Linear and Nonlinear Predictability in Investment Style Factors: Multivariate Evidence

Abstract

This paper studies the predictive performance of multivariate models at forecasting the (excess) returns of portfolios mimicking the Market, Size, Value, Momentum, and Low Volatility factors isolated in asset pricing research. We evaluate the accuracy of the point forecasts of a number of linear and regime switching models in recursive, out-of-sample forecasting experiments. We assess the accuracy of the models using several measures of unbiasedness and predictive accuracy, and, using Diebold and Mariano’s approach to test whether differences in expected losses from all possible pairs of forecast models are statistically significant. We fail to find evidence that complex statistical models are uniformly more accurate than a naive constant expected return model for factor-mimicking portfolio (excess) returns. However, we show that it is possible to build simple portfolio strategies that profit from the higher out-of-sample predictive accuracy of forecasting models with Markov switching in conditional mean coefficients. These results appear to be independent of the forecasting horizon and robust to changes in the loss function that captures the investors’ objectives.