Decoding the Unique Price Behavior in the Japanese Stock Market with Convolutional Neural Networks

Abstract

Technical analysis charts contain a wealth of information in a two-dimensional space, depicting price, volume, and moving averages, which include relational attributes that are challenging to discern using one-dimensional time series methods. In this study, we investigate return predictability in the Japanese stock market, a unique setting where the common momentum effect remains unobservable. We employ a statistical learning approach to uncover the predictive patterns underlying the data, using a Convolutional Neural Network (CNN) designed to automatically extract a large number of features from chart images. Our findings suggest that the features extracted from past stock charts possess predictive power for subsequent returns, particularly in larger, more liquid stocks. This result reveals the distinctive characteristics of return predictability in the Japanese stock market, highlighting that factors other than the common momentum effect play a significant role. It suggests that two-dimensional historical data may uncover valuable information about the future, offering insights into the unique price behavior observed in this market. The implications of our findings extend to the development of sophisticated trading strategies and the reevaluation of market inefficiency in different contexts.

Keywords: convolutional neural network (CNN), image classification, machine learning, technical analysis, return prediction, Japanese stock market

JEL Classification: G12, G14