Predicting the Equity Premium: The Advantages of Combination Forecasts

While a host of economic variables have been identified in the literature with the apparent in-sample ability to predict the equity premium, Goyal and Welch (2007) find that these variables fail to deliver consistent out-of-sample forecast gains relative to the historical average. In this paper, we demonstrate that despite the failure of individual predictive regression model forecasts to outperform the historical average, combinations of individual model forecasts deliver statistically and economically significant out-of-sample gains relative to the historical average on a consistent basis over time. Why do combination forecasts deliver consistent out-of-sample gains? We argue that the data-generating process for the equity premium is highly complex and subject to substantial structural instability, and Hendry and Clements (2004) show that combining forecasts across individual models can lead to improved forecast accuracy in such an environment. To provide further insight into the relevance of this, we test for potentially multiple structural breaks in individual predictive regression models of the equity premium. We find extensive evidence of breaks, suggesting that the usefulness of combination forecasts stems in large part from their ability to improve equity premium forecasts in the presence of widespread structural instabilities. In addition, we find evidence of structural breaks in predictive regression models that use the same set of economic variables to predict real output or corporate profit growth, and there are a number of cases where the breaks in predictive regression models of real output or corporate profit growth occur relatively near the breaks in predictive regression models of the equity premium. Structural breaks in the dynamic behavior of the real economy thus help to provide an economic explanation for the structural instabilities in the dynamic behavior of the equity premium. Overall, our results indicate that the underlying processes driving the equity premium cannot be adequately captured by a single, relatively parsimonious predictive regression model. JEL classifications: C22, C53, G11, G12