Decomposing Price Chart Signals in Return Prediction

Abstract

Image models trained on stock charts predict returns out of sample, though a chart adds no information beyond the OHLCV data it renders. We decompose the image model's advantage over tabular benchmarks from the same daily window into three sources: function-class capacity, chart-specific encoding, and how each model smooths noisy returns. Capacity contributes little. Under a distillation accounting, roughly 60% of the apparent gap is tied to learning from noisy realized returns rather than chart-specific content. The residual localizes to recent bars, intraday extremes, and temporal order, and is largely recoverable with explicit predictors.