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

For the purpose of stock return prediction, we propose LASSO methods augmented by further penalties related to differences in coefficient estimates at t and t+1. The economic motivation is that the coefficient for a characteristic should not change too much from t to t+1, i.e., the market's valuation model should exhibit some degree of persistence.We find that Fused LASSO with a penalty on the absolute value of this difference provides a useful tool for truly dynamic model selection. We analyze the method with respect to the properties of the nonsparsity set and the associated trading strategies. p, li { white-space: pre-wrap; }