The literature on exchange rate forecasting and the out of sample evaluation has basically started with the work by Meese and Rogoff (1983). They were the first to show that the basic random walk model outperforms other economic models of exchange rate in terms of forecasting.
Their results have been revisited by a recent large studies due to Cheung et al. (2007) – “Empirical models of the nineties: are any fit to survive?”. What this study brought new was that it considered a wider range of exchange rate models (the reference ones as they emerged in the nineties). Five such different models were considered, namely:
– The power purchasing parity
– The sticky price monetary model of Dornbusch and Frankel
– A productivity based model
– A composite model (which includes features from the Behavioral Equilibrium Exchange rate model and from the Real Equilibrium Exchange Rate)
– The uncovered interest parity
The forecasting exercise considered two different samples and different forecast horizons. Several evaluation criteria were used, like the mean square error, the direction of change and the consistency criterion.
A few interesting results emerged from this exercise. First, the structural models have a rather poor performance even in the long run. Second, the exchange rate models can beat the random walk but only in the long run. Third, there is no model which is the best in terms of forecasting accuracy of the exchange rate.
Their research, although well executed and addressing the most important models at that time, could have been somewhat improved by considering a wider choice of test statistics for out-of-sample forecasts. I underlined this issue in an earlier post about DSGE model – based forecasts. Such statistics would include the Diebold Mariano test, test for forecast encompassing, and so on.
At the same time, it would appear interesting to see what the recent crisis has brought new with respect to this topic. A good piece of research is the one by Molodtsova and Papell (2012) who discussed the performance of a Taylor rule – based model to forecast the exchange rate before, during and after the financial crisis. Among their main findings was that all specifications of the Taylor rule models outperformed the random walk before the beginning of the crisis. However, only a certain specification outperformed the random walk for the sample that included the financial and economic crisis. I rather find very significant their result that the performance of the different models in terms of exchange rate forecasting differs when the sample includes the crisis years.