基於長短期記憶神經網路以新聞事件進行股價預測

Abstract

近年來，資訊技術慢慢滲透進傳統金融業，運用科技手段使得金融服務更有效率，並且發展許多創新的服務，其中機器人理財為自動化的投資顧問，自動提供客戶理財相關的資訊及建議，為了能提供機器人理財更精確的資訊，推播給客戶更完善、適合的投資組合，本篇研究將對股價進行預測，尋找適合的買點及賣點，基於Eugene F Fama (1965)所提出之金融市場為資訊有效的( Informationally Efficient )，因此使用消息面之資訊，取得美國路透社以及華爾街日報之新聞資料，對台灣加權指數進行股價漲跌預測。為從龐大之新聞資料中擷取重要事件，引進開放資訊擷取( Open Information Extraction )概念，並基於Divide-and-Conquer演算法之概念，提出全新的事件擷取方式，更完整的擷取出新聞事件。爾後訓練自動編碼器( AutoEncoder )模型，學習事件參與者與事件關係間的語義關聯，根據此模型從事件矩陣中萃取出特徵向量，達到事件嵌入( Event Embedding )之效果。之後依據事件向量建立深度學習模型，考量新聞事件對股價市場之影響並非短短一兩天，因此使用過往一個月之新聞事件對股價漲跌進行預測，為要處理時間序列之資料，因此建立長短期記憶神經網路( Long Short-Term Memory Network, LSTM )。此外，為衡量事件嵌入以及長短期神經網路之表現，建立文字嵌入以及極限梯度提升模型( XGBoost )進行比較，實驗結果顯示不論是在牛市亦是熊市之市場趨勢下，本篇研究所提出之方法表現皆較佳，且只需使用最簡單之交易策略進行模擬交易，便可得到不錯之報酬。

In recent years, information technology has slowly penetrated into the traditional financial industry, using technology to make financial services more efficient and develop more innovative services. Among them, Robo-advisor is an automated investment consultant, automatically providing customers with financial information and advice. In order to provide more accurate information with Robo-advisor, this research will build a deep learning model, based on the news resources obtained from Reuters and the Wall Street Journal, to predict the stock price rise and fall of the Taiwan Weighted Index, hoping to find suitable buying and selling points. In order to extract important events from huge news data, the concept of Open Information Extraction is introduced. We propose a new event extraction method, based on the concept of Divide-and-Conquer algorithm, to more accurately extract news event. Afterwards, an AutoEncoder model is trained to learn the semantic pattern between event participants and event relationships. According to this model, feature vectors are extracted from the event matrix to achieve the effect of event embedding. After that, considering the impact of news events on the stock market as a long-term effect, a deep learning model, a long short-term memory neural network, is built using news events from the past month. In addition, in order to measure the performance of event embedding and long short-term neural networks, a word embedding and extreme gradient boosting model (XGBoost) was established for comparison. The experimental results show whether in the bull market or bear market, the methods we proposed perform well, and only need to use the simplest trading strategy for market simulation, we can get a good reward.