實數分類元系統於股票型基金投資組合建構之運用

Abstract

金融市場呈現非線性與半結構化的特性。靜態的方法難以掌握市場瞬息萬變的走勢。然而，人工智慧因具有彈性，可搜尋複雜求解空間，近年來人工智慧在求解金融問題上的應用相當廣泛。人工智慧方法中的分類元系統因具有機器學習與加強式學習的機制，可與問題環境互動。因此，本研究應用實數分類元系統建置股票基金投資組合建構模型，研究方向以投資標的與時點選取為主，以期建構非時間落後性的投資組合決策模型。

投資標的範圍選定台灣50指數之成份股，研究中以技術指標作為輸入因子，持有特定天期後之漲跌作為輸出，建構股票型基金之選股模型；並依模型特性設計一套交易機制，依照模型所建議之標的建構投資組合進行樣本外模擬交易測試，實驗結果顯示依此模型建構出的基金投資組合之報酬率顯著優於市場報酬。本研究結論為採用具有動態環境適應能力的實數分類元系統結合技術指標，可有效掌握市場變動，適於基金經理人於管理股票型基金投資組合之用。

The characteristics of financial market are nonlinear and semi structure. Thus, the behavior of the dynamic market is difficult to catch by using static approaches. Furthermore, artificial intelligence was widely applied to solve financial problem due to its flexible. Extended Classifier Systems (XCS) is a novel methodology in artificial intelligence, which consists of machine learning and reinforcement learning technique that can be used to interact with a given environment. By the generalization and online learning ability of XCS, it can generate initial rules from training data and keep evolving rules in testing environment. In this study, XCS with continuous-valued inputs (XCSR) is applied to develop a portfolio construction model which can adapt to the dynamic financial market.

The experiment is designed to demonstrate the predictive ability of XCSR. The investing targets of this research are Taiwan 50 index constituents. Five technical indicators are taken as input factors. The simulation and statistical results show that XCSR portfolio construction model is able to achieve a positive excess return in out-of-sample simulated trading. The performance of XCSR model is obviously superior to other investing strategies and that market return, and XCSR portfolio construction model can be concluded which is suitable for fund managers to manage the portfolio of equity fund.