Forecasting stock returns: A time-dependent weighted least squares approach

We improve the performance of stock return forecasts using predictive regressions with ordinary least squares (OLS) estimates weighted by a class of time-dependent functions (TWLS). To address the structural breaks in predictive relationships, these functions assign heavier weights to more recent observations. We find return predictability that is statistically and economically significant using a forecast combination of univariate TWLS models. TWLS estimates lead to much stronger return predictability than OLS estimates. The forecast improvement from TWLS is also found when forecasting characteristic portfolio returns and when using newly proposed predictor variables. These findings survive a series of robustness checks.