國際外匯市場自從1973年布列登森林協定時代結束後,各主要工業國家貨幣之兌換價格不再維持固定。近年來,匯率波動的幅度加劇。本研究期望比較出最適合預測匯率價格走勢的模型,以進一步來為台灣的進出口貿易廠商提供更佳的匯率價格走勢預警模型,降低台彎進出口廠商因匯率波動而產生的風險。 本研究為了解不同類神經網路模型,探討總體經濟變數時,對於匯價的變動是否具預測能力,分別就三種常用的預測模型來預測匯率價格走勢,倒傳遞類神經網路、回饋式類神經網路、回饋式模糊類神經網路,並比較此三種類神經網路於預測匯率價格時的學習效果與預測效果。 在研究後得知: 1.在預測匯率價格時,倒傳遞類神經網路由於沒有資料的回饋效果因此其隱藏層的神經元個數明顯的較回饋式類神經網路來的多,顯示回饋式類神經網路的資訊回饋效果在預測外匯價格上還不錯。 2.在回饋資訊的加強後,可以提高學習速率來加快收歛的速度。 3.加入模糊後,資訊的處理似乎因此而較為複雜,需增加隱藏層的神經元個數來加強類神經網路的運算效果。且學習速度率也因此而需降低。
Since the collapse of Bretton Woods system in 1973, the floating exchange rate has become a popular exchange rate system. During the past three decades, there are dramatic fluctuations in the foreign exchange rate. For those countries that international trade plays an important role such as Taiwan, had encountered serious problems because of the fluctuations of foreign exchange rate. This study attempts to examine the optimum model which can more accurately predicts the tendency of the exchange rate than conventional models in the current global economic environments in order to provide the companies which involved in international trade have a good forecasting power on exchange rate trend and reduce the foreign exchange rate risk of the exporters and importers. This study explores whether the overall macro economic parameters have significant impacts on the forecasting ability of the trend movement of foreign exchange rate. In order to achieve this goal, this work employs three neural network models, namely, the Backpropagation Neural Network (BPN), Recurrent Neural Network (RNN), Recurrent Fuzzy Neural Network (RFNN) models to predict the trend of foreign exchange rate. This syudy also investigates which Neural Network model has the best forecasting power among the three neural networks to predict the foreign exchange rate. The empirical results are summarized as follows: 1. Owing to the fact that BPN does not have the recurrent ability, the forecasting power on foreign exchange rate by BPN is inferior to that by RNN. 2. The convergent speed of learning improves by strengthening the recurrence of information. 3. The RFNN’s learning speed is slower than RNN after the fuzzy logic is added to RNN. However, RFNN’s forecasting power on foreign exchange rate outperforms RNN and BPN. In summary, this study finds that the RFNN has the best predictive power, the RNN is ranked the second, and the BPN has the least predictive power among the three neural networks.