From vision to value: Stock chart image-driven factors and their pricing power

초록

This study investigates whether visual information embedded in stock charts contains economically meaningful signals for asset pricing. Using deep learning models trained on candlestick images, we construct predictive factors via convolutional neural networks (CNN) and vision transformers (ViT). The models generate monthly probabilities of future returns, transformed into double-sorted, value-weighted portfolios, and evaluated within a high-dimensional asset pricing framework. Employing the double-selection LASSO, we assess whether these image-driven factors contribute incremental pricing power beyond 161 established risk factors. The ViT-derived factor exhibits robust and statistically significant stochastic discount factor loadings, indicating that it captures a novel priced dimension of risk associated with visual market heuristics and trend saliency. It consistently delivers positive and economically meaningful return spreads across firm sizes, with superior Sharpe and Sortino ratios relative to standard benchmarks. In contrast, the CNN-based factor shows weaker performance, particularly among large-cap stocks, suggesting limited economic relevance despite statistical significance. Factor selection patterns reveal that ViT signals are linked to multiple economic channels (low-risk, profitability, and financing), while CNN signals primarily align with quality and safety characteristics. Overall, the results demonstrate that advanced deep learning can extract interpretable, orthogonal source of priced risk from chart images, enriching the empirical asset pricing landscape and highlighting the potential of computer vision as a tool for understanding market dynamics.

키워드