Abstract (eng)
In this study I apply Bayesian Model Averaging to out-of-sample exchange rate forecasting. Following Eklund and Karlsson (2007) I employ predictive likelihoods rather than marginal likelihoods to determine posterior model probabilities. Forecasting results are evaluated relative to the standard random walk benchmark, as well as the best and median models. The latter two are the models with the highest posterior model probability and the model including all variables with a posterior inclusion probability above 50%, as suggested by Barbieri and Berger (2004), respectively. The main aim of this paper is to test various model specifications and find out which work best at what horizons. Specifications include level, difference, and error correction models. Furthermore, I examine the merit of using rolling model weights in forecast combination, and compare results for models using one and two lags, as well as models that include country specific variables versus cross-country differentials. Forecasts are performed for 8 different horizons and a set of 20 exchange rates. This extensive empirical analysis allows for a more reliable evaluation of various model specifications and forecasting performance over different horizons, reducing the influence of idiosyncratic movements in certain exchange rates on overall results. I also check for robustness of the results by performing forecasts over two different time windows. The main results are the following: one, Bayesian Model Averaging clearly outperforms single models (best or median) across all model specifications, time windows, and forecasting horizons in terms of RMSFE. Two, Bayesian model Averaging outperforms the random walk benchmark in terms of RMSFE significantly at longer horizons (9 and 12 months) but in general tends to do worse at shorter horizons. Three, level specifications appear to work somewhat better at longer horizons while difference specifications tend to do better at shorter horizons.
Four, alternative specifications point to potential improvement. This is particularly
true for the cointegration specification. However, it remains to be seen whether
these results hold when estimating the cointegration relationship on a rolling basis,
using only the data at hand at the time of the forecast. The rolling model weights
specification offers some improvement at longer horizons for the difference specification.
However, since level models tend to do better than difference models at these
horizons, the approach does not appear particularly promising. Finally, the results
do not indicate any systematic improvement in including country-specific variables
compared to cross-country differentials. At the same time it entails using twice the
number of variables and longer estimation time. From this standpoint it is therefore
clearly advisable to use differentials. Using one lag instead of two leads to better
results and is therefore also to be preferred, especially considering the reduction in
estimation time.