Exchange rate prediction under model uncertainty

Conventional exchange rate models are well-known for their poor out-of-sample prediction relative to random walk. Different models seem to perform differently over different subsamples, horizons, and exchange rates; and hence, there is intrinsic uncertainty about which forecasting model should be adopted.;Chapter 1 uses factor combining technique for nesting alternative models into a general one that outperform each individual model. Factor combined forecasts demonstrate that all bilateral exchange rate models are misspecified due to the omission of common factors. Chapter 1 proposes factor analysis to combine alternative, competing models. Factor Combined Model captures the omitted common factors. Moreover, its prediction across horizons is outperforming random walk; and, it is as good as the performance of best individual model for each horizon, with persistent outperformance relative to random walk model as we extend the forecast horizon. In conclusion, we confirm that all conventional bilateral exchange rate models are misspecified due to the omission of common exchange rate factors.;Chapter 2 provides evidence that including three empirical factors, identified as Euro/Dollar, Japanese Yen/Dollar, and Swiss Franc/Dollar, improves forecast accuracy especially for the long-horizon forecasts. However, Chapter 2 does not find evidence that the three empirical factors are uniquely defined; it suggests that there exist some other missing common factors. Chapter 3 adds average nominal interest rate on the Euro and U.S dollar as additional common factor along with the other three empirical factors. Adding the average short term interest rate factor slightly improves the stability of best models across forecast horizons. However, there is no strong evidence that the four empirical factors sufficiently explain all the variations in the unknown common factors.;On the other hand, we find adding interest rate factor significantly improves forecast accuracy and stability for longer horizons. We find some evidence that the set of four empirical common factors sufficiently correspond to the unknown common factors for long-term prediction. In long-horizon prediction, the performance of best augmented models with the new set of empirical factors is significantly stable across horizons. The uncertainty about which model is optimal for long term prediction is reduced significantly.