This study develops a statistical procedure that is able to improve the ability of mutual fund performance models to identify funds that are most likely to generate positive risk adjusted returns in the future. Virtually every mutual fund study done by academic researchers uses a specific performance model (such as the Carhart model) to examine every fund in the sample being studied. The authors of this study argue that the wide variety of investment strategies, objectives, and styles employed by fund managers makes it unreasonable to believe that any one model can accurately capture the variation in performance across a large group of funds. Thus, parameter estimates (i.e. a fund’s alpha) from the models may not reflect the true performance of the fund but reflect the best estimate that an inappropriate model can make. Conclusions drawn using these estimates may understate (or overstate) the true value being provided by managers, and the error in the estimates reduces the predictive power of the model. The authors’ solution to this issue is to back test a model before using it to examine the performance of a group of funds. In other words, before a model is used to forecast the future performance of a fund, there needs to be evidence that the model has generated reasonable abnormal return forecasts using recent past data for the specific fund.     

            The back testing procedure suggested by the author’s works in the following fashion. Once a model has been selected, the model is estimated using sixty months of historical data (i.e. from time t-60 to time t-1). This first step of the estimation procedure provides an estimate of alpha which serves as the alpha forecast for time t. Next, the fund’s excess return at time t, defined as the fund’s realized return minus the return on the stock market, is calculated. If the fund’s excess return and the alpha forecast have the same sign, the model is considered to be appropriate for that fund, and the fund is included in the active pool of funds that will be examined using the model. Then for funds in the active pool, the model is estimated again using data from t-60 to time t (i.e. one additional period from the previous estimation). If a fund’s alpha estimate is between -.02 and +.02 and all other estimated coefficients are between zero and two, the fund remains in the active pool. This final step is to ensure that not only does the model provide reasonable forecasts, but that the estimates of the model fall within ranges that are considered plausible by academic researchers.   

            For fund investors, the most important result of the study is that this back testing procedure of the performance model improves the predictive power of the model. By focusing only on funds in the active pool, the model is better able to identify the mutual funds that are most likely to generate positive risk adjusted returns in the future. The main results of the paper focus on two specific models: the traditional Carhart model and a more technically sophisticated specification referred to as the Kalman filter model. Using all equity funds for which return data is available from January 1970 to December 2002, the authors identify the active pool of funds for a specific model. Each month, they sort funds into deciles based on their alphas. They then calculate the return to each decile the following month. The authors find strong evidence that the funds with the highest alphas (i.e. the top decile) generate consistent positive abnormal returns of over 4% annually. The authors conduct a variety of tests to ensure that their results are not due to statistical anomalies, and the results continue to hold. 

            Overall, the study provides strong evidence that future fund performance can be predicted when performance models are applied only to the appropriate set of funds. Investors can use this information to select the appropriate funds in which to invest.