Thai Stock Chart Pattern via Deep Learning-Based Multi-Indicator Object Detection

Abstract

Currently, stock market investment strategies are categorized into fundamental analysis and technical analysis. Technical analysis seeks to understand market behavior by examining historical price and volume data, employing methods such as price pattern analysis, candlestick interpretation, and support–resistance levels. Among these, one of the particularly interesting approaches is chart pattern recognition, which consists of bullish and bearish patterns that may indicate potential price reversals to support investment decisions. However, the complexity and uncertainty of price movements, coupled with the presence of multiple technical indicators on a single chart, pose significant challenges to human pattern recognition. This study proposes a deep learning–based object detection framework to automatically recognize bullish and bearish patterns. Daily stock chart images from the Stock Exchange of Thailand were collected, incorporating price graphs alongside RSI, MACD, and STO indicators. The labeled images were used to train and validate models through K-fold cross-validation, evaluating seven state-of-the-art detection architectures using the mAP@50 metric. Experimental results revealed that YOLOX achieved the best performance, with an average mAP@50 of 0.887 and a variance of 0.057, particularly excelling on charts combining price graphs with MACD indicators. These findings suggest that deep learning models may assist technical analysis by supporting more systematic chart pattern recognition.