Forecasting returns using image-based convolutional neural networks: Evidence from Korea

Abstract

This study employs a chart image-based convolutional neural network (CNN) to predict stock returns in the Korean stock market, following Jiang et al. (2023). We transform historical price and volume data into chart images and utilize CNN to extract predictive patterns. Our findings demonstrate that the CNN-based models outperform traditional benchmarks, particularly for short-term return forecasts. Additional double-sort and panel logistic regression analyses with firm characteristic variables, buy-sell imbalance analysis of investor groups, and subsample tests confirm the robustness of CNN-based predictors. This study represents the first application of a chart image-based deep learning model to the Korean stock market, providing new insights into the potential of deep learning models for stock return forecasting in emerging markets.