Forecasting Exchange Rates using Bayesian Threshold Vector Autoregressions

In this paper we assess the predictive abilities of a Bayesian threshold vector autoregression (B-TVAR) to forecast the EUR/USD exchange rate. By introducing stochastic search variable selection priors (SSVS), we account for the inherent model uncertainty when it comes to modeling exchange rates. Our results suggest that, by applying Bayesian methods to the TVAR, it is possible to improve upon the random walk forecast. Surprisingly, we even managed to outperform the naive benchmark model in short-term forecasting, where the gains in terms of predictive ability are substantial.