Bayesian switching model for forecasting GDP growth rates

Abstract

The forecast properties of econometric models for gross domestic product (GDP) have been of great interest since Nelson (1972) criticized large-scale macroeconometric models. Various approaches have been proposed, yet forecasting over the irregular duration such as recessions remains a bottleneck. The well known Markov switching models take into account the different phases of business cycles, but they are generally unable to provide more accurate forecasts than the conventional autoregressive-moving-average (ARMAX) model. We present a new Bayesian multi-country model, with a random intercept process. The new model features nonstationarity occurring from stochastic breaks in the process and accommodates the multi-country variation with country-specific parameters. The Markov switching, ARMAX and the proposed Bayesian models are compared in an empirical forecast analysis for 13 countries. The proposed Bayesian model provided the most accurate forecasts in terms of RMSFE (root mean squared forecast errors). The developed Bayesian model offers a flexible foundation for further analysis of the model parameters and reliable prediction interval estimates.