台灣股票市場趨勢預測月營收策略機器學習系統

Abstract

股票市場趨勢預測是許多學者在研究的議題，有許多人依據技術分析、基本面分析、籌碼面分析以及消息面分析來構建交易策略，也因為近幾年電腦硬體設備的進步使得運算速度的提升，越來越多的研究者使用機器學習來做金融相關交易策略的研究。 本研究希望建置一個交易頻率不要太過頻繁而且能夠依據市場動態調整持股的交易策略，而使用月營收策略來研究，每當月營收公布後，該月的11日買，次月的10日賣。使用的訓練特徵有：股價技術指標、成交量技術指標、基本面指標、風險指標以及月營收指標，使用的標籤為該月11日至次月10日股價漲跌幅的排名。使用九種不同的機器學習演算法進行訓練：Random Forest、AdaBoost、GBDT、LightGBM、SVR、DNN、ResNet、CNN、LSTM+Attention。將九個訓練出來的模型合成為集成模型系統，提升預測的準確度，回測時分別計算做多、做空以及多空皆做，且使用不同數量的股票分析模型績效。以獲利的觀點來看是做多前1檔表現最好，年化獲利率可以到達401.48%、每月交易勝率為90.90%。 最後探討月營收指標在集成模型系統中的重要性，以及用集成模型系統和一般月營收策略、市售策略、元大台灣50 ETF的績效做比較。

Stock market trend prediction is the research topic in which academic interests reside. Many people build trading strategies based on technical analysis, fundamental analysis, chip analysis, and message analysis. In addition, given the advancement of computer hardware equipment has increased the speed of computing in recent years, more and more researchers use machine learning to research financial trading strategies. With the aim to build a trading strategy that does not have overly high trading frequency and can adjust the holdings contingent on the market dynamics, this study employs monthly revenue strategy, i.e., following the announcement of the current monthly revenue the stocks are purchased on the 11th of the current month and sold on the 10th of the next month. The training features used include: stock price technical indicators, trading volume technical indicators, fundamental indicators, risk indicators, and monthly revenue indicators. The labels used are the rankings of stock price changes from 11th of the current month to 10th of the next month. This study uses nine different machine learning algorithms for training: Random Forest, AdaBoost, GBDT, LightGBM, SVR, DNN, ResNet, CNN, LSTM+Attention. The nine trained models are combined into an ensemble model system to improve the accuracy of the prediction. When backtesting, this study respectively calculates go long, go short, and go both of long and short with the different number of stocks. From the profits perspective, go long in the first rank has better performance, with an annualized interest rate of 401.48% and a monthly trading win rate of 90.90%. Finally, this study discusses the importance of monthly revenue indicators in the ensemble model system; compares the performance of the ensemble model system with the general monthly revenue strategy, the strategy sold from market, and Yuanta Taiwan Top 50 ETF.