Research
Quantifying Observable High-Frequency Swapping on Arbitrum
Layer-2 rollups like Arbitrum replaced Ethereum's mempool race with a single sequencer, deterministic ordering, ~250 ms block times, and swap fees typically below $0.05, which largely removes…

- arXiv
- 2609.14481
- Published
- 2026-09-13
- Authors
- Shijian Chen, Ya Chen, Jing Cai, Catherine Liu
AI summary
Overview
Research area: Blockchain measurement and decentralized finance market microstructure, specifically on-chain trading behavior on Ethereum Layer-2 rollups (Arbitrum). It sits at the intersection of crypto-security/MEV research and empirical market-microstructure analysis.
Technical level: Intermediate. The paper is readable without deep mathematics, but familiarity with DEXs, automated market makers, MEV, and basic statistics helps.
Scope (one sentence): An 18-month, address-level empirical measurement of "High-Frequency Swapping" on Arbitrum — identifying 477 swappers responsible for 29.9M single-swap transactions — and characterizing their timing, size, directionality, token/venue concentration, and case-level behavior.
What This Paper Is About
Layer-2 rollups like Arbitrum replaced Ethereum's mempool race with a single sequencer, deterministic ordering, ~250 ms block times, and swap fees typically below $0.05, which largely removes classical MEV such as frontrunning and sandwich attacks. The authors argue this creates a different behavioral regime they name High-Frequency Swapping (HFS): addresses performing continuous single-hop DEX swaps at machine cadence, where the edge comes from speed and persistence rather than transaction ordering. The paper's goal is to systematically define, detect, and quantify this behavior on Arbitrum over 18 months of full-chain data (Jan. 2023–Jun. 2024), since prior work had not identified or measured it.
Key Contributions
- First characterization of HFS. The authors introduce and empirically validate High-Frequency Swapping as a distinct category of on-chain behavior, showing it accounts for a substantial share of Arbitrum's DEX activity (about 22% of observed swaps in the period).
- A systematic detection methodology. They build a reproducible, scalable pipeline over full-chain traces with three stages: transaction filtering (single-swap only, no liquidations, CEX-listed tokens), swapper attribution (recipient field of the
Swap()event, excluding routers/aggregators), and high-frequency classification from inter-arrival gap statistics. - Multi-level statistical and behavioral analysis. They report aggregate scale, monthly temporal dynamics, identity/labelling coverage, token-pair and venue concentration, inter-swap gaps, trade sizes, directional runs, and the Fano factor of trade counts.
- Case-level behavioral signatures. They document stablecoin arbitrage, CEX–DEX execution edges, round-trip trading, and groups of apparently linked addresses, plus a released dataset intended to support follow-up research.
Main Findings
- Scale of HFS: 29.9M single-swap transactions by 477 distinct swappers, exceeding $1.02 × 10^11 in aggregate notional volume, or roughly 22% of all observed swaps in the same period.
- Classification thresholds: an address is an HFS swapper if it has N ≥ 150 swaps, median inter-arrival gap ≤ 120 seconds, 90th-percentile gap ≤ 900 seconds, and no more than 20% of gaps longer than one hour. Bi-directionality is deliberately not required.
- Heavy-tailed trade sizes: median swap about $1.75k, p90 about $7.9k, p99 about $22.4k (the trade-size subsection reports $7.95k and $22.5k), with a maximum single transaction above $1.0M.
- Heavy-tailed swapper activity: the median swapper executes about 2,223 swaps and $2.91M notional, while the 90th-percentile swapper reaches about 134,222 swaps and over $337.35M.
- Cyclical but rising volume: monthly trades grew from 291,742 in Jan. 2023 ($1.17B) to 1,322,581 in Apr. 2023 ($7.20B), bottomed at 424,923 trades ($1.70B) in Aug. 2023, then rose to 2,760,604 trades ($9.10B) in Jan. 2024 and 3,483,929 trades ($12.86B) in Mar. 2024; Apr. 2024 posted the sample-period peak of 4,242,892 trades.
- Single-day peak: May 23, 2024 recorded over 314,842 transactions and more than $1.08B notional, concentrated across only 57 active swappers.
- ETH pairs dominate: USDC/WETH and WETH/USDC alone contributed 52.29% of daily volume, and USDT0/WETH plus WETH/USDT0 added another 19.56%; the authors link the peak to the May 2024 spot Ethereum ETF news, when ETH reportedly moved from roughly $3,150 to above $3,800 in a day.
- Machine-level cadence: median inter-arrival gap is 4 seconds, p90 is 101 seconds, p99 is 881 seconds (about 15 minutes); the longest observed gap is roughly 24.7 days.
- Burstiness: across swappers the median Fano factor of trade counts is 150.93 and the 90th percentile 885.50, far above the Poisson benchmark of 1, indicating clustered bursts rather than uniform arrivals.
- Persistent core of actors: about 42 unique HFS addresses per day (range 17–70). Against Arkham Intelligence labels, only 8 addresses (1.68%) map to professional firms — Wintermute (2), Flow Traders (1), Selini Capital (1), Manifold Trading (4) — 27 (5.66%) have weak identifiers such as ENS names, and 442 (92.66%) remain unlabeled.
- Venue and token concentration: Uniswap accounts for over 21.98M trades and more than $85.06B notional; Camelot, PancakeSwap, SushiSwap, and Ramses each host between 1M and 3M trades and $2.5B–$6.5B volume, while Arbswap, Zyberswap, and Chronos together are under 0.2% of activity. USDC-WETH alone shows over 9,813,743 trades and more than $50.44B; ARB-USDC and WETH-ARB together exceed 6,086,221 trades and about $23.35B.
- Directional persistence, not rapid alternation: median alternation rate is 0.25 (about four consecutive same-direction trades before switching), p90 stays below 0.45, the bottom decile never switches, median run length is three trades and p90 is eight.
- Stablecoin micro-arbitrage: 16,321 same-peg stablecoin swaps executed at a rate strictly above 1, generating an estimated gross profit of about $122.7K on $168.1M volume; one Jun. 15, 2023 example swapped $750K USDC to USDC.e at 1.000065 for a $48.7 gain.
- CEX–DEX execution edge: across 124 filtered swapper–pool combinations in USDC/WETH, USDT0/WETH and DAI/WETH (Jun. 2023–Jun. 2024), the average mean execution edge is +5.4 bp, with 69.6% of seconds showing a DEX advantage and a standard deviation of about 28 bp; address 0x98f989 reached +26.9 bp over 2,011 observed seconds with 97.7% DEX-favorable seconds.
- Round-trip profitability: 161 swapper-pair samples covering 109 distinct swappers show a mean round-trip return of +5.7 bp, median +8.0 bp, weighted mean +9.6 bp, with roughly 75% of combinations positive.
- Notable cases: 0x2d512b executed 127,296 swaps in a single Uniswap V3 ARB/USDC pool (49.2% of that pool's 258,897 swaps), spent 5.846 ETH on swap costs, and showed about $20,000 overall profit; 0x44c6c4 is the most active swapper with 4,469,893 swaps across 37 pairs and $11.37B over 380 days (about 11,762 transactions and $29.9M per day); 0x51c728, tagged by Arkham as Wintermute, executed 2,473,343 trades across 16 pairs for over $9.1B.
- Apparent linked wallets: 0xabdba4, 0x769d6a, and 0x2d512b exchanged 23 direct ERC-20 transfers (mostly 10,000–171,000 USDC), including three negligible transfers of 0.001, 0.001 and 1 USDC interpreted as connectivity tests, and each maintained buy-shares between 45% and 60% over nearly their entire operational period.
Methodology in Plain English
The authors start from raw Arbitrum execution traces and parse Swap() events from major DEXs (Uniswap v2/v3, SushiSwap, Camelot, PancakeSwap, Ramses, Trader Joe and others). They then keep only transactions containing exactly one swap, which removes multi-hop routing and cyclic arbitrage; drop liquidation-triggered trades as involuntary; and restrict to tokens that were listed on Binance spot or perpetual futures by Jul. 2024 and also had live ERC-20 contracts on Arbitrum. Each swap is attributed to the address recorded as the output recipient, with routers and aggregators excluded so that infrastructure contracts are not mistaken for traders.
To separate professionals from casual users, they sort each address's swaps by timestamp, compute the gaps between consecutive swaps, and apply fixed thresholds (at least 150 swaps, median gap of at most 120 seconds, 90th-percentile gap of at most 900 seconds, and at most 20% of gaps longer than an hour). These cutoffs were set by hand from the observed distribution, favoring sustained automated activity over occasional bursts.
Beyond detection, they compute per-swapper summary statistics and a Fano factor (variance of trade counts in fixed windows divided by the mean) to measure burstiness. For case analysis they build a per-second volume-weighted DEX price and compare it to Binance spot quotes to get a signed execution edge in basis points, and they match each buy to the earliest subsequent sell in the same pool within 24 hours to compute round-trip profit and return in basis points.
Why This Matters
Research impact. The paper shifts attention from ordering-based MEV on Ethereum L1 to timing-based behavior on L2 rollups, and provides a definition, detection rule, and 18-month dataset that others can reuse. It complements AMM theory such as Loss-versus-Rebalancing by supplying large-scale address-level empirics rather than new theory.
Real-world applications:
- Exchange and market-maker monitoring: the cadence, run-length, and buy-share signatures give a practical fingerprint for spotting algorithmic market-making or delegated trading wallets.
- Compliance and attribution: the finding that 92.66% of HFS addresses are unlabeled, alongside transfer patterns between related wallets, is directly relevant to clustering and entity-resolution work.
- Protocol and DEX design: knowing that over 21.98M trades and $85.06B notional concentrate on Uniswap, and that ETH pairs dominate, informs liquidity-incentive and pool-design decisions.
- Regulatory and market-integrity analysis: measuring how fast off-chain news (the May 2024 ETH ETF rumor) transmits into on-chain flow helps frame disclosure and surveillance questions for decentralized venues.
Industry relevance. Arbitrum is described as the largest L2 by total value secured ($19.4B per L2Beat as of Aug. 27, 2025), and HFS represents about 22% of its observed swap flow. That makes these actors material to fee revenue, liquidity provision, and price formation on the venue where much L2 trading occurs.
Future Directions
- Breaking the observability ceiling. The authors note that Arbitrum's single-sequencer model exposes only finalized swaps, not bidding, bundling, or ordering, so the CEX–DEX edge they measure remains a plausible but unverifiable explanation; methods that recover intent without mempool visibility are an open problem.
- Entity resolution and coordination detection. With 442 of 477 addresses unlabeled, linking wallets through transfer graphs, funding patterns, and shared execution signatures (as suggested by the 0xabdba4/0x769d6a/0x2d512b cluster) is a natural extension.
- Cross-chain replication. Whether similar HFS regimes exist on other rollups and how they differ from Arbitrum's sequencer-based environment is untested here.
- Extending the PnL and risk analysis. The round-trip analysis is truncated mid-sentence in the available text at the point of discussing return volatility, and the work covers a fixed 18-month window (Jan. 2023–Jun. 2024), leaving longer horizons and fuller profitability/risk decompositions open.
Target Audience
Researchers in blockchain measurement, DeFi market microstructure, and MEV; data engineers and analysts building on-chain attribution or entity-clustering systems; market makers and trading firms active on L2s who want a public benchmark of peer behavior; and DEX/protocol teams, compliance analysts, and regulators seeking an empirical picture of machine-cadence trading in decentralized markets.
Authors’ abstract
Layer-2 rollups have reshaped Ethereum's transaction economy by replacing mempool competition with deterministic sequencing, sub-second block times, and negligible gas fees. While these properties suppress classical Miner/Maximal Extractable Value (MEV), they give rise to a new and previously unrecognized behavioral regime. In this paper, we introduce the concept of High-Frequency Swapping (HFS), referring to continuous single-hop swaps at machine cadence within…
The full abstract on arXiv