基於混合注意力機制與長短期記憶之股票趨勢預測

Combining Hybrid Attention Networks and LSTM for Stock Trend Prediction

本研究結合長短期記憶（LSTM）中股價時間序列的特徵，以及混合注意力模型（HAN）中模擬人判讀新聞影響股票重要性，有效地學習股價時間序列和新聞報導中的訊息順序，藉此訓練新聞文章與歷史股票交易資料之間的關聯，建構股票漲跌趨勢之模型。根據實驗結果，相較於僅單獨使用新聞或是股價資訊的模型，加入時間序列的新聞文章訊息使模型更能精準的預測股價市場趨勢，在兩種資訊的結合中，與HAN、LSTM模型相比，最佳準確度為80%，整體最高可提升40%的準確度。

Our research merge two different models of the Hybrid Attention Networks (HAN) and the Long Short-Term Memory (LSTM) to improve the stock trend prediction. The combination of the two algorithms helps leverage the advantages of both models to learn sequential information in time series and news articles. The experimental results show that the best accuracy score, combined with news and stock prices, is 80 %. The performance of the proposed model compared to HAN and LSTM model increased by up to 40%.