Artificial Intelligence and Systemic Risk: A Unified Model of Performative Prediction, Algorithmic Herding, and Cognitive Dependency in Financial Markets

Battle Count: 70

Highly relevant to quantitative trading as it directly addresses how AI-driven trading strategies create systemic risk through correlated signals, performative feedback, and cognitive dependency. The model provides a framework for understanding algorithmic herding, flash crash amplification, and the 'calm before the storm' pattern. The empirical findings on portfolio convergence (cosine similarity +12% post-AI), tail-loss amplification (18-54%), and the distinguishing cross-sectional signature (within-AI-group dispersion ratio ≈1) are directly actionable for quant traders managing correlated strategy risk. The saddle-node bifurcation and hysteresis results inform regime-switching risk models. However, the paper is primarily a theoretical/econometric contribution rather than a trading strategy paper.

Very high complexity. The theoretical model involves a three-layer structure with Kyle equilibrium, supermodular adoption games, saddle-node bifurcations, and endogenous state accumulation. The empirical strategy requires processing 99.5M 13F holdings, constructing Bartik IVs with leave-one-out jackknife, EDGAR NLP analysis, and agent-based simulations with 500 institutional agents, 1000 noise traders, and 5040 periods. The closed-form derivations require solving coupled fixed-point problems. Reproducing the full analysis requires access to SEC structured data, EDGAR API, and substantial computational resources for ABM simulations. The theoretical proofs span multiple online appendices.

Simulation code available upon request. Empirical analyses use publicly available SEC Form 13F-HR structured data sets, EDGAR full-text search API, and FRED economic data. Data processing scripts and replication code documented in replication package. However, the theoretical model proofs are in Online Appendix A, and some event study microstructure metrics are model-generated rather than directly measured from proprietary TAQ data. The ABM calibration is moment-matching rather than structural estimation. Key parameters (ρ, β, κ) are indirectly identified rather than directly estimated.