The poor predictive performance of asset pricing models

This paper examines the time-series forecast errors of conditional and unconditional asset pricing models for both portfolio and individual firm equity returns. The mean squared forecast error of models is decomposed into separate variance and bias components. A new result concerning model specification and forecasting is also introduced. The results indicate that conditional versions of the models generally produce lower mean squared errors than do unconditional versions for in-sample but not for step-ahead prediction. This holds true even for individual firm data when the instruments are firm specific. Time varying parameter specifications of conditional asset pricing models typically produce mean square forecast errors that are larger then those of fixed parameter specifications. While the models produce relatively unbiased predictions, their one step ahead predictive ability rarely exceeds that of simple benchmarks.