結合獨立成份分析與類神經網路於財務時間序列預測模式之建構

Abstract

財務時間序列預測一直是財務決策之重要議題，但由於財務時間序列資料具有高頻率、雜訊、非定態與混沌等性質，使得其在時間序列預測領域中，向來被視為是一極具挑戰性的應用領域。本研究提出一結合獨立成份分析(independent component analysis, ICA)與倒傳遞類神經網路(back-propagation neural network, BPN)之二階段模式建構程序，先利用ICA具有將混合訊號分離出個別獨立來源訊號之能力，從預測變數中估計出獨立成份，在使用測試接受法找出獨立成分中代表資料中雜訊的獨立成分後，將代表雜訊的獨立成份去除，並利用其餘不代表雜訊的獨立成份重建時間序列資料，得到濾除雜訊後之預測變數，最後再使用BPN以濾除雜訊後之預測變數建構預測模式，期望可以使BPN於建構模式時不受雜訊影響，進而提升預測結果的準確度。為驗證所提方法之有效性，本研究以日經225(Nikkei 225)現貨開盤指數進行實證研究，並與直接使用BPN、隨機漫步(random walk)模式及其他常用之去除雜訊方法－簡單移動平均(simple moving average)及小波框架(wavelet frame)之預測結果進行比較。實證結果顯示，所提之方法不論在預測誤差或是趨勢預測準確度的表現上均較直接使用BPN、隨機漫步、簡單移動平均及小波框架為佳。

The characteristics of time series financial data are inherently high frequency, noisy, non-stationary and deterministically chaotic, which render the time series financial forecast extremely challenging. Owing to advantages in building non-parametric and non-linear models, artificial neural networks (ANN) have also been applied to time series predictions, especially for modeling financial time series forecasting. In ANN based financial time series modeling, one of the primary issues is the inherent high noise. It is an important but difficult task to identify and alleviate the noise in order to build a reliable ANN forecasting model. To minimize the influence of noise, we propose to conduct financial time-series forecasting by combining the approaches of both independent component analysis (ICA) and back-propagation neural network (BPN). The combined approach first applies ICA to deduce independent components (ICs) from the forecasting variables. After identifying and removing each IC contained noise, the filtered IC signals are then used to reconstruct forecasting variables and applied to the BPN forecasting model. In order to validate the performance advantage of the proposed approach, the Nikkei 225 opening cash price index has been used as an illustrative example. The experimental results show that the proposed model outperforms in forecasting accuracy the conventional BPN model with non-filtered forecasting variables, the random walk model, the simple moving average model, and the wavelet frame model.