基於合成經驗模態解構法建立之類神經網路的匯率預測表現

Abstract

本文嘗試以合成經驗模態解構法為基礎建立前饋型類神經網路以對匯率進行預測。首先將原始匯率價格時間序列取自然對數並一階差分, 然後利用EEMD 方法將序列拆解成為數個不同頻率的本質模態函數(IMF)。接著, 對每一個被選出的IMF, 我們使用三層前饋型類神經網路來模擬它們的動態規則並作預測。最後, 我們將被選出的IMF 的預測結果結合起來成為我們對對數差分後之序列的最終預測。我們在樣本外的預測測試結果顯示此方法產生的均方誤差顯著比隨機漫步模型的均方誤差高。然而, 方向預測準確性檢定卻顯示此方法具有顯著的捕捉匯率時間序列方向變動的能力, 而由方向預測產生的交易平均報酬亦確認此一結論。

In this study, an ensemble empirical mode decomposition (EEMD) based feedforward neural network framework is proposed for exchange rate forecasting. For this purpose, the original exchange rate series is first decomposed into a finite (and often small) number of intrinsic mode functions (IMFs). Then a 3-layer neural network is used to model each of the selected IMFs, so that the tendencies of these IMFs could be accurately predicted. Finally,the prediction results for all IMFs are combined to formulate an aggregate output of the predicted exchange rate movement. Our empirical results show that this modeling procedure has significantly larger root mean squared errors (RMSE) than the random walk model. However, sign tests and trading strategy returns suggest that this method indeed has superior predictive ability for directional change.