Factor return predictability: A comparison of multivariate forecasts of the Size, Value, Momentum, and Low Volatility premia

Abstract: This paper studies the predictive performance of multivariate models in forecasting joint returns of portfolios tracking the Size, Value, Momentum, and Low Volatility premia. We run recursive out-of-sample forecasting experiments on a number of linear and regime switching models, and compare the accuracy of their point predictions in a qualitative and quantitative fashion. We find evidence that a pure random walk model is at least as accurate as more complex statistical models in forecasting returns of the Size, Value, and Momentum factors. We also conclude that the introduction of Markov switching regimes, heteroskedasticity, and autoregressive terms provides a significant improvement in the forecast of the Low Volatility anomaly only. These results appear to be independent of the forecasting horizon and robust to changes in the loss function used in statistical evaluations.