應用多重類神經網路整合總體經濟因素驗證CAPM行為之有效性研究

Abstract

台灣股市是一個淺碟型市場，存在高度系統風險，股價極易受外力影響，而導致劇烈波動。本研究提出以多重類神經網路為架構，搭配總經基本分析月資料(長期)、技術分析週資料(短中期)，來對台灣個股波段股價趨勢行為進行學習。主網路將針對長期、短中期單一網路做總和評判，使類神經網路的輸出更具可靠性，以建立一個有效度的個股波段趨勢預測模型。 由實驗結果得知，多重類神經網路模型在預測準確率以及獲利能力上，表現均較單一技術面類神經網路模型、隨機漫步模型優異。證實經多重類神經網路統整長、短期物理力量總和評判後，對於個別股票波段投資的預測效果確實會優於單一網路。同時也與CAPM所提的個別公司價值會受整體市場因素(即總體經濟指標)影響之理論相呼應。

The stock market in Taiwan is a shallow-plate market, which is full of high level of systemic risk. Stock price in this market goes up and down seriously because it affects by lots of economic factors. This study proposes a multi-neural network model with monthly data on fundamental analysis (long-term) and weekly data on technical analysis (short-term), and tries to find the knowledge rules of the trends in stock price behavior. By using multi-neural network, we make integrated evaluation of long-term and short-term sub-network together, and we analyze the result to increase reliability of the neural network’s output. After that, a Taiwan Stock Market trend of stock wave forecast model is established.

The results show that multi-neural network is significantly more effective than single neural network and random walk model in forecasting accuracy and trading profitability. This thesis built the multi-neural network model of integrating long-term and short-term factors with physical quantity in stock wave. It prove multi-neural model has the better prediction result than single neural network. In addition, we prove the theory by CAPM which claims the value of individual company will be affected by the whole market.