訊號分解對於集成學習預測股價準確率之影響—以台灣加權股價指數為例

Abstract

股價為一高噪音、非線性和非平穩的時間序列資料，因此股價預測長期以來均是一項具有挑戰性之熱門研究。本文提出一基於變分模態分解 (Variational Mode Decomposition, VMD)和經驗模態分解 (Empirical Mode Decomposition, EMD)之二次分解技術，結合極限學習機 (Extreme Learning Machines, ELM)和改良式和弦搜尋優化演算法 (Improved Harmony Search Algorithm, IHS)之二階段混合模型，並利用此混合模型預測台灣加權股價指數之股價。本文將VMD分解技術應用於分解台灣加權股價指數之收盤價，取得多個子序列和噪音項後，再將EMD分解技術應用於分解噪音項，最後將子序列和由台灣加權股價指數衍生出之技術指標透過ELM模型得出初步預測結果，再以IHS演算法整合並優化最終結果。而為驗證模型的有效性，本文將此混合模型和單一ELM模型以及單一VMD分解技術之混合模型進行比較，並比較預測一日、三日和五日之結果。實證結果顯示，本文所提出之混合模型無論在短天期或是長天期，均具有較好的預測效果，其中二次分解技術優於一次分解技術之結果亦說明：深入分析噪音項所含之有效資訊，不僅更完善的捕捉原始序列的特徵，亦更有效地提升模型的預測能力。

As stock data is characterized by high-noise, non-linear, and non-stationary, predicting stock price is usually subject to a main challenge. In this study, to enhance the predictive performance, we proposed a new two-stage hybrid model by combining with extreme learning machine (ELM) and improved harmony search algorithm (IHS) which based on the secondary decomposition technique of variational mode decomposition (VMD) and empirical mode decomposition (EMD), named VMD-EMD-ELM-IHS model. The hybrid model applies VMD techniques to the original closing price of TAIEX to obtain different subsequences and the residual term, then applies EMD techniques to the residual term, then predicts all subsequences and technical analysis indicators by ELM models, and then applies IHS to integrate the prediction results of ELM models to obtain the final prediction results. To verify the performance and robustness of the hybrid model, the results were compared with other models, including single ELM model, and VMD-ELM-IHS model, and respectively, tested by one-step, three-step, and five-step forward forecasting. The empirical results show that the hybrid model we proposed achieves the best prediction performance in other models and all prediction scenarios. Also, the secondary decomposition technique superior to the single decomposition technique shows that fully considering the residual term not only captures the characteristics of the original sequence but also effectively improves the prediction accuracy.