Forecasting Exchange Rates with a Large Bayesian VAR

Abstract

Models based on economic theory have serious problems forecasting exchange rates better than simple univariate driftless random walk models, especially at short horizons. Multivariate time series models suffer from the same problem. In this paper, we propose to forecast exchange rates with a large Bayesian VAR (BVAR), using a panel of 33 exchange rates vis-a-vis the US Dollar. Since exchange rates tend to co-move, a large set of them can contain useful information for forecasting. In addition, we adopt a driftless random walk prior, so that cross-dynamics matter for forecasting only if there is strong evidence of them in the data. We produce forecasts for all 33 exchange rates in the panel, and show that our model produces systematically better forecasts than a random walk for most of the countries, and at all forecast horizons, including 1-step-ahead.