Study Finds Widespread Signs of Wash Trading in NFT Markets
Using machine-learning models and blockchain data, BU researcher Gerry Tsoukalas and collaborators identified suspicious patterns across major exchanges.
By Brendan Galvin
Non-fungible tokens, or NFTs, became a symbol of the digital asset boom, with buyers spending huge sums on blockchain-based art, collectibles, and virtual items. Yet new research suggests that much of that activity may have been artificially inflated.
While headlines mention million-dollar sales of digital apes and pixelated figures, a new study published in Research Policy reveals that a staggering amount of this activity is “fake.”
Approximately 38% of NFT trades and 60% of the total traded value across several major exchanges showed patterns consistent with wash trading – a form of market manipulation that makes it difficult to determine whether transactions reflect legitimate market activity or coordinated self-trading.
NFT markets were built on Web3 infrastructure. Unlike traditional stock markets, where your identity is tied to an individual brokerage, Web3 allows anyone to create unlimited anonymous wallets.
“One of the core problems is that blockchain markets make it extremely easy to create the illusion of demand,” says Gerry Tsoukalas, Dean’s Research Scholar, Professor of Information Systems at Questrom School of Business, and Hariri Institute Faculty Affiliate. “What looks like genuine market activity may in many cases be the same actors trading with themselves.”
Teaching Machines to Spot Wash Trades
Tsoukalas, along with collaborators from University of Pennsylvania and University of California, Los Angeles, developed a machine-learning framework that uses public blockchain data to identify likely wash trades at the transaction level. The researchers first built a set of on-chain detection filters designed to identify suspicious trading patterns, including:
- Self-Trades: Trading an NFT between wallets controlled by the same user.
- Back-and-Forth: Wallet A sells to Wallet B, then Wallet B sells it right back.
- Circular Trades: A loop where an NFT passes through three or more wallets and ends up back to the original owner.
- Common Funding: Both the “buyer” and “seller” were funded by the same source.
The outputs from those filters were then used to train machine-learning models capable of identifying broader behavioral patterns associated with wash trading.
“The transparency in NFT markets is real, but it is transaction-level transparency, not identity-level. You see every trade, every wallet address, every transfer,” says Tsoukalas.
“What you don’t see is who actually owns those wallets in the real world. One person can spin up ten wallets and trade between them, and on-chain it looks like ten distinct users interacting.”
Why go to all this trouble?
In many cases, the marketplaces themselves may unintentionally encourage the behavior. Platforms like LooksRare and Blur rose to prominence in part by offering “rewards” tied to trading volume. If the rewards earned through trading outweigh the transaction costs, wash trading can become highly profitable.
While wash trades accounted for about 31% of transactions on OpenSea, they accounted for a massive 95% of trade volume on LooksRare, where volume-based rewards were especially aggressive.
“Anywhere you bolt a per-trade reward or royalty on top of an anonymous trading layer, you should expect wash trading to follow. The mechanism doesn’t care what the underlying asset is,” says Tsoukalas.
“Platforms that wrap a real-world asset in a token and then bolt on Web3-style reward programs to bootstrap liquidity—that’s exactly the design that broke NFTs.”
Previous attempts to study this problem often relied on statistical methods or leaked private data. These approaches often failed with NFT markets, where digital assets behave differently than traditional financial securities.
The researchers bridged this gap by training machine-learning models on 16 transactional-level “behavioral fingerprints,” such as how quickly an NFT is flipped, how much the sale price deviates from previous trades, and how active the wallets are involved over time..
This machine-learning approach was so effective that it reduced estimation errors to below 1.3% and even successfully flagged wash trading in old Bitcoin data.
“To evade a model built around things like a buyer’s 7-day trade count, time since last trade, and price deviation – a wash trader has to space trades out, break their patterns, and slow down,” says Tsoukalas.
“All of that costs time, gas, and risk. At some point, the economics of the wash stop working. We aren’t claiming the detection method we developed will end wash trading. We’re saying it could raise the floor on what it costs to do it convincingly.”
The data needed for this kind of surveillance is already available to regulators, but Tsoukalas says a one-size-fits-all enforcement approach is unlikely to work because exchanges operate with very different incentives.
To protect honest collectors, Tsoukalas suggests that marketplaces should publish “clean volume” metrics alongside raw trading data to distinguish legitimate market activity from manipulation.
From Human Wash Traders to AI Agents
As we move toward a world where AI increasingly participates in trading, there is growing concern that these bots might learn to collude with each other to artificially pump prices without explicit human instruction.
Fortunately, the implications of this study extend beyond NFTs. Because blockchain transactions are publicly recorded, similar machine-learning approaches could eventually be used for large-scale market surveillance across other tokenized financial systems. In his working paper “On the Fragility of AI Agent Collusion,” Tsoukalas explores how large language model (LLM) pricing agents affect market prices in repeated competitive settings.
“The bottleneck on enforcement isn’t really the math anymore,” says Tsoukalas. “It’s more about institutional will.”
“The data is sitting there, the methods are published, and almost nobody is running this kind of surveillance at scale. The realistic ask isn’t a new technology, it’s actually using the one that exists.”
A replication package, including the data and code used in this study, are available at Falk et al. (2025).
Citation: Brett Hemenway Falk, Gerry Tsoukalas, Niuniu Zhang, “Detecting crypto wash trades via machine learning” Research Policy, 2026, 105508, ISSN 0048-7333, https://doi.org/10.1016/j.respol.2026.105508.