From Pixels to Signals: Hierarchical Vision Transformers for Return Prediction

Abstract

Recent evidence suggests that candlestick-chart image predictability weakens in more recent data and remains hard to interpret. Using CRSP U.S. equities (1993-2024), we revisit this claim by comparing a CNN benchmark with a hierarchical vision transformer (Hiera) trained on candlestick-and-volume images. Hiera produces a stronger and more stable cross-sectional return ordering that persists in a post-2019 holdout and survives implementability checks. We develop two diagnostics - controlled synthetic probes and occlusion maps - paired with spanning regressions to characterize the learned signal. The diagnostics constrain interpretation and localize predictive content to patterns near the end of the lookback window.