Forecasting Exchange Rate in a Large Bayesian VAR Model: The Case of Taiwan

Abstract

We study the out-of-sample forecasting performance of 32 exchange rates vis-a-vis the New Taiwan Dollar (NTD) in a 32-variable vector autoregression (VAR) model. The Bayesian approach is applied to a large-scale VAR model (LBVAR) and its forecasting performance is compared to the random-walk model in terms of both Diebold-Mariano and the Giacomini-Rossi fluctuation tests. Several results are found in the paper when we pay attention to the top three trading partner for Taiwan, particularly the China, U.S. and Japan, in which the corresponding bilateral exchange rates forecasts are denoted as CNY-NTD, USD-NTD and JPY-NTD respectively. First, a LBVAR model has a relatively better forecasting performance of CNY-NTD exchange rate in both medium-run and long-run. Second, a LBVAR model performs better than the random-walk model only in the short-run when forecasting USD-NTD exchange rate. Lastly, the random-walk model outperforms a LBVAR model all the time on forecasting JPY-NTD exchange rate.