Forecasting inflation remains an intriguing research topic among academics and practitioners alike. Recent studies by Stock and Watson (1999 JME, 2003 JEL) have documented the limited usefulness of Phillip curve type models using unemployment or other macroeconomic and financial variables for the purpose of forecasting inflation. In particular, it is found that the predictive value of such variables is highly unstable. In this paper, we employ the recently developed approach of Bayesian Averaging of Classical Estimates (BACE) for forecasting inflation. Based upon the idea of model averaging, the BACE forecasting model effectively is a weighted average of every possible combination of candidate predictors. The weights for each individual model are based on measures of in-sample fit or previous out-of-sample forecasting performance. Applying the BACE methodology to 3-, 6- and 12-months ahead forecasts of US inflation, we find that indeed the importance of different macroeconomic and financial variables varies substantially over time. Allowing for this feature results in improved out-of-sample forecast performance compared to fixed linear specifications, both in terms of the magnitude of inflation as well as in terms of its direction
