![[2602.11020] When Fusion Helps and When It Breaks: View-Aligned Robustness in Same-Source Financial Imaging](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.11020] When Fusion Helps and When It Breaks: View-Aligned Robustness in Same-Source Financial Imaging
Summary
This paper explores the robustness of same-source multi-view learning in financial imaging, focusing on the effectiveness of early versus late fusion techniques for next-day direction prediction.
Why It Matters
Understanding the conditions under which fusion methods improve or degrade prediction accuracy is crucial for developing reliable financial forecasting models. This research highlights the importance of model robustness in the presence of adversarial attacks, which is vital for financial applications where accuracy is paramount.
Key Takeaways
- Early fusion can lead to negative transfer in noisier environments.
- Late fusion is generally more reliable once labels stabilize.
- Robustness is significantly affected by adversarial attacks, particularly under constrained conditions.
Computer Science > Machine Learning arXiv:2602.11020 (cs) [Submitted on 11 Feb 2026 (v1), last revised 25 Feb 2026 (this version, v2)] Title:When Fusion Helps and When It Breaks: View-Aligned Robustness in Same-Source Financial Imaging Authors:Rui Ma View a PDF of the paper titled When Fusion Helps and When It Breaks: View-Aligned Robustness in Same-Source Financial Imaging, by Rui Ma View PDF HTML (experimental) Abstract:We study same-source multi-view learning and adversarial robustness for next-day direction prediction using two deterministic, window-aligned image views derived from the same time series: an OHLCV-rendered chart (ohlcv) and a technical-indicator matrix (indic). To control label ambiguity from near-zero moves, we use an ex-post minimum-movement threshold min_move (tau) based on realized absolute next-day return, defining an offline benchmark on the subset where the absolute next-day return is at least tau. Under leakage-resistant time-block splits with embargo, we compare early fusion (channel stacking) and dual-encoder late fusion with optional cross-branch consistency. We then evaluate pixel-space L-infinity evasion attacks (FGSM/PGD) under view-constrained and joint threat models. We find that fusion is regime dependent: early fusion can suffer negative transfer under noisier settings, whereas late fusion is a more reliable default once labels stabilize. Robustness degrades sharply under tiny budgets with stable view-dependent vulnerabilities; late fus...
