應用分類元系統於台灣的指數期貨價差交易之實證研究

Abstract

股市交易為一個動態市場，具有學習能力的資訊方法學可以為動態環境的變化找到最佳的合理解。Wilson於1995年第一個提出的XCS分類元系統。它是一個以正確性為基礎的分類元系統，強調正確的資訊應該得到所應得的報酬相反的錯誤的資訊型為掌握則應該降低其所帶來的行為分析效果。該方法學搭配人工智慧的基因演化，針對已經染色體化的規則進行交配與突變，以達到知識上的演化。知識規則於演化後再被放入環境中競爭，以得到最終發現最適合當下環境的規則。 因此本研究以台灣期貨交易所所發行的指數期貨商品為研究標的，採用XCS分類元系統來探討任兩個不同的指數期貨商品價差交易的內在行為知識規則。於本實驗模擬中將在早上九點到下午一點三十分間對台灣指數期貨進行一買一賣相同口數的交易策略並於當天平倉出場。 本研究實證結果顯示在準確率方面，本研究所進行的三個實證測試月份中，三種分別應用XCS的交易策略皆遠優於傳統基於價差基本交易理論的隨機交易策略。每個月表現最佳的投資策略各為電子期貨與金融期貨(八月)、台指期貨與金融期貨(九月)、台指期貨與電子期貨(十月)，代表市場上的最佳解會隨著時空環境而改變。此外在績效方面，於本研究所應用的XCS交易策略在盤中進行日內交易皆可獲得淨獲利的交易績效，也亦優於傳統基於價差基本交易理論的隨機交易策略。其中以採取電子期貨與金融期貨對沖的基差交易策略績效最佳，其次為台股期貨與金融期貨，最後為台指期貨與電子期貨。

The stock market is a dynamic system. Only an information methodology with learning ability can find optimal solutions for a dynamically changing environment. Wilson was the first to present the Extended Classifier Systematic at 1995. It is an accuracy-based classifier system. It emphasizes that correct information should be rewarded and incorrect information be suppressed to reduce its effects. Combined with the artificial intelligence in genetic algorithms, it evolves trading rules by operating on digital chromosomes using genetic operators such as mutation and crossover. Evolved rules are then released into the world to compete with each other. This method is applied over time to discover best-suited candidates for surviving the current environment. Therefore, this study focuses on the research of future goods issued by the Taiwan Future Exchange. The system applies XCS classifiers to explore the rules of information behavior for spread trade with any two future’s goods. The simulated experiments will process the trade strategy that together buys and sells the same lots is the goods of Taiwan index future from nine o'clock in the morning to thirty past one in the afternoon. It settles this trade at the same trading day. As illustrated in the accuracy study of our experiments, all three trading strategies of this study that utilize XCS perform better than random trade that bases on the basis of traditional theory in spread for three experiment months. The best investment strategy of this study for each month is TE-TF (August), TX-TF (September) and TX-TE (October). Results indicate that the best solution dynamically changes according to the environment. In the profit experiments, Intraday Trading by XCS also beats random trade. Best results were found in TE-TF, which trades the hedgy strategy on this basis, second is TX-TF, and last is TX-TE.