{"id":"eaad8a3a-7f18-4643-b5e9-4fec64d17621","arxiv_id":"2506.08992","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"An informed broker's optimal policy is to conceal the drift until a deterministic critical time, then disclose it fully, with an explicit piecewise control that is C/sqrt(N) optimal for finite trader populations.","lead":"An informed broker facing many uninformed traders should hide the price-drift signal until a critical time and then reveal it all at once. The paper derives this optimal externalization strategy using mean-field game theory and proves it stays nearly optimal when the number of traders is finite.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The bang-bang disclosure theorem rests on an unverified equivalence between the relaxed filtration optimization and the filtration actually generated by the candidate control; nondegeneracy of the disclosure coefficient is not checked.","rationale":"The reader's weakest assumption focuses on the open-loop broker restriction and the small-b contraction, both of which are explicitly acknowledged modeling limitations rather than internal gaps. The unproved Lemma 5.3 is a lesser issue because the estimates in (5.16) and (5.19) make the concentration argument routine. The genuinely load-bearing internal step is the passage from the relaxed filtration problem to the implementable candidate control. The scalar reduction in (4.19) is valid as an upper bound, but the attainability claim is compressed into an 'easy to see' verification of properties 1-4. Whether the candidate \\nu^B generates exactly the two-point filtration is a quantitative nondegeneracy question that the paper does not address. This does not show the theorem is false, but it leaves a real gap in the central claim, so the CONDITIONAL verdict from the reader is appropriate and no change is needed.","tokens_in":27367,"tokens_out":18942,"duration_ms":188576,"concrete_test":"Fix the Section 6 parameters with a concrete small b (for example b = 10^{-4}) and a two-point μ. Compute t_c as the argmin of A. Then solve the linear ODE for Q^B_t under the candidate \\nu^B from (4.20) and compute the total sensitivity c(t) = ∂\\nu^B_t/∂μ for t > t_c. Verify that c(t) ≠ 0 for all t in (t_c, T] and that \\nu^B_t is constant on [0, t_c]. If c(t) has a zero, simulate the traders' filter to identify the sub-σ-field of σ(μ) actually generated; if it is strictly smaller than σ(μ), the attainability step of Theorem 4.3 fails for this parameter set.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 4.3 the broker's problem is relaxed to an optimization over arbitrary filtrations. The objective is reduced to a linear functional of V_t = E[μ_t^2], and the authors infer that the optimal V_t jumps from E[μ]^2 to E[μ^2] at the critical time t_c determined by the minimum of A (equation (4.19) and Theorem 4.3). This relaxation gives an upper bound. For Theorem 4.3 to be true, the candidate control (4.20) must attain this bound, which requires that the filtration generated by the candidate \\nu^B_t is exactly trivial before t_c and equals σ(μ) after t_c. The paper only says that 'it is easy to see that such a filtration indeed satisfies properties 1-4'. The delicate point is that \\nu^B_t is an affine function of μ for t>t_c only after solving the coupled ODE for Q^B_t; the explicit coefficient of μ in formula (4.20) (involving \\hat D^B_t and the integral of \\hat C^B_{s,t}) could in principle vanish on a set of positive measure, in which case the generated filtration would not reveal μ at those times. No nondegeneracy or genericity check is provided. This is load-bearing because if the coefficient vanishes, the value delivered by \\nu^B is lower than the relaxed upper bound, and the claimed optimality of the critical-time disclosure policy is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a Stackelberg mean-field game between an informed broker and a continuum of uninformed traders in a linear-quadratic model. The broker knows the drift μ of the asset price, commits to an open-loop execution rate ν^B that is independent of traders' initial inventories, and traders observe the filtration generated by ν^B. The main theoretical result, stated informally in Section 2.3 and proved in Section 4.3, is that the broker's optimal information-leaking strategy is bang-bang: no information about μ is revealed before a critical time t_c, and all information is disclosed at t_c. The broker's optimal control is given by the explicit piecewise formula (4.20), where t_c is determined by the negative minimum of an explicitly computed function A. Section 5 proves that, for finitely many traders, this mean-field control is C/√N-optimal. A numerical example in Section 6 exhibits a nontrivial t_c and illustrates the resulting trader and broker behavior.","tokens_in":27604,"tokens_out":14737,"duration_ms":178079,"significance":"If the proof is completed, the paper gives a crisp and non-obvious qualitative conclusion: in this LQ Stackelberg mean-field game, delayed full disclosure dominates gradual information leakage. The strength of the paper is its explicit, closed-form construction: the critical time is computed from the derived function A in (4.19), not calibrated to data, and the finite-N approximation rate in Theorem 5.1 is stated with explicit dependence on the model parameters. This makes the result a useful benchmark for the literature on externalization, internalization, and information disclosure by liquidity providers. The main caveats are that Theorem 4.3 relies on a relaxed filtration optimization whose attainment is not fully verified, and the key propagation-of-chaos lemma used in the finite-N result is essentially unproved. These are load-bearing but plausibly fixable.","major_comments":[{"comment":"The relaxed optimization over filtrations gives only an upper bound. The statement immediately before Theorem 4.3 that \"it is easy to see that such a filtration indeed satisfies properties 1-4\" is not sufficient: property 1 requires σ(ν^B_s; 0≤s≤t) = σ(μ) for t>t_c. For the candidate control (4.20), the pre-t_c part is visibly deterministic, but after t_c the coefficient of μ is c_t := ∫_{t_c}^t Ĉ^B_{s,t} ds + D̂^B_t. The generated filtration reveals μ at time t only if c is not identically zero on every interval (t_c, τ), and the paper cites only real-analyticity, which does not exclude the possibility that c_t ≡ 0 on (t_c,T) (or on an initial interval after t_c) for special parameter values. The authors should prove that this coefficient is not identically zero under the standing small-b condition (4.14), or introduce an explicit nondegeneracy/genericity condition, and then verify properties 2-4 for the constructed filtration.","section":"Section 4.3, Theorem 4.3 and Eq. (4.20)"},{"comment":"Lemma 5.3 is the quantitative propagation-of-chaos estimate used in Theorem 5.1, but its proof is a single sentence referring to Cauchy-Schwarz, independence, (4.13), (5.16), and (5.19). The estimates in (5.4) involve a supremum over t∈[0,T] of empirical averages of the traders' equilibrium controls and their squares, where the controls depend on μ_s and on the initial inventories Q0^j. The appearance of the supremum, the second-moment inequality, and the role of the fourth-moment assumption on m0 all require a detailed proof. As written, this is a gap in the proof of the finite-N optimality claim.","section":"Section 5.4, Lemma 5.3"},{"comment":"The reduction of the broker's objective to the single function A' and the conclusion that E[μ_t^2] jumps only at t_c are the technical heart of the bang-bang theorem, but the displayed symmetrization identities are only sketched and contain apparent index errors. For example, after conditioning on F_u, an inner integral is printed as ∫_s^u while the surrounding variables suggest it should be over r from 0 to u or a related range. Because the sign and shape of A determine t_c, these identities should be written out carefully and in correct notation. If the printed formulas are not typos, the subsequent expression for A' needs a rigorous derivation.","section":"Section 4.3, derivation of A' in Eq. (4.19)"}],"minor_comments":[{"comment":"The sentence \"Note that β^B_0 and γ^R_0 only depend on μ_0\" seems to contain a typo: γ^R_0 should be γ^B_0, and γ^B_0 is deterministic rather than dependent on μ_0.","section":"Section 4.3, preamble to optimization over filtrations"},{"comment":"The notation C^ω([0,T]×[0,T], R^4) is inaccurate because the four functions have different domains: Â^B_t and B̂^B_t are functions of t only, Ĉ^B_{s,t} is a function of (s,t), and D̂^B_t is a function of t only. Please state the separate domains explicitly.","section":"Theorem 4.3 statement"},{"comment":"The notation ar Q_0 is used before it is defined; it should be introduced as ar Q_0 := E[Q_0] at its first appearance.","section":"Section 4.2, after Eq. (4.13)"},{"comment":"There is a typo in the admissibility condition: \"ν_t is independ of (Q^{N,j}_0)\" should read \"ν_t is independent of (Q^{N,j}_0)_{1≤j≤N}\".","section":"Section 5.2, definition of A^{N,B}"},{"comment":"Figure 1 and the formulas for A and A' are computed only to leading order in b. Since the theorem requires b sufficiently small, the text should state explicitly that the displayed curves correspond to b=0, or that the O(b) remainder is uniformly small enough to preserve the existence and location of the negative minimum.","section":"Section 6, numerical example"}],"recommendation":"major_revision","confidential_remarks":"This is a potentially strong paper with an explicit and surprising conclusion. I would not recommend rejection because the central construction is explicit and the gaps look repairable. However, the proof of Theorem 4.3 needs a genuine nondegeneracy argument for the disclosure coefficient, and Lemma 5.3 needs a real proof before the finite-N result can be accepted. I also suggest the authors double-check the symmetrization calculation leading to A', since any error there would propagate to the critical time."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is worth a serious look. It derives an explicit optimal information-disclosure policy for a broker with private drift information facing a mean field of uninformed traders: stay silent until a deterministic critical time t_c, then fully reveal. The broker's control is piecewise analytic, and the same policy is C/sqrt(N)-optimal in the finite-N game. This is genuinely new relative to the authors' earlier relaxed open-loop framework [10], and it is derived rather than fitted. No free parameters, no calibrated constants. The citation pattern is normal for this group, with appropriate acknowledgement of the closed-loop master-equation limitation.\n\nWhat is good: the linear-quadratic structure is solved in closed form; the critical-time characterization via a deterministic function A is explicit; the numerical section gives an economic interpretation (let traders unwind first, then reveal to profit from spread) that matches the math. The finite-N approximation theorem is a real contribution if the proof holds up.\n\nSoft spots, in order of softness:\n1. Lemma 5.3 is the backbone of the C/sqrt(N) result, and its proof is a one-sentence reference to Cauchy-Schwarz. That is too thin for a main lemma. The estimates are probably standard, but they need to be written.\n2. The step from the relaxed filtration optimization to attainment (Section 4.3) is compressed. The claim that the candidate control generates the trivial filtration before t_c and sigma(mu) after is 'easy to see' but relies on a nondegeneracy condition: the coefficient of mu in (4.20) must be non-zero on a right-neighborhood of t_c. Since that coefficient is analytic and equals D^B_{t_c} at t_c, a zero value or an interval of zeros would break attainment. The paper does not check this. I am not saying the result fails; I am saying the proof needs either a genericity assumption or a short argument.\n3. The numerical example is asymptotic in b: no concrete b satisfying (4.14) is given, and the critical-time plot is for the O(b) leading term. Minor for a theory paper.\n\nThe central argument is sound in spirit; the two gaps are fillable. This deserves peer review. I would send it to a rigorous referee and expect heavy-ish revision but a strong paper after.\n\nRecommendation: accept for review; ask for a full Lemma 5.3 proof and the nondegeneracy check.","headline":"A genuinely new explicit bang-bang disclosure policy for an informed broker in a mean-field Stackelberg game, solidly derived but with two proof gaps that need closing.","tokens_in":28184,"tokens_out":5089,"would_cite":true,"duration_ms":55588,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91A16","91A65","91G10","49N10"],"pacs":[],"model":"deepseek-v4-flash","headline":"A broker with private drift information should hide it until a critical time, then disclose it all at once.","keywords":["market making","algorithmic trading","externalization","mean field games","Stackelberg equilibrium","information leakage","optimal control","hedging"],"falsifier":"Solve the finite-$N$ Stackelberg game numerically with a broker who is allowed to condition the execution rate on the aggregate trader inventory: if a gradual-leakage policy then beats the bang-bang policy (4.20), the paper's optimality claim fails in that relaxation. A cheaper check is to find parameters violating the small-$b$ condition (4.14) and test whether the single-disclosure-time policy is still optimal.","tokens_in":27111,"feed_emoji":"🤫","tokens_out":7955,"duration_ms":86503,"temperature":0.7,"pith_summary":"This paper studies a broker who knows the drift of an asset price while many traders do not, and whose own trading rate is observed by those traders. The paper claims that in the mean-field limit of infinitely many traders there is a critical time $t_c \\in [0,T]$ such that the broker optimally leaks no information before $t_c$ and discloses all information exactly at $t_c$, provided the traders' permanent market impact $b$ is small. It gives an explicit, piecewise analytic formula for the broker's execution rate and proves that this infinite-trader policy remains a $C/\\sqrt{N}$-optimal Stackelberg equilibrium when there are $N$ traders. The reason to care is that the result turns a vague trade-off between hiding and exploiting private information into a sharp policy: gradual leakage is dominated by a single, timed revelation.","feed_headline":"Informed broker's best leak: nothing, then everything","feed_subtitle":"A mean-field model says the optimal policy is to hide the drift until a critical time, then disclose it all.","key_machinery":"The carrying object is the critical-time function $A(t)$, whose derivative $A'_s$ is assembled in (4.19) from the kernel functions (4.11) and (4.18); the optimal disclosure time $t_c$ is a point where $A$ attains a negative minimum. These kernels come from an explicit solution of a linear-quadratic backward stochastic Hamilton-Jacobi equation, whose Riccati component is solved in closed form and whose linear component is expressed through conditional expectations of the drift. That explicit solution turns the broker's optimization over filtrations into a one-dimensional variational problem over the increasing process $\\mathbb{E}[\\mu_t^2]$, whose optimum is to keep $\\mathbb{E}[\\mu_t^2]$ flat at $\\mathbb{E}[\\mu]^2$ until $t_c$ and then jump to $\\mathbb{E}[\\mu^2]$. The same formulas feed Theorem 5.1, where propagation of chaos shows the mean-field control is approximately optimal in the finite-$N$ game.","core_discovery":"The central claim is that the optimal information-leaking strategy for an informed broker facing a continuum of uninformed traders is bang-bang: the broker's execution rate reveals nothing about the drift $\\mu$ until a critical time $t_c$, and reveals $\\mu$ completely at $t_c$. The optimal broker control is the explicit piecewise formula (4.20), with $t_c$ determined by the negative minimum of a deterministic function $A$ built from the model parameters; the numerical section exhibits a nontrivial case with $t_c \\simeq 0.32$ in a two-day horizon. For a finite game with $N$ traders, the same control is a $C/\\sqrt{N}$-optimal Stackelberg equilibrium, meaning the mean-field strategy loses at most of order $1/\\sqrt{N}$ in value compared with the true finite-player optimum.","pith_inferences":["Beyond the paper: if the broker is allowed to condition on traders' inventories, the formal master-equation problem is unsolved and the bang-bang structure could fail; testing this numerically is a direct next step.","Beyond the paper: the critical time $t_c$ gives a testable signature in observable order flow—flat execution rates until a date and then a jump—that could be looked for in OTC or automated-market-maker data.","Beyond the paper: the same variational argument over the martingale $\\mathbb{E}[\\mu_t^2]$ may extend to multi-signal private information, predicting a sequence of disclosure dates rather than a single one."],"forward_implications":["In the infinite-trader limit, the broker should act as if uninformed until $t_c$ and then disclose the drift fully; any gradual leakage policy is strictly suboptimal.","For a large but finite number of traders, the mean-field bang-bang policy is within $C/\\sqrt{N}$ of the true Stackelberg value, so the recommendation survives finite population effects.","In the pre-disclosure phase traders unwind their inventories, letting the broker collect transaction fees $\\eta$ without taking market risk; after disclosure, traders build positions in the direction of the drift and the broker profits by executing at a cheaper cost $\\eta^B<\\eta$.","The optimal policy depends on the drift only through its expectation and realized value, not on the full distribution of $\\mu$.","For observable liquidity providers such as automated market makers, the model predicts that the optimal response to private drift information is a single scheduled information event rather than continuous leakage."],"supporting_citations":[{"why":"Supplies the explicit solution method for the linear-quadratic backward stochastic Hamilton-Jacobi equation used to compute both the trader's and the broker's optimal controls.","marker":"[14]"},{"why":"Introduces the Stackelberg mean-field-game framework with an informed major player that this paper adapts to an open-loop broker and closed-loop traders.","marker":"[10]"},{"why":"Defines mean-field games with a major player, the modelling device for a dominant broker interacting with a population of small traders.","marker":"[36]"},{"why":"Provides the bridge between finite-player Stackelberg games and mean-field major-player limits that underpins the C/sqrt(N) approximate-optimality statement.","marker":"[28]"},{"why":"Analyses mean-field games with a dominating player in a leader-follower form, giving the equilibrium concept this paper uses for the Stackelberg game.","marker":"[8]"}],"fun_headline_variants":["Optimal leak: hide the drift, then spill it all","Broker's best play: stay quiet, then reveal everything","Mean-field optimal: broker flips from silence to full disclosure","Critical time reveals all: informed broker's bang-bang leak","Hide until t_c, then leak all: broker's mean-field edge"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the broker can commit, for the whole horizon, to an execution plan that does not depend on traders' inventories (and that the traders' permanent impact $b$ is below an explicit small threshold), so the only information traders receive is the timing of that precommitted plan.","fun_headline_variants_meta":{"raw":{"variants":["Optimal leak: hide the drift, then spill it all","Broker's best play: stay quiet, then reveal everything","Mean-field optimal: broker flips from silence to full disclosure","Critical time reveals all: informed broker's bang-bang leak","Hide until t_c, then leak all: broker's mean-field edge"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000192,"raw_usage":{"total_tokens":1281,"prompt_tokens":816,"completion_tokens":465,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":432,"completion_tokens_details":{"reasoning_tokens":377}},"tokens_in":432,"tokens_out":465,"duration_ms":5527,"temperature":1.0,"reasoning_tokens":377,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:57:52.991107+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Solve the finite-$N$ Stackelberg game numerically with a broker who is allowed to condition the execution rate on the aggregate trader inventory: if a gradual-leakage policy then beats the bang-bang policy (4.20), the paper's optimality claim fails in that relaxation. A cheaper check is to find parameters violating the small-$b$ condition (4.14) and test whether the single-disclosure-time policy is still optimal.","supporting_citations":[{"cited_title":"Linear quadratic optimal stochastic control with random coefficients.SIAM Journal on Control and Optimization, 14(3):419–444, 1976","cited_arxiv_id":null,"evidence_quote":"Supplies the explicit solution method for the linear-quadratic backward stochastic Hamilton-Jacobi equation used to compute both the trader's and the broker's optimal controls."},{"cited_title":"Mean field games in a stackelberg problem with an informed major player.SIAM Journal on Control and Optimization, 62(3):1737–1765, 2024","cited_arxiv_id":null,"evidence_quote":"Introduces the Stackelberg mean-field-game framework with an informed major player that this paper adapts to an open-loop broker and closed-loop traders."},{"cited_title":"Large-population lqg games involving a major player: the nash certainty equivalence principle","cited_arxiv_id":null,"evidence_quote":"Defines mean-field games with a major player, the modelling device for a dominant broker interacting with a population of small traders."},{"cited_title":"Mean field games with a dominating player.Applied Mathematics & Optimization, 74:91–128, 2016","cited_arxiv_id":null,"evidence_quote":"Analyses mean-field games with a dominating player in a leader-follower form, giving the equilibrium concept this paper uses for the Stackelberg game."}],"review_version":1}