflected in changes of the beta, then inference on fund strategy is made possible even though information on the portfolio structure is not provided. The explanatory power of a conditional model depends on the researcher selecting a representative benchmark for the funds in the sample and indicators of economic conditions that fund managers rely on in reality.

The structure of this paper is the following: chapter 2 builds the theoretical foundation of conditional models and presents their two forms; chapter 3 relates this study to previous literature in the area; chapter 4 employs conditional models to evaluate strategies and performance of German fund managers; chapter 5 sums up the findings and draws conclusions.

Table of Contents:

Text Sample:

Chapter 4.3, The Hypotheses:

In this section, I will concretize the objective of this study to relate it to the models presented in the previous section. As previously stated, the objective is to identify dynamic strategies by looking at responses to public information. If it is proved that there are dynamic strategies within the presented models, accounting for their effects on performance would be useful in order to obtain more realistic and more persistent measures.

Public variables reflect the state of the market and are known even to the wide investment public. Since investors use the same variables when they form their expectations regarding returns and risks, using the same variables is plausible when evaluating the performance of fund managers. Both the existence of dynamic strategies and their effects will be substantiated in the hypotheses. These are based the informative value of the used proxies, as discussed in section 2.4.

Hypothesis 1: Funds employ dynamic beta strategies.

The coefficients on the interaction variables, consisting of excess returns of market indices multiplied by public information variables will be statistically significant. This is interpreted as evidence of dynamic strategy.

Hypothesis 2: The responses to information variables will be different for equity and bond fund managers.

A manager who wants to reduce his exposure to market risk can hold cash or deposit it e.g. for a month. The short-term rate is therefore an alternative to the investment strategy that the fund manager employs. The Euribor doesn’t exactly match the fixed-deposit rate available to every fund manager, but serves as its proxy. Bond markets are more directly linked to the short end of the yield curve, which is why the bond manager will respond to the effect consisting of (expectedly) positively correlated variables, the bond index and the lagged Euribor, more strongly. This is why I expect a positive coefficient for the bond fund groups. The same coefficient resulting from regressions on equity fund groups will probably be negative and not as large as the one obtained from the regressions on bond funds. I expect negative coefficients because high short-term rates, which are associated with less risk than the stock market is, make equities less attractive. The expectation on the magnitude is derived from the existence of both high and low interest rate phases during bull market periods.

The dividend yield is a proxy for the expected performance of publicly listed companies i.e. a stock market performance predictor. As described earlier, the dividend yield captures long-term expectations. A positive effect on equities and long-term bond portfolios is expected. Bond managers who can switch their portfolio to hold short-term bonds will probably have more possibilities to respond to changes in the dividend yield and so achieve benefits.

The term spread is a proxy for the preferences of long- vs. short term investments. Stock and long-term bond portfolios will profit from larger spreads, which should be reflected in a positive coefficient. Since bond funds are not sorted by maturity, a consentaneous effect will probably not be found.

The other interest rate based variable, the default spread, carries similar information to the dividend yield. However, the used grouping doesn’t provide insight into which fund managers would be able to make use of its changes.

Hypothesis 3: Performance measures based on the conditional model are more significant and more persistent.

Conditional alpha will be closer to zero than its unconditional counterpart. Also, alphas from conditional models will be accompanied by higher significance. This hypothesis is based on the expectation that manager skills are better accounted for, paired with the belief that the average manager neither outperforms nor underperforms the market.

If the information is chosen well, different levels of market dynamics (based on expectations for different horizons) will be modeled. This would lead to a more realistic value of the alpha, which is why I expect it to be significant and close to zero. It is further plausible that an alpha, which better incorporates manager skills, is more persistent. Section 4.6.2 will be dedicated to the analysis of persistence.

Hypothesis 4: Conditional timing coefficients will be positive.

Negative timing coefficients will be removed after controlling for time-variation in beta risk associated with public information.

Excursus: Panel Data Estimation:

I provide an excursus on panel data estimation procedures used later in this study. The basis for this excursus is provided in chapters on panel data analysis from textbooks of Wooldridge and Greene .

Panel Data Specifics:

I estimate the empirical models for the four fund groups which I treat as four datasets. Deciding which estimation method should be used relies on appropriate tests from the econometric toolbox. These tests are performed separately for each group.

The four datasets are unbalanced panels with 192 monthly panel waves (I use the term ‘unbalanced’ to emphasize that there aren’t observations for each fund in each panel wave). Fund identification numbers are used as the panel variable and the variable ‘period’ with 192 different values as the time variable. Since funds with missing observations within the examination period have already been removed, the sample contains only those funds that start later than January 1991 or seize to exist before December 2006. This is why Stata regards the used data sets as ‘strongly balanced’.

Panel data are data which have both time-series and cross sectional properties. As in time series, the returns of a certain fund in two consecutive months could be correlated. This is called serial correlation. It is induced either by unknown characteristics which are constant over time or unknown time-dependent characteristics which themselves are serially correlated.

OLS estimation works with a set of assumptions, some of which may be wrong when dealing with a set of panels. One of the most problematic assumptions is that of constant variance of the error term, namely homoscedasticity. Independence of the error terms is assumed across observations, although many can be traced back to the same fund. This fact should be accounted for by using a model that allows the homoscedasticity assumption to be relaxed. More precisely, the model needed is one that takes into account that there are repeated observations within the same unit that is fund returns are generated by the same manager.

Treating panel data as a cross section is called pooling, a method commonly used in literature on asset pricing and performance. I employ a pooled OLS regression to provide basis for the analysis using more appropriate methods.