Image‐Based Deep Learning Models for Stock Predictions: Combining Line, Candlestick, and Bar Charts

ABSTRACT Recently, numerous studies have appeared in the literature on the use of deep learning techniques to analyze the image patterns for stock prediction. Although there are several different types of imaging techniques that can be used to represent related technical indicators, including line charts, candlestick charts, and bar charts, no one has yet examined the effect of using these types of image visualization techniques on the prediction performance of deep learning models. In this paper, three types of image patterns are compared, specifically, line charts with trading volume information represented by a bar chart, candlestick charts with trading volume information, and a mixed type of image with two other related technical indicators, that is, MACD and RSI. The experimental results that are based on data for six companies from different industries and with different scales of stock price fluctuation show that the mixed image pattern type allows 2‐D CNN and VGG16 to perform better than the other two image pattern types in terms of predicting the stock prices for the next day, week, and month. In addition, they outperform the LSTM and 1‐D CNN baseline models when using the time series data representing historical stock prices. Furthermore, three ensemble deep learning models are constructed for performance comparison, including VGG16‐LSTM, 2‐D CNN‐LSTM, and the stacking model, in which the VGG16‐LSTM and the stacking models perform the best for the prediction of short‐ to mid‐term and mid‐ to long‐term stock prices, respectively.