Overcoming the Forecast Combination Puzzle: Lessons from the Time-Varying Effciency of Phillips Curve Forecasts of U.S. Inflation

This paper proposes a new dynamic forecast combination strategy for forecasting inflation. The procedure draws on explanations of why the forecast combination puzzle exists and the stylized fact that Phillips curve forecasts of inflation exhibit significant time-variation in forecast accuracy. The forecast combination puzzle is the empirical observation that a simple average of point forecasts is often the best forecasting strategy. The forecast combination puzzle exists because many dynamic weighting strategies tend to shift weights toward Phillips curve forecasts after they exhibit a significant period of relative forecast improvement, which is often when their forecast accuracy begins to deteriorate. The proposed strategy in this paper weights forecasts according to their expected performance rather than their past performance to anticipate these changes in forecast accuracy. The forward-looking approach is shown to robustly beat equal weights combined and benchmark univariate forecasts of inflation in real-time out-of-sample exercises on U.S. and New Zealand inflation data.