# The Origins of MEV: Systematic Attribution of Arbitrage Opportunity Creation at Scale

Crypto.world

> Almost all MEV research asks how value is extracted (arbitrage, sandwich attacks, liquidations) or how to mitigate it. This paper asks the reverse question: which on-chain transactions create the…

**arXiv**: 2604.27979

**Published**: 2026-04-30

**Authors**: Andrei Seoev, Dmitry Belousov, Anastasiia Smirnova, Ksenia Kurinova, Aleksei Smirnov, Denis Fedyanin, Yury Yanovich

## AI summary

## Overview

**Research area:** Blockchain systems and decentralized finance (DeFi) — specifically Maximal Extractable Value (MEV), arbitrage, and on-chain causal analysis on EVM-compatible networks.

**Technical level:** Advanced. The paper combines a formal blockchain state-machine model, counterfactual replay simulation, cooperative game theory (Shapley values), and a reinforcement-learning searcher bot.

**Scope (one sentence):** The paper formalizes the problem of attributing an extracted atomic arbitrage opportunity back to the transaction(s) that created it, and compares four attribution methods at scale on Polygon data.

## What This Paper Is About

Almost all MEV research asks how value is *extracted* (arbitrage, sandwich attacks, liquidations) or how to *mitigate* it. This paper asks the reverse question: which on-chain transactions *create* the arbitrage opportunities in the first place, and how much of the extracted profit should be credited to each one? The authors formalize this as an attribution problem, build a modular pipeline with four different attribution methods, and test it retrospectively over more than one million Polygon blocks to see whether opportunities can reliably be traced back to a single source transaction.

## Key Contributions

1. **Formalization of MEV opportunity attribution.** The problem is defined inside a deterministic blockchain state-machine model, with explicit inputs (blockchain state from an archive node, ordered transaction traces, the flagged arbitrage event) and outputs (source transaction identity, attributed MEV value, and aggregation by sender/protocol).
2. **Four complementary attribution methods.** The authors design and implement a bot-data-driven method (using a production searcher's bidding behavior as a proxy), a simulation-based counterfactual replay method, a coefficient-based "K-value" method that tracks marginal changes in a pool price-multiplier coefficient, and a Shapley-based method using cooperative game theory with Monte Carlo approximation.
3. **A working system and large-scale retrospective evaluation.** The pipeline was run over 1,050,000 Polygon blocks (83,770,001 through 84,820,000, March 2–29, 2026), covering 360,026 identified atomic arbitrage transactions with a total extracted MEV volume of $334,799, plus a focused February 2026 dataset for head-to-head method comparison.
4. **An empirical analysis of MEV *creation* distribution.** The paper positions itself as the first systematic analysis of which protocols and users generate arbitrage opportunities, reporting a highly concentrated distribution of opportunity creation.

## Main Findings

- **The single-source hypothesis holds:** 96.7% of atomic arbitrage opportunities can be attributed to a single source transaction, supporting the assumption that competitive searchers extract value immediately rather than waiting for multi-transaction conditions.
- **MEV creation is highly concentrated:** the abstract and contributions state that "a small subset of protocols generates most opportunities." The specific protocol-level percentages and named protocols are not present in the available excerpt of the paper (the text is truncated at Section 5.3), so they are not reported here.
- **Source transactions are usually very close to the arbitrage:** across an evaluation spanning 2M blocks, 99.3% of attributable opportunities had their edge transaction (`T_edge`) within 7 blocks of the arbitrage transaction. The authors therefore set the search depth to `D_max = 100` blocks as substantial headroom.
- **Method trade-offs are explicit:** bot-data runs in O(1) after model loading and suits validation only; coefficient-based attribution is O(1) and suitable for candidate sets up to roughly 1,000; simulation is O(log|C| + m) and is recommended as the primary method for candidate sets up to about 200; exact Shapley is O(2^|C|) and practical only below 20 candidates; Monte Carlo Shapley is O(N·|C|) and usable up to about 50 candidates.
- **Validation coverage differs sharply by source:** bot consensus covered 38.4% of the 2,526 February 2026 events (limited to the bot's real-time mempool visibility and missing last-in-block atomic arbitrage by design); exact Shapley covered 12.0%; Monte Carlo Shapley with 1,000 samples covered 100%. Manual expert review covered 200 stratified events, with a method counted correct if it matched at least 2 sources.
- **Pre-existing opportunities are rare in value terms:** 23% of events were attributed to the pre-block state `S_0` (no responsible transaction inside the search window), but these account for less than 4% of total profit. A further 2% of events had candidate sets of 20 or more, making exact Shapley intractable.
- **Bot data measures intent, not causality:** bot attributions reflect what searchers *believed* was causal at submission time and systematically diverge from retrospective methods that analyze finalized blocks; the authors use it as an external validation signal rather than primary attribution.
- **Monte Carlo Shapley converges quickly:** 1,000 samples suffice for estimates to stabilize within 5% of asymptotic values for typical candidate-set sizes in the dataset.
- **The searcher bot's design:** an RL agent with a graph neural network encoder over pool topology and a multi-layer perceptron value head, trained with proximal policy optimization on historical Polygon data, rewarded by realized arbitrage profit minus gas and priority bids, targeting under 10 ms inference per candidate.
- **Concrete example of an arbitrage event:** in block 58,329,504, a transaction exploited a price imbalance between Uniswap V3 (WMATIC/USDC) and Uniswap V2 (USDC/WMATIC), executing N = 2 swaps for a net profit of 0.3121 WMATIC ($0.24 at execution time).

## Methodology in Plain English

The researchers treat the blockchain as a deterministic machine: the same transactions in the same order always produce the same state. That makes a "what if this transaction hadn't happened?" experiment possible — you can replay history on an archive node and see how the profits would change.

The pipeline processes blocks in order, flags atomic arbitrage transactions using three criteria from prior work (at least two swaps, no asset ends up with a negative net balance change, and profit stays strictly positive after fees and priority bids), and then builds a candidate set of earlier transactions that touched any of the same liquidity pools, searching back within the block and up to 100 blocks earlier.

Each of the four attribution methods then picks a culprit:

- **Bot data:** if a live searcher bot bid on an arbitrage route triggered by a specific pending transaction, treat that transaction as the source.
- **Simulation:** binary-search backwards to find the "edge" transaction where arbitrage profit collapses to 5% or less of the observed value, then walk backwards computing each transaction's marginal impact on profit and pick the largest.
- **Coefficient (K-value):** compute how much each transaction nudges a price-multiplier coefficient for the arbitrage cycle, using only pool reserve data from logs, and pick the biggest nudge.
- **Shapley:** treat the candidates as players in a cooperative game whose value is the profit achievable after executing exactly that subset, then average each transaction's marginal contribution over random orderings (Monte Carlo for large sets).

For ground truth, the authors triangulate: bot consensus, exact Shapley where feasible, Monte Carlo Shapley everywhere, and 200 manually reviewed events. Experiments ran on a cluster of 32-core Intel Xeon Platinum machines (2.5–3.5 GHz) with 128 GB RAM, with the pipeline written in Rust and Python wrappers for analysis.

## Why This Matters

**Impact on research:** The paper shifts MEV analysis from the extraction side to the creation side, supplying a formal problem statement, an open design space of methods with quantified trade-offs, and a validation protocol for a setting where causal ground truth is not directly observable on-chain. It also makes an explicit claim about the structure of competitive MEV markets — that opportunities are overwhelmingly single-source — which other work can test.

**Real-world applications:**

- **Protocol designers** can identify which contracts and pool designs leak the most value to arbitrageurs and reduce that MEV leakage.
- **Validators and block builders** can use source attribution to inform transaction ordering and prioritization decisions.
- **Analysts and ecosystem researchers** can measure ecosystem health through opportunity *creation* rather than only through extracted volume.
- **Searchers and MEV infrastructure operators** can use the retrospective methods to audit whether their real-time bot predictions were actually correct.

**Industry relevance:** The paper comes from a collaboration spanning MEV-X, the Moscow Institute of Physics and Technology, HSE University, and the Skolkovo Institute of Science and Technology, and it evaluates a deployed searcher bot's bidding data, making the results directly relevant to practitioners running MEV infrastructure rather than purely theoretical.

## Future Directions

- **Extending attribution beyond atomic arbitrage.** The authors explicitly exclude liquidations and sandwich attacks, noting they require distinct causal models; building those models is a natural next step.
- **Making exact attribution tractable for large candidate sets.** Roughly 2% of events had 20 or more candidates, where exact Shapley is intractable, and the paper's recommended method is limited to candidate sets of about 200 for simulation.
- **Resolving and reducing `S_0` attributions.** 23% of events had no in-block source within the search window, leaving open how far back these pre-existing opportunities actually originated and whether deeper search or cross-chain analysis would recover them.
- **Porting the framework to other execution models.** The formal framework is claimed to apply to any deterministic blockchain state machine with replayable history, and the paper discusses parallel-execution networks such as Solana and TON and sharded designs, but evaluation is limited to Polygon.

## Target Audience

This paper is best suited to MEV researchers and graduate students in blockchain systems, quant and protocol engineers at DEXs and block-building/relay infrastructure, and data engineers who build on-chain analytics pipelines. Readers will get the most from it with prior familiarity with AMM mechanics, EVM transaction traces, and basic cooperative game theory; the formal model and complexity analysis make it a poor fit for newcomers seeking a conceptual introduction to MEV.

## Authors’ abstract

Maximal Extractable Value (MEV) represents billions of dollars in extracted value that fundamentally shapes blockchain network dynamics and participant incentives. While research has focused on MEV extraction and mitigation, we lack systematic methods to attribute MEV opportunities to their on-chain origins. This paper formalizes the MEV opportunity attribution problem and introduces a systems framework for identifying which transactions create arbitrage opportunities and…

[The full abstract on arXiv](https://arxiv.org/abs/2604.27979)

[Read the original paper](https://arxiv.org/abs/2604.27979)

---

Canonical: https://crypto.world/paper/the-origins-of-mev-systematic-attribution-of-arbitrage-opportunity-2604.27979
