Liquidity-Based Audit of AI and Algorithmic Trading Strategies

Battle Count: 66

The paper is directly and fundamentally relevant to quantitative trading. It provides a model-free method to classify any algorithmic strategy (including black-box neural networks) as a liquidity consumer or provider from trade and price data alone. The Roll-spread identity gives a real-time illiquidity proxy computable from the same trajectory data. The N-agent fire-sale extension quantifies systemic risk from correlated AI strategies. The O(T·nd) computational complexity makes it practical for daily monitoring. The empirical calibration to CRSP 2016–2025 with specific Sharpe ratios for momentum, reversion, bias-corrected, and minimum-variance strategies provides actionable benchmarks. The neural-network audit section directly addresses the practical problem of auditing proprietary quant fund strategies.

The core trajectory estimator (sum of per-period covariances) is straightforward: compute cross-products of cost vectors and position vectors, sum over T periods, apply Newey-West HAC correction. This is O(T·nd) and computable in real-time. The AR(1) correction requires estimating the autocorrelation coefficient and computing the geometric series correction. The N-agent systemic extension requires constructing the N×N strategy-covariance matrix (O(N²·n) per day). The neural-network audit requires training a small MLP with walk-forward validation. The main complexity lies in understanding the theoretical framework (multi-period stochastic dynamic programming, covariance decomposition, Kyle embedding) rather than in implementation. For a typical institutional setting (T=252, d=n=100), the audit requires ~2.5×10⁶ floating-point operations per cycle.

The paper uses publicly available CRSP daily equity data (2016–2025, n=21,183,373 stock-days) accessible via WRDS. Computational complexity is fully specified (O(T·nd) for the trajectory estimator). The neural-network audit uses a standard 2-hidden-layer MLP (16→8 ReLU) with walk-forward training. All theoretical results are proven with full derivations in the appendix. However, no code repository is provided, and some empirical details (e.g., exact WRDS query parameters, network hyperparameters beyond architecture) would need to be inferred.