| Summary: | 碩士 === 國立成功大學 === 財務金融研究所 === 92 === This thesis uses short-term prediction models such as ARIMA, Genetic Algorithm (GA) and Back-Propagation Neural Network (BPN) model and long-term prediction models such as Econometrics model, GA and BPN not only to predict exchange rates but also to find out which model has the best forecasting ability. The thesis uses MAPE to measure each model’s precision, and it also uses moving direction which is forecasted by models to measure each model’s validity.
The thesis finds out the fact that, in the short-term period, the GA has the best precision ability, and there is not apparent validity difference among ARIMA, GA and BPN model. On the other hand, this thesis also finds out the fact that, in the long-term period, the GA also has the best precision ability. When an exchange rate shows random walk style (such as Swiss Franc/US dollar and Yen/US dollar), the GA has the best validity ability; when an exchange rate doesn’t show random walk style (such as NT dollar/US dollar and British Pound/US dollar), the BPN has the best validity ability.
The reasons why the GA has the best precision ability are that BPN model uses the same parameters as ARIMA or Econometrics model, which causes to give restriction to the BPN and restrains its ability from forecasting exchange. The thesis also suggests future researcher to find out appropriate learning rule, hidden layers and transfer function when using the BPN to predict exchange rates. The further researcher can also use rolling regression by increasing out-of-the-sample period in order to acquire more samples, which can improve model’s precision and validity.
|
|---|
