A Study on CNN Feature Extraction for Stock Price Prediction

Abstract

The return on investment for investors in the stock market is highly dependent on the investor's timing strategy, that is, the decision of what time to buy or sell a stock. A successful timing strategy requires investors to accurately identify the price movement of a company. As a result, some investment professionals have created technical analysis analytical methodologies to forecast the short-term trend of a stock. However, technical analysis approaches are prone to subjectivity, such as the selection of technical indicators and indicator periods. This essay attempts to utilize a convolution layer in deep learning to extract features as an alternative to technical indicators and to reduce subjective elements' effect on prediction bias. Several stock predictions are evaluated between a standard LSTM model and an LSTM model with convolution layers (CNN-LSTM model) in this research. The experimental results show that the CNN-LSTM model outperforms the standard LSTM model in predicting the price of certain stocks with a big market capitalization and high liquidity.