On the short-term predictability of exchange rates: A BVAR time-varying parameters approach

In this paper we propose a Bayesian vector autoregressive model with time-varying parameters (BVAR-TVP) to examine the short-term predictability of exchange rates. An important contribution of the paper is the application of the BVAR-TVP model, for the first time, to daily data using information from financial markets. Another contribution is the production of forecasts in real time at the very short horizon of one-trading day-ahead typically used by traders and investors in financial markets. We employ financial criteria and recently developed statistical tests to assess the exchange rate predictability. We find that the BVAR-TVP model outperforms the random walk for all exchange rates. These forecast gains are due primarily to the time-variation of coefficients, and secondly to information from other financial markets. It is shown that international investors could have made statistically significant excess profits if they had followed an inter-day trading strategy based on the buy/sell signals generated by the model's one-day-ahead exchange rate forecasts, even after allowing for transaction costs and risk factors.