---
title: Wash Trading Definition, Types, and Detection
url: https://www.emergentmind.com/topics/wash-trading
type: topic
---

# Wash Trading Definition, Types, and Detection

Wash trading is the coordinated purchase and sale of an asset by the same economic actor or by colluding actors to create the appearance of genuine market activity without a meaningful change in beneficial ownership, market position, or economic exposure. It can inflate reported volume, apparent liquidity, prices, rankings, investor attention, and marketplace-reward eligibility. In pseudonymous blockchain markets, wash trading is investigated through transaction graphs, wallet relationships, payment flows, temporal repetition, statistical irregularities, and behavioral anomalies; these methods generally identify suspicious patterns rather than conclusively establish identity, intent, or legal liability.

## 1. Definition, mechanisms, and economic purposes

Traditional legal formulations describe wash trading as entering into, or purporting to enter into, transactions that give the appearance of purchases and sales without incurring market risk or changing the trader’s market position. The Commodity Futures Trading Commission and related market guidance emphasize two elements: intent to execute fictitious trades not subject to market risk, and a “wash result” in which the same instrument is bought and sold at the same or a similar price for accounts having common beneficial ownership.

In cryptoasset markets, the operational focus is usually behavioral. Suspicious activity may involve:

- **Self-trading**: the same address, account, or effective owner appears on both sides of a transaction.
- **Reciprocal trading**: two accounts repeatedly trade in opposite directions while approximately balancing their positions.
- **Circular trading**: an asset moves through a cycle such as \(A\rightarrow B\rightarrow C\rightarrow A\).
- **Repeated-token circulation**: the same NFT repeatedly traverses a small group of addresses.
- **Common funding or exit**: several addresses are funded by, or return funds to, a common source.
- **Artificially timed activity**: trading intensifies during thin markets, promotional periods, token-reward programs, or periods of heightened attention.

The intended objectives include inflating reported volume, increasing apparent liquidity, manipulating price histories, improving exchange or collection rankings, attracting uninformed buyers, satisfying marketplace verification thresholds, and obtaining rewards distributed according to trading volume. The distinction between genuine and artificial activity is economically important because volume-based indicators are often used as proxies for demand, liquidity, market quality, and popularity.

A blockchain address is not equivalent to a natural or legal person. Multiple addresses may be controlled by one actor, while several addresses may belong to unrelated actors who happen to interact. Accordingly, address-level circularity or common funding provides evidence of possible coordination, but does not by itself establish common ownership or intent. This distinction is central to the interpretation of nearly all blockchain-based wash-trading studies.

## 2. Market structure and empirical manifestations

Decentralized exchanges and NFT marketplaces expose transaction data that are unavailable or difficult to obtain on many centralized exchanges. Smart-contract settlement can reveal buyer and seller addresses, token identifiers, timestamps, payment amounts, and ownership histories. This enables construction of token-specific or collection-specific directed graphs.

For fungible-token markets, a trade graph can be represented as \(G(V,E)\), where \(V\) is the set of trader addresses and \(E\) is the set of token-flow trades. If account \(a_j\) sells \(V\) tokens to account \(a_i\), the trade can be written as:

\[
T=(+a_i,-a_j)_V.
\]

A self-trade is a directed self-loop:

\[
T=(+a,-a)_V.
\]

For NFTs, the graph is generally constructed separately for each token because NFT identity is economically significant. An edge \(u\rightarrow v\) represents transfer or sale of a specific token from one address to another, with annotations such as timestamp, transaction hash, marketplace contract, and payment value. Positive-value edges are candidate sales, while zero-value edges are transfers; however, marketplace APIs may misclassify sales and transfers, and marketplace contracts may appear as intermediate nodes.

NFT wash trading can manifest through rapid closed cycles, repeated two-address transactions, multi-address circulation, stable group-level holdings, private-sales sequences, and repeated funding relationships. Some methods also flag rapid paths that do not close into a cycle when an NFT moves through several addresses within 12 hours with no more than 5% price deviation. Other methods classify a sale followed by repurchase of the same NFT by the original seller within 30 days as a wash-sale pattern. These are operational rules, not universal legal definitions.

Market design strongly affects incentives. Volume-based token rewards on LooksRare and Rarible, for example, can make artificial trading economically attractive. One study found that reward exploitation was substantially more reliable than attempting to inflate an NFT’s price and resell it to an external buyer. On LooksRare, more than 84% of observed marketplace volume was classified as artificial under one graph-and-funding methodology, while OpenSea’s corresponding share was 0.49% [2212.01225]. Another study comparing four Ethereum marketplaces reported wash-volume shares of 94.5% for LooksRare and 84.2% for X2Y2, compared with 25.1% for OpenSea and 29.9% for Blur [2403.10361]. These figures are method-dependent classifications rather than independently verified market-wide measurements.

## 3. Graph-based and payment-flow detection

### Fungible-token exchanges

A major graph-based procedure for decentralized exchanges first identifies repeatedly recurring strongly connected components (SCCs). A directed graph is strongly connected when every vertex can reach every other vertex through a directed path. SCCs are useful because circular trading requires reciprocal reachability, while the SCC representation can accommodate overlapping cycles, parallel edges, nested subcycles, and complex multi-account structures.

The procedure simplifies multiple trades between the same account pair into weighted edges, repeatedly computes SCCs, decreases edge weights, and removes exhausted edges. SCCs observed at least 100 times are retained as candidates. Within each candidate, trades are examined in one-hour, one-day, and one-week windows. A subset is classified as position-neutral when each account’s net token balance is approximately zero. With tolerance margin \(m=0.01\), the residual position must satisfy the conceptual condition:

\[
|p_i|\leq m\cdot v_{\mathrm{mean}}.
\]

The procedure deliberately avoids exhaustive subset enumeration. It iteratively removes trades and tests remaining prefixes rather than searching the full power set. This makes computation tractable but can miss valid non-prefix combinations. Applied to IDEX and EtherDelta, the method detected a lower bound of approximately \$159 million in wash-traded volume from on-chain records covering approximately September 2017 through May 4, 2020. More than 30% of traded tokens on each exchange had detected wash activity; approximately 10% of EtherDelta tokens were almost entirely or entirely wash traded [2102.07001].

A related study of seven Ethereum NFT collections integrates NFT ownership traces with the broader Ethereum normal-transaction network. It constructs a collection-specific Linkability Network from shortest paths among NFT-owner addresses in the complete history of ordinary ETH transfers. A maximum breadth-first-search depth of \(D=3\) is used in the final analysis. NFT transfers without recorded payment provide initial address clusters, which are merged when members are linked through the Ethereum transaction network. Positive-value NFT trades between addresses within an inferred cluster are then classified as potential wash trades. Excluding Meebits, the reported suspicious volume ratio was 25.24%; Meebits was an extreme outlier, with approximately 93–94% of volume classified as suspicious [2312.16603].

### NFTs and address funding

NFT-specific graph methods use closed ownership paths, repeated token circulation, and group-level position accounting. One study of 52 large ERC-721 collections detected 36,385 suspicious sale transactions, representing 2.04% of sale transactions and up to \$149.5 million, or 2.17% of sample volume. Most flagged patterns were cycles among a few addresses; rapid sequential patterns accounted for approximately 0.3% of all sale transactions [2202.03866].

Another Ethereum-wide study used SCCs, service-account removal, smart-contract filtering, zero-volume-component removal, common funders, common exits, self-trades, and reuse of confirmed address patterns. It identified 12,413 confirmed or suspected wash-trading activities and approximately \$3.4 billion of artificial volume, with 97.41% of detected volume concentrated on LooksRare [2212.01225]. Common funding and common exits were especially important because they provide evidence that funds circulated through several addresses without independent economic exposure.

A broader multi-dimensional study separated NFT wash trading into three categories:

1. **Round-trip trading**: repeated buying and selling of the same NFT within short time windows.
2. **Unprofitable trading**: the apparent buyer is funded by the seller, or sale proceeds are returned shortly afterward.
3. **Hidden trading**: repeated private or reserved sales in which the seller designates the buyer.

Using 285 popular Ethereum collections, the study reported 5,330 suspicious sale events and 824 suspicious transfer events with a combined value of \$8,857,070.41. It estimated a minimum loss to subsequent benign users of \$3,965,247.13 [2312.12544].

### Alternative graph formulations

Other approaches deliberately relax the requirement for an explicit cycle. A 30-day repurchase method constructs transfer and sale graphs for each NFT and uses breadth-first search to identify paths through intermediary wallets that return the token to the original seller. The method reported 0.136% of transactions, 0.157% of tokens, and 0.109% of wallets as involved in detected wash-sale patterns, with aggregate price-manipulation, sale, and repurchase quantities of approximately \$930,494, \$1,110,423, and \(-\$1,586,365\), respectively [2305.01543].

A human-centered visualization system, NFTDisk, combines a radial collection-level overview with token-level flow inspection. Its suspicious score for an address pair is:

\[
S=1-\frac{N}{M},
\]

where \(M\) is the number of transactions between the pair and \(N\) is the number of unique NFTs involved. A high score indicates many transactions involving relatively few NFTs. The system then examines group holdings, address inventories, and individual token paths. The score is a heuristic: collectors, arbitrageurs, market makers, and users managing a small inventory may also obtain high values [2302.05863].

## 4. Statistical, anomaly-based, and machine-learning methods

### Distributional tests

When account-level information is unavailable, researchers may test aggregate trade distributions. Cong, Li, Tang, and Yang use three regularities as benchmarks for genuine financial-market activity:

- first-significant-digit distributions approximating Benford’s law;
- human preference for rounded transaction sizes;
- fat-tailed trade-size distributions approximating a power law.

For the first significant digit \(d\), Benford’s law predicts:

\[
\Pr(D=d)=\log_{10}\left(1+\frac{1}{d}\right).
\]

Power-law tails are represented as:

\[
\Pr(X>x)\sim x^{-\alpha},
\]

with a conventional Pareto–Lévy range of \(1<\alpha<2\). Roundedness tests examine whether trades cluster around multiples of specified base units.

Using data from 29 cryptocurrency exchanges, the study estimated average wash-trading shares above 70% on unregulated exchanges, with estimates of 70.85% under equal weighting and 77.50% under volume weighting without controls. With controls, the corresponding figures were 60.96% and 71.43%. Tier-2 unregulated exchanges had higher estimates than Tier-1 exchanges. The paper extrapolated more than \$4.5 trillion of spot and \$1.5 trillion of derivatives wash trading in the first quarter of 2020, while emphasizing that these were extrapolations from reported volume rather than directly observed fake transactions [2108.10984].

Distributional methods have important limitations. Benford’s law can fail for legitimate reasons, including heterogeneous asset populations, discrete pricing conventions, constrained datasets, and small samples. Algorithmic trading can generate unrounded transactions legitimately. Power-law exponents can vary with collection structure, market segmentation, price floors, and cutoff selection. These methods are therefore better interpreted as screening indicators than as direct identification of individual wash trades.

### Roundedness in NFT markets

A direct comparison of on-chain NFT classifications with indirect statistical tests found that Benford and power-law tests poorly distinguished suspected wash trades from legitimate trades. Price roundedness was more informative but remained weak. At a 1% threshold, roundedness classified 21.12% of LooksRare trades and 16.21% of Blur trades as non-round, compared with direct wash classifications of 29.42% and 26.94%, respectively. Precision and recall relative to the direct benchmark were 38.31% and 27.51% for LooksRare, and 35.09% and 21.11% for Blur [2311.18717].

The study did not contain the AI estimator described in some accounts of the work. It reported no neural-network architecture, training loss, train–validation split, AI AUC, calibration analysis, or integrated machine-learning model. Its indirect analysis consisted of rule-based filters, statistical tests, regressions, and roundedness-threshold analysis.

### Unsupervised clustering

An OpenSea study represented each wallet using 26 network, monetary, and temporal features and applied K-means clustering. Seven clusters were selected using elbow, Davies–Bouldin, and silhouette diagnostics. Two clusters, comprising 13,607 of 252,924 wallets, were interpreted as potential wash-trader groups. This corresponds to approximately 5.38% of sampled wallets, but not to 5.38% of confirmed wash trades. The method detects wallet-level behavioral groups rather than individual suspicious transactions and has no ground-truth labels, precision, recall, or independently verified false-positive rate [2306.04643].

### Local outlier detection

For a Solana NFT audit, Local Outlier Factor (LOF) was applied with \(k=20\). LOF compares the local reachability density of an observation with that of its neighbors:

\[
LOF_k(y_i)=
\frac{1}{|N_k(y_i)|}
\sum_{y_j\in N_k(y_i)}
\frac{lrd_k(y_j)}{lrd_k(y_i)}.
\]

The study reported 138 affected collections, 2,175 suspicious nodes, and 46,612 associated transactions. Eight collections had reported wash rates above 50%, but the paper did not specify the LOF threshold, feature vector, distance metric, scaling procedure, or ground-truth validation. Its reported “wash rate” also mixed token-count and volume-based definitions [2403.10879].

### Complexity measures

Complexity-based detection has been applied to centralized-exchange data where trader identities are unavailable. Measures include tail distributions, autocorrelation functions, multifractal detrended fluctuation analysis, approximate entropy, sample entropy, and detrended cross-correlations among returns, volume, and transaction counts.

A study of Binance, Bitget, KuCoin, and Kraken during April–June 2025 found an anomalous regime on Bitget for BTC and ETH after May 21. Transaction counts rose sharply, while volume and return fluctuations did not increase proportionally. The regime exhibited many small trades, weaker transaction-count autocorrelation, reduced multifractal organization, higher short-pattern irregularity, and weaker cross-correlations between transaction counts, volume, and returns. The authors described this as consistent with a noise-like component and possible artificial transaction generation, while explicitly noting that it did not prove wash trading [2607.13916].

A related liquidity framework defines liquidity jump and liquidity diffusion:

\[
\beta_T^r=\left|\frac{R_T}{R_T^\ell}\right|,
\qquad
\beta_T^\sigma=\frac{\sigma_T}{\sigma_T^\ell}.
\]

The first measures the magnitude of a daily liquidity-related return distortion; the second measures intraday liquidity volatility. The proposed screening condition is:

\[
\beta_T^r\geq 1
\quad\text{and}\quad
\beta_T^\sigma\geq 1.
\]

The framework interprets high liquidity diffusion followed by high liquidity jump as more consistent with manipulative activity than high liquidity jump alone. In a sample of eight large cryptoassets, joint high-jump/high-diffusion days ranged from 0.86% for BTC to 41.50% for ETC before treatment. A simulated treatment that reduced third-quartile trading amounts by 50% and fourth-quartile amounts by 75% sharply reduced diffusion while leaving part of the jump component intact [2411.05803].

## 5. Incentives, timing, profitability, and market effects

Wash trading is not necessarily profitable through the same mechanism in every market.

### Price inflation and resale

An operator may circulate an NFT or fungible token to construct a higher apparent price, then attempt to sell to an independent buyer. This strategy bears genuine risk because the external buyer may not appear, the price may decline, and fees may exceed the eventual gain. In one Ethereum NFT study, approximately half of price-resale operations lost money after transaction and marketplace fees. A documented example involved an NFT whose internal price rose from 0.66 ETH to 12.5 ETH before an external sale at 14.85 ETH; the reported post-fee profit exceeded \$44,000 [2212.01225].

### Reward farming

Marketplace rewards can make wash trading profitable without an external buyer. If rewards are allocated according to trading volume,

\[
R_A=\frac{a}{b}\times c,
\]

where \(a\) is the user’s volume, \(b\) is marketplace volume, and \(c\) is the reward allocation, an operator can trade against controlled accounts and receive tokens. LooksRare and Rarible reward-exploitation operations had success rates of approximately 80% and 95%, respectively, among analyzed activities [2212.01225].

The relationship between incentives and manipulation is also visible over time. On LooksRare, wash volume remained above 92% during early reward phases and declined after reductions in daily LOOKS rewards. A cross-market study similarly found that reward-based marketplaces had much larger detected wash-volume shares than non-incentivized venues [2403.10361].

### Timing relative to market conditions

Historical Mt. Gox data indicate that wash trading intensified when legitimate Bitcoin volume was low and diminished when legitimate volume was high. Machine-learning prediction models, vector autoregressions, Granger tests, and impulse responses found predictive relationships between lagged legitimate volume, volatility, and subsequent wash activity. The study also reported spillovers across exchanges and associations with activity in stocks, gold, foreign exchange, media attention, and online rumors [2411.08720].

The economic interpretation is that artificial activity is more salient in thin markets. When genuine volume is low, a fixed amount of wash volume represents a larger share of observed activity. Around a Bitcoin-demand event associated with Silk Road’s “Pot Day,” the estimated effect of wash trading weakened after legitimate demand increased. This is consistent with strategic timing, although the event does not constitute a randomized experiment.

### Prices and losses to subsequent users

Wash trading can generate misleading price histories even when the wash trader does not ultimately profit. A study of NFT markets estimated that subsequent users incurred a minimum loss of \$3,965,247.13 in selected cases where buyers paid more than the last price in a wash-trading window [2312.12544]. Other studies found that suspicious activity often occurred near collection launches, when artificial volume might increase attention and perceived demand. Such timing is suggestive of promotional or visibility-related motives, but does not establish who ordered or financed the trades.

### Market measurement

Reported volume can be contaminated at several levels:

- exchange-reported spot and derivatives volume;
- NFT collection volume and floor-price histories;
- marketplace rankings;
- liquidity and market-depth measures;
- empirical datasets used in research on price discovery, volatility, and market integration.

The magnitude of the problem varies substantially across methods and markets. A decentralized-exchange graph analysis produced a conservative lower bound of \$159 million. A statistical study of unregulated centralized exchanges estimated average suspicious volume above 70%. NFT studies ranged from approximately 0.14% of transactions in one OpenSea sample to approximately 25.24% of non-Meebits volume under a broader wallet-linkability method. These figures are not directly comparable because they use different samples, definitions, denominators, and detection rules.

## 6. Limitations, countermeasures, and research directions

### Identification limits

No single observed pattern proves wash trading. Common false positives include legitimate arbitrage, market making, custody transfers, treasury operations, marketplace escrow, professional collecting, loans, gifts, refunds, operational wallet movements, and unrelated users funded by the same exchange. Common false negatives include activity distributed across many addresses, trades separated by long periods, off-chain agreements, mixers, bridges, unrelated funding sources, and schemes deliberately designed to avoid cycles or repeated routes.

The most important methodological limitations are:

- wallet addresses do not reveal beneficial ownership;
- intent is not directly observable;
- many API datasets misclassify transfers and sales;
- exchange-level statistical tests cannot identify individual trades;
- rule-based algorithms are usually conservative and incomplete;
- indirect methods may confuse legitimate market heterogeneity with manipulation;
- weakly supervised and unsupervised models lack verified ground truth;
- high accuracy can be misleading under class imbalance;
- causal claims about price effects, incentives, and regulation are often not identified.

A BSC meme-token warning system illustrates both the potential and limitations of supervised approaches. It constructed 12 token-level features from Self, Matched, and Circular patterns and trained Logistic Regression and Random Forest models on seven tokens and 33,242 transfer records. Random Forest achieved AUC \(=0.9098\), PR-AUC \(=0.9185\), and \(F_1=0.7429\), with \(FP=1\) and \(FN=8\). Trade-level features were the principal performance driver, while address-level features provided complementary information. The system was positioned as a high-precision screener rather than a high-recall autonomous alarm engine [2603.13830]. Because labels were weak and the sample was small, these metrics measure high-risk prediction under the study’s labeling scheme rather than validated wash-trade detection.

### Countermeasures

Potential countermeasures include:

- **Self-trade prevention**: prevent an account from filling its own orders.
- **Counterparty-frequency monitoring**: identify repeated interactions between a small number of accounts.
- **Graph surveillance**: monitor SCCs, self-loops, reciprocal cycles, repeated token paths, and linked-wallet clusters.
- **Funding-flow analysis**: examine common funders, common exits, and circular payment flows.
- **Reward-program controls**: exclude suspicious volume from token rewards, weight rewards toward unique counterparties or holding duration, and avoid linear incentives based solely on gross volume.
- **Identity and beneficial-ownership controls**: use KYC or related identity mechanisms where appropriate.
- **Volume-quality reporting**: distinguish reported volume from organic or adjusted volume.
- **User warnings**: identify collections or assets whose price histories and volume are materially affected by suspicious patterns.
- **Human review**: require analyst investigation before sanctions, public accusations, or irreversible marketplace actions.

Self-trade prevention alone is insufficient because several addresses can create cycles such as \(a\rightarrow b\rightarrow c\rightarrow a\). Effective surveillance therefore requires network-level and behavioral analysis rather than only same-address execution controls.

### Research directions

Future systems are likely to combine:

- token-specific ownership graphs;
- strongly connected components and repeated-counterparty motifs;
- common funding and payment-flow networks;
- transaction timing, price, fee, and gas features;
- marketplace metadata and reward eligibility;
- complexity measures for centralized exchanges;
- cross-marketplace and cross-chain activity;
- supervised and probabilistic models with verified labels;
- human-centered visual investigation;
- causal designs for reward changes, regulation, and demand shocks.

A robust system should report separate quantities for transaction count, token count, address count, and monetary volume. It should also provide uncertainty estimates, sensitivity to thresholds, out-of-sample validation, and explicit treatment of legitimate high-frequency trading.

Wash trading is therefore best understood as a multi-layer market-integrity problem. Its observable manifestations range from exact self-trades to large-scale statistical anomalies in transaction counts and liquidity. Blockchain transparency makes address-level investigation possible, but pseudonymity prevents direct attribution. The strongest evidence arises when several independent signals coincide: repeated circular flows, near-zero net position changes, common funding, unusual concentration, temporal coordination, weak integration with price formation, and market-design incentives that make artificial activity profitable. Even then, detected activity should ordinarily be described as suspicious, likely, or wash-trading-consistent unless identity, intent, and economic exposure have been independently established.

Source: https://www.emergentmind.com/topics/wash-trading