由於傳統的時間數列預測方法會因財務時間序列資料的非線性及非穩態的特性而有所限制,因此本研究採用黃鍔等學者在1998年所提出的希爾伯特-黃轉換,透過經驗模態分解將資料分解為一組頻率由高至低的本質模態函數,以呈現時間序列中不同頻率空間中的資訊。本研究再以測試接受法將本質模態函數進行重要程度的排序,藉此篩選出雜訊。最後,以排除雜訊後的本質模態函數組合做為輸入變數,利用倒傳遞類神經模式建置預測模型。為了驗證本研究所提方法的可行性,本研究以台灣加權指數、台灣摩根指數以及台灣電子指數的收盤價格歷史資料作為實證資料。實證結果顯示,經由經驗模態分解後的本質模態函數可呈現財務時間序列資料在不同頻率空間上的重要特徵,而測試接受法也可有效排序重要程度去除雜訊,進而提升預測模型的正確率。因此本研究結果可以有效提供投資者判斷未來趨勢方向時的重要參考資訊。
The time series prediction method was usually limited by the non-linearity and non-stationary of the financial time series data. As a result, Hilbert-Huang transform (HHT) was adapted in this paper. Through the processes of the empirical mode decomposition (EMD), the time series data could be decomposed into intrinsic mode function (IMF) components. Therefore, important patterns in different frequency spaces could be shown. Further, Testing-and-Acceptance method was used to sort the IMF components according to their importance to filter out the noise. Finally, the IMF components which are not noises were used to be the input variables of the back-propagation neural network method. The empirical results show that Hilbert spectrum analysis could be effectively used to explain the important characteristic of the financial time series. Further, the empirical results also demostrate that the back-propagation neural network forecasting model can accurately predict with Hilbert-Huang transform.