Predicting Returns Out of Sample: A Naïve Model Averaging Approach

Abstract

We propose a naïve model averaging (NMA) method that averages the OLS out-of-sample forecasts and the historical means and produces mostly positive out-of-sample R2s for the variables significant in sample in forecasting market returns. Surprisingly, more sophisticated weighting schemes that combine the predictive variable and historical mean do not consistently perform better. With unstable economic relations and a limited sample size, sophisticated methods may lead to overfitting or be subject to more estimation errors. In such situations, our simple methods may work better. Model misspecification, rather than declining return predictability, likely explains the predictive performance of the NMA method.