A New Model Every Month? — Dynamic Model Selection for Stock Return Prediction

Abstract

In this paper we introduce a new class of approaches to empirical asset pricing research, namely LASSO methods augmented by further penalties related to differences in adjacent coefficient estimates (at t and t+1) for a given characteristic. The economic motivation for this is that the coefficient for a given characteristic should not change too much from one point in time to the next, i.e., the valuation model used by the representative agent should exhibit some degree of persistence.

We find that Fused LASSO (FL) with a penalty on the absolute value of this difference provides a very useful tool for dynamic model selection without having to rely on rolling window estimates. We analyze the properties of FL with respect to the size and the composition of the nonsparsity set, i.e., the set of characteristics with nonzero coefficient estimates, as well as with respect to the trading strategy based on the predictions.

Keywords: Sparsity, LASSO, fusion penalty, uncertainty

JEL Classification: C52, C55, G0, G1, G17