Cross-Market Robustness of CNN + PPO for Multi-Stock Trading: Evidence from the United States, China, and India

This study examines the external validity of a convolutional neural network (CNN) feature extractor trained with Proximal Policy Optimization (PPO) for multi-stock daily trading across three equity universes: S&P50 (United States), SSE50 (China), and Nifty50 (India). This model was trained on 2023 data and evaluated for out-of-sample performance in 2024, using the same preprocessing, features, hyperparameters, and training protocol. It studied the sensitivity to lookback (10–50 trading days) with a window-expansion technique, and benchmark the CNN+PPO approach against a multilayer perceptron (MLP) baseline. The U.S. sample showed a clear advantage for CNN in shorter windows (particularly 20 days), as evidenced by increased risk-adjusted returns and more localized salience on recent channels. The magnitude of these gains is mixed on SSE50 and non-existent or negative on Nifty50. Complementary gradient saliency and permutation-importance diagnostics suggest that these cross-market differences in performance and saliency can be traced to a mismatch between the convolutional inductive bias and the presence or absence of short, cross-sectional motifs. Results suggest that architecture choice and deployment should be data-driven and informed by market diagnostics and feature engineering.