pith. sign in

arxiv: 2606.08791 · v1 · pith:MRXTSSGDnew · submitted 2026-06-07 · 💰 econ.EM · cs.AI· q-fin.PM· q-fin.RM· q-fin.ST

Evaluating AI Investment Strategies

Irene Aldridge This is my paper

Pith reviewed 2026-06-27 17:27 UTC · model grok-4.3

classification 💰 econ.EM cs.AIq-fin.PMq-fin.RMq-fin.ST
keywords regret decompositionalgorithmic auditingdynamic programmingcovarianceblack-box evaluationstochastic policiesreinforcement learningmechanism design
0
0 comments X

The pith

Cumulative regret of a dynamic policy equals the sum of per-period cost-decision covariances under i.i.d. costs and mean-unbiased Markov policies.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes an exact decomposition for the regret incurred by sequential decision policies observed only through their inputs and outputs. It proves that total regret across periods equals the sum of the covariances between the realized cost vector and the action chosen in each period. The identity holds precisely when costs arrive independently and identically distributed and the policy is a mean-unbiased Markov process; closed-form corrections are supplied for departures from these conditions. The result extends an earlier single-period identity to the full multi-period stochastic dynamic programming setting and supplies a Bellman recursion that ties the covariance functional to standard reinforcement-learning methods. The decomposition yields a tractable, model-free estimator for auditing black-box algorithmic systems in mechanism design, repeated games, and procurement settings.

Core claim

Under i.i.d. costs and mean-unbiased Markov policies the cumulative regret of a dynamic policy equals the sum of per-period covariances between the cost vector and the policy's decision. The identity extends the single-period case to stochastic dynamic programming, admits closed-form bias corrections for non-stationary and time-varying environments, and possesses a discounted-horizon analog. A Bellman recursion for the covariance regret functional connects the decomposition to reinforcement-learning algorithms, while the associated trajectory estimator is consistent, asymptotically normal with HAC variance, and computable in O(T · nd) time.

What carries the argument

The covariance-regret decomposition, which equates cumulative regret exactly to the sum of per-period covariances under the i.i.d. and mean-unbiased conditions.

If this is right

  • The decomposition supplies a welfare-based audit metric for platform mechanisms that does not require the agent's private type.
  • Covariance reduction is a sufficient condition for policy improvement in repeated games.
  • Bias corrections quantify welfare loss from strategic misreporting in procurement and ad auctions.
  • The estimator remains consistent and asymptotically normal with HAC variance for rolling-window policies with bias O(d/w).
  • The Bellman recursion for the covariance functional links the identity to standard reinforcement-learning value iteration.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same observable-inputs-only audit could be applied directly to portfolio rebalancing rules or automated trading systems without inspecting their internal models.
  • If the i.i.d. assumption is relaxed, the bias-corrected versions might still serve as practical lower bounds on regret in slowly varying environments.
  • The O(T · nd) estimator could be embedded in online monitoring dashboards that flag rising covariance between costs and actions in real time.
  • The decomposition might be combined with existing regret bounds from online learning to produce parameter-free performance guarantees for any Markov policy.

Load-bearing premise

Costs must be i.i.d. and the policy must be mean-unbiased and Markovian; without these the exact equality fails.

What would settle it

A counter-example in which a mean-unbiased Markov policy faces i.i.d. costs yet the observed cumulative regret differs from the summed per-period covariances.

read the original abstract

We study the problem of auditing a black-box algorithmic decision-maker from observable inputs and outputs alone. Our main result is an exact decomposition: under precisely characterized conditions, the cumulative \emph{regret} of a dynamic policy equals the sum of per-period covariances between the cost vector and the policy's decision. This extends the single-period identity of Aldridge~(2026) to the full multi-period setting of stochastic dynamic programming. We prove the identity holds exactly under i.i.d. costs and mean-unbiased Markov policies, derive closed-form bias corrections for non-stationary and time-varying cases, and establish the discounted-horizon analog. A Bellman recursion for the covariance regret functional connects the result to standard reinforcement learning algorithms; for rolling-window policies, the estimation-error bias is $O(d/w)$. The decomposition has direct implications for algorithmic auditing in strategic environments: in platform mechanism design, it provides a welfare-based audit metric without access to the agent's private type; in repeated games, covariance reduction is a sufficient condition for policy improvement; in procurement and ad auctions, the bias correction quantifies welfare loss from strategic misreporting. The associated trajectory estimator is consistent, asymptotically normal with HAC variance, and computable in $O(T \cdot nd)$ time. This makes the proposed approach a tractable, model-free audit tool for platform mechanisms, algorithmic portfolio strategies, and any sequential decision system subject to external performance review.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 0 minor

Summary. The paper claims an exact decomposition in which, under i.i.d. costs and mean-unbiased Markov policies, the cumulative regret of a dynamic policy equals the sum of per-period covariances between the cost vector and the policy's decision. This extends the single-period identity of Aldridge (2026) to the multi-period stochastic dynamic programming setting. The manuscript derives closed-form bias corrections for non-stationary and time-varying cases, establishes the discounted-horizon analog, provides a Bellman recursion for the covariance regret functional, and shows that the associated trajectory estimator is consistent, asymptotically normal with HAC variance, and computable in O(T nd) time. Applications to algorithmic auditing in platform mechanisms, repeated games, and procurement/ad auctions are discussed.

Significance. If the central identity holds under the stated conditions, the result supplies a model-free, welfare-based audit metric for black-box sequential decision systems that does not require access to private types. The O(d/w) estimation-error bound for rolling-window policies, the explicit bias corrections, and the link to standard RL via the Bellman recursion are concrete strengths that would make the approach immediately usable for evaluating dynamic AI investment strategies and mechanism performance.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive assessment of the manuscript, the accurate summary of its contributions, and the recommendation for minor revision. No specific major comments were raised in the report.

Circularity Check

0 steps flagged

Minor self-citation for single-period base; multi-period identity derived independently

full rationale

The paper's central claim is an exact identity equating cumulative regret to the sum of per-period covariances, which it states is proved under explicitly listed conditions (i.i.d. costs and mean-unbiased Markov policies) and extended via a Bellman recursion for the covariance functional. The reference to Aldridge (2026) is limited to the single-period starting point; the multi-period decomposition, bias corrections, and RL connection are presented as new content derived in this manuscript. No equation or step is shown to reduce by construction to the prior work, and the result is framed as falsifiable under the stated assumptions rather than forced by self-citation. This qualifies as a normal, non-load-bearing self-reference.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The claim depends on these domain assumptions for the exact equality; no free parameters are explicitly fitted in the abstract description.

axioms (2)
  • domain assumption Costs are i.i.d.
    Required for the exact identity to hold in the multi-period setting.
  • domain assumption Policies are mean-unbiased Markov policies
    Necessary condition stated for the decomposition.

pith-pipeline@v0.9.1-grok · 5785 in / 1402 out tokens · 30202 ms · 2026-06-27T17:27:10.458531+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

59 extracted references · 2 canonical work pages · 1 internal anchor

  1. [1]

    Regret Equals Covariance: A Closed-Form Characterization for Stochastic Optimization

    Aldridge, Irene , title =. arXiv preprint arXiv:2605.14019 [econ.EM] , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [2]

    , title =

    Chen, Hao and Didisheim, Antoine and Somoza, Luis A. , title =. NBER Working Paper No.\ 34965 , year =

  3. [3]

    Bellman, Richard , title =

  4. [4]

    , title =

    Bertsekas, Dimitri P. , title =

  5. [5]

    McLeish, D. L. , title =. Annals of Probability , year =

  6. [6]

    and West, Kenneth D

    Newey, Whitney K. and West, Kenneth D. , title =. Econometrica , year =

  7. [7]

    , title =

    Puterman, Martin L. , title =

  8. [8]

    Journal of Finance , year =

    Markowitz, Harry , title =. Journal of Finance , year =

  9. [9]

    Financial Analysts Journal , year =

    Black, Fischer and Litterman, Robert , title =. Financial Analysts Journal , year =

  10. [10]

    and Grigas, Paul , title =

    Elmachtoub, Adam N. and Grigas, Paul , title =. Management Science , volume =

  11. [11]

    http://www.nber.org/papers/w35153

    Carlin, Bruce I and Israelsen, Ryan D and Wazzan, Christopher F. AI Managed Household Portfolios: A Preliminary Report. 2026. doi:10.3386/w35153 , URL = "http://www.nber.org/papers/w35153", abstract =

    work page doi:10.3386/w35153 2026

  12. [12]

    Contributions to the Theory of Games , volume =

    Hannan, James , title =. Contributions to the Theory of Games , volume =. 1957 , publisher =

    1957

  13. [13]

    Proceedings of the 20th International Conference on Machine Learning (ICML) , pages =

    Zinkevich, Martin , title =. Proceedings of the 20th International Conference on Machine Learning (ICML) , pages =

  14. [14]

    2016 , volume =

    Hazan, Elad , title =. 2016 , volume =

    2016

  15. [15]

    2012 , volume =

    Shalev-Shwartz, Shai , title =. 2012 , volume =

    2012

  16. [16]

    Working Paper , year =

    Aldridge, Irene , title =. Working Paper , year =

  17. [17]

    and Vohra, Rakesh V

    Foster, Dean P. and Vohra, Rakesh V. , title =. Games and Economic Behavior , volume =

  18. [18]

    Econometrica , volume =

    Hart, Sergiu and Mas-Colell, Andreu , title =. Econometrica , volume =

  19. [19]

    Roughgarden, Tim , title =

  20. [20]

    , title =

    Myerson, Roger B. , title =. Journal of Mathematical Economics , volume =

  21. [21]

    Laffont, Jean-Jacques and Tirole, Jean , title =

  22. [22]

    and Yilankaya, Okan , title =

    Dekel, Eddie and Ely, Jeffrey C. and Yilankaya, Okan , title =. Review of Economic Studies , volume =

  23. [23]

    American Economic Review , volume =

    Edelman, Benjamin and Ostrovsky, Michael and Schwarz, Michael , title =. American Economic Review , volume =

  24. [24]

    , title =

    Varian, Hal R. , title =. International Journal of Industrial Organization , volume =

  25. [25]

    Econometrica , volume =

    Gibbard, Allan , title =. Econometrica , volume =

  26. [26]

    , title =

    Satterthwaite, Mark A. , title =. Journal of Economic Theory , volume =

  27. [27]

    Proceedings of the 35th International Conference on Machine Learning (ICML) , pages =

    Kearns, Michael and Neel, Seth and Roth, Aaron and Wu, Zhiwei Steven , title =. Proceedings of the 35th International Conference on Machine Learning (ICML) , pages =

  28. [28]

    Working Paper , year =

    Roth, Aaron , title =. Working Paper , year =

  29. [29]

    Econometrica , volume =

    Hansen, Lars Peter , title =. Econometrica , volume =

  30. [30]

    Andrews, Donald W. K. , title =. Econometrica , volume =

  31. [31]

    Watkins, Christopher J. C. H. and Dayan, Peter , title =. Machine Learning , volume =

  32. [32]

    and Barto, Andrew G

    Sutton, Richard S. and Barto, Andrew G. , title =

  33. [33]

    , title =

    Howard, Ronald A. , title =

  34. [34]

    Journal of Machine Learning Research , volume =

    Jaksch, Thomas and Ortner, Ronald and Auer, Peter , title =. Journal of Machine Learning Research , volume =

  35. [35]

    Minimax Regret Bounds for Reinforcement Learning , booktitle =

    Azar, Mohammad Gheshlaghi and Osband, Ian and Munos, R. Minimax Regret Bounds for Reinforcement Learning , booktitle =

  36. [36]

    Journal of Machine Learning Research , volume =

    Ernst, Damien and Geurts, Pierre and Wehenkel, Louis , title =. Journal of Machine Learning Research , volume =

  37. [37]

    Proceedings of the European Conference on Machine Learning (ECML) , pages =

    Riedmiller, Martin , title =. Proceedings of the European Conference on Machine Learning (ECML) , pages =

  38. [38]

    Review of Financial Studies , volume =

    DeMiguel, Victor and Garlappi, Lorenzo and Uppal, Raman , title =. Review of Financial Studies , volume =

  39. [39]

    Proceedings of the National Academy of Sciences , volume =

    Brodie, Joshua and Daubechies, Ingrid and De Mol, Christine and Giannone, Domenico and Loris, Ignace , title =. Proceedings of the National Academy of Sciences , volume =

  40. [40]

    Journal of Multivariate Analysis , volume =

    Ledoit, Olivier and Wolf, Michael , title =. Journal of Multivariate Analysis , volume =

  41. [41]

    Annals of Statistics , volume =

    Ledoit, Olivier and Wolf, Michael , title =. Annals of Statistics , volume =

  42. [42]

    , title =

    Bai, Zhidong and Silverstein, Jack W. , title =

  43. [43]

    Journal of Risk , volume =

    Almgren, Robert and Chriss, Neil , title =. Journal of Risk , volume =

  44. [44]

    Quantitative Finance , volume =

    Gatheral, Jim , title =. Quantitative Finance , volume =

  45. [45]

    Welford, B. P. , title =. Technometrics , volume =

  46. [46]

    Journal of Finance , volume =

    Jegadeesh, Narasimhan , title =. Journal of Finance , volume =

  47. [47]

    Journal of Finance , volume =

    Roll, Richard , title =. Journal of Finance , volume =

  48. [48]

    , title =

    Lehmann, Bruce N. , title =. Quarterly Journal of Economics , volume =

  49. [49]

    and French, Kenneth R

    Fama, Eugene F. and French, Kenneth R. , title =. Journal of Financial Economics , volume =

  50. [50]

    , title =

    Carhart, Mark M. , title =. Journal of Finance , volume =

  51. [51]

    Journal of Finance , volume =

    Jegadeesh, Narasimhan and Titman, Sheridan , title =. Journal of Finance , volume =

  52. [52]

    and Miller, Merton H

    Grossman, Sanford J. and Miller, Merton H. , title =. Journal of Finance , volume =

  53. [53]

    , title =

    Engle, Robert F. , title =. Econometrica , volume =

  54. [54]

    Review of Financial Studies , volume =

    Chang, Yen-Cheng and Hong, Harrison and Liskovich, Inessa , title =. Review of Financial Studies , volume =

  55. [55]

    Review of Financial Studies , volume =

    Jegadeesh, Narasimhan and Luo, Jiang and Subrahmanyam, Avanidhar and Titman, Sheridan , title =. Review of Financial Studies , volume =

  56. [56]

    and Bhattacharya, Suman , title =

    Anderson, Robert M. and Bhattacharya, Suman , title =. Working Paper, University of California Berkeley , year =

  57. [57]

    Review of Financial Studies , volume =

    Nagel, Stefan , title =. Review of Financial Studies , volume =

  58. [58]

    Journal of Forecasting , year =

    Mamais, Panagiotis , title =. Journal of Forecasting , year =

  59. [59]

    Financial Analysts Journal , volume =

    Dai, Wei and Medhat, Mamdouh and Novy-Marx, Robert and Rizova, Savina , title =. Financial Analysts Journal , volume =