Signal Collision in Open Finance: An Auditable Common-Trade Penalty for Learned Portfolios

Abstract

Open-finance platforms can combine models from different vendors while still concentrating capital in the same trade. Shared data, objectives, and representations make decision diversity the relevant object, rather than diversity of model names. We call loading on a shared direction a signal collision. A linear-impact game shows that each allocator chooses excessive common exposure because it does not bear the price effect imposed on others; an overlap toll implements the social choice. A feedback model gives the exact gain at which return-responsive allocation becomes unstable, and a spectral result states the restrictive conditions under which a leading eigenportfolio can proxy for the collision direction. These results motivate a transparent post-model penalty. In a frozen panel of ten liquid cross-asset ETFs over 760 completed weeks from January 2012 through July 2026, the penalty changes maximum drawdown from 23.6% to 14.1% and volatility from 12.6% to 10.7%, while excess-return Sharpe moves from 0.664 to 0.667. A paired stationary-bootstrap 95% interval for drawdown improvement is [-3.3, 18.4] percentage points. The result is an auditable risk-control association, not causal evidence of AI crowding; holdings and signed flows remain necessary for structural validation.