運用圖像辨識模型於股票趨勢預測－以台股市場為例

Abstract

許多投資人會利用技術分析來尋找股市的進出場時機藉此賺取報酬，儘管技術分析目前已經廣為使用，但其本身仍有些限制，因此本論文將深度學習中的圖像辨識技術應用在技術分析上，希望能改善技術分析的限制，同時又保有穩定的超額報酬。我們嘗試藉由週、月、季的技術線圖來預測台股市場個股的隔月價格趨勢，其中的技術線圖包含K棒、成交量及移動平均線，並透過卷積神經網路模型來實現價格趨勢預測。實證結果發現，短天期模型提供較優的預測能力，此外，模型在預測市值小個股的效果優於市值大個股，這顯示圖像辨識模型仍有其限制，儘管如此，我們相信深度學習於金融市場的應用依舊有很大的發展空間。

After the invention of technical analysis, many investors try to find out the excess return through technical analysis. However, despite the fact that technical analysis has been used generally, there still exists some restrictions. In order to deal with those restrictions, we try to apply image recognition technology into technical analysis. We hope we can improve the restriction and find out the excess return at the same time. We use the weekly, monthly and quarterly chart pattern and convolutional neural network to predict the price trend of next month in Taiwan stock market. The information in the chart pattern includes OHLC chart, volume and moving average. In our empirical result, we find that the weekly model provides the best performance on prediction. In addition, the performance on predicting the stocks with low market capital is better than those with large market capital. These results show our models still have some restrictions. Nevertheless, we believe the development of using deep learning in the financial field is still huge.