Wash trading detectionanswers one question: did this trade change anything real? A genuine trade transfers an asset between independent decision-makers who each face real economics. A wash trade only rearranges an actor's own holdings while emitting the public signals of a market. Detection is the craft of telling the two apart from data, without the power to subpoena anyone.
In regulated equities and futures, surveillance teams see account ownership directly, so matched trades between related accounts are flagged mechanically. On public blockchains ownership is pseudonymous, and the work shifts to inference. That inference is the methodological backbone of my paper The Economics of Wash Trading, and this page describes the toolkit in the order a practitioner would apply it.
1. Graph structure: who trades with whom
Build the trade graph: wallets as nodes, sales as directed edges. Wash trading produces shapes organic trading rarely does:
- Self-loops and round trips.A token that returns to its origin wallet, especially quickly and repeatedly, is the canonical flag. In NFT data the token's full custody chain is visible, so round trips are directly enumerable.
- Closed cliques. Small sets of wallets that trade overwhelmingly with each other, at volumes wildly out of proportion to their interaction with the rest of the market.
- Improbable reciprocity. Counterparty pairs that alternate buyer and seller roles far more often than random matching in a large market would predict.
2. Funding forensics: who pays for the buyer
Graph shapes can be defeated by using fresh wallets for every trade. Funding analysis catches that: cluster wallets by the origin of their gas and purchase capital. When the buyer's ETH arrived from the seller, from a common parent wallet, or from the same exchange withdrawal minutes earlier, independence fails no matter how new the address is. Funding-source clustering is the single highest-yield technique in NFT forensics, because creating wallets is free but capitalising them leaves a trail.
3. Statistical fingerprints
Manufactured activity has different statistics from organic activity. Useful tests include first-significant-digit distributions against Benford's expectation, unnatural roundness in trade sizes, mechanical regularity in inter-trade times, and a volume-to-unique-funders ratio that explodes when few actors print many trades. Cong, Li, Tang and Yang applied this family of tests to centralised exchanges in Crypto Wash Trading and attributed the majority of reported volume on many unregulated venues to wash trading. The same logic transfers to NFT markets with the advantage that individual trades, not just aggregates, are observable, as in von Wachter et al.'s transaction-level NFT study.
4. Incentive context: what the venue pays for
The cheapest detector is knowing where to look. Wash trading concentrates where a reward keys off volume: token emissions per traded dollar, fee rebates, leaderboard placement, airdrop eligibility. In my data, flagged activity clustered overwhelmingly on venues with token-based incentives, and thinned out where no such subsidy existed. If you must triage a market with limited compute, sort venues by reward design before running a single graph query.
Failure modes and honest caveats
- False positives.Legitimate behaviour mimics wash patterns: moving tokens between one's own hot and cold wallets, market-maker inventory shuffling, collateral migrations. Filters need allowlists for known custodial and protocol flows.
- False negatives. A determined manipulator with patient capital, many exchange accounts, and randomised behaviour can stay under any threshold. Detection bounds the phenomenon from below.
- Aggregation hides everything. Volume-level statistics from venues that do not publish trades cannot be audited at all. Transparency is the precondition; blockchains provide it, most centralised venues do not.
For what detection implies at market scale, see NFT wash trading: scale and motives. For what happens to those who are caught, see Is wash trading illegal?