多因子股票價格預測之回饋式類神經網路模型

證券市場上有眾多的金融商品可以選擇，股票因為具有較高的變現性和投資報酬率，且進入門檻相對低，但因報酬率較高伴隨著相對高的風險，因此預測股票是股票市場中很重要的議題。影響股價變動的因素非常多且複雜，主要影響因素如企業營收、投資大眾的預期心理、股票市場動態等等。股票價格主要是透過股票市場的相關資訊進行分析，來判斷股票的走勢，以提高報酬率或避開風險。因此本研究應用回饋式類神經網路建立預測模型與分類模型，包含平均收盤價、最高價、最低價進行預測，以及三個分類模型-平均收盤價報酬率模型、最高報酬率模型、最低報酬率模型，考量三大法人買賣超和技術指標等其他因子進行模型實驗。實驗結果顯示三個價格模型，選擇的輸入變數比僅使用收盤價訓練模型的預測效果較佳，而以LSTM和倒傳遞網路相比，三個價格模型中都為使用LSTM運算的損失值較低。在分類模型中，三個模型預測的準確率皆大於80%。

關鍵字

股價預測；回饋式類神經網路

並列摘要

There are many financial products to choose from in the securities market. The stocks have high liquidity and return on investment, and the entry threshold is relatively low. Because the high rate of return is accompanied by relatively high risks, it is important to predict stocks in the stock market issue. The factors affecting stock price changes are many and complex, such as corporate revenue, investor expectations, stock market dynamics, and so on. The stock price is mainly analyzed through the relevant information of the stock market to judge the trend of the stock to increase the rate of return or avoid the risk. Therefore, this study uses a recurrent neural network to build predictive models and classification models, including the average closing price, the highest price, and the lowest price for forecasting, and three classification models - the average closing price return rate model, the highest return rate model, and lowest return rate model. Consider other factors such as the Net Buy/Sell of Three Institutional Investors and technical indicator. The results show that the three price models, the selected input variables are better than the prediction model using only the closing price training model. Compared with the LSTM and the BP, the loss values for using the LSTM operation in the three price models are lower. In the classification model, the accuracy of the predictions of the three models are greater than 80%.