應用基因演算法與倒傳遞類神經網路於匯率預測模型之開發

Integrating Genetic Algorithm with Back-Propagation Neural Network to Build the Forecasting Model of Exchange Rate

在此研究中，我們利用倒傳遞類神經網路結合基因演算法（GABPN）來發展新台幣對美元的匯率（exchange rate NT$/US$）預測模型。一般類神經網路（Artificial Neural Network）對於輸入層資料是採用嘗試錯誤（try-and-error）的方式，也因此造成最後收斂誤差過大等的結果不穩定現象。我們期望對於輸入層資料可以進行一定程度的評估，進而達到改善類神經網路模型穩定性的目的。因此，先利用基因演算法（CA）的特性，找出最具代表性的輸入資訊與訊息個數；再使用三層倒傳遞類神經網路訓練其匯率預測模型。在此具備輸入層最佳化的機制下，我們發現只要使用10個代表性輸入資訊即可成功建立出更好的匯率預測模型，而不必使用全部的27個輸入變數同時，多餘的輸入變數也在實驗結果中呈現反效果，而破壞預測模型的穩定與準確性。實驗中我們以移動視窗的方式將每5年之歷史匯率值為訓練資料，並選擇與其部份重疊之一年為測試資料（i.e., 1997~2007）。結果顯示，在我們所選的6組資料中，此模型的預測表現皆優於單獨使用倒傳遞類神經網路（BPN）開發之模型。

In this thesis, we proposed the genetic algorithm back propagation network （GABPN）, which is applied on forecasting exchange rate NT$/US$. The GABPN model replaces the traditional neural network that used the try-and-error to find the input layer neurons. We find out the optimal Input layer neurons for back propagation network （BPN） which rely on the means of genetic algorithm （GA） which gave optimal solve. The forecasting performance of GABPN obtains the best performance. We also find that the performance of GABPN, which has only ten variables, achieve the better performance than BPN which has twenty-seven variables. We infer that too many variables might interfere with the forecast performance. After the experiment, we propose a set of variables and weight that could be the consult when investors or managers forecasting the exchange rate. In the meanwhile, investor could not only know that which variables but how many lagged period are explainable.