時間序列分析 : 應用經驗動態建模於股市交易

Abstract

經驗動態建模（Empirical Dynamic Modeling，簡稱 EDM），該方法透過吸引子重建（attractor reconstruction）、狀態空間重建（state space reconstruction）等數據驅動的方法，在生態學中被廣泛應用並取得巨大的成功，但至今尚未被廣泛使用於其他領域尤其在金融領域上，而EDM應用在金融資料上是否也能有良好的預測效果著實令人好奇。故本研究首度嘗試將EDM應用於股市交易之中，除了利用EDM原先所擅長的數值預測對未來股價進行預測。還自MDR S-map延伸發展出一個新的分類預測器MDR S-map Classifier對股票的未來漲跌進行分類預測。實驗中我們透過收集到的台灣股市資料，比較了MDR S-map Classifier與其他常被使用的深度學習模型，在許多方面上EDM方法都優於深度學習模型，甚至在對未來的漲跌預測上，MDR S-map Classifier的獲利表現遠遠的好過於AE-MLP模型。最後，我們還對這個新方法進行了更深入的研究，發現了當某些特定的資料特性存在時，如交易標的尚未過熱以及自營商介入較少等，MDR S-map Classifier有機會取得較優秀的預測結果。

Empirical Dynamic Modeling (EDM) utilizes data-driven techniques such as attractor reconstruction and state space reconstruction. This method has been widely applied and highly successful in the field of ecology. However, it has not been broadly adopted in other fields, particularly in finance. The potential effectiveness of EDM in predicting financial data is an intriguing question. Therefore, this study represents the first attempt to apply EDM to stock market trading. In addition to using EDM's original numerical prediction capabilities to forecast future stock prices, we have developed a new classifier predictor, the MDR S-map Classifier, derived from the MDR S-map, to classify future stock price movements.

Through experiments with Taiwanese stock market data, we compared the MDR S-map Classifier with commonly used deep learning models. The EDM method outperformed deep learning models in several aspects, and the MDR S-map Classifier significantly outperformed the AE-MLP model in profitability for predicting future stock movements. Additionally, further investigation into this new method revealed that the MDR S-map Classifier could achieve superior prediction results under certain specific data characteristics.