The Externality of Learning: Adaptive AI and Market Resilience

Abstract

Adaptive financial AI creates a system-level question that single-model evaluation cannot answer: what happens when multiple agents learn from the same experience? I identify this externality with a marginal-preserving experiment that varies exact cross-agent overlap in sequential learning histories while holding each agent's marginal learning problem fixed. Learned policies enter matched limit-orderbook markets under Deep/Thin liquidity and Shock/No-shock counterfactuals. In a pre-specified 240block architecture screen, KNN5 produces a +12.25 bp-round market-resilience effect and remains significant after family-wise correction across Ridge, CART, and KNN5 (max-T adjusted p = 0.023), while Ridge and CART remain near zero. KNN5 is then frozen and run on a fresh 240-block seed domain; the second estimate has the same sign and is statistically compatible with the first. Pooling the two independent KNN samples yields +9.32 bp-rounds [2.74, 15.91], and the result survives exclusion of the shock round itself. Shared histories increase policy dependence without materially changing heldout loss, while KNN converts that dependence into common local reference sets. Common learning therefore creates joint dependence; the architecture and locality with which experience is reused determine its market pass-through under stress.

Author abstract from the research manuscript. Typography and line breaks have been normalized; the linked source may show a different manuscript version.