應用分類元-類神經網路模型於台灣加權指數趨勢預測之研究

Abstract

傳统經濟學者利用計量經濟學模型討論股票市場的趨勢變動時，往往是預先建立相當的嚴密假設，例如回歸誤差項服從均值為零、方差為常數的分佈，且不同期的誤差項是不相關的等等。礙於這些經濟學研究預設的假設，傳統經濟學對於股市行為、趨勢的研究也因而發生實驗室所規劃或證明的數學模式往往未能合理的解釋現實實際行為。有鑑於此，本研究嘗試提出以演化理論產生適應股市行為的規則庫，並以再學習的方式對規則作最佳化，期能發現隱藏在股票市場內的知識。

本研究嘗試整合分類元系統與倒傳遞類神經網路，構成：分類元-類神經網路股票交易系統(CNTS)。研究總取樣資料為2002年3月1日至2007年3月14日台灣加權指數 1256 筆日資料，並算出每日各項技術指標。CNTS實證之第一部份是將訓練資料及測試資料劃分為加權指數上漲時機的訓練資料集、加權指數下跌時機的訓練資料集、加權指數上漲時機的測試資料集、加權指數下跌時機的測試資料集。隨後CNTS實證之第二部份以知識純化模組分別對各資料集進行訓練及測試。

實驗結果顯示，CNTS在二份測試資料上的測試準確率超過50 %，足以顯示市場趨勢並非隨機漫步，且人工智慧技術可幫助投資人對股市進行更準確的預測。

Econometricians build precise hypotheses in advance when they use econometric models to discuss the changing trends in the stock market. But baffled by these unreasonable hypotheses, economics usually can’t explain real behaviors of stock markets very well with mathematical models. Therefore, this research tries to use genetic theories to produce the rule base adapting to the behaviors of stock markets, and then re-learn it to refine those rules, so that hopefully knowledge hidden in the stock market could be discovered.

Artificial intelligence models are frequently used in financial analysis in recent years. Compared with the use of many hypotheses and limitations in econometric models, artificial intelligence models are more flexible, able to solve any nonlinear problems, and more suitable to analyze dynamic environments like stock markets. This research combines two Artificial intelligence technologies: extended classifier system and backpropagation neural network to construct a XCS-Neural-network Based Trading System, and we use this system to learn patterns from the environment and then predict values of the test set later. Experiments reveal that all test data in this research have accuracy rate 50% above. Therefore, we are confident to conclude that this system could help investors to make more precise investment decisions.