The Inference Problem in Financial AI: Forecasting, Portfolios, Market Quality, and Systemic Resilience

Abstract

Financial AI is commonly evaluated at the model or forecasting level, while its economically consequential effects arise through portfolio construction, market interaction, and stress-state behavior. Evidence at these layers is not interchangeable. This paper develops a cross-layer inference framework that separates five economic objects—AI capability, volatility forecasting, portfolio performance, market quality, and systemic resilience—and treats each transition as a distinct empirical question. Recent evidence reveals three recurring divergences: greater general model capability does not guarantee superior volatility forecasts; attractive AIgenerated portfolios can arise without superior forecast skill; and lower price variation can coexist with deteriorating liquidity. The framework shows why these divergences are economically meaningful rather than merely terminological: optimization can alter the value of forecasts, aggregation can transform private decisions into market-wide order flow, and liquidity or funding constraints can change normal-market relationships under stress. A structured evidence audit further reveals an identification gradient, with comparatively dense predictive evidence and much thinner evidence linking adaptive AI to systemic outcomes during real market stress. The paper does not presume that AI is stabilizing or destabilizing. It specifies where inference stops and what additional evidence is required before an upstream performance gain can support a downstream financial conclusion.

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