📈 economics

Evaluating AI Investment Strategies

This paper establishes an exact decomposition of cumulative regret into per-period covariances between cost vectors and policy decisions, providing a tractable, model-free audit framework for evaluating black-box algorithmic strategies in dynamic environments such as platform mechanisms and portfolio management.

Original authors: Irene Aldridge

Published 2026-06-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Irene Aldridge

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a food critic trying to judge a chef's cooking, but you aren't allowed into the kitchen. You can't see the recipes, the ingredients being chopped, or the chef's internal thought process. All you can see is the order the customer gave (the input) and the plate that came out (the output).

For years, regulators and investors have struggled with this exact problem regarding AI algorithms that manage money. They can see what the AI bought and sold, and they can see the market prices, but they can't "open the box" to see how the AI made its decisions.

This paper by Irene Aldridge offers a clever, mathematical "magic trick" to solve this. It proves that you can accurately measure how well (or poorly) an AI is doing just by looking at the relationship between the orders it received and the plates it served.

Here is the breakdown of the paper's main ideas using everyday analogies:

1. The Core Idea: Regret is Just "Covariance"

In the world of finance, "Regret" is the gap between what an algorithm actually earned and what it could have earned if it had perfect foresight (or at least, perfect knowledge of the average market).

The paper's main discovery is a simple equation: Regret = Covariance.

Think of it like a dance partner:

The Cost (The Music): This is the market condition (e.g., prices going up or down).
The Decision (The Dance Move): This is what the AI decides to do (buy or sell).

If the AI is a good dancer, it moves with the music. If the music goes up, it steps up. If the music goes down, it steps down. In this case, the "covariance" (how well they move together) is high, but the Regret is low because the AI is doing exactly what the situation demands.

However, if the AI is a bad dancer, it might step up when the music goes down. It is moving against the flow. The paper shows that you can measure exactly how much "bad dancing" (Regret) happened just by calculating how often the AI's moves clashed with the music. You don't need to know the choreography; you just watch the dance.

2. The "Perfect" vs. The "Real"

The paper proves that this "Regret = Covariance" rule works perfectly under specific conditions:

The Music is Random: The market moves are independent of each other (like flipping a coin every day).
The Dancer is Honest: The AI's average decision matches the mathematically perfect decision for that day.

If these conditions are met, the math is exact. You can sum up the daily "clashes" between the market and the AI's decisions, and that total sum is exactly how much money the AI lost compared to the best possible strategy.

3. What Happens When the AI is "Biased"?

In the real world, AI isn't always perfect. Sometimes it has a "habit" or a "bias."

The Momentum Trap: Imagine an AI that always tries to chase the trend. If the market went up yesterday, it buys today. The paper shows that in a market that naturally "reverses" (goes up, then down), this habit creates a specific type of error.
The Correction: The paper provides a "correction factor." It's like adding a footnote to your review: "The dance looked bad, but the music was tricky today, so we adjust the score."
- If the AI is biased toward high-cost decisions, the raw math might make it look worse than it is.
- If the AI is biased toward low-cost decisions, the raw math might make it look better than it is.
- The paper gives a formula to fix this, allowing auditors to get the true score even if the AI has a bad habit.

4. Why This Matters for Auditing

The paper is essentially a toolkit for regulators.

No Secrets Needed: You don't need the AI's source code. You just need a log of what happened (the market data) and what the AI did (the trades).
Fast and Cheap: The math is simple enough to run on a standard computer in real-time. It's not a heavy, complex simulation; it's a straightforward calculation of how much the AI's moves "clashed" with the market.
The "Black Box" is Open: Even though the AI is a "black box" (you can't see inside), this method lets you see the results clearly. It turns the mystery of "Is this AI cheating or just unlucky?" into a statistical test with a confidence interval (a range of certainty).

5. Real-World Test: The Momentum Strategy

The authors tested this on real stock market data from 2016 to 2025.

The Finding: They looked at "Momentum" strategies (AI that buys stocks that went up yesterday).
The Result: The math showed that these strategies consistently "danced against the music." The market tends to reverse quickly (what goes up often comes down the next day), but the momentum AI kept chasing the trend.
The Audit: By using their formula, they could quantify exactly how much money these strategies lost due to this mismatch. They found that while momentum strategies seemed okay in some years, once you applied the "bias correction," the true regret was often higher.

Summary

This paper is a guide for how to audit a mysterious AI manager without ever asking it to explain itself. It says: "Don't look at the brain; look at the dance."

By simply measuring how well the AI's decisions matched the market's movements over time, you can calculate its exact "Regret" (performance loss). If the AI has a bad habit (like chasing trends in a reversing market), the paper gives you the math to correct for it. This allows regulators to say, "We know exactly how much this algorithm cost you, and we know it's not just bad luck—it's a structural flaw in how it dances with the market."

Technical Summary: Evaluating AI Investment Strategies

Problem Statement
The paper addresses the critical challenge of auditing "black-box" algorithmic decision-makers in sequential environments, such as hedge funds, robo-advisors, and execution algorithms. Regulators (e.g., under SR 11-7 and SR-26-2) and platform designers require performance metrics to assess algorithmic strategies without access to internal source code or private agent types. While single-period auditing is possible, extending these methods to multi-period stochastic dynamic programs—where managers reoptimize based on new data each period—remains an open problem. The core difficulty is determining whether the cumulative regret (the gap between realized performance and a theoretical optimal benchmark) can be decomposed into observable quantities (inputs and outputs) alone.

Methodology and Theoretical Framework
The author formulates a general finite-horizon stochastic dynamic program where a policy $\hat{\pi}_t$ maps observed states and cost vectors $c_t$ to decisions $z_t$ . The benchmark is defined as the conditional-mean optimal policy $\pi^*_t$ , which minimizes expected costs given current information.

The central methodological contribution is the derivation of an exact decomposition for cumulative regret. The paper establishes that, under specific conditions, the total regret over $T$ periods equals the sum of per-period covariances between the cost vector and the policy's decision, plus a bias correction term:
$\text{Regret}^{(T)}(\Pi) = \sum_{t=1}^T \text{Cov}(c_t, \hat{\pi}_t(c_t)) + \sum_{t=1}^T \bar{c}_t^\top b_t$
where $b_t = E[\hat{\pi}_t(c_t)] - \pi^*_t$ represents the per-period policy bias.

The analysis proceeds through several theoretical extensions:

i.i.d. Costs: Under independent and identically distributed costs and mean-unbiased Markov policies (where $E[\hat{\pi}_t] = \pi^*_t$ ), the bias term vanishes, yielding an exact identity where regret equals the sum of covariances.
Non-Stationarity: For time-varying cost distributions, the paper derives a closed-form bias correction bounded by the cross-period covariance of cost increments.
Infinite Horizon: A discounted-horizon analog is established, showing that discounted regret scales linearly with the effective horizon $1/(1-\gamma)$ .
AR(1) Processes: For mean-reverting execution costs (modeled as AR(1)), the paper derives a specific auto-covariance correction term, proving that raw covariance sums can overestimate or underestimate true regret depending on the sign of the autocorrelation.
Bellman Recursion: A Bellman-style recursion for the covariance regret functional is derived, structurally analogous to the Q-function in reinforcement learning. This allows for the estimation of multi-period regret using standard dynamic programming algorithms without solving the original optimization problem.

Key Contributions
The paper makes eight specific contributions:

Exact Decomposition: Proving that total expected regret decomposes exactly into the sum of per-period input-output covariances under i.i.d. costs and mean-unbiased policies.
Bias Characterization: Deriving closed-form bias corrections for non-stationary costs and time-varying policies, specifically quantifying the bias in momentum strategies as $O(|\hat{\rho}|T\sigma_c^2)$ .
Algorithmic Auditing Tool: Proposing a trajectory covariance estimator that is consistent, asymptotically normal with HAC (Heteroskedasticity and Autocorrelation Consistent) variance, and computable in $O(T \cdot nd)$ time. This enables model-free auditing using only observable sequences of costs and allocations.
Financial Applications: Applying the theory to rolling-window portfolio rebalancing (showing estimation-error bias is $O(d/w)$ ) and mean-reverting execution costs (providing a closed-form AR(1) correction).
Reinforcement Learning Connection: Demonstrating that standard RL infrastructure (e.g., fitted-Q) can be repurposed to estimate multi-period regret via the derived Bellman recursion.

Empirical Results
The theory is validated using CRSP daily return data from January 2016 to December 2025.

AR(1) Coefficients: Daily AR(1) coefficients were persistently negative ( $\hat{\rho} \in [-0.144, -0.004]$ ) across all years and NBER regimes, indicating that short-horizon reversal dominated momentum.
Strategy Performance: Reversion strategies outperformed momentum strategies in seven of ten years.
Bias Correction: The bias-corrected reversion strategy provided the best risk-adjusted performance in 2019 and 2022, years where raw reversion strategies failed. The bias correction acts as "insurance," incurring a slight cost in favorable years but delivering significant gains when momentum dynamics are distorted.
Robustness: The paper identifies five confounds (size effects, bid-ask bounce, volatility clustering, cross-sectional vs. time-series momentum, and liquidity provision) that likely bias the estimated $\hat{\rho}$ upward. Consequently, the calculated regret figures are conservative upper bounds on momentum-attributable losses.

Significance and Claims
The paper claims to provide a "tractable solution" for regulatory oversight by reframing algorithmic auditing as a statistical testing problem.

For Regulators: Compliance officers can construct confidence intervals for cumulative regret using only observable inputs and outputs, satisfying mandates like SR 11-7 without requiring access to proprietary optimization engines.
For Mechanism Design: The identity allows principals to bound an agent's regret without access to private types, connecting black-box auditing to covariance estimation problems in game theory.
For Policy Improvement: The Bellman recursion implies that pointwise covariance reduction is a sufficient condition for multi-period regret improvement, offering a computationally tractable alternative to full strategy enumeration.

The author maintains a modest stance regarding the empirical magnitude of regret, noting that while the theoretical decomposition holds exactly under stated assumptions, the empirical estimates serve as upper bounds due to unobserved microstructure noise and factor exposures. The primary value proposition is the establishment of a rigorous, model-free metric for performance attribution in sequential decision systems.