以人工智慧方法及時間序列模型建構日曆效應價差交易策略

Abstract

本研究以S&P 500指數成分股探討市場中日曆效應異常現象，並以「週一日報酬低於前週五」為主要星期效應之探討現象。本論文有別於傳統趨勢預測與買賣訊號判定，而改以人工智慧方法與時序建模，找尋強化星期效應現象之特徵因子與二元分類器，並藉由五階段實驗建構星期效應是否發生的預測模型，包含「二元分類演算法比較」、「相異特徵組合分析」、「特徵萃取區間分析」、「時間序列穩定性測試」、以及「建模資料擷取區間檢驗」，最後再以此設計雙邊交易策略以驗證本研究在金融市場上的實際應用價值。 實證研究得出S&P500指數成分股確實存在星期效應，且以LDA-Linear Uniform Prior為核心之二元分類器，相較於SVM與倒傳遞類神經網路擁有較佳的預測力及時效性。此外，針對相異特徵因子萃取方式之分析與完整對比實驗，證明本研究以特徵因子「多天期技術指標」與「週五日內資訊」建構星期效應預測模型的準確率可達到七成以上。然而「週五日內資訊」雖存在較高獲利表現，但投資績效穩定性卻不如「多天期技術指標」，反而可能導致投資人的風險提高。總結來說，以本研究所提之「多天期技術指標」及LDA-Linear Uniform Prior建構決策模型，確實能掌握星期效應現象之有效成因，以此為預測基礎之雙邊交易策略，由於隱含金融工程物理學意義，因此能掌握價格的趨勢與動能，能具金融市場的解釋能力、帶來穩定小額報酬，提供投資人較穩健、低風險之投資策略參考。

This paper focuses on analyzing and modeling the calendar anomalies in the 502 component stocks of S&P 500 Index. The research target of Day-of-the-week effect is defined as “daily return on Mondays could be lower than it on previous Friday”. Compared to the previous learning-based approaches which focus on market trend prediction and trading signal extraction, we use artificial intelligence methods and time-series modeling to improve a well-known financial phenomenon, Day-of-the-week effect. The experimental results evaluate the system performance with multiple binary classifications techniques, various feature extraction combinations, and different parameter settings of learning system design. We found that the proposed system with LDA-Linear Uniform Prior algorithm may outperforms SVM and Artificial Neural Networks kernels. In addition, the accuracy evaluation can achieve 70% when adopting two discriminant feature representation method, including “long-term technical indicators” and “intra-Friday information.” The first one may lead to higher prediction accuracy and the second one could produce lower investment risks. In summary, the proposed artificial intelligence methods can effectively predict the Day-of-the-week effect in the component stocks of S&P 500 index. The classification based on LDA-Linear Uniform Prior learning kernel and long-term technical indicators may lead to the robust experimental results. We also develop an extended trading strategy based on the proposed classifier to demonstrate the practical value of this research in real investing applications of financial markets.