利用深度學習圖形辨識技術建置最適投資策略-以台灣股票市場為例

Abstract

近年來，隨著電腦技術的革新，人工智慧在各領域皆有所突破。其中，圖像辨識可說是人工智慧運用的相當廣泛的一個領域，因此，本研究希望透過深度學習中圖像辨識相關技術，來預測股票線圖在未來的走勢，進一步選出預期報酬較高之股票作為投資組合。\n本研究針對股票線圖一共進行兩階段處裡，第一階段採用自動編碼器（Autoencoder）技術，訓練出可將股票蠟燭圖、成交量圖降維之模型；第二階段則使用多層感知機(Multiple Perception Layer)模型對降為後資料進行學習，預測未來股票報酬率，建置投資組合。\n最後，本文透過實證分析，回測模型績效，回測期間從2012至2019共8年，回測結果平均年化報酬率達22.69%，平均年化夏普比為1.49，明顯優於台灣加權指數表現。

In recent years, with the innovation of computer technology, artificial intelligence has made lots of breakthroughs in various fields. Among them, image recognition can be said to be a really successful one. Therefore, this paper hopes to predict the trend of stock charts through the image recognition skill in deep learning in order to construct the optimal portfolio.\nThis paper applies two models to predict stock charts. First, an AutoEncoder is used to reduce the candlesticks charts and volume charts from three dimensions to one dimension. We then take these 1D data as input to our second model - Multiple Layer Perception（MLP, supervised learning）. We apply MLP model to predict stocks’ future returns, thereby constructing the portfolio.\nFinally, this paper evaluates the investment strategy through the empirical analysis. In conclusion, the strategy deliver an average annualized return of 22.69% and an average annualized Sharpe Ratio of 1.49, which all outperform than Taiwan Capitalization Weighted Stock Index（TAIEX）.