{"id":"0333c2f8-1510-4082-9e19-03d01bbde1ed","arxiv_id":"2608.00885","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Under a mean-reverting order-book model, the optimal strategy is a symmetric band with half-width solving theta*(theta* - phi) = s_G^2 and earning rate alpha s_G sqrt(2/pi) exp(-theta*^2 / (2 s_G^2)).","lead":"This paper builds a mathematical model of an order book in which the displayed mid price drifts back toward a hidden true price, then solves for the trading rule that earns the most from that pull. A generalist might read it because it reduces a high-frequency trading problem to a simple formula telling how far to let the price move before trading.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Timing surrogate's O(δ/θ) error is unproved, and its regime δ≪θ may be violated exactly where the large-tick calibration gives s_G∼δ, leaving θ* and R* without controlled accuracy for the target asset class.","rationale":"The reader's weakest assumption is exactly the timing heuristic, and I agree that it is the load-bearing point. My stress-test adds a sharper observation: not only is the O(δ/θ) bound unproved, but the paper's own qualitative calibration for large-tick assets (s_G 'of the order of the tick') places θ* at only about 1.3δ, so δ/θ is of order one, outside the regime δ≪θ in which the paper claims to work. This turns an open heuristic into a potential first-order error in the target setting. The paper is otherwise transparent: the surrogate solution is derived rigorously, the exact model's rate is given by the renewal-reward formula (3.3), and the authors explicitly label the timing error and the band optimality as open. The Monte Carlo evidence in Figure 4 is reassuring but not reproducible without code or parameter sets. My proposed check would settle whether the timing approximation holds beyond its asymptotic regime; if it does, the concern is mitigated and the CONDITIONAL verdict remains appropriate. I therefore do not change the reader's verdict, and I mark my agreement as partial because I emphasize a regime mismatch that the reader's statement did not explicitly flag.","tokens_in":24761,"tokens_out":10710,"duration_ms":94427,"concrete_test":"Run the exact-model Monte Carlo with explicit parameters satisfying the large-tick condition s_G between 1δ and 2δ (choose σ_X, α, μ, and ramp slopes so that Remark 2.1's bounds give s_G≈1.5δ and p≈1%). For θ in a grid around θ_D and θ*, estimate m(θ) by simulation and compare to the closed-form m̃(θ) of Proposition B.1; report the ratio m(θ)/m̃(θ) and the true optimal θ within the band class. If the relative timing error exceeds the paper's claimed O(δ/θ), or if the true peak lies more than about 20% from θ_D, the closed-form θ* is not validated for the target regime. As a secondary check at one parameter point, compare the band rate against a strategy that rests flat inside an inner band to test the band-class optimality conjecture on the exact process.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline closed forms θ* and R* are exact only for the Gaussian surrogate; for the jump-process model of Definitions 2.1–2.2 they require two unverified steps: (i) the timing approximation |m̃(θ)−m(θ)|/m(θ)=O(δ/θ), which Section 3 error (c) and Section 5 explicitly leave open, and (ii) optimality of the symmetric band class on the exact process, stated as a conjecture. Step (i) is load-bearing: if the true mean inter-fill time differs from the OU passage time by more than O(δ/θ), the optimal threshold and the profit rate for the exact model are not θ* and R*. The concern is sharpened by an apparent regime mismatch. The paper works in 'δ≪θ−φ' (Section 3), but its own target is liquid large-tick assets, where it says the sampled range has |G| of a few s_G, 'itself of the order of the tick' (Section 2.3). With s_G∼δ and φ=δ/2, the band equation θ*(θ*−φ)=s_G^2 gives θ* roughly 1.28δ for s_G=δ, so δ/(θ−φ)≈1.3 and δ/θ≈0.8: these are not small parameters, and the asymptotic error bounds (a) and (c) do not apply. The Monte Carlo in Figure 4 suggests the surrogate works even in this regime, but no code or parameter sets are provided, so the support is not independently checkable. Thus the paper's central claim, as a statement about the exact book model, rests on an unproved timing heuristic whose claimed small-parameter regime may not cover the empirically motivated calibration.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper constructs a limit order book model for liquid large-tick assets in which the spread is always one or two ticks and coincides with the parity of the mid on the half-tick grid, so the only continuous state is the gap G between the mid and a latent efficient price X, an exogenous Brownian martingale. Six event intensities are affine in the positive and negative parts of G. Under the balanced-response condition 2α_s+α_o=α_c=α, the paper proves that the gap's conditional mean decays as e^{-αh} and its stationary autocovariance is exactly that of an OU process with rate α and variance s_G^2. The trading problem maximizes long-run average profit net of half-spread; a pathwise layer decomposition reduces the search to positions in {-1,0,1}, and a symmetric threshold band is studied. The exact jump-process rate has a renewal-reward form, and passage times are evaluated on the OU surrogate with matched moments. On the surrogate, the band is optimal, with exact maximizer θ_D from a Dawson equation and Kramers approximation θ* solving θ*(θ*−φ)=s_G^2, with rate R*=αs_G√(2/π)e^{−θ*^2/(2s_G^2)}. The paper explicitly states that the timing error O(δ/θ) is heuristic and that optimality of the band on the jump process is a conjecture; Figure 4 reports Monte Carlo support.","tokens_in":25201,"tokens_out":8563,"duration_ms":78812,"significance":"If the advertised results are taken as statements about the Gaussian surrogate, the paper is a clean and useful contribution: exact moment matching, a rigorous layer reduction, exact renewal-reward accounting, and closed-form switching thresholds with a transparent separation of proved and heuristic parts. The book-level exact reversion theorem (Proposition 2.1) and the pathwise layer reduction (Proposition 2.2) are genuine and merit credit. The main formulas are simple and falsifiable, and the derivation introduces no fitted constants. The weakness is that the headline closed forms and the claimed optimal rule are not theorems for the jump-process model of Definitions 2.1–2.2; they depend on an unproved timing approximation and on a small-parameter regime that is not met by the paper's own large-tick calibration. These issues are acknowledged in the text, but they are precisely what must be resolved or explicitly disclaimed for the central claim to hold.","major_comments":[{"comment":"The passage from the exact rate (3.3) to the surrogate rate (3.5) and hence to the closed forms (3.7)–(3.8) rests on the timing approximation |m̃(θ)−m(θ)|/m(θ)=O(δ/θ), which the paper explicitly leaves open (error (c)). Because this approximation is load-bearing, the statements in the abstract ('solve for the trading rule') and in Section 5 ('the result') overstate the status of θ* and R* as properties of the jump-process model: they are exact only for the Gaussian surrogate. The revision should either prove a bound on the timing error under stated conditions, or consistently present the exact-process claims as conjectures in the abstract, introduction, and conclusion, with the theorem-level results for the surrogate clearly separated.","section":"§3, after Eq. (3.4); §5"},{"comment":"The claimed regime δ≪θ−φ is not satisfied in the target calibration. The paper argues in Section 2.3 that for liquid large-tick assets the sampled |G| is of order a tick, s_G∼δ, and it sets φ=δ/2. At s_G=δ, the band equation (3.7) gives θ*≈1.28δ, so δ/(θ−φ)≈1.28 and δ/θ≈0.78 rather than ≪1. Consequently the proved reward-side error of order δ/(θ−φ) and the heuristic timing error of order δ/θ are both uncontrolled exactly in the empirically motivated regime. The Figure 4 Monte Carlo is suggestive, but without code or parameter sets the applicability of the closed forms to the stated asset class is not established.","section":"§2.3 and §3 (regime after error (c))"},{"comment":"The Monte Carlo experiment is the only evidence that the surrogate performs well outside the asymptotic regime, but the manuscript gives no code, no seed, and no parameter values, and reports only 'one standard error' bands without a table of the baseline intensities and ramp slopes used. Since the sweep in Figure 4 is precisely where the asymptotic bounds are not available, the experiment should be reproducible: include the full parameter set (or code), the number of paths, and the standard-error construction.","section":"§3.1, Figure 4"}],"minor_comments":[{"comment":"The displayed root θ* is the Kramers/large-threshold approximation, while the exact surrogate maximizer is u_D of Appendix B; Proposition B.2 should be cited at the first occurrence of (3.8) so that readers do not read R* as exact at θ*.","section":"§3, Eq. (3.7)–(3.8)"},{"comment":"The sentence 'Three assumptions carry the result' should also state that the two further open questions (timing error and band optimality) mean the headline formulas are surrogate theorems; as written, 'the result' is ambiguous.","section":"§5"},{"comment":"The captions say 'illustrative parameters; nothing is calibrated,' but no parameter values are listed; providing them would help readers assess the claimed openness fraction of about 18 percent in Figure 2 and the visual behavior in Figure 5.","section":"Figures 2 and 5"},{"comment":"The parity lock relies on the two-valued spread assumption; a sentence noting that assets with three or more spread values fall outside the lock would make the scope of the state-space reduction clearer.","section":"Fact 1, Section 2.2"}],"recommendation":"major_revision","confidential_remarks":"The exact results in the paper are solid and the paper is unusually candid about its limitations, but the headline formulas are not yet theorems about the jump-process model. A major revision that either controls the timing error or consistently restricts the optimality claims to the Gaussian surrogate would make the contribution publishable; the regime mismatch with the large-tick calibration should be addressed head-on."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This one is worth your time. The paper's genuinely new pieces are the parity lock (spread as a parity bit of the mid on the half-tick grid) and the closed-form band equation theta*(theta* - phi) = s_G^2, with the myopic rule earning exactly zero. The author builds a queue-reactive book around a latent efficient price and solves the optimal band problem on a Gaussian surrogate matched to the gap's exact conditional mean and stationary covariance. The parity lock is a real structural simplification, and the exact reversion theorem under the balanced-response condition is a neat result, with proofs in the appendix that look internally consistent. The layer reduction to {-1,0,+1} positions is pathwise and exact, not a surrogate statement. On the Gaussian diffusion, the Dawson-root optimization is rigorous, and the paper is unusually candid about what is theorem and what is conjecture.\n\nThe soft spots are real but openly acknowledged. The load-bearing timing approximation |m_tilde(theta)-m(theta)|/m(theta) = O(delta/theta) is left unproved, and the stress-test note is correct that in the paper's own large-tick calibration (s_G ~ delta), delta/theta is around 0.8, not the delta << theta regime the asymptotic claims assume. The Monte Carlo in Figure 4 suggests the surrogate still works there, but no code or parameter sets are given, so that support is not independently checkable. Similarly, the optimality of the symmetric band class on the exact jump process is stated as a conjecture. These are limitations the paper names directly, which is a mark in its favor, but they do mean the headline formulas should be read as exact for the surrogate and heuristic for the book model.\n\nThis paper deserves a serious referee. The core contribution is solid and the author has done the hard work of separating proved claims from heuristic ones. A revision should either prove the timing estimate under explicit conditions or replace it with reproducible numerical evidence that covers the regime the paper actually targets. I would bring it to a reading group, and I would cite it if I worked on microstructure trading.\n\nSend it to peer review.","headline":"Novel parity-lock and closed-form band for trading microstructure mean reversion; the surrogate math is rigorous, but exact-process claims rest on an unproved timing heuristic that the paper itself flags.","tokens_in":25655,"tokens_out":3931,"would_cite":true,"duration_ms":36875,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60G55","60J25","60J60","93E20","91G80"],"pacs":[],"model":"deepseek-v4-flash","headline":"A symmetric band with half-width $\\theta^*$ solving $\\theta^*(\\theta^*-\\phi)=s_G^2$ is the optimal strategy for trading mid-price mean reversion in a large-tick book, earning $R^*=\\alpha s_G\\sqrt{2/\\pi}\\,e^{-\\theta^{*2}/2s_G^2}$.","keywords":["market microstructure","limit order book","large-tick assets","mean reversion","optimal switching","ergodic control","renewal-reward","parity lock"],"falsifier":"Simulate the exact jump model with known parameters and measure the stationary mean time between opposite fills $m(\\theta)$ alongside the surrogate passage time $\\tilde{m}(\\theta)=\\frac{\\pi}{\\alpha}\\operatorname{erfi}(u/\\sqrt{2})$ for several $\\theta/\\delta$. If $|m(\\theta)-\\tilde{m}(\\theta)|/m(\\theta)$ does not shrink like $\\delta/\\theta$ as $\\delta/\\theta\\to 0$, or if the exact rate peak deviates from $\\theta_D$ by much more than the simulated ~20% inward shift, the paper's closed-form optimum is not the true book's optimum. Alternatively, on real tick data, estimate $\\alpha$ and $s_G$ from the mid's autocovariance, implement the band, and check whether realized P&L exceeds the myopic rule's near-zero rate.","tokens_in":24551,"feed_emoji":"📈","tokens_out":6293,"duration_ms":50627,"temperature":0.7,"pith_summary":"The paper builds a limit-order-book model whose own order flow creates the stationary, mean-reverting gap between the displayed mid and a latent efficient price, and solves the optimal trading rule for that gap. In a liquid large-tick asset the spread is either one tick or two, and the paper's \"parity lock\" shows the spread is just the parity of the mid, so the gap is the only continuous state variable. Under a balanced-response condition the gap's conditional mean and stationary covariance are exactly those of an Ornstein-Uhlenbeck process, and the paper establishes that a symmetric no-churn band---buy when the gap hits $-\\theta$, sell at $+\\theta$---is optimal on the Gaussian surrogate with those two moments. The optimal half-width solves $\\theta^*(\\theta^*-\\phi)=s_G^2$ and earns $R^*=\\alpha s_G\\sqrt{2/\\pi}\\,e^{-\\theta^{*2}/2s_G^2}$; trading as soon as the gap covers the spread earns zero. The argument is exact on the surrogate; on the actual jump process the band's optimality and the timing approximation remain conjectures.","feed_headline":"One equation sets the optimal band: θ(θ−φ)=σ²","feed_subtitle":"In large-tick books, the best strategy waits for deeper mean reversion; trading as soon as the spread is covered earns zero.","key_machinery":"The machinery is the parity lock, the balanced-response condition, and the Gaussian-surrogate passage-time evaluation. The parity lock (Fact 1) makes the spread a deterministic function of the mid's parity, leaving $G$ as the only continuous state variable. The balanced-response condition $2\\alpha_s+\\alpha_o=\\alpha_c=\\alpha$ equalises the book's corrective drift across tight and open books, turning gap reversion into an exact theorem (equations (2.12)--(2.14)). The surrogate (3.4) is the unique Gaussian diffusion with the gap's exact conditional mean and stationary covariance; renewal--reward with Kramers' law on that surrogate yields the band equation and the closed-form rate. The layer decomposition (Proposition 2.2) reduces the full inventory problem to $\\{-1,0,+1\\}$ paths exactly, and the switching literature supplies threshold optimality on the surrogate.","core_discovery":"The central discovery is that microstructure mean reversion reduces to a one-dimensional switching problem with a closed-form solution. The parity lock collapses bid, ask, spread, and mid into the gap $G$ plus a parity bit; Definition 2.2's linear intensities, under $2\\alpha_s+\\alpha_o=\\alpha_c=\\alpha$, make the gap's conditional mean and stationary covariance identical to an OU process with rate $\\alpha$ and variance $s_G^2$ (Proposition 2.1), although paths jump. Passage times are evaluated on that Gaussian surrogate, and on it the symmetric band is optimal among all admissible strategies (Proposition 3.1 plus switching literature). The optimal band's half-width is $\\theta^*(\\theta^*-\\phi)=s_G^2$, the rate is $R^*=\\alpha s_G\\sqrt{2/\\pi}\\,e^{-\\theta^{*2}/2s_G^2}$, and the myopic $\\theta=\\phi$ rule earns exactly zero: all profit is the option value of waiting. The paper states clearly that on the jump process the band-class reduction and the $O(\\delta/\\theta)$ timing error are heuristic, and simulations show the exact-model peak sits about a fifth inside the surrogate optimum with small rate loss.","pith_inferences":["The surrogate timing approximation could be tested directly by simulating the exact jump book at several tick sizes and comparing measured mean inter-fill times $m(\\theta)$ with the OU passage time $\\tilde{m}(\\theta)$ over a range of $\\theta/\\delta$; if the relative error is not $O(\\delta/\\theta)$, the closed forms still describe the surrogate but not a book with those parameters.","The same moment-matching logic might extend to other cost structures---for example quadratic impact or inventory risk---where the optimal policy would no longer be a pure flip band; the paper's Proposition 2.2 already flags that interior positions can win under such costs.","If the gap is not observable, the paper's state-space structure (Brownian state observed through linear point-process intensities) suggests a filter in which both moves and silences inform the gap estimate; quantifying the performance loss from filtering is an open continuation.","The balanced-response condition is a genuine restriction, but the paper's bound via open-book occupancy suggests a testable prediction: in books with very small $p$, the parity-averaged rate $\\alpha_{\\mathrm{eff}}$ should approximate the fitted $\\alpha$ well; a book violating (2.8) with large $p$ would split the reversion rate and break the closed form."],"forward_implications":["If correct, a trader needs only the gap's autocovariance parameters $(\\alpha, s_G)$ to set the band; the optimal half-width is $\\theta^*=(\\phi+\\sqrt{\\phi^2+4s_G^2})/2$.","The myopic benchmark---trade as soon as the gap covers the half-spread---is worthless on the surrogate; the entire profit rate comes from waiting for a deeper reversion.","Because the rate is flat at its maximum, a relative error $\\varepsilon$ in the threshold costs only $O(\\varepsilon^2)$ in rate, so the strategy is robust to estimation error in $\\alpha$ and $s_G$.","All assets with the same spread-to-dispersion ratio $\\gamma=\\phi/s_G$ share the same optimal band in gap units, so calibration reduces to one dimensionless number.","Open-book fills are rare (occupancy $p$), so the tight-book cost convention loses $O(p)$; venue rules that act only on open books are second-order for this strategy."],"supporting_citations":[{"why":"Defines the large-tick asset class with spread pinned at one or two ticks and deep queues at the touch, the empirical regime the model targets.","marker":"[14]"},{"why":"Identifies the efficient price as the permanent component of the observed price, which the paper proves its latent $X$ is via equation (2.1).","marker":"[21]"},{"why":"Supplies the verification result that threshold entry-exit rules are optimal for mean-reverting assets, used for the surrogate band's optimality.","marker":"[36]"},{"why":"Extends threshold optimality to general one-dimensional diffusions, another pillar of the band-class reduction on the surrogate.","marker":"[35]"},{"why":"Gives the complete analysis for the OU asset, the third verification source for threshold optimality.","marker":"[26]"},{"why":"Provides the renewal--reward criterion for long-run trading-cycle rates adopted as the paper's objective.","marker":"[9]"},{"why":"Is the reference idealisation of the mid as Brownian motion, cited for the diffusion-approximation spirit of the surrogate.","marker":"[6]"},{"why":"Supplies the regenerative ratio theorem used to compute the band's rate exactly on the surrogate.","marker":"[5]"},{"why":"Provides the Foster--Lyapunov criteria used to prove ergodicity and the unique invariant law of $(G,S)$.","marker":"[28]"}],"fun_headline_variants":["Optimal mean-reversion band: θ(θ−φ)=σ²","All profit is in waiting: optimal band for mean reversion","Trading too soon earns zero: wait for the mean-reversion edge","One equation rules microstructure mean reversion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the Gaussian-surrogate mean inter-fill time matches the true jump-process inter-fill time to relative error $O(\\delta/\\theta)$; the paper leaves this as a heuristic, and if it fails the closed-form $\\theta^*$ and $R^*$ describe only the surrogate, not the real book.","fun_headline_variants_meta":{"raw":{"variants":["Optimal mean-reversion band: θ(θ−φ)=σ²","All profit is in waiting: optimal band for mean reversion","Trading too soon earns zero: wait for the mean-reversion edge","One equation rules microstructure mean reversion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000986,"raw_usage":{"total_tokens":4293,"prompt_tokens":1167,"completion_tokens":3126,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":783,"completion_tokens_details":{"reasoning_tokens":3055}},"tokens_in":783,"tokens_out":3126,"duration_ms":20974,"temperature":1.0,"reasoning_tokens":3055,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:15:22.695731+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the exact jump model with known parameters and measure the stationary mean time between opposite fills $m(\\theta)$ alongside the surrogate passage time $\\tilde{m}(\\theta)=\\frac{\\pi}{\\alpha}\\operatorname{erfi}(u/\\sqrt{2})$ for several $\\theta/\\delta$. If $|m(\\theta)-\\tilde{m}(\\theta)|/m(\\theta)$ does not shrink like $\\delta/\\theta$ as $\\delta/\\theta\\to 0$, or if the exact rate peak deviates from $\\theta_D$ by much more than the simulated ~20% inward shift, the paper's closed-form optimum is not the true book's optimum. Alternatively, on real tick data, estimate $\\alpha$ and $s_G$ from the mid's autocovariance, implement the band, and check whether realized P&L exceeds the myopic rule's near-zero rate.","supporting_citations":[{"cited_title":"Dayri and M","cited_arxiv_id":null,"evidence_quote":"Defines the large-tick asset class with spread pinned at one or two ticks and deep queues at the touch, the empirical regime the model targets."},{"cited_title":"Hasbrouck","cited_arxiv_id":null,"evidence_quote":"Identifies the efficient price as the permanent component of the observed price, which the paper proves its latent $X$ is via equation (2.1)."},{"cited_title":"Zhang and Q","cited_arxiv_id":null,"evidence_quote":"Supplies the verification result that threshold entry-exit rules are optimal for mean-reverting assets, used for the surrogate band's optimality."},{"cited_title":"Zervos, T","cited_arxiv_id":null,"evidence_quote":"Extends threshold optimality to general one-dimensional diffusions, another pillar of the band-class reduction on the surrogate."},{"cited_title":"Leung and X","cited_arxiv_id":null,"evidence_quote":"Gives the complete analysis for the OU asset, the third verification source for threshold optimality."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the renewal--reward criterion for long-run trading-cycle rates adopted as the paper's objective."},{"cited_title":"Avellaneda and S","cited_arxiv_id":null,"evidence_quote":"Is the reference idealisation of the mid as Brownian motion, cited for the diffusion-approximation spirit of the surrogate."},{"cited_title":"Asmussen.Applied Probability and Queues","cited_arxiv_id":null,"evidence_quote":"Supplies the regenerative ratio theorem used to compute the band's rate exactly on the surrogate."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Foster--Lyapunov criteria used to prove ergodicity and the unique invariant law of $(G,S)$."}],"review_version":2}