Research
Where Does MEV Really Come From? Revisiting CEXDEX Arbitrage on Ethereum
MEV — the profit extractable by reordering, inserting, or censoring transactions in a block — is one of the central economic questions in the Ethereum ecosystem, and a key policy question is how much…

- arXiv
- 2604.15973
- Published
- 2026-04-17
- Authors
- Bence Ladóczk, Miklós Rásonyi, János Tapolcai
AI summary
Overview
Research area: Blockchain economics and decentralised finance — specifically Maximal Extractable Value (MEV) on Ethereum, automated market makers (AMMs), and stochastic modelling of centralised-exchange–decentralised-exchange (CEX–DEX) arbitrage.
Technical level: Advanced. The paper combines a rigorous discrete-time jump-diffusion framework (ergodic Markov chains, invariant densities, function iteration) with an empirical calibration pipeline, so it assumes comfort with stochastic processes and AMM mechanics.
Scope: The paper builds and calibrates a discrete-time price model with stochastic jumps to explain why earlier Black–Scholes-based theory drastically understated CEX–DEX arbitrage volumes and profits on Ethereum, comparing its predictions against one month of data from a Uniswap V2 ETH–USDT pool.
What This Paper Is About
MEV — the profit extractable by reordering, inserting, or censoring transactions in a block — is one of the central economic questions in the Ethereum ecosystem, and a key policy question is how much of it comes from arbitrage (considered benign and necessary) versus attacks on ordinary users. Earlier theoretical work modelled CEX–DEX arbitrage using continuous-path Black–Scholes-style price processes and predicted only modest volumes and profits, yet newer empirical studies found volumes orders of magnitude larger than atomic (DEX–DEX) arbitrage and sandwich attacks. The authors argue the gap is caused by the continuous-path assumption, which ignores price jumps — precisely the events that create arbitrage opportunities — and they build a model that includes jumps to close the gap.
Key Contributions
-
An extended discrete-time AMM/arbitrage model with jumps. The CEX log-price evolves as a discrete-time general stochastic difference equation consisting of a drift term, a volatility-scaled noise term with arbitrary (possibly non-Gaussian) distribution, and a jump term formed from a Bernoulli jump indicator times an arbitrarily distributed jump amplitude. The authors show this nests the Black–Scholes / geometric Brownian motion model, the Merton jump-diffusion SDE, and the Ornstein–Uhlenbeck SDE as special parameter choices.
-
A convergence and ergodicity theory for the mispricing process. Under constant drift and volatility and a mild assumption that the noise density is bounded away from zero on a finite interval, the paper proves the existence of an invariant density for the log-mispricing process, establishes geometric-speed convergence in L1 (a bound of the form C·ρⁿ), proves a law of large numbers for the process, and derives closed-form expressions for the stationary arbitrage trade probability, the expected arbitrageur profit per time step, and the expected arbitrage trading volume per time step.
-
A numerical solver plus an empirical calibration pipeline. The authors implement the stationary-distribution computation in C++ (grid of 201 points, run for roughly 100 iterations), fit the model's seven parameters to real spot-price and on-chain data, and compare the resulting stationary distribution and implied trading statistics against a histogram of empirical mispricing.
-
Quantitative reconciliation with empirical MEV measurements. For a Uniswap V2 ETH–USDT pool, the fitted jump model predicts arbitrage swap counts, volumes and profits far closer to observed values than the Black–Scholes baseline, and the authors argue that much of the remaining observed swap volume is also arbitrage executed by hidden arbitrageurs.
Main Findings
-
Black–Scholes badly understates the pool's activity. For the Uniswap V2 pool
0x0d4a11d5eeaac28ec3f61d100daf4d40471f1852(TVL $86M) over 13–29 September 2025, the Black–Scholes/LVR estimate is 86 swaps and $208,844 volume in the USDT→WETH direction and 75 swaps and $177,961 in WETH→USDT, with profit estimated at $48. -
The jump model lands much closer to observation. The authors' model yields 256 swaps and $1,881,277 volume (USDT→WETH) and 251 swaps and $1,838,985 (WETH→USDT), with profit $1,454 — versus empirical totals of 2724 swaps / $3,369,073 and 1907 swaps / $3,072,624 across all swap transactions in the pool, and an empirical MEV-based profit estimate of about $2,030.
-
Heuristic CEX–DEX classification captures a large but partial share. A published heuristic attributed 496 swaps / $970,436 and 241 swaps / $1,002,020 to CEX–DEX arbitrage, while sandwich attacks averaged 0.2 swaps per day per direction ($25,345 and $64,443 volume, $414 profit) and atomic arbitrage averaged 3 and 2 swaps ($10,515 and $34,562 volume, $136 profit) — confirming that sandwich and atomic arbitrage are small in aggregate.
-
Observed price jumps dwarf diffusive moves. From two months of Binance Spot price data (November–December 2024), the authors report jumps whose amplitude can be two orders of magnitude larger than instantaneous volatility; the concrete example given is a move from 3542 to 3640 within one second at 2024-12-09 22:06:29.
-
Fitted parameters. With jump threshold τ = 2.0, the estimated daily volatility is 1.9%, drift 0.023, jump probability q = 0.045 per one-second step, with jump mean −24.9% and standard deviation 15.7% in log-return units. Refitting under pure diffusion (q = 0) gives daily volatility 2.1% and drift 0.046. A threshold of τ = 2.0 implies roughly 4.55% of genuine GBM observations are misclassified into the jump sample under a Gaussian assumption.
-
Sensitivity to the CEX bid–ask spread. Because the effective fee γ_CEX is not precisely known, the authors sweep it: at γ_CEX = 0 bp the model's arbitrage volume is 75% of empirically measured total pool volume and profit is 93% of the MEV-based estimate (volume-to-profit ratio 1312); at 10 bp these are 58% / 72% (1264); at 20 bp 46% / 57% (1264); at 30 bp 39% / 48% (1267). A linear-combination fit with weights w₀ = 19%, w₁₀ = 19%, w₂₀ = 28%, w₃₀ = 34% gives 51% / 63% (1274).
-
Swap counts are over-predicted relative to a one-arbitrageur assumption. Observed swap counts exceed the model's prediction, which the authors attribute to many arbitrageurs each seeing slightly different CEX prices, and to widening CEX bid–ask spreads after jumps, so a gap is closed by a sequence of transactions rather than one.
-
Noise traders appear marginal in volume terms. Since the model excludes noise traders entirely yet still matches or over-explains observed volume, the authors conclude that most daily pool volume is driven by the pool tracking the CEX price, and that noise traders play only a minor volumetric role.
-
Real crypto returns are far from Gaussian. Hourly log-return analysis of major cryptocurrency pairs (Appendix 0.F) shows heavy tails, skewness, and uniformly low Kolmogorov–Smirnov p-values, undermining Gaussian-based models.
-
Numerical convergence is fast. Function iteration reaches a distribution visually indistinguishable from the limit after approximately 100 iterations; the authors also verify in Appendix 0.E that the framework reproduces the numerical values reported in Table 1 of the earlier reference [23].
Methodology in Plain English
The authors start from a simple idea: the price of ETH on a centralised exchange and the price quoted by an on-chain AMM drift apart, and arbitrageurs close the gap for profit. The gap can only exist between the AMM's fee thresholds, so its long-run behaviour determines how often and how much arbitrageurs trade.
They model the CEX log-price in discrete one-second steps as three pieces: a drift, a random shake proportional to volatility, and — occasionally — a discrete jump with its own random size. An arbitrageur arrives each step with probability p; if the gap exceeds the fee threshold, the arbitrageur pushes it back to the threshold and pockets the difference. This makes the mispricing a Markov chain: its next state depends only on its current state and fresh random noise.
They then prove that, under mild technical conditions, this chain forgets its starting point at a geometric rate, so a stationary distribution exists and time averages converge to it. That stationary distribution, computed by repeatedly applying the transition operator to a density (function iteration), is what lets them read off expected arbitrage frequency, volume and profit.
For the empirical side, they run their own Ethereum execution client to collect swap events from a large Uniswap V2 ETH–USDT pool over 13–29 September 2025, and pull one-second exchange rates from multiple CEXs (best asks sourced 84% Binance, 9% KuCoin, 7% Gate.io; best bids 88% Binance, 5% KuCoin, 7% Gate.io). They use a 12-second window with an assumed ~8-second block-arrival delay to decide when an arbitrageur could have acted. Parameters are fitted by separating diffusive moves from jumps using a threshold of τ = 2.0 times volatility, then fitting Gaussian curves to the GBM and jump components of the empirical mispricing histogram. Feeding the fitted parameters plus observable pool data (TVL and fee) into the stationary distribution yields predicted swap counts, volumes and profits, which are then compared to the on-chain record.
Why This Matters
Impact on research. The paper challenges a modelling assumption that underpins much of the existing MEV and loss-versus-rebalancing (LVR) literature: that price paths are continuous. If jumps drive most arbitrage, then closed-form Black–Scholes-derived formulas systematically understate the cost of providing liquidity and the value captured by arbitrageurs — and any protocol design or empirical correction built on those formulas inherits the error. The paper also supplies a reusable theoretical toolkit (ergodicity, geometric convergence, closed-form stationary profit and volume expressions) for a much broader class of price processes.
Real-world applications:
- Liquidity provider economics: a sharper estimate of how much of a pool's volume is arbitrage-driven, and therefore what fraction of fee revenue is really compensation for adverse selection rather than genuine flow.
- MEV measurement and accounting: a principled way to estimate how much "dark" MEV is hidden inside swaps that heuristics cannot classify as arbitrage.
- Block builder and searcher economics: understanding why builders pay so much for block-building rights, and how much of that value derives from benign arbitrage versus user-harming strategies.
- Pool metrics and comparison: the authors note that the one-day volume/TVL metric depends on the pair's exchange-rate dynamics and the DEX fee, and appears similar across Uniswap V2 pools pairing a currency with various stablecoins regardless of TVL — relevant to how users and aggregators compare pools.
Industry relevance. The results are directly usable by DEX designers setting fee tiers, by MEV researchers and analytics firms producing arbitrage estimates, by searchers modelling competition, and by anyone reasoning about the "darkness" of MEV for policy or protocol-governance purposes.
Future Directions
-
Estimating the CEX bid–ask spread endogenously. The model is "highly sensitive" to the effective fee γ_CEX, which is treated as constant even though the paper notes it corresponds to an unknown and time-varying CEX spread. Making it stochastic or jointly estimated would remove the largest source of uncertainty in the predictions.
-
Modelling multiple competing arbitrageurs. The model assumes each opportunity is taken by a single arbitrageur and trades with high probability when the gap is large; observed swap counts exceed this, motivating an extension with heterogeneous arbitrageurs observing different CEX prices.
-
Explaining the residual volume gap. At γ_CEX = 0 bp the model still accounts for 75% of total pool volume rather than all of it; the paper's interpretation that the remainder is hidden arbitrage is a hypothesis that empirical work would need to test directly.
-
Generalising beyond constant drift, volatility and Gaussian jumps. The fitted jump-size distribution is Gaussian, the empirical return analysis shows heavy tails and skewness, and the theory assumes constant μ and σ — a natural next step is state-dependent coefficients and non-Gaussian jump laws.
Target Audience
Researchers working on MEV, AMM design and DeFi market microstructure; quantitative finance practitioners interested in jump-diffusion models applied to crypto; applied probabilists interested in ergodicity and stationary distributions of controlled Markov chains; and blockchain data analysts who need a theoretical benchmark for interpreting arbitrage heuristics and pool volume statistics. Readers without a stochastic-processes background will find the empirical comparison in Table 1 and Figure 2 accessible, but the theoretical sections require graduate-level probability.
Authors’ abstract
A central question of the Ethereum ecosystem is where Maximal Extractable Value (MEV)revenue originates and to what extent it stems from harming unsuspecting users. It is acceptable if MEV arises from arbitrages between centralised and decentralised exchanges (CEX-DEX). Yet theoretical models have significantly underestimated the scale of these arbitrages, while empirical studies have highlighted their importance - though these remain conservative estimates, constrained by…
The full abstract on arXiv