Do Return Prediction Models Add Economic Value?

Abstract:     
This paper shows that statistical and economic measures of forecasting performance weight forecast errors very differently and that return forecasts from models with time-varying mean and variance, when used to guide the portfolio choice of an investor with power utility, can lead to significant improvements over the forecasts from a model that assumes a constant return distribution. Specifically, models with constant mean and volatility tend to overestimate the right tail of the return distribution and so lead to stock allocations that are on average too large with resulting lower average utility for risk averse investors. Our results demonstrate that return prediction models can add economic value even when they fail to produce accurate forecasts of mean returns and suggest the need for focusing on broader measures of distributional accuracy when evaluating the economic value of return prediction models.

Number of Pages in PDF File: 38

Keywords: predictability of stock returns, mean squared forecast error, portfolio selection, probability distribution forecasts

working papers series