When Fusion Helps and When It Breaks: View-Aligned Robustness in Same-Source Financial Imaging

Battle Count: 76

Directly relevant to quantitative trading in terms of: (1) next-day direction prediction as a fundamental trading signal, (2) evaluation protocol design (leakage-resistant splits, embargo) critical for backtesting reliability, (3) adversarial robustness analysis relevant for understanding model fragility in live trading, (4) fusion design choices affecting signal quality. However, the paper focuses on methodology and robustness analysis rather than proposing a deployable trading strategy. The pixel-space threat model is a representation-level stress test rather than a realistic market manipulation scenario.

Moderate complexity. Requires: (1) custom image rendering of OHLCV charts and indicator matrices, (2) dual-encoder late fusion architecture with consistency regularization, (3) FGSM/PGD attack implementation with budget conversion for standardized inputs, (4) leakage-resistant time-block splits with embargo, (5) min_move filtering protocol. The CNN backbones are standard, but the multi-view construction, view-channel mapping, and adversarial evaluation protocol add implementation overhead. PyTorch-based with ImageNet pretrained weights.

Code promised upon acceptance at GitHub. Detailed appendices cover implementation (Appendix C), training protocol (Appendix E), attack hyperparameters (Appendix F), filtering/split protocol (Appendix D), and normalization (Appendix A). Fixed random seeds {1-8} for deep models. Full architecture specifications provided. However, code is not yet publicly available.