Optimizing Time-series Forecasts for Inflation and Interest Rates Using Simulation and Model Averaging

Motivated by economic-theory concepts—the Fisher hypothesis and the theory of the term structure—we consider a small set of simple bivariate closed-loop time-series models for the prediction of price inflation and of long- and short-term interest rates. The set includes vector autoregressions (VAR) in levels and in differences, a cointegrated VAR, and a non-linear VAR with threshold cointegration based on data from Germany, Japan, UK, and the U.S. Following a traditional comparative evaluation of predictive accuracy, we subject all structures to a mutual validation using parametric bootstrapping. Ultimately, we utilize the recently developed technique of Mallows model averaging to explore the potential of improving upon the predictions through combinations. While the simulations confirm the traded wisdom that VARs in differences optimize one-step prediction and that error correction helps at larger horizons, the model-averaging experiments point at problems in allotting an adequate penalty for the complexity of candidate models.