Local Predictability in High Dimensions

We propose a time series forecasting method designed to effectively handle large sets of predictive signals, many of which may be irrelevant or short-lived over time. Our method transforms predictive signals into candidate density forecasts via time-varying coefficient models, and subsequently combines them into an aggregate density forecast via time-varying subset combination. The approach is computationally efficient because it uses online prediction and updating. Through extensive simulation analysis, we find that our approach outperforms competitive benchmark methods in terms of forecast accuracy and computing time. We further demonstrate the capabilities of our method in applications to forecasting aggregate daily stock returns and quarterly inflation.