Predicting Commodity Returns Through Image‐Based Price Patterns

We examine the predictability of commodity futures returns using image‐based price patterns extracted from open‐high‐low‐close (OHLC) charts. Applying convolutional neural networks (CNNs) to US commodity futures data, we extract predictive signals without predefined patterns such as momentum or mean reversion. Empirical results demonstrate that image‐based predictions enhance predictive accuracy, particularly over short‐ and medium‐term horizons, with 20‐day OHLC images yielding the most robust performance. Compared with traditional financial predictors, CNNs capture nonlinear dependencies while retaining unique explanatory power. Panel regressions confirm that image‐based predictions are correlated with established return factors. However, transfer learning—from the US to the Chinese markets—proves ineffective in commodity futures markets, highlighting the necessity of market‐specific adaptation.