| This dissertation investigated the predictability of five currency exchange rates by comparing a nonlinear approach based on the feedforward neural network model to three traditional linear models. The predictability comparison results generally favor the neural network model, in that a neural network model outperforms the traditional linear models in most major aspects, although it fails in other areas.;For accuracy in point prediction, the empirical result shows that the neural network model provides a smaller average forecast error as measured by the mean-squared-error. However, the neural network model generally underperforms the traditional linear models in terms of the Theil's U-statistic criteria. Concerning market trend predictability, the neural network model demonstrates a significant advantage over all three traditional linear models. Finally, the results from a trading simulation over a two-year trading period show an apparent superiority in the neural network model's trading profitability.;The forecasting comparison between a neural network model and three traditional linear models revealed some evidence inconsistent with the market efficiency hypothesis.;First, the forecasts of the neural network model using two or more lagged spot rates as inputs are generally better than those from the neural network model using only one lagged spot rate as the input, i.e., the latest market price of a currency does not necessarily reflect all the information incorporated in its price history.;Second, all five sample currencies follow a typical random walk type movement that suggests the unpredictability in foreign exchange market according to traditional linear approaches. However, the forecasts of the neural network model outperform the forecasts from a random walk model in all major comparison aspects except Theil's U-statistic. Therefore, the outperformance of the neural network model over other traditional linear models may come from the successful detection of the nonlinearities existing between the spot exchange rate and its lagged values. |
