Learning Robust Time-Varying Forecast Combinations for Output and Inflation

State-of-the-art forecast combination methods performs poorly against the simple mean, noise and structural breaks. Time-varying weights have been suggested as a way forward. Here we propose an online machine learning approach to learn robust time-varying forecast combinations to forecast output, inflation or any macroeconomic time series of interest. Further, the proposed procedure also provides an adaptive hedge against poor performance against the mean, while also providing strong theoretical guarantees. Robust performance is demonstrated on a standard macroeconomic dataset for predicting output and inflation.