運用機器學習預測台股股價走勢

Abstract

機器學習為人工智慧的範疇，其應用層面不斷擴大，也包括了金融投資領域。運用機器學習預測股價，有效協助投資決策，為眾多研究努力的方向；但是金融資訊具有小樣本、高雜訊、及非穩態時間序列的特性，使得以往研究的預測準確率不易提高；雖然將預測期間縮短可以增加預測的準確率，但是過短期的預測同時也降低了實務上運用的可行性。 本研究以台灣股市投資為例，結合領域知識，將金融資料做基本面、趨勢面、動能面、及籌碼面等多面向的特徵轉換，再運用不同的機器學習技術和模型來克服上述股價預測的困難。 本研究結果顯示，我們利用長短期記憶模型(Long short-term memory, LSTM)加上整合學習(Ensemble learning)，預測上市櫃14檔權值個股未來一個月股價漲跌的準確率為83.23%。若以台積電(2330.TT)為例，預測未來一至三個月的股價漲跌的準確率高達84.57%。我們進一步利用此模型預測結果做2016/1/4~2018/12/28三年期間的回溯測試，所獲得的模擬投資報酬率為341.66%，遠勝過同期間台積電的77.92%漲幅。模型對於國巨(2327.TT)的模擬回測報酬率更高達2,888.36%，大幅超越該股票同期間364.74%的漲幅。 本研究未來發展方向為：1)廣度發展。將研究、測試、及應用範圍擴大至全台灣股市個股、海外股市、或是其他金融商品市場。2)深度發展。以本研究為基礎，利用強化學習(Reinforcement learning)實現人工智慧投資，取代經理人投資操作，並獲取更優異的報酬。

One of the category of artificial intelligence, machine learning and its applications continue to expand, including the financial investment field. Using machine learning to predict stock prices in order to effectively assist investment decisions is the top goal of many research efforts; However, the financial data has the characteristics of small sample size, high noise, and nonstationary time series dependency, which make the prediction accuracy too low to be reliable. Although previous researches indicate that shortening the forecast horizon can increase the accuracy of predictions, it also reduces the feasibility of practical application. This study takes the Taiwan equities as examples by combining feature transformations of the domain knowledge, together with fundamental, technical, momentum, trend, trade factors and a variety of machine learning techniques so as to overcome the difficulties that pertain to the stock price prediction. The outcome of our study is encouraging. Among the five candidate machine learning models we use, the designed LSTM generates highest predictive accuracy of 84.57% for predicting TSMC(2330.TT)’s price movement in the next 1~3 months. We also attain comparable results of 83.23% accuracy rate for the total 14 large cap names. Moreover, the 3-year back-testing simulations we conduct on TSMC(2330.TT), Yageo(2327.TT), Largan(3008.TT), and Fubon(2881.TT) shows astonishing returns of 341.66%, 2,888.36%, 371.13%, and 131.03%, compared with their same period returns of 77.92%, 364.74%, 13.41%, and 34.82%, respectively. There are two aspects for our future study. First, we will increase the coverage to all stocks in Taiwan equity market and expand the application to overseas and other financial markets if possible. Second, based on our finding, we can then use the model predictions as policies of reinforcement learning so that we can truly realize artificial intelligence investment.