Evaluating AI Investment Strategies

Battle Count: 50

Directly addresses the core problem of auditing algorithmic trading strategies without access to internal models. Provides a tractable, model-free metric (input-output covariance) for evaluating momentum, reversion, and minimum-variance strategies. The regret decomposition quantifies exactly how much momentum strategies lose relative to optimal benchmarks. Computational complexity O(T·nd) makes it practical for real-time monitoring. The paper explicitly addresses hedge funds, execution algorithms, and robo-advisors. Empirical validation on 10 years of CRSP data demonstrates practical applicability. The connection to SR 11-7/SR-26-2 regulatory requirements makes it directly relevant to institutional quantitative trading compliance.

The core trajectory covariance estimator is straightforward: compute sample means, form cross-products, sum over T periods. HAC variance estimation with Newey-West bandwidth h=O(T^(1/3)) is standard. Online updates via Welford's algorithm are O(nd) per observation. The Bellman recursion requires state-space discretization or function approximation (fitted-Q). The main complexity lies in: (1) correctly specifying the benchmark policy π*, (2) handling non-stationary cost distributions with regime detection, (3) implementing bias corrections for momentum strategies, (4) ensuring w > d for covariance matrix invertibility in portfolio applications. For a 100-asset portfolio over 252 days, computation is ~2.5×10^6 FLOPs per cycle, well within standard hardware budgets.

The paper uses CRSP daily return data from WRDS (January 2016 - December 2025), which is available to academic subscribers. The theoretical framework is fully specified with proofs. However, no code repository is explicitly provided. The trajectory covariance estimator is well-defined and computable in O(T·nd) time. Empirical analysis is confined to momentum/reversion strategies with a simple if-then signal. Key parameters (window sizes, asset dimensions) are specified.