Modelling Exchange Rate Returns Using Non-linear Models

Forecasting exchange rate movements is challenging, as they exhibit high volatility, complexity and noise. Most traditional models cannot forecast exchange rates, with significantly higher accuracy, than a random walk model. In this study, a non-linear model called artificial neural network (ANN) is used to forecast short-term (daily and weekly) movement of United States dollar (USD)/Japanese yen (JPY). ANN’s out-of-sample performance is benchmarked against the traditional Autoregressive Integrated Moving Average (ARIMA) model. Performance of both models is rigorously evaluated using three different penalty-based criteria: Directional Accuracy (DA), Correct Upward (CU) and Correct Downward (CD) trends and two non-penalty-based criteria: mean square error (MSE) and normalised mean square error (NMSE). Moreover, the robustness of the two models is tested for different sampling periods. Empirical results show that ANN per-forms better than ARIMA and delivered consistent results across all periods tested. This supports ANN’s robustness and also the fact that it can be used to formulate a strategy for trading in USD/JPY.