{"id":"f2dc9895-b1b4-4e99-a6ab-1cf1e8be7577","arxiv_id":"2504.16542","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"An LP interval-width optimization model, solved with sample average approximation on Uniswap v3 data, finds that full-range liquidity provision would have been the most profitable choice in the analyzed USDC/ETH period.","lead":"This paper builds a stochastic optimization model that chooses the width of a Uniswap v3 liquidity position, balancing trading fees, divergence loss, and reallocation costs. Applied to two years of USDC/ETH data, the model finds that full-range, non-concentrated liquidity provision beats any concentrated interval, and that providing liquidity was less profitable than simply holding the tokens.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported optimum α≈4 is the imposed upper bound, not an interior trade-off solution: Fig 3 is monotone increasing in α, Table II optima attach to the bound with seed-dependent spread (mins ≈1.01, std up to 0.95), and the validation 'alignment' compares the widest allowed choice with the widest…","rationale":"The paper is transparent about parameters, data window, and limitations, reports honestly that providing liquidity lost to holding tokens in this period, and links code; it deserves credit for that. My concern is about whether the demonstration actually supports the method claim. The reader's weakest assumption—constant median cf with linear interpolation of the active fraction (Section III-C, eq. 3p)—is real, and the Section V-C caveat is apt. But it is not the most load-bearing issue for the central claim: the interpolation error overestimates rewards mainly for small α, which biases the optimizer toward small α; the optimizer nonetheless returns the widest allowed α, so the headline conclusion survives that error directionally. The load-bearing issue is that the reported optimum is the imposed bound. On the realized path, profit rises monotonically with α (Fig. 3), and the authors' own comparison ('full range ... 71%, compared to 65% for α=4') places the true optimum beyond the constraint. In the SAA, Table II shows the argmax at the bound in the median with huge seed spread (T=10: std ≈0.93–0.95, minima ≈1.02–1.04), so the objective is flat or regime-dependent rather than a robust trade-off surface. The Section IV-B validation—model output 'aligns' with the retrospective best—is consequently a comparison of the widest allowed choice with the widest possible choice, and cannot establish that the model balances the three costs. The model's actual message is 'as wide as feasible,' which is a useful, honestly-scoped empirical finding but does not validate the optimization approach as a tool for computing optimal concentration. I did not find a decisive inconsistency in the appendix wealth equation (3k)–(3n): the reallocation conservation constraint is internally coherent on its face, although an author clarification is warranted. I kept the verdict CONDITIONAL/UNCHANGED: if the bound-removal test returns an interior optimum near 4, the method is largely vindicated and the verdict could move toward ACCEPT; if the argmax tracks the new bound, Section IV-B's validation claim should be reframed or withdrawn. Neither outcome supports REJECT given the paper's transparency and reproducibility.","tokens_in":13769,"tokens_out":23411,"duration_ms":231761,"concrete_test":"Re-solve formulation (3) with identical inputs (sample σ, median cf, c_g = $109.8, c_tr = 0.05%, S=30, T=10) but with the upper bound on α removed or raised to ≥10^4, across ≥50 seeds; report the distribution of SAA-optimal α and the objective value at α=4 versus at the unbounded argmax and at the full-range limit (α→∞ in eq. (1)). If the argmax tracks the new bound, or if the objective strictly increases past α=4, then the published 'optimal α≈4' (Table II) is a constraint artifact: the model's actual recommendation is 'as wide as feasible,' and the Section IV-B alignment with the hindsight-optimal interval is not a validation of the trade-off model. If the argmax remains interior near α=4 across seeds, the bound concern is refuted and the CONDITIONAL verdict can be relaxed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that the stochastic optimization in eqs. (2)–(3) computes an interval width α balancing liquidity rewards, divergence loss, and reallocation costs, and that the Uniswap v3 demonstration validates it—rests on the reported 'optimal' output α≈4. That output is the solver's upper bound, imposed by the authors in Section IV-A ('We set the upper bound to 4...'), not a point where the modeled trade-offs balance. For the realized path, Fig. 3 is monotonically increasing in α over [1.01, 4] for all γ, and the authors note that 'going full range' yields 71% versus 65% at α=4—so the true hindsight optimum lies beyond the constraint. For the SAA, Table II shows the optimum attaching to the bound in the median but spanning the feasible range across seeds (T=10: std ≈0.93–0.95, minima ≈1.02–1.04), so the objective is flat or regime-dependent, not a robust trade-off surface. The claimed validation in Section IV-B—'the optimal solution was consistently set to ≈4; equivalent to the value for α that an LP should have chosen in retrospect'—therefore compares the widest interval the solver was allowed to select with the widest interval possible. It does not show the model resolves a rewards-versus-divergence-loss-versus-reallocation trade-off, because over the tested range wider is always better. The demonstration's honest content is that for this pool and period, fees never compensate divergence loss plus reallocation costs at any tested α; that is a market-specific empirical finding (consistent with Section V-A) and does not require or validate the optimization machinery. The zero-drift GBM (Section III-C) with 5–10-step SAA paths (σ≈0.006/hr) also cannot probe the 23-month, nearly-tripling realization, and the assertion that the trend 'does not substantially bias the results' is untested. Whether the model computes a meaningful interior optimum therefore remains open: the reported α≈4 cannot distinguish a genuine optimum from a truncation artifact.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper models a single Uniswap v3 liquidity provider's choice of interval width α as a stochastic optimization problem. The objective is expected terminal wealth in USD, comprising token holdings and accumulated fees, under a reallocation rule triggered when the price leaves the interval. The model accounts for liquidity rewards, divergence loss, and reallocation costs. The authors solve a sample-average approximation with Gurobi using hourly on-chain data from the 0.05% USDC/ETH pool (January 2023 to November 2024) and report an optimal α of approximately 4, which they interpret as full-range provision. The empirical conclusion is that full-range provision would have been most profitable but that liquidity provision still underperformed simply holding the tokens.","tokens_in":14198,"tokens_out":6705,"duration_ms":64200,"significance":"If the formulation were correct and the empirical result robust, the paper would offer a useful decision-support tool for LPs and a surprising counterpoint to the common view that concentrated ranges are the most profitable. The authors provide an explicit mixed-integer nonlinear formulation, an SAA scheme, a real dataset, and open-source code, which are notable strengths. However, the central numerical claim is weakened by a dimensional inconsistency in the appendix formulation and by the fact that α≈4 is the solver's imposed upper bound, not an interior trade-off point. The validation is in-sample and compares the widest allowed range with the widest possible range. Consequently, the paper's main claim—that the optimization resolves the rewards-versus-divergence-loss-versus-reallocation-cost trade-off—is not currently demonstrated.","major_comments":[{"comment":"Equation (3k) is dimensionally inconsistent with the wealth definition used in the objective (3a). The objective values terminal wealth as π_T^2 \\tilde{x}^r_T + \\tilde{y}^r_T + Σ^f_T, but (3k) writes the wealth balance at reallocation times using π_t/2 x^r_t + y^r_t and π_t/2 \\tilde{x}^r_t + \\tilde{y}^r_t. Since π_t is the square root of the price, π_t/2 has units of sqrt(price), not price, so the equality cannot hold as a USD wealth conservation law; the rebalancing cost term ctr |y_t^r − \\tilde{y}_t^r| also lacks the appropriate price conversion. This is not a notational slip: the SAA results in Table II are obtained by solving this formulation, so the numerical results must be recomputed with a corrected eq. (3k).","section":"Appendix, eq. (3k)"},{"comment":"The reported optimum α≈4 is the imposed upper bound, not an interior solution. Section IV-A states that the upper bound was set to 4 because values above 1.73 result in no reallocation and that α=4 can be interpreted as going full range. Figure 3 shows terminal profit increasing monotonically for α≥1.73 on the realized path, and the text states that full-range provision gives 71% profit versus 65% at α=4, so the hindsight optimum lies outside the feasible set. In Table II, all medians are 4.00 but minima range from 1.01 to 1.04 with standard deviations up to 0.95, indicating a flat or regime-dependent SAA objective rather than a robust trade-off surface. The validation in Section IV-B therefore compares the widest allowed interval with the widest possible interval and does not establish that the optimization balances rewards, divergence loss, and reallocation costs.","section":"Section IV-A, IV-B, Fig. 3, Table II"},{"comment":"The activity-fraction interpolation used in eq. (3p) can materially overestimate expected liquidity rewards for small α, as the authors themselves acknowledge in Section V-C. Since the objective is precisely what determines the optimal α, and since the finding that α=4 is optimal depends on the relative magnitude of rewards, the paper should quantify the bias (for example, with Brownian bridge estimates or intra-hour price data) or show that the optimum is insensitive to this modeling choice. Without such an analysis, the empirical demonstration is not robust.","section":"Section V-C and eq. (3p)"},{"comment":"The price process is assumed to be GBM with zero drift, but the sample period has ETH rising from approximately $1,350 to $3,725. The claim in Section IV-B that this trend does not substantially bias the results is not adequately supported: both terminal wealth and the comparison with holding assets depend directly on the realized appreciation, while the zero-drift SAA paths make such appreciation unlikely. The authors should either fit a drift, allow time-dependent α_t, or provide a concrete counterfactual or robustness check for the zero-drift assumption before the empirical conclusion can be accepted.","section":"Section III-C and Section IV-A"}],"minor_comments":[{"comment":"The second case of eq. (1) states the condition as π_l ≤ π ≤ π_l, which appears to be a typo for π_l ≤ π ≤ π_u.","section":"Section III-A, eq. (1)"},{"comment":"The main text refers to 'eq. (3) in section VI-A' for the tractable formulation, but the formulation is in the Appendix, not in Section VI-A; the cross-reference should be corrected.","section":"Section III-C and Appendix"},{"comment":"The GitHub URL contains spaces ('AMM Liquidity Concentration Optimization') and should be given as a proper hyperlink or with URL encoding.","section":"Section IV-A"},{"comment":"The text says α=4 'can be interpreted as going full range' but then reports a distinct profit of 71% for going full range versus 65% for α=4; this internal inconsistency should be clarified.","section":"Section IV-B"},{"comment":"The wealth variable Y_t is used in formulation (2) but is not explicitly defined in the main text; the definition appears only in the appendix objective (3a), and it would help to define Y_t in Section III-C.","section":"Section III-C"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern is accurate: the reported optimum is a boundary artifact, and the appendix formulation contains a dimensional error. The paper's empirical finding that full-range provision is best for this pool and period is interesting, but the manuscript should be revised to correct eq. (3k), extend or justify the α bounds, and provide a genuinely out-of-sample or at least sensitivity-based validation before the central claim can be considered established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nShort version: the paper adds a real piece to the LP interval-selection literature—an SAA-based MINLP that optimizes interval width α under a discrete reallocation rule with costs—but its headline result α≈4 is an artifact of the imposed upper bound, not a demonstrated trade-off. On the data shown, profits rise monotonically with α up to the bound, and the authors themselves note that full-range provision would have done better (71% vs 65%). So the central validation doesn't land.\n\nWhat's genuinely new and good: unlike Cartea et al.'s continuous-time model or Fan et al.'s fixed-width SAA, this formulation makes α a decision variable inside a discrete-reallocation problem, and it accounts for reallocation costs. The appendix gives a solvable form, code is on GitHub, and the empirical work uses real on-chain USDC/ETH data. The paper is also honest about limitations: it states that LPs were better off holding, that fee interpolation can overestimate rewards, and that the zero-drift GBM assumption is a simplification. The related-work coverage looks fair.\n\nThe soft spots are not small. The upper bound α≤4 is chosen from the same data (α>1.73 implies no reallocation), and the solver returns the bound in most seeds. Fig. 3 shows terminal profit increasing across the entire feasible range, so there is no interior optimum—wider is better up to the constraint. The validation then compares the SAA solution to the hindsight-optimal choice under the same model, which is in-sample consistency, not external confirmation. A zero-drift GBM cannot probe a 23-month period where ETH nearly tripled, and the claim that the trend doesn't bias results is asserted, not tested. I'd also flag the appendix eq. (3k) as looking dimensionally off with π_t/2 instead of π_t^2, though the rendering makes this hard to confirm from the text.\n\nThat said, the empirical caution—that concentrating liquidity in this pool over this period would have lost to simply holding—is worth having, and the model structure can be built on. Who is this for? People working on LP strategies and AMM optimization; they'll want to fix the validation and test whether an interior α exists when the bound is removed.\n\nFor peer review: yes, send it out—it's a serious formulation with real data—but major revision is needed. The authors should re-run without the α≤4 cap, or at least report whether the objective turns over, and should do an out-of-sample or at least out-of-sample-period check before claiming the model recovers the hindsight optimum.\n\nYours,","headline":"A useful LP-interval formulation undercut by a bound artifact: the reported optimal α is the imposed upper bound, not a resolved trade-off.","tokens_in":14848,"tokens_out":3480,"would_cite":true,"duration_ms":32170,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that on Uniswap v3's USDC/ETH pool, the optimal liquidity interval is effectively full range (α ≈ 4), and even that best case underperformed simply holding the tokens.","keywords":["automated market makers","Uniswap v3","concentrated liquidity","liquidity provision","stochastic optimization","divergence loss","reallocation costs","sample average approximation"],"falsifier":"On the same 0.05% USDC/ETH pool, replace hourly linear price interpolation in equation (3p) with Brownian-bridge occupation-time estimates of the active fraction and re-solve the sample average approximation; if the optimal α falls well below 4, or if any finite α beats both full-range provision and the passive holding benchmark, the paper's central conclusion is falsified.","tokens_in":13593,"feed_emoji":"📉","tokens_out":8303,"duration_ms":76951,"temperature":0.7,"pith_summary":"This paper tries to turn an automated market maker's concentrated-liquidity interval choice into a solvable stochastic optimization problem. It models a single liquidity provider who reallocates only when the token price leaves the chosen interval, and maximizes terminal wealth over an interval-size parameter α while accounting for liquidity rewards, divergence loss, and reallocation costs. Applied to hourly on-chain data from Uniswap v3's 0.05% USDC/ETH pool for 2023–2024, the model's optimal α saturates at the upper bound 4, which the authors equate with full-range provision; they conclude that wide, full-range intervals would have been most profitable and that even the optimal interval would have earned less than passively holding the tokens. The contribution is a tractable formulation that can be adapted to other price processes, fee tiers, and cost structures, not a general claim that concentration is useless.","feed_headline":"Even best Uniswap v3 liquidity lost to holding tokens","feed_subtitle":"A stochastic optimization model says full-range positions beat narrow ones on USDC/ETH—yet still trail holding tokens.","key_machinery":"The load-bearing object is a mixed-integer nonlinear stochastic program (equations (2) and (3)) whose single decision variable is α, the half-width of the square-root-price interval [π/α, απ]. A binary variable z_t records whether the price has left the interval and forces a reallocation; the new interval is then recentered at the current price. Liquidity rewards are computed as a constant reward rate c_f times the linearly interpolated fraction of each hourly step spent inside the interval (equation (3p)), terminal wealth is the value of the real token reserves from equation (1) plus accumulated unclaimed fees, and sample average approximation converts the expectation into an average over simulated geometric Brownian motion paths with zero drift. This machinery lets a commercial mixed-integer solver return a single number α that balances the three competing effects.","core_discovery":"The paper's central claim is that optimal liquidity provision on a concentrated-liquidity constant-product AMM can be stated as a finite-horizon stochastic optimization over the interval scale α, with bounds [π/α, απ] on the square-root price; the objective is expected terminal wealth, and the trade-off is driven by three competing terms: liquidity rewards grow with narrowness, divergence loss grows as the interval narrows, and reallocation costs grow with how often the price exits the interval. Solving the resulting mixed-integer nonlinear program by sample average approximation on the Uniswap v3 0.05% USDC/ETH pool (Jan 11, 2023 to Nov 30, 2024) gives an optimal α ≈ 4 across seeds and sample sizes. Because an interval of that width covers price moves from 1/16 to 16 times the starting price, the authors identify α ≈ 4 with full-range provision and claim that full-range intervals would have been the most profitable choice in retrospect. They further claim that liquidity provision even at this optimal width would have returned less than holding the initial token basket, because fee income did not compensate divergence loss.","pith_inferences":["Correcting the paper's flagged overestimation of small-α rewards—by using Brownian-bridge occupation times instead of linear interpolation—would likely push the optimal α even closer to full range, since narrow intervals would lose their overstated reward advantage.","Re-running the model with the empirical upward drift of ETH would be a natural test: the passive basket captures that drift directly, while the LP's exposure is path-dependent, so the holding benchmark and the optimal α could both shift.","In a range-bound or mean-reverting market, narrow intervals might plausibly beat full-range provision; the paper's demonstration period has a near-tripling price path, so the full-range conclusion should be read as period-specific until other regimes are tested."],"forward_implications":["If the model is right, narrower is not automatically better: the optimal interval sits where the marginal reward from concentration equals the marginal divergence loss plus expected reallocation cost, and for this pool that point is the full range.","The same formulation can be rerun for other pools, fee tiers, gas prices, or price processes by changing inputs or the simulated paths, so the method yields a pool-specific α rather than a universal rule.","Delaying reallocation until the price is a threshold distance outside the interval (γ > 0) can raise profits for narrow intervals by cutting reallocation frequency; the data does not settle whether γ helps when α is already optimal.","The model's forward-looking optimum matched the retrospectively best interval on the same data, which is the main evidence that the tractable formulation captures the reward–loss–cost trade-off well enough to guide liquidity providers.","Even at the optimum, the demonstration found liquidity provision less profitable than not providing liquidity, implying that fee rewards alone did not compensate divergence loss over a period in which ETH roughly tripled."],"supporting_citations":[{"why":"Defines Uniswap v3 concentrated-liquidity positions, active-position fee accrual, and the lower bound on interval width used in the model.","marker":"[2]"},{"why":"Supplies the continuous-time optimal-liquidity framework, the GBM price assumption, the α parameterization, and the transaction-cost estimate used in the demonstration.","marker":"[6]"},{"why":"Establishes that narrower concentrated-liquidity intervals amplify divergence loss, a key term in the trade-off.","marker":"[9]"},{"why":"Provides the sample-average-approximation approach to LP profit optimization that the paper adapts.","marker":"[10]"},{"why":"Supplies the ULRA reallocation rule (reallocate when price exits the interval) that the paper implements and the fixed-interval baseline it extends.","marker":"[11]"},{"why":"Provides the real-reserve formulas (the paper's equation (1)) linking token holdings to the interval bounds.","marker":"[15]"},{"why":"Provides the sample average approximation method that replaces the expected terminal wealth with a sample average over simulated paths.","marker":"[19]"},{"why":"Supplies the Big-M linearization used to reformulate the mixed-integer nonlinear program for commercial solvers.","marker":"[20]"}],"fun_headline_variants":["Optimal Uniswap v3 range? Still lose to HODLing","Best AMM strategy: hold tokens, not provide liquidity","Stochastic model: full-range wins, but HODL still better","Uniswap v3 LPs: even optimal intervals underperform holding","Full-range liquidity optimal? Paper says HODL still better"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model treats the hourly reward rate as a constant and interpolates price linearly between hourly observations to decide when the position was active; if fee income is bursty or price moves nonlinearly within the hour, the optimal interval size could be different.","fun_headline_variants_meta":{"raw":{"variants":["Optimal Uniswap v3 range? Still lose to HODLing","Best AMM strategy: hold tokens, not provide liquidity","Stochastic model: full-range wins, but HODL still better","Uniswap v3 LPs: even optimal intervals underperform holding","Full-range liquidity optimal? Paper says HODL still better"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000699,"raw_usage":{"total_tokens":3147,"prompt_tokens":925,"completion_tokens":2222,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":541,"completion_tokens_details":{"reasoning_tokens":2129}},"tokens_in":541,"tokens_out":2222,"duration_ms":13932,"temperature":1.0,"reasoning_tokens":2129,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:01:29.697949+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the same 0.05% USDC/ETH pool, replace hourly linear price interpolation in equation (3p) with Brownian-bridge occupation-time estimates of the active fraction and re-solve the sample average approximation; if the optimal α falls well below 4, or if any finite α beats both full-range provision and the passive holding benchmark, the paper's central conclusion is falsified.","supporting_citations":[{"cited_title":"Uniswap v3 Core,","cited_arxiv_id":null,"evidence_quote":"Defines Uniswap v3 concentrated-liquidity positions, active-position fee accrual, and the lower bound on interval width used in the model."},{"cited_title":"Decentralized finance and auto- mated market making: Predictable loss and optimal liquidity provision,","cited_arxiv_id":null,"evidence_quote":"Supplies the continuous-time optimal-liquidity framework, the GBM price assumption, the α parameterization, and the transaction-cost estimate used in the demonstration."},{"cited_title":"Differential Liquidity Provision in Uniswap v3 and Implications for Contract Design,","cited_arxiv_id":null,"evidence_quote":"Provides the sample-average-approximation approach to LP profit optimization that the paper adapts."},{"cited_title":"Strategic Liquidity Provision in Uniswap V3,","cited_arxiv_id":null,"evidence_quote":"Supplies the ULRA reallocation rule (reallocate when price exits the interval) that the paper implements and the fixed-interval baseline it extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Big-M linearization used to reformulate the mixed-integer nonlinear program for commercial solvers."}],"review_version":1}