Research
Pools as Portfolios: Observed arbitrage efficiency & LVR analysis of dynamic weight AMMs
A dynamic-weight AMM runs a portfolio strategy on-chain: its target weights shift over time, and arbitrageurs who trade the pool back toward those weights effectively perform the rebalancing.

- arXiv
- 2602.22069
- Published
- 2026-02-25
- Authors
- Matthew Willetts, Christian Harrington
AI summary
Overview
Research area: Decentralised finance market microstructure — arbitrage dynamics, MEV, and execution-cost benchmarking for automated market makers (AMMs), specifically dynamic-weight / Temporal Function Market Makers (TFMMs).
Technical level: Intermediate. The empirical results are readable with basic AMM knowledge, but the appendices use G3M invariant mathematics, no-arbitrage band geometry, and the LVR/RVR benchmark framework.
Scope: A block-level empirical study of arbitrage on two live QuantAMM pools (Ethereum mainnet and Base L2) across two short windows six months apart, plus a longer-horizon comparison of pool performance against Loss-versus-Rebalancing (LVR) and Rebalancing-versus-Rebalancing (RVR) benchmarks.
What This Paper Is About
A dynamic-weight AMM runs a portfolio strategy on-chain: its target weights shift over time, and arbitrageurs who trade the pool back toward those weights effectively perform the rebalancing. The pool therefore pays for rebalancing not through an exchange, but through a continuous sequence of small, predictable mispricings that the authors describe as a Dutch reverse auction. The paper measures how well that auction actually works on live pools — how much value arbitrageurs extract, how the pattern changes over six months, and whether the resulting execution beats simulated centralised-exchange (CEX) rebalancing benchmarks.
Key Contributions
- First per-block empirical study of live TFMM pools. The authors scrape per-block reserves, weights, prices and gas for two QuantAMM pools, match every reserve change to its on-chain transaction, and characterise the arbitrage process trade by trade.
- A calibrated, time-varying profitability threshold. Rather than assuming a fixed cost, they build a per-block threshold from the EIP-1559 base fee (burned), the priority fee (to the builder) and optional builder tips, using 450,000 gas units (the median for observed arb trades) and a calibrated markup μ = 1.50 derived from July open-market arbitrage trades.
- Identification of "incidental routing" as a distinct rebalancing regime. On Base, they document pools being rebalanced as a side effect of ecosystem-wide multi-hop arbitrage at essentially zero per-trade extraction, which they argue is a separate mechanism from targeted arbitrage.
- Live-pool benchmarking against LVR and RVR. They compare cumulative pool value against a frictionless CEX rebalancer (LVR) and a realistic CEX rebalancer paying spreads and commissions (RVR), rather than relying on simulation alone.
Main Findings
- Auction compression on Ethereum mainnet. In the Safe Haven pool, a two-hour window on 22 July 2025 (10:00–12:00 UTC, 606 blocks) shows 20 arbitrage trades, roughly one every 30 blocks (~6 minutes), with allocation drift building to 0.5–0.65% between trades. In a two-hour window on 8 January 2026 (01:00–03:00 UTC, 591 blocks) there are 78 trades — roughly 4× more — with drift rarely exceeding 0.45% and trades about every 90 seconds.
- Falling extraction per trade, and falling total extraction. Empirical extraction per trade falls from $2.58 to $0.28, and total empirical extraction over the window falls from $51.55 to $22.00 despite nearly 4× more trades. The theoretical optimum for the two windows is $58.19 and $35.82 respectively.
- High efficiency against the theoretical optimum in July. On mainnet in July, empirical extraction of $51.55 versus a theoretical maximum of $58.19 gives an efficiency ratio of 88.6%. The authors attribute the gap mainly to conservative trade sizing: arbitrageurs execute roughly 60% of the optimal trade size while capturing almost all available profit, which is rational because the G3M profit function is flat near its maximum.
- More trades fall below the standalone threshold over time. All 20 July mainnet trades exceed the realistic threshold; in January, 32% of trades fall below it, which the authors read as evidence that arbitrageurs amortise gas across multi-venue (multi-leg) strategies.
- Routing landscape shifted between windows. By January 2026, Uniswap V4 accounts for 37% of swap events (up from zero), complex MEV bundles with 4+ legs account for 24% of trades, and mean all-in cost per transaction fell from $1.08 to $0.20. The base fee fell roughly 12× to about 0.047 gwei from about 0.55 gwei, with cost share shifting from 82% base fee / 10% priority fee in July to 39% / 59% in January. The authors associate this with the Fusaka upgrade in December 2025.
- Incidental routing on Base L2. In the Base Macro window on 14 January 2026 (08:00–10:00 UTC), the authors observe 202 block-to-block balance changes from 225 on-chain transactions. Mean per-trade empirical profit is $0.0004, the median is below $0.001, and 98% of trades extract less than a cent — many extract less than zero, effectively adding a tiny amount of value to the pool. January transactions average 7 swaps across Uniswap V2, V3 and V4, all to route a $20–80 trade through the pool.
- Base outperforms both benchmarks. Over August 2025 to January 2026, Base Macro shows cumulative outperformance of approximately 27 percentage points versus LVR and 43 percentage points versus RVR. The authors caution that this reflects the combined effect of the strategy's weight changes and execution efficiency, not the rebalancing cost alone.
- Both pools beat the realistic benchmark. Both Safe Haven and Base Macro outperform the RVR benchmark over their respective periods, and the L2 pool additionally outperforms the frictionless LVR benchmark — consistent with per-trade costs that are near or below zero.
- No retail flow detected. In both mainnet windows, every transaction originates from MEV market participants, identified by contract-mediated execution and multi-hop routing through 2+ DEXs. The pool therefore earns no swap fee revenue from non-arbitrage flow; returns derive entirely from the weight strategy net of arbitrage extraction.
- Consolidation did not increase extraction. Unique arbitrageurs on Base Macro fell from 65 in July to 4 dominant bots by January, yet per-trade extraction fell rather than rose — the authors attribute this to the winners being sophisticated multi-venue routers.
Methodology in Plain English
The authors treat each pool as a portfolio whose target weights move block by block. Because the pool's quoted prices drift as the weights interpolate, a gap opens between the pool price and the external market price. That gap grows until it is large enough to cover an arbitrageur's costs, at which point someone trades and the pool snaps back — then the cycle repeats. To measure this, the authors:
- Scrape on-chain data. For each pool and window they collect per-block reserves, weights, prices and gas statistics, and match each balance change to its transaction. Because QuantAMM sits on Balancer V3's vault, which keeps 50% of the pool's swap fee, balance deltas are scaled before calculation.
- Measure mispricing with allocation drift. For multi-token pools (3 or 4 tokens) they use the sum of absolute differences between actual allocation and target weight, which is zero at equilibrium and produces a clean "sawtooth" as drift builds and then resets on a trade — avoiding the need to plot pairwise price comparisons.
- Build a realistic cost threshold. A block's threshold combines the base fee, priority fee and any builder tip, scaled by 450,000 gas units and a markup of 1.50 that captures how competitive the arbitrage market is. Trades above the threshold are plausibly standalone; trades below it are viable only as part of a larger multi-venue route.
- Compare against benchmarks. They compute what a frictionless CEX rebalancer following the identical weight trajectory would have earned (LVR), and what a realistic one paying commissions and crossing spreads would have earned (RVR), then plot cumulative pool value relative to each.
The paper deliberately favours LVR and RVR over impermanent loss, arguing that IL-type measures conflate the choice of strategy with rebalancing efficiency.
Why This Matters
Impact on research. Prior work on dynamic-weight AMMs characterised the mechanism theoretically and in simulation, with some of that work predicting the design could beat CEX-style rebalancing. This paper is the first block-level empirical test on live pools, and it both supports the prediction and identifies a mechanism — incidental routing — that simulation-based studies did not model. The authors also report that simulations in the prior literature appear to have over-estimated retail volume relative to arbitrage volume at the low TVLs studied here, while over-estimating effective gas costs.
Real-world applications:
- Tokenised funds and index products. The results support treating an AMM pool as a viable execution layer for a managed portfolio, where weight changes are implemented by paying arbitrageurs rather than by placing orders.
- Choosing where to deploy a strategy. The mainnet-versus-L2 contrast suggests that lower data costs and dense routing on an L2 can compress arbitrage spreads enough to beat even "perfect" rebalancing benchmarks.
- Cost budgeting for pool designers. The share of trades falling below the standalone profitability threshold gives a simple, observable diagnostic for how much a pool's rebalancing is being subsidised by the surrounding DEX ecosystem.
- MEV and routing analysis. The transaction-structure breakdown (leg counts, DEX venue shares, base-fee versus priority-fee cost decomposition) is useful to anyone modelling searcher behaviour or routing on mainnet versus an L2 sequencer.
Industry relevance. The finding that rebalancing costs can fall over six months with no change to pool fee or design — purely because the routing ecosystem matured — is directly relevant to anyone pricing AMM-based execution, and suggests that strategy infrastructure should be evaluated against a moving baseline rather than a static one.
Future Directions
- Broaden the sample. The authors state explicitly that generalisation would require studying more pools under more market conditions; this study covers two pools and two short windows per pool.
- Model noise trading more accurately. The mismatch between simulated retail-to-arbitrage volume ratios and what was observed at these TVLs is flagged as an open issue, with the possibility that noise-trade assumptions hold at higher TVLs.
- Test non-linear interpolation. The pools studied use linear weight interpolation only; the authors note that non-linear methods exist that could further lower the barrier for dynamic-weight AMMs to beat both realistic and perfect CEX rebalancing.
- Re-examine benchmark design. Since LVR assumes frictionless rebalancing and RVR adds spreads and commissions, and since an L2 pool beat even the frictionless benchmark, the paper implicitly raises the question of what the right benchmark is for execution mechanisms as on-chain costs continue to fall.
Target Audience
Protocol designers and quantitative researchers working on AMM design, dynamic-weight or index-style on-chain strategies, and MEV/arbitrage economics. It is also relevant to market microstructure researchers interested in empirical measurements of auction mechanisms on blockchains, and to portfolio managers evaluating whether on-chain execution can substitute for centralised rebalancing. Readers need comfort with AMM invariants and arbitrage reasoning to get the most from the appendices, though the main results are accessible with a general DeFi background.
Authors’ abstract
Dynamic-weight AMMs (aka Temporal Function Market Makers, TFMMs) implement algorithmic asset allocation, analogous to index or smart beta funds, by continuously updating pools' weights. A strategy updates target weights over time, and arbitrageurs trade the pool back toward those weights. This creates a sequence of small, predictable mispricings that grow until taken, effectively executing rebalances as a series of Dutch reverse auctions. Prior theoretical and simulation…
The full abstract on arXiv