Author(s)

Yuting Zhang, Shuang Diao

Corresponding Author

Shuang Diao

Abstract

This paper investigates how AI algorithmic bias distorts financial resource allocation across credit, investment, and market-making, and proposes a regulatory design that aligns model performance with allocative efficiency, fairness, and stability. A governance-ready framework integrates causal-robust bias diagnostics, efficiency-oriented evaluation, and supervisory tooling across model validation, disclosure, and impact audits. A five-layer architecture-data, features, forecasting, optimization, and risk-and-compliance-links micro decisions to macro allocative outcomes via counterfactual measurement and distributionally robust optimization. A semi-synthetic evaluation calibrated to real financial data indicates that bias in data, models, and interactions amplifies mispricing, rations credit to solvent but underrepresented segments, and concentrates liquidity, lowering the marginal product of capital and elevating tail risk. A regulatory bundle with pre-trade fairness constraints, post-trade outcome monitoring, model change governance, and sandboxes with graduated obligations improves allocative efficiency with limited performance cost.

Keywords

Algorithmic bias; Financial resource allocation; Allocative efficiency; Causal robustness; Fairness; Robust optimization; Model risk management; Conduct regulation; Supervisory technology; Systemic risk