Alternative regime switching models for forecasting inflation

US inflation appears to undergo shifts in its mean level and variability. W e evaluate the performance of three useful models for capturing such shifts . The models studied are the Markov switching models, state space models wi th heavy-tailed errors, and state space models with compound error distribu tions. Our study shows that all three models have very similar performance when evaluated in terms of the mean squared or mean absolute forecast error s. However, the latter two models are considerably more parsimonious, and e asily beat the more profligately parameterized Markov switching models in t erms of model selection criteria, such as the AIC or the SEC. Thus, these m ay serve as useful continuous alternatives to the popular discrete Markov s witching models for capturing shifts in time series. Copyright (C) 2001 Joh n Wiley & Sons, Ltd.