Pith. sign in

REVIEW 3 major objections 4 minor 34 references

When should one stop the most exciting game? Sequential Inference for win-martingales

T0 review · 3 major / 4 minor · reviewed 2026-08-16 · deepseek-v4-flash

Pith's one-line read For a broad class of win-martingales, the optimal decision time is the first exit from a symmetric interval around 1/2.

desk verdict A genuine free-boundary theory for a broad class of win-martingales under explicit assumptions; the advertised examples need verification, but the core proof is coherent. read the letter →

arxiv 2608.12291 v1 pith:LWVSCNM5 submitted 2026-08-12 math.PR

classification math.PR MSC 60G4062L1562L1060G3535R35
keywords sequentialinferencewin-martingalespredictionmarketsoptimalstoppingfree-boundaryproblemsnonlinearintegralequationssmoothfittimechange
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Prediction markets, sports games, and elections all pose the same practical question: once beliefs evolve according to a win-martingale, when should you stop watching and declare an outcome? This paper claims that, for a large class of such martingales with volatility that splits into a time factor and a state factor, the optimal rule is always to stop when the posterior leaves a symmetric interval around 1/2. The value function is shown to be once continuously differentiable, the stopping boundary is shown to be smooth before the horizon, and the boundary is uniquely characterized by a nonlinear integral equation. If correct, this gives a common solution theory for the Aldous, Bass, and binary sequential-inference martingales, and it works even when the running cost is nonmonotone, producing boundaries that can shrink, expand, or oscillate.

What carries the argument

The load-bearing object is the two-sided free boundary $b_T(t)$ that defines the continuation interval. It is produced by combining a deterministic time change that absorbs time-dependent volatility into the running cost, the Lamperti transform (a coordinate change making the diffusion coefficient equal to one), probabilistic derivative representations of the value function, the smooth-fit principle, and a change-of-variable formula with local time on curves, which turns the optimal stopping problem into the integral equation (4.2).

What would settle it

Solve the nonlinear integral equation (4.2) for the paper's sinusoidal-cost Aldous example and compare the first-exit value with a direct PDE or numerical solution of the optimal stopping problem; if the values differ, the characterization is wrong. A sharper test is to take the same costs with a martingale that can be absorbed at 0 or 1 before the horizon: if the boundary is no longer $C^1$ or the integral equation loses uniqueness, the Feller condition (3.4) is doing essential work.

Watch

Extended reading notes

Core claim

Under the paper's structural assumptions, the Bayes risk of a stop-and-declare problem for an autonomous win-martingale reduces to an optimal stopping problem whose continuation region at every time is the interval $(1-b_T(t), b_T(t))$. The first exit from this region is an optimal stopping time, the value function satisfies smooth fit and is $C^1$ globally, the boundary $b_T$ is $C^1$ on $[0,T)$, and $b_T$ is the unique solution of the nonlinear integral equation (4.2). The same structure holds for finite and infinite horizons without discounting and requires no monotonicity of the running cost.

Load-bearing premise

The load-bearing assumption is that the win probability never reaches 0 or 1 before the horizon, so absorption happens only in the limit; if early absorption is allowed, the theory's regularity results and integral equation would need to be rebuilt.

Editorial extensions

If this is right

  • For every model in the class, the optimal decision rule is known once $b_T$ is computed, so no general search over stopping times is needed.
  • The nonlinear integral equation gives a constructive numerical route to the boundary, for example by Picard iteration, with uniqueness guaranteeing that the computed boundary is the right one.
  • Monotone running costs in calendar time need not imply monotone boundaries; the paper's sinusoidal-cost example has a boundary with no limit as the horizon is approached.
  • The value functions provide a decision-theoretic ordering of games: under equal costs, the illustrated Aldous martingale is more costly to call optimally than the Bass martingale.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • One could test the same machinery on asymmetric terminal costs by dropping the symmetry assumption; the paper indicates this needs extra bookkeeping, and the boundary would become two independent curves rather than a symmetric pair.
  • The absence of a boundary limit near the horizon suggests that finite-horizon numerical approximations of infinite-horizon problems may converge non-uniformly in time, so numerical schemes should track oscillatory boundaries carefully.
  • A natural extension would allow early absorption at 0 or 1 before the deadline; the paper explicitly leaves this open, and one can conjecture the continuation region remains an interval but requires an additional exogenous termination boundary and rederived regularity.
  • The difference in value functions between two win-martingales under the same cost could serve as a decision-theoretic measure of excitement, complementing purely probabilistic distances between martingale laws.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper studies a sequential decision problem in which a decision maker observes a binary-outcome win-martingale, modeled as a [0,1]-valued martingale posterior, and chooses a stopping time and terminal declaration under a running cost and a terminal loss. After reducing the Bayes risk to an optimal stopping problem for the posterior, the authors apply a deterministic time change to convert separable volatility into a time-homogeneous diffusion with a time-inhomogeneous running cost. Under structural Assumptions A and B, they characterize the continuation region as a symmetric interval with a free boundary, prove C^1 regularity of the value function and C^1 regularity of the boundary on [0,T), derive a nonlinear integral equation uniquely characterizing the boundary, and treat finite and infinite horizons without imposing monotonicity of the running cost. The claimed applications include the Aldous, Bass, and binary sequential-inference martingales, with an example where the optimal boundary oscillates and has no limit in calendar time.

Significance. If the main theorems hold under their stated hypotheses, the paper would provide a substantial and unified free-boundary theory for a broad class of posterior martingale stopping problems, extending earlier Bayesian sequential testing results and accommodating nonmonotone running costs and nonmonotone boundaries. The proof strategy is generally careful and self-contained: the reduction to a free-boundary problem, the use of De Angelis–Peskir and De Angelis–Lamberton regularity results, and the derivation of the integral equation are coherent. The paper also makes a useful decision-theoretic connection between martingale-distance models of "exciting games" and concrete sequential prediction. The central theorems are conditional on Assumption B, however, and the advertised applicability to the canonical examples is asserted rather than demonstrated, which is a significant scope gap. The paper is honest about its main limitation, the exclusion of finite-time absorption at 0 and 1, but the internal verification of the assumptions for the motivating examples is missing.

major comments (3)
  1. [§8 and §3.2, Assumption B.(v)] The paper claims that the Aldous, Bass, and binary sequential-inference martingales satisfy Assumption B, but the verification of Assumption B.(v) for the Bass martingale is not supplied. The text in §3.2 states that linear growth of the Lamperti drift μ(l) suffices via a Beneš-type argument and that this indeed holds for the Bass win-martingale, but μ is never displayed and the growth or Lipschitz check is not carried out. Since Assumption B.(v) is used in Lemma 6.2, in the derivative representations (6.6) and (6.21), and in the smooth-fit arguments of Propositions 6.22 and 6.23, the reader cannot currently verify that the Bass example is inside the hypotheses of Theorems 4.1–4.5.
  2. [§8.5, Assumption B.(ii)] Example 8.5 uses the cross-entropy terminal cost g_CE with the Aldous martingale and states that Assumption B holds, but Assumption B.(ii) is not checked. For g_CE one has Lg = -σ^2/(2x(1-x)), so the required sign change of ∂_x Lg on (0,x0) and (x0,1) is not immediate for σ_A or σ_B and is not demonstrated in the paper. Because B.(ii) is load-bearing for the interval structure of the continuation region in Proposition 6.13 and for the strict positivity used in Lemma 6.15, the oscillatory-boundary example of §8.5 is not rigorously supported as stated.
  3. [§8.4, Examples 8.4 and 8.5] The text says "It can be verified that all of the examples satisfy Assumption B" and "It can easily be verified that Assumption B holds," but no verification or reference to supplementary material is provided. This is a load-bearing point because the abstract and introduction claim a common solution theory for the Aldous, Bass, and binary sequential-inference martingales, while the main theorems are conditional on Assumption B. I would ask the authors to either provide explicit checks of B.(ii), B.(iii), and B.(v) for each claimed example, or state clearly which examples are only heuristic.
minor comments (4)
  1. [§3.2, Assumption B.(v)] The definition of the stochastic flow derivative ∂_x X^x_t is deferred to equation (3.14), which appears after Assumption B.(v) is stated; please add a forward cross-reference at the assumption.
  2. [Eq. (4.2), infinite-horizon case] In Theorem 4.5, for T=∞ the term g(X^x_{T-t}) should be explicitly written as g(X^x_∞), since T-t is infinite; the current notation is formally ambiguous even though the subsequent sentence clarifies it.
  3. [Proposition 6.23] In the proof of Proposition 6.23, the interval notation J=[a,a] appears degenerate because the underline and overline on the endpoints are lost; please typeset the endpoints as a_* and a^* or as a and overline{a} to make the interval nondegenerate.
  4. [§8.5 and Figure 3] The captions in the examples refer to dashed curves as γ and 1-γ, but in the printed text the dashed and solid styles are not always clearly distinguished; please ensure all curves in Figures 1–3 are labeled consistently.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular reduction: the free-boundary derivation is self-contained and the examples are scope-conditional, not circular.

full rationale

The core derivation is not circular. The paper reduces the Bayes-risk problem to an optimal stopping problem (Section 2), performs a deterministic time change, and then proves the free-boundary structure, smooth fit, boundary regularity, and the integral equation under Assumptions A and B. The main regularity results are obtained by applying external results [10], [11], and the change-of-variable formula [27], not by citing a fitted parameter or reusing the conclusion. The homogeneous infinite-horizon case in Section 5 extends prior work [9] and invokes a named external geometric theorem, Bisztriczky's theorem, cited to [9]; this is a legitimate external mathematical fact and not a self-citation chain that supplies the target conclusion. Theorem 4.5's integral equation is derived from the optimal stopping value via the change-of-variable formula and its uniqueness is established by a comparison argument against the already-constructed optimal value; this is a standard characterization proof, not a tautology. The paper explicitly states limitations: the Conclusion acknowledges that the theory excludes processes absorbed at 0 or 1 in finite time, such as Wright-Fisher and absorbed Brownian motion. The claimed coverage of the Aldous, Bass, and binary sequential-inference examples rests on assertions that Assumption B is satisfied—e.g., Section 3.2 states 'This indeed holds for the Bass win-martingale' without displaying the Lamperti-drift verification, and Section 8 says 'It can easily be verified that Assumption B holds.' These are unverified scope checks and correctness risks, not circular reductions: no equation is being reused as its own input, and no fitted quantity is renamed as a prediction. Therefore no significant circularity is present.

Assumptions & free parameters 0 free parameters · 6 assumptions · 0 invented entities

The statements are fully conditional on the listed structural assumptions. No free parameters are fitted to data; the constants k, k1, k2, alpha, and c0 in Section 8 are illustrative example inputs, not fitted quantities. The framework introduces no new postulated entities such as particles, forces, or dimensions.

assumptions (6)
  • standard math Standard Itô calculus, Dynkin's formula, the change-of-variable formula with local time on curves, Girsanov's theorem, and maximum principles are valid.
    Used throughout Sections 3 through 7; these are well-established external results, not proved in the paper.
  • domain assumption The belief process has separable volatility: dPi_s = rho(s) sigma(Pi_s) dW_s, with rho continuous and positive on [0,S) and sigma such that a strong [0,1]-valued solution exists.
    Section 2, equation (2.4); this is the class over which the deterministic time change and the free-boundary theory are developed.
  • domain assumption The time-changed diffusion does not hit 0 or 1 in finite time: sigma > 0 on (0,1), sigma = 0 only at endpoints, and the Feller condition (3.4) holds.
    Assumption A.(i); used in Proposition 3.1 and throughout. The Conclusion explicitly acknowledges this excludes early-termination models such as Wright-Fisher and absorbed Brownian motion.
  • domain assumption The running cost c is C^1 and satisfies |c'(t)| / c(t) <= K with c > 0.
    Assumption A.(ii); the exponential growth bound (3.6) is used for continuity, admissibility, and derivative estimates.
  • domain assumption The terminal cost g is continuous on [0,1], C^2 in the interior, concave, zero exactly at 0 and 1, symmetric, and Lg is strictly unimodal.
    Assumptions A.(iii)-(iv) and B.(i)-(ii), B.(iv); these shape conditions produce the interval continuation region (1-b,b) and are needed in the uniqueness proof of Theorem 4.5.
  • domain assumption The stochastic flow derivative is a true martingale and either g' is bounded or c is bounded below.
    Assumption B.(iii), B.(v); the derivative representations in Proposition 6.4 and Corollary 6.5 and the uniform bounds in the regularity analysis require these conditions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of When should one stop the most exciting game? Sequential Inference for win-martingales." pith.science (2026). https://pith.science/paper/LWVSCNM5

@misc{pith2026260812291,
  author       = {Pith},
  title        = {Pith review of: When should one stop the most exciting game? Sequential Inference for win-martingales},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LWVSCNM5}},
  note         = {Machine review of arXiv:2608.12291}
}
abstract

Prediction markets have become a prominent way of aggregating beliefs about binary future events, and their price processes are often interpreted as evolving win probabilities, or ``win-martingales.'' Motivated by this perspective and recent work on Aldous' ``most exciting game,'' we study when a decision maker should stop observing a win-martingale and make a decision about the outcome. In particular, we allow the true outcome to be revealed at a fixed finite horizon, as in a sports game or election. Under a general terminal loss and running cost, we reduce the Bayes risk to an optimal stopping problem for the win probability process. When the win-martingale is a diffusion and its volatility separates into a deterministic time factor and a state-dependent factor, a deterministic time change transforms the problem into one for a time-homogeneous diffusion with a generally time-inhomogeneous running cost. Under explicit structural assumptions, we obtain a complete free-boundary characterization of the solution to the stopping problem in both finite and infinite horizons without discounting. We prove smooth-fit and $C^1$ regularity of the value function, $C^1$ regularity of the optimal stopping boundaries before the horizon, and derive a nonlinear integral equation that characterizes the boundaries uniquely. Taken together, these results yield a common decision-theoretic framework and solution theory for a broad class of posterior dynamics that includes the Aldous, Bass, and binary sequential-inference martingales as special cases. Our analysis requires no temporal monotonicity of the running cost and therefore accommodates highly nonmonotone stopping boundaries. In particular, we exhibit an example in which the optimal boundary has no limit as the calendar-time horizon is approached.

Figures

Figures reproduced from arXiv: 2608.12291 by the authors.

Figure 1
Figure 1. Numerical plot of the value function x 7→ V (x) for the Aldous (dotted) and Bass (dashed) win-martingales with L 2–loss and constant transformed running cost c0 = 0.075. The vertical lines indicate the optimal stopping boundaries from Theorem 5.2, and the solid curve is x 7→ gL2 (x). If T = ∞ and c is decreasing, the reverse comparison gives W∞(t2, x) − W∞(t1, x) ≤ E "Z τ ∗∞(t1,x) 0 [PITH_FULL_IMAGE:figures/full_fi… view at source ↗
Figure 2
Figure 2. Numerical plot of the optimal stopping boundaries for the classical se￾quential testing win-martingale with L 2–loss and S = 1. The three panels corre￾spond to transformed running costs that are increasing, nonmonotone, and decreas￾ing, respectively. Solid curves are the optimal stopping boundaries while the dashed lines are γ and 1 − γ. Example 8.5. Finally, we consider an oscillatory transformed running cost for t… view at source ↗
Figure 3
Figure 3. Numerical plot of the optimal stopping boundary for the Aldous win￾martingale with cross-entropy loss and sinusoidal transformed running cost c(t) = k1 + k2 sin(2πt). The dashed curves show γ and 1 − γ. Proposition 8.6. Consider an optimal stopping problem (2.3) such that the time transformation obeys A(s) → ∞ as s → S. Assume, furthermore, that the transformed optimal stopping problem satisfies Assumption B and the… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 33 canonical work pages

  1. [9]

    Campbell and Y

    S. Campbell and Y. Zhang. A bayesian sequential soft classification problem for a brownian motion’s drift.arXiv preprint arXiv:2501.11314, 2025

  2. [10]

    De Angelis and D

    T. De Angelis and D. Lamberton. A probabilistic approach to continuous differentiability of optimal stopping boundaries.Electronic Journal of Probability, 31:1–46, 2026

  3. [11]

    De Angelis and G

    T. De Angelis and G. Peskir. GlobalC 1 regularity of the value function in optimal stopping problems.Ann. Appl. Probab., 30(3):1007–1031, 2020

  4. [27]

    G. Peskir. A change-of-variable formula with local time on curves.J. Theoret. Probab., 18(3):499–535, 2005

  5. [1]

    D. Aldous. What is the max-entropy win-probability martingale?https://www.stat.berkeley.edu/ ~aldous/ Research/OP/max_ent_mg.html

  6. [2]

    Backhoff, G

    J. Backhoff, G. Pammer, and W. Schachermayer. TheL 2 gradient flow of the Bass functional in martingale optimal transport.Math. Financ. Econ., 19(4):665–686, 2025

  7. [3]

    Exciting games and Monge-Amp\`ere equations

    J. Backhoff, Z. Wang, and X. Zhang. Exciting games and Monge-Amp` ere equations.arXiv preprint arXiv:2412.01995, 2024

  8. [4]

    Backhoff-Veraguas and M

    J. Backhoff-Veraguas and M. Beiglb¨ ock. The most exciting game.Electronic Communications in Probability, 29:1–12, 2024

Show all 34 references
  1. [5]

    Backhoff-Veraguas, W

    J. Backhoff-Veraguas, W. Schachermayer, and B. Tschiderer. The Bass functional of martingale transport.Ann. Appl. Probab., 35(6):4282–4301, 2025

  2. [6]

    Backhoff-Veraguas and X

    J. Backhoff-Veraguas and X. Zhang. Specific Wasserstein divergence between continuous martingales.Mathe- matics of Operations Research, 2026

  3. [7]

    Brigo and F

    D. Brigo and F. Vrins. On the boundaries, asymptotic law and Bernoulli-Doob representation of homogeneous bounded martingales.arXiv preprint arXiv:2607.12365, 2026

  4. [8]

    Campbell and K

    S. Campbell and K. K. Engelund. Embedding martingale diffusions as binary posteriors in sequential inference. arXiv preprint arXiv:2607.20373, 2026

  5. [12]

    De Angelis and G

    T. De Angelis and G. Stabile. On Lipschitz continuous optimal stopping boundaries.SIAM J. Control Optim., 57(1):402–436, 2019

  6. [13]

    Ekstr¨ om, I

    E. Ekstr¨ om, I. Karatzas, and J. Vaicenavicius. Bayesian sequential least-squares estimation for the drift of a Wiener process.Stochastic Process. Appl., 145:335–352, 2022

  7. [14]

    Ekstr¨ om and J

    E. Ekstr¨ om and J. Vaicenavicius. Bayesian sequential testing of the drift of a Brownian motion.ESAIM Probab. Stat., 19:626–648, 2015

  8. [15]

    Ekstr¨ om and Y

    E. Ekstr¨ om and Y. Wang. Multi-dimensional sequential testing and detection.Stochastics, 94(5):789–806, 2022

  9. [16]

    Gantert.Einige grosse Abweichungen der Brownschen Bewegung

    N. Gantert.Einige grosse Abweichungen der Brownschen Bewegung. PhD thesis, Rheinische Friedrich-Wilhelms- Universit¨ at Bonn, 1991

  10. [17]

    P. V. Gapeev and G. Peskir. The Wiener sequential testing problem with finite horizon.Stoch. Stoch. Rep., 76(1):59–75, 2004

  11. [18]

    P. V. Gapeev and A. N. Shiryaev. Bayesian quickest detection problems for some diffusion processes.Adv. in Appl. Probab., 45(1):164–185, 2013

  12. [19]

    P. V. Gapeev and Y. I. Stoev. On the sequential testing and quickest change-point detection problems for Gaussian processes.Stochastics, 89(8):1143–1165, 2017

  13. [20]

    G. Guo, S. D. Howison, D. Possama¨ ı, and C. Reisinger. Randomness and early termination: what makes a game exciting?Probab. Theory Related Fields, 192(1–2):135–162, 2025

  14. [21]

    Ikeda and S

    N. Ikeda and S. Watanabe. A comparison theorem for solutions of stochastic differential equations and its applications.Osaka Math. J., 14(3):619–633, 1977

  15. [22]

    Jaillet, D

    P. Jaillet, D. Lamberton, and B. Lapeyre. Variational inequalities and the pricing of American options.Acta Appl. Math., 21(3):263–289, 1990

  16. [23]

    Karatzas and S

    I. Karatzas and S. E. Shreve.Brownian motion and stochastic calculus, volume 113 ofGraduate Texts in Math- ematics. Springer-Verlag, New York, second edition, 1991

  17. [24]

    Karatzas and S

    I. Karatzas and S. E. Shreve.Methods of mathematical finance, volume 39 ofApplications of Mathematics (New York). Springer-Verlag, New York, 1998

  18. [25]

    Klebaner and R

    F. Klebaner and R. Liptser. When a stochastic exponential is a true martingale. Extension of the Beneˇ s method. Theory Probab. Appl., 58(1):38–62, 2014

  19. [26]

    D. I. Lisovskii. Bayesian sequential testing problem for a Brownian bridge.Theory of Probability & Its Applica- tions, 63(4):556–579, 2019

  20. [28]

    Peskir and A

    G. Peskir and A. Shiryaev.Optimal stopping and free-boundary problems. Lectures in Mathematics ETH Z¨ urich. Birkh¨ auser Verlag, Basel, 2006

  21. [29]

    M. H. Protter and H. F. Weinberger.Maximum principles in differential equations. Springer-Verlag, New York,

  22. [30]

    P. E. Protter.Stochastic integration and differential equations, volume 21 ofStochastic Modelling and Applied Probability. Springer-Verlag, Berlin, second edition, 2005. Corrected third printing

  23. [31]

    R. T. Rockafellar and R. J.-B. Wets.Variational analysis. Springer-Verlag, Berlin, 1998

  24. [32]

    A. N. Shiryaev. Two problems of sequential analysis.Cybernetics, 3(2):63–69, 1967

  25. [33]

    Wolfers and E

    J. Wolfers and E. Zitzewitz. Prediction markets.Journal of Economic Perspectives, 18(2):107–126, 2004

  26. [1984]

    52 WHEN SHOULD ONE STOP THE MOST EXCITING GAME?

    Corrected reprint of the 1967 original. 52 WHEN SHOULD ONE STOP THE MOST EXCITING GAME?

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.