{"id":"122b2ad5-3383-40ee-afb8-b9869d22a3bc","arxiv_id":"2411.19637","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Under linear price impact and Poisson liquidation arrivals, the optimal DeFi liquidation policy is to trade a constant fraction sqrt(phi/k) of inventory per unit time, yielding ergodic reward 2*r*lambda*eta*S0 - lambda*eta^2*b - 2*lambda*eta^2*sqrt(k*phi).","lead":"An exchange liquidating positions it took over from defaulted DeFi traders should sell a fixed fraction of its remaining inventory per unit time. The paper derives this rule and the resulting long-term profit formula from a stylized stochastic model, with all parameters except risk aversion calibrated from market data.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The gamma formula uses η² where the HJB jump term requires E[ζ²]; unless ζ is degenerate, Theorem 2.1's closed form is not a consequence of the stated assumptions.","rationale":"The reader's conditional verdict is essentially right. The optimal control result is solid: both finite-horizon and discounted limits give feedback ν*=√(ϕ/k)q, and that conclusion does not depend on Eζ² versus η². The central closed-form γ, however, is derived from h0, and h0 is wrong under the stated random-size assumption. The proof of Theorem 2.1 also omits the final ergodic identification (delegated to [7]), but that is a less concrete problem than the algebraic mismatch in (22). This concern is testable and fixable: replace η² by Eζ² throughout Sections 5.2–5.4 and augment the calibration in §3 with an estimate of the second moment. Because the policy recommendation and the qualitative comparative statics are unchanged, the appropriate verdict remains conditional acceptance subject to this correction rather than rejection.","tokens_in":15144,"tokens_out":11512,"duration_ms":101149,"concrete_test":"Analytically: substitute u(q)=h0+h2q² into HJB (19)/(26) with ζ± drawn from a distribution with mean η but second moment M≠η² (e.g. Gamma(shape=1, scale=10), η=10, M=200). Solve for h0 and take β→0; the ergodic constant is 2rληS0 - λM(b+2√(kϕ)), whereas (8) gives 2rληS0 - λη²(b+2√(kϕ)). The residual in (26) under the paper's gamma (8) is then λ(M-η²)(b+2√(kϕ)), non-zero, so the claimed closed form is inconsistent with the stated stochastic dynamics. Numerically, simulate Q under (3) with this size distribution and check which γ the achieved long-run average matches.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing issue is not the acknowledged S0 simplification in (7) but an unflagged substitution of random liquidation sizes by their means. The reduced HJB (19) retains random jumps: it contains λE[u(t,q+ζ⁺)-u(t,q)] + λE[u(t,q-ζ⁻)-u(t,q)]. Substituting the ansatz u=h0+h2q² gives a constant term 2λh2E[ζ²] (with symmetric ζ±), not 2λη²h2. Yet the solved ODE (22), and the analogous discounted h0 in (24), use 2λη²h2 and hence give γ = 2rληS0 - λη²b - 2λη²√(kφ). The derivation is internally inconsistent unless Varζ=0. The corrected constant would be γ = 2rληS0 - λE[ζ²]b - 2λE[ζ²]√(kφ) (for equal second moments). The paper's numerical ζ~N(10,0.5) has Eζ²=100.25 ≈ η²=100, so the error is invisible there; with realistic heavy-tailed liquidation sizes the reward formula and the calibration recipe in §3 (which estimates only η) would be materially wrong. The optimal feedback ν*=√(ϕ/k)q survives, because the q-linear parts of the two jump expectations cancel.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper models the gradual liquidation of positions accumulated by a decentralized derivatives exchange as an ergodic stochastic control problem. The setup includes Brownian mid-price dynamics, linear temporary and permanent price impact, Poisson liquidation events with random sizes, and a simplified cash balance dynamics in which liquidation cash inflows are priced at the initial mid-price. Under these simplifications the paper derives candidate closed-form solutions: the ergodic constant γ = 2rληS0 − λη²b − 2λη²√(kϕ) and the optimal Markov control ν*(q)=√(ϕ/k)q. The proofs proceed through finite-horizon and discounted HJB equations with a quadratic ansatz and then pass to the limits T→∞ and β→0. Numerical simulations compare the proposed strategy with a heuristic, study parameter sensitivity, and propose a calibration method from order-book and liquidation data.","tokens_in":15411,"tokens_out":11341,"duration_ms":98299,"significance":"The paper addresses a timely and practically relevant problem, and if the formulas are correct the proportional-inventory policy ν*(q)=√(ϕ/k)q is a clean, implementable benchmark for DeFi liquidation desks. The HJB ansatz and verification structure are mostly transparent, the admissibility checks for the candidate controls are explicit, and the calibration section gives a concrete route to apply the model. A notable strength is that the optimal feedback control is robust to the second-moment issue identified below. However, the quantitative content of the main theorem depends on an unstated substitution of E[ζ²] by η², and the proof of the key ergodic identification is outsourced to a preprint. The central closed-form reward formula and the calibration recipe therefore need correction before the results can be used as stated.","major_comments":[{"comment":"The jump term in the HJB equation (21) is λE[u(t,q+ζ+)−u(t,q)] + λE[u(t,q−ζ−)−u(t,q)]. With the quadratic ansatz u=h0+h2q² and h1=0, this term equals λh2(E[(ζ+)²]+E[(ζ−)²]) after the linear q-terms cancel under η+=η−, not 2λη²h2. The paper's ODE (22) and the subsequent formulas for h0 in (17) and (24) replace the second moment E[ζ²] by η² without any assumption such as Var(ζ)=0. Consequently Theorem 2.1's closed form γ=2rληS0−λη²b−2λη²√(kϕ) is not a consequence of the stated assumptions. For equal second moments m2=E[ζ²] the corrected constant is γ=2rληS0−λm2b−2λm2√(kϕ). This is load-bearing because the formula is the paper's main quantitative result and because the calibration recipe in §3 estimates only η and would be materially wrong for heavy-tailed liquidation sizes. The numerical section uses ζ∼N(10,0.5), for which m2=100.25 is very close to η²=100, so the error is invisible in the reported simulations.","section":"§5.2, Eq. (21)–(22); §5.4, Theorem 2.1; §3"},{"comment":"The proof of Theorem 2.1 computes the limits lim β→0 βvβ(q) and lim T→∞ (1/T)v(0,q;T) and then states that the ergodic constant γ of problem (12) equals this common limit, but the justification is deferred to '[7, Theorem 5]' and the argument is omitted. Since the state process and jump structure here are not identical to those in [7], and since [7] is an unpublished preprint by the authors, this leaves a gap in the proof of the paper's central identification result. The same reliance appears in the proof of Corollary 2.1, which refers to '[7, Appendix A.3]' for the ergodic HJB equation. Please provide a self-contained verification argument, or at minimum state the precise ergodic theorem used and verify its hypotheses for the present model.","section":"§5.4–§5.5, Theorem 2.1 and Corollary 2.1"}],"minor_comments":[{"comment":"The caption contains the typo 'left penal' instead of 'left panel'.","section":"§2.2, Figure 1 caption"},{"comment":"The symbol M is used both for the number of fixed trade sizes Qj and for the number of subintervals in the average defining k; this should be disambiguated, and the formula k=(1/M)Σ k(i) should use the number of subintervals, with a different symbol.","section":"§3, parameter calibration"},{"comment":"The condition 'α ≫ b, k' for the terminal penalty is imprecise. The formulas require the denominator in ξ to be positive, i.e. α > b/2 + √(kϕ), so that ξ>1 and the solution has no singularity on [0,T].","section":"§5.2, Proposition 5.1"},{"comment":"In the dimension reduction, the replacement of the random jump sizes ζ by their means η in the compensated Poisson terms should be explained as taking compensators; as written, it may suggest that the jump sizes are deterministic.","section":"§5.1, Eq. (16)"},{"comment":"The conclusion that σ has no effect on γ under the original dynamics (4) is based on a single parameter configuration and a finite number of Monte Carlo trajectories; the text should hedge accordingly.","section":"§2.2, Figure 3 and surrounding text"},{"comment":"The in-text citation 'Bertsimas and Lo [6]' does not match the three-author reference [6]; please check the bibliographic details.","section":"References, [6]"}],"recommendation":"major_revision","confidential_remarks":"The central proof step γ=γ̂ is outsourced to [7], a preprint by the same research group, and the second-moment substitution affects the main formula in a way that is easy to fix mathematically but pervasive in the manuscript. I would ask the editor to require a self-contained proof of the ergodic identification and a correction of η² to the appropriate second moment (or an explicit deterministic-size assumption) throughout the abstract, Theorem 2.1, the h0 formulas, and the calibration section before considering publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: the paper's main quantitative result, the closed-form ergodic constant in Theorem 2.1, is not correct under the stated assumptions. The optimal policy, however, is right. The error is in the jump terms: the HJB contains λE[u(q+ζ⁺)-u(q)] + λE[u(q-ζ⁻)-u(q)], and with the quadratic ansatz that yields 2λh2·E[ζ²], not 2λη²h2. The paper silently uses η² everywhere, which is only valid if the liquidation sizes are degenerate. Equation (16) makes the same substitution in the martingale integrands. So the gamma formula should read γ = 2rληS0 - λE[ζ²]b - 2λE[ζ²]√(kφ), not (8). The policy ν*(q) = √(φ/k)q survives because the q-linear parts of the jump expectations cancel, but the reward formula and the calibration recipe in Section 3 (which estimates only η) are wrong for non-degenerate sizes. The numerical example uses N(10,0.5), so E[ζ²]=100.25 vs η²=100, and the error is invisible; with realistic heavy-tailed liquidation sizes it would be material.\n\nNow the credit. The ergodic formulation of a liquidation problem with random incoming positions is genuinely new, as far as I know. The problem is well posed, the finite-horizon and discounted solutions are worked out carefully with verification theorems, and the control intuition—always trade a fixed fraction of inventory—is clean and practically useful. The paper is honest about many of its simplifications; it flags the S0 approximation and the linear-impact choice explicitly. The exposition is clear and the proof structure is easy to follow, apart from the final step in Theorem 2.1, which is outsourced to a self-cited preprint; that is a minor issue compared with the E[ζ²] slip.\n\nThe paper deserves a serious referee: the problem is relevant, the method is standard but competently executed, and the main error is fixable. I would send it back to the authors with a clear request to replace η² with E[ζ²] throughout and to extend the calibration section to estimate the second moment. If that is done, the central formula becomes correct as stated. As it stands, I would not cite it.","headline":"The paper gets the optimal policy right but the headline ergodic-reward formula wrong: it substitutes E[zeta]^2 for E[zeta^2] in the HJB jump terms, so the closed-form gamma in Theorem 2.1 needs correction.","tokens_in":15938,"tokens_out":3470,"would_cite":false,"duration_ms":30252,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93E20","91G80"],"pacs":[],"model":"deepseek-v4-flash","headline":"Under a simplified market model, a DeFi exchange's optimal liquidation policy is to sell the fixed fraction $\\sqrt{\\phi/k}$ of its current inventory per unit time, yielding the closed-form average profit $\\gamma = 2 r \\lambda \\eta S_0…","keywords":["ergodic control","optimal liquidation","decentralised finance","price impact","liquidation risk","closed-form solution","Hamilton-Jacobi-Bellman","optimal execution"],"falsifier":"Run the same controlled dynamics (4) with $S_0$ varied over a wide range (say from $1$ to $100$) while holding other parameters fixed, compute the long-run average reward under the control $\\nu^*(q) = \\sqrt{\\phi/k}\\, q$, and check whether it equals $\\gamma$ from (8) within Monte-Carlo error; a systematic drift away from (8), or a clear dependence of the optimal control's coefficient on $S_t$, would refute the central claim's practical validity. A sharper test is to solve the HJB for the original dynamics (4) numerically and compare the optimal feedback map to the linear map $\\sqrt{\\phi/k}\\, q$.","tokens_in":14935,"feed_emoji":"📉","tokens_out":4709,"duration_ms":35029,"temperature":0.7,"pith_summary":"Decentralised derivatives exchanges must dispose of positions acquired when traders fail to maintain margin. This paper treats that disposal problem as an ergodic stochastic control problem and proves, under linear price-impact and simplified cash-balance dynamics, that the optimal long-run strategy is to sell a constant fraction $\\sqrt{\\phi/k}$ of the current inventory per unit time. It also derives a closed-form expression for the associated optimal long-run average profit, which depends only on market parameters and not on initial conditions or the current price. The authors argue this gives exchanges a simple, implementable rule that balances immediate execution against temporary and permanent price impact, and they show how to calibrate all parameters from order-book and liquidation data. The practical value is a liquidation policy that maximises insurance-pool profitability while dampening price moves caused by forced selling.","feed_headline":"DeFi liquidation: always sell the same fraction of inventory","feed_subtitle":"Closed-form ergodic control yields the optimal trading speed and average profit, calibrated from market data.","key_machinery":"The argument is carried by the ergodic control / Hamilton–Jacobi–Bellman framework. Under the simplifications, the value function decouples from the price level and the problem reduces to a one-dimensional controlled jump-diffusion for inventory $q$. The paper solves the finite-horizon and discounted infinite-horizon versions with a quadratic ansatz $u(t,q) = h_0(t) + h_2(t) q^2$ (and similarly $u(q) = h_0 + h_2 q^2$ in the discounted case), then lets the discount $\\beta$ tend to $0$ to obtain the ergodic constant. The optimal feedback control from the first-order condition is $\\nu^* = -(b q + \\partial_q u)/(2k)$, which with the chosen root $h_2 = -\\sqrt{k\\phi} - b/2$ collapses to $\\nu^*(q) = \\sqrt{\\phi/k}\\, q$. The key mechanism is therefore the quadratic structure of the value function plus the vanishing-discount limit, which turns a stochastic control problem into simple algebra.","core_discovery":"The paper claims that the ergodic optimal liquidation problem in DeFi, after three simplifications (fixed-price cash inflows, symmetric long/short liquidation flows, linear price impacts), has a complete closed-form solution. Theorem 2.1 identifies the ergodic constant, the optimal long-run average reward, as $\\gamma = 2 r \\lambda \\eta S_0 - \\lambda \\eta^2 b - 2 \\lambda \\eta^2 \\sqrt{k\\phi}$, and Corollary 2.1 gives the optimal Markov control $\\nu^*(q) = \\sqrt{\\phi/k}\\, q$. In words, whatever inventory $q$ the exchange holds, it should trade the same fraction $\\sqrt{\\phi/k}$ of that inventory per unit time, and the resulting average profit per unit time is the formula above. The strategy depends only on the risk-aversion parameter $\\phi$ and the temporary price-impact coefficient $k$, so it is robust to misspecification of the other parameters; the average reward, meanwhile, is independent of initial state and of volatility $\\sigma$.","pith_inferences":["If the proportional-inventory policy survives contact with the original price-dependent dynamics beyond the single parameter set tested here, it would give liquidation desks a rule that is essentially parameter-light: only the temporary price impact and risk aversion matter.","The fixed-price simplification is most strained in a fast crash, where $S_t$ can be far from $S_0$; a natural stress test would re-solve or simulate the original dynamics with $S_0$ varying over a wide range, or with jumps in the price, to see when the closed-form $\\gamma$ breaks down.","The model's symmetry assumption ($\\lambda^+ = \\lambda^-$, $\\eta^+ = \\eta^-$) is convenient but may be violated in practice, e.g. in a one-sided market; an asymmetric extension would likely keep the quadratic ansatz but change the constants $h_0$ and $h_2$, and would be a feasible next step.","The same framework could be applied to lending-protocol liquidators by reinterpreting $r$ as the liquidation bonus, which the paper itself notes; one could then compare the ergodic reward of a protocol's auction design against this continuous benchmark."],"forward_implications":["An exchange that follows the policy will, in the model, sell a fraction $\\sqrt{\\phi/k}$ of its inventory per unit time at all times, never waiting for better prices.","The ergodic constant $\\gamma$ gives an explicit forecast of the insurance pool's long-run average earnings, letting an exchange set margin levels (inverse leverage $r$) to keep that balance healthy.","Because the optimal control is independent of $\\lambda$, $\\eta$, $b$, and $S_0$, an exchange that misestimates liquidation frequency, position sizes, permanent impact, or price level still executes the optimal trading speed, as long as $k$ and $\\phi$ are right.","Volatility $\\sigma$ does not appear in $\\gamma$, and the numerical experiments indicate this remains approximately true under the original dynamics, meaning higher volatility raises tail risk (VaR/ES) but not average profit.","The calibration section gives a data pipeline (MLE for $\\lambda$, $\\eta$; order-book walk for $k$; regression for $b$) so the model can be fitted to real exchange data."],"supporting_citations":[{"why":"Supplies the ergodic stochastic control method, specifically the vanishing-discount and long-horizon limits used to define the ergodic constant.","marker":"[3]"},{"why":"Provides the proof template for showing that the ergodic constant equals the limit of the discounted value, used directly in the proof of Theorem 2.1.","marker":"[7]"},{"why":"Justifies the linear temporary and permanent price-impact assumptions and the terminal penalty choice adopted in the model.","marker":"[9]"},{"why":"Supplies the martingale theorems used to show that stochastic integrals with respect to compensated Poisson processes and Brownian motion have zero expectation under the admissible controls.","marker":"[26]"},{"why":"Provides the maximum-likelihood estimation methodology used to calibrate the liquidation intensity and average position size.","marker":"[10]"}],"fun_headline_variants":["DeFi liquidation's optimal strategy: sell a constant fraction of inventory","Ergodic optimal control yields a fixed-fraction liquidation rule","Closed-form DeFi liquidation: always trade the same fraction of holdings","Optimal DeFi liquidation: a single fraction governs all trades","DeFi liquidation simplified to a constant sell-speed rule"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing simplification is that cash inflows from liquidations are valued at the initial price $S_0$ instead of the current market price $S_t$; the paper supports this only with a single-parameter-set numerical comparison, so if that approximation fails over long horizons the closed-form $\\gamma$ and the proportional policy may not be optimal in the original model.","fun_headline_variants_meta":{"raw":{"variants":["DeFi liquidation's optimal strategy: sell a constant fraction of inventory","Ergodic optimal control yields a fixed-fraction liquidation rule","Closed-form DeFi liquidation: always trade the same fraction of holdings","Optimal DeFi liquidation: a single fraction governs all trades","DeFi liquidation simplified to a constant sell-speed rule"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000838,"raw_usage":{"total_tokens":3627,"prompt_tokens":892,"completion_tokens":2735,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":508,"completion_tokens_details":{"reasoning_tokens":2649}},"tokens_in":508,"tokens_out":2735,"duration_ms":17690,"temperature":1.0,"reasoning_tokens":2649,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:59:10.757582+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same controlled dynamics (4) with $S_0$ varied over a wide range (say from $1$ to $100$) while holding other parameters fixed, compute the long-run average reward under the control $\\nu^*(q) = \\sqrt{\\phi/k}\\, q$, and check whether it equals $\\gamma$ from (8) within Monte-Carlo error; a systematic drift away from (8), or a clear dependence of the optimal control's coefficient on $S_t$, would refute the central claim's practical validity. A sharper test is to solve the HJB for the original dynamics (4) numerically and compare the optimal feedback map to the linear map $\\sqrt{\\phi/k}\\, q$.","supporting_citations":[{"cited_title":"Arisawa and P.-L","cited_arxiv_id":null,"evidence_quote":"Supplies the ergodic stochastic control method, specifically the vanishing-discount and long-horizon limits used to define the ergodic constant."},{"cited_title":"Cartea, S","cited_arxiv_id":null,"evidence_quote":"Justifies the linear temporary and permanent price-impact assumptions and the terminal penalty choice adopted in the model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the martingale theorems used to show that stochastic integrals with respect to compensated Poisson processes and Brownian motion have zero expectation under the admissible controls."},{"cited_title":"Casella and R","cited_arxiv_id":null,"evidence_quote":"Provides the maximum-likelihood estimation methodology used to calibrate the liquidation intensity and average position size."}],"review_version":1}