應用分類元股票交易系統於台灣加權指數趨勢預測之研究

Abstract

人工智慧模型近年來在財金領域上的應用相當廣泛，也常與傳統的統計模型相互比較。傳統時間序列數學模型存在許多假設與限制，但人工智慧模型較具有彈性，可解決非線性問題，較適合應用於像股票市場這種動態環境。

本研究為應用人工智慧方法之延伸分類元系統（Extended Classifier System, XCS）於台灣加權指數趨勢之預測。分類元是一種以基因演算法為基礎的學習模式，它擁有一個規則集，而且會動態對環境進行調整。本研究以民國七十八年九月二十六日至民國九十四年三月七日台灣加權指數4200筆日資料，運用不同天期移動平均線（MA）、隨機指標（KD）、平滑異同移動平均線（MACD）等技術指標當作輸入因子，加權指數之買賣訊號作為輸出因子。實驗期間前1500天為訓練期，後2700天為測試期，經由分類元系統動態學習買賣規則，每次交易以加權指數為標的。實證結果顯示十年測試期間分類元股票交易系統(CSTS)三十次模擬結果之平均報酬率為165.38%，平均交易勝率為60.31%。統計分析顯示CSTS系統之交易報酬率及交易準確率皆顯著優於傳統迴歸模式及隨機交易策略。本研究結論為分類元系統可較傳統策略準確掌握加權指數之趨勢，非常適合投資者作為交易決策系統。

Stock market is nonlinear and semi-structured. In other words, there are many different situations and the environment states are changing quickly. In Taiwan, stock market is always affected by political factors. So the fluctuation of stock price is always larger than other country. Therefore, to predict the trend of stock index is more important. Traditional trading strategy like regression model and random walk are limited in fixed time interval and can not perform well. Other learning models like genetic algorithm result in stable trading rules which are generated from specific training time period without being adapted when the environment state is changed. This paper adopts Extended Classifier System (XCS) technique to design an XCS-based Security Trading System (CSTS), which makes continuous on-line learning while making decision and generate trading rules to adapt environment state. The simulation results showed that this system could get an outstanding trading profit and accuracy rate.