以群體學習為基礎的知識再強化智慧型財務決策支援系統建置—以台灣加權指數內含行為的知識萃取為例

Abstract

人工智慧歷經多年發展，已有為數眾多的研究將其應用在股市趨勢的預測。然而股市所具有如莫名波動、非線性或近似渾沌的特質長久以來一直是股市投資者及財務研究學者最有興趣的課題。由於過去多數研究所採用的人工智慧模型均以試誤法為基礎進行學習，學習結果不僅效率與準確性不盡理想外，也只能針對變異性不大的封閉型環境進行應用。隨著電腦運算能力不斷的大幅提昇，人工智慧的學習模式應該重新檢視，以充分應用現有的電腦科技、財工數學和計量經濟等理論基礎與相關資源，俾產生最適宜的財務投資決策，來降低人為錯誤、提高學習與決策效率、並擴大獲利機率。

本研究嘗試由傳統學習理論的角度出發，以人類學習行為中群組學習的概念與精神建構一個智慧型的群組學習模型，對台灣股價指數的漲跌趨勢進行分析預測。在資訊科技的應用上本研究採用類神經網路(neural network, NN)以及分類元系統(eXtended Classifier System, XCS)之動態知識學習。由於XCS系統是一個以規則為基準的主動式機械學習（machine learning）系統，並具適應動態環境學習的特性，故可以更貼近人類學習、決策方式及瞭解股票脈動。配合類神經網路進行強化式學習，將系統所發掘的知識規則予以強化確認。此外，研究中還採用多重代理人機制，結合群體學習理論中知識分享的概念提升決策準確率。

本研究在系統效能評估上，以未採用類神經網路強化學習的XCS單一技術指標、單一分類元系統、Buy & Hold，以及銀行6年期定存等操作策略做為對照組進行分析比較，實證結果顯示，本研究所提出的系統不論在預測的準確率、投資累積報酬率等模擬績效均優於對照組的績效表現。

Artificial intelligence has been applied in numerous studies to predict stock trends after years of development. The fluctuating and chaotic nature of the stock market has long been a topic of interest for investors and financial researchers. As most learning models form the past studies were based on trial and error, they produced unsatisfactory performances in terms of efficiency and accuracy, or could only be applied in a closed form environment. Since the power of computers has been improved tremendously in recent years, the learning model of artificial intelligence should also be re-examined to improve decision quality.

This study built an intelligent group learning model based on the group learning concept in human behavior in traditional learning theories. Cooperative learning is widely defined as the process through which a group of individuals interact to achieve their goal. In the fluctuating stock market, investors often have various decision making approaches. This work integrates eXtended Classifier System (XCS) and neural network modules incorporating features such as dynamic learning and group decision making. An empirical study is conducted by comparing the profitability of the proposed system with that of investment strategies based on simple rules with single technical indices, individual learning XCS, buy and hold, and six-year term deposits based on the Taiwan Index. The proposed system demonstrates superior performance in terms of accuracy and the rate of cumulative return.