# Detecting unusual trading patterns on cryptocurrency exchanges by means of complexity measures

Crypto.world

> Public trade feeds from centralised exchanges report prices, volumes, and timestamps but not trader identities, so individual wash trades cannot be labelled directly.

**arXiv**: 2607.13916

**Published**: 2026-07-15

**Authors**: Jakub Zwydak, Marcin Wątorek, Jarosław Kwapień, Stanisław Drożdż

## AI summary

## Overview

**Research area:** Econophysics / market microstructure — detecting artificial trading activity (including possible wash trading) on centralised cryptocurrency exchanges using complexity and statistical-structure measures of high-frequency trade data.

**Technical level:** Advanced. The paper assumes familiarity with multifractal detrended fluctuation analysis, detrended cross-correlation analysis, and entropy-based regularity measures, though the intuition behind each tool is explained.

**Scope (one sentence):** The paper builds a diagnostic framework from tail distributions, autocorrelation functions, multifractal spectra, approximate/sample entropy, and scale-dependent cross-correlations, and applies it to 1-minute BTC, ETH, and XRP data from Binance, Bitget, KuCoin, and Kraken between April 1 and June 30, 2025.

## What This Paper Is About

Public trade feeds from centralised exchanges report prices, volumes, and timestamps but not trader identities, so individual wash trades cannot be labelled directly. The authors ask a different question: whether the *internal statistical organisation* of trading activity — how transaction counts, volume, and returns co-vary across time scales — departs persistently from the pattern seen on other exchanges and other assets. Their goal is a diagnostic framework that flags exchange-specific anomalies hidden from price-based indicators.

## Key Contributions

1. **A multi-measure diagnostic framework** combining complementary tools — complementary cumulative distribution functions (CCDFs) with power-law and stretched-exponential references, autocorrelation functions, multifractal detrended fluctuation analysis (MFDFA), multifractal detrended cross-correlation analysis (MFCCA) with the *q*-dependent detrended correlation coefficient ρ(q,s), and approximate entropy (ApEn) plus sample entropy (SampEn) — applied jointly to log-returns, volume, and transaction counts.
2. **Empirical benchmarking against cross-exchange consistency:** because price dynamics are largely exchange-independent at 1-minute resolution, the authors use agreement across Binance, Bitget, KuCoin, and Kraken as an empirical definition of "usual" behaviour, rather than a single parametric null model.
3. **Identification of a specific regime change:** a pronounced, sustained anomaly on Bitget for BTC and ETH beginning May 21, 2025, in which transaction counts decouple from volume and return fluctuations.
4. **A demonstration that complexity measures detect what price data cannot:** return distributions and return multifractal spectra are nearly identical across exchanges, so the anomaly is only visible in the transaction-count and cross-correlation structure — while the authors explicitly stop short of claiming proof of wash trading.

## Main Findings

- **Returns are exchange-invariant at 1-minute resolution.** CCDFs of standardised absolute log-returns overlap closely across all four exchanges for each asset, with tails far heavier than Gaussian and close to a power law with exponent γ = 3, consistent with the inverse cubic law; the authors attribute this to arbitrage equalising rates.
- **Volume distributions differ mainly in the tails.** Volume CCDFs are non-Gaussian and better described by a power law with γ = 2 or a stretched exponential with β = 0.4; KuCoin shows a slightly thicker tail, while Binance's is thinnest.
- **Transaction counts separate the exchanges sharply.** For BTC and ETH on Bitget, the distribution of 1-minute transaction counts is much narrower and closer to Gaussian than on the other venues, whereas Binance, Kraken, and KuCoin are heavy-tailed and bursty.
- **The Bitget transaction-count autocorrelation is exceptionally persistent.** For BTC and ETH, elevated autocorrelations on Bitget persist to lags of order 10⁴ minutes, far longer than on other exchanges; the effect is absent for XRP, whose Bitget curve resembles the others.
- **A dated regime shift on Bitget.** From May 21, 2025, through the end of the sample, BTC and ETH transactions per minute on Bitget do not fall below 150 except in a few cases; three of the moments with fewer than 5 transactions per minute occur simultaneously for both assets, suggesting a systematic rather than incidental effect.
- **Liquidity and trade size rankings are exchange-specific.** Binance is the most liquid venue (highest average transaction count and volume, and no 1-minute intervals without transactions). Kraken has the largest average transaction sizes — several times those of other exchanges — and ranks second only to Binance in average volume despite far fewer trades. The smallest average trade sizes occur on Bitget for BTC and ETH and on KuCoin for XRP. KuCoin trades least frequently, with the highest share of empty 1-minute intervals.
- **Multifractal spectra are similar for returns, differentiated for volume.** Return spectra are left-sided asymmetric and nearly identical across exchanges, with BTC showing the widest right arm (stronger organisation of small fluctuations). For volume, Bitget and KuCoin show weaker scaling for q < 0 (poorly developed or absent right arms), Binance has the shortest left arm and longest right arm, and KuCoin's long left arm matches its near-power-law volume tails.
- **The anomaly's signature, per the abstract:** numerous low-volume trades, weaker autocorrelations, reduced multifractal organisation, higher short-pattern irregularity, and weaker cross-correlations involving the transaction-count series — described as "consistent with a noise-like component in trading activity" and possibly indicating artificially increased transaction counts. Note that the supplied text is truncated partway through the analysis of the transaction-count multifractal properties, so the detailed cross-correlation, ApEn, and rolling-window results are not visible in the provided content.

## Methodology in Plain English

The authors download tick-by-tick trades and bin them into 1-minute buckets, producing three parallel series per asset-exchange pair: log-returns, traded volume, and number of transactions. They then ask whether these three series behave the way normal market activity behaves.

- **Tails:** they plot how often unusually large values occur (the complementary cumulative distribution) and compare the curves to power-law and stretched-exponential shapes, plus a Gaussian benchmark for contrast.
- **Memory:** autocorrelation functions measure how long a spike in volatility, volume, or trade counts leaves a trace.
- **Multifractality:** MFDFA removes local polynomial trends (degree 2 here) from the integrated series, computes fluctuation functions at many time scales and many fluctuation-amplitude orders *q*, and extracts generalised Hurst exponents and singularity spectra. A spectrum width indicates how many different scaling behaviours coexist; a well-formed spectrum means small and large fluctuations are both organised.
- **Cross-correlations:** MFCCA extends the same detrending idea to pairs of series, yielding ρ(q,s) — a scale- and amplitude-dependent correlation coefficient that still works when scaling is imperfect, which matters because the anomaly itself may destroy scaling. The main analysis uses q = 2.
- **Irregularity:** approximate entropy and sample entropy count how often short patterns of length *m* recur when extended by one point. Higher values mean less repeatable, more noise-like local dynamics. Parameters are m = 2, τ = 1, r = 0.2σ, with robustness checks at m = 3, 4, 5; SampEn is added to remove ApEn's self-matching bias, and undefined values (zero matches) are discarded rather than interpreted.
- **Time-localisation:** cross-correlations and entropy are computed in rolling windows to pinpoint when a regime changes.

The comparison logic is cross-sectional: a given asset's returns look the same everywhere, so any measure where one exchange stands apart from the other three is a candidate anomaly.

## Why This Matters

**Impact on research.** The paper reframes manipulation detection as a problem about the *statistical organisation* of market activity rather than about individual suspicious trades. It shows that price-based stylised facts (heavy tails, volatility clustering) are nearly useless as venue-level diagnostics because arbitrage makes them universal, while activity-based measures retain exchange-specific information. It also gives econophysics tools a concrete forensic application with a dated, replicable case.

**Real-world applications:**

- **Exchange due diligence and ranking integrity** — data aggregators and institutional allocators could screen venues for trade-count inflation that would otherwise improve an exchange's apparent liquidity ranking.
- **Market-quality monitoring by regulators** — persistent decoupling of transaction counts from volume and returns can serve as a triage signal for further investigation, without requiring account-level data.
- **Token-listing and market-making decisions** — projects and liquidity providers can check whether a venue's apparent depth reflects genuine flow before committing inventory.
- **Data-vendor quality scoring** — historical datasets could be flagged where reported activity is unlikely to represent real economic risk transfer.

**Industry relevance.** The framework uses only publicly available trade-level archives, is asset-agnostic, and does not require order-book or identity data, which is exactly the constraint practitioners and regulators face on centralised venues. The authors are explicit that the indicators are diagnostic, not conclusive — a distinction that matters for anyone acting on such signals.

## Future Directions

- **Ground-truth validation.** Because the framework identifies statistical anomalies rather than proven manipulation, the natural next step is testing it against cases where wash trading has been confirmed (for example through exchange enforcement actions or blockchain-linked account data).
- **Extending to the truncated analyses.** Full reporting of the transaction-count multifractal spectra, the ρ(q,s) cross-correlation results, and the rolling-window ApEn/SampEn timelines would allow calibration of thresholds for automated alerting.
- **Coverage beyond three assets and four venues.** Whether the Bitget-style signature generalises across smaller-cap tokens, derivatives markets, and other exchanges with heavy retail or copy-trading user bases is untested here; the authors note the effect was absent for XRP on the same venue.
- **Causal attribution.** The paper cannot distinguish coordinated automated strategies, excessive order fragmentation, and exchange-side liquidity-provision mechanisms. Linking the statistical signature to specific order-generation mechanisms, or to exchange design features, remains open.

## Target Audience

Quantitative researchers and graduate students in econophysics and complex systems; market-microstructure and market-integrity analysts at regulators, exchanges, and data aggregators; and quantitative traders or risk managers who need to assess whether a venue's reported activity is trustworthy. Readers without a background in detrended fluctuation analysis will find the methods sections demanding, but the results and the Bitget case study are accessible to anyone comfortable with time-series statistics.

## Authors’ abstract

Artificial transaction generation remains an important source of potential market manipulation on cryptocurrency exchanges, as it may distort reported liquidity and reduce market transparency. This study proposes a diagnostic framework for detecting unusual trading patterns based on complexity and statistical-structure measures derived from high-frequency trade-level data. The analysis considers log-returns, trading volume, and transaction counts, using tail distributions,…

[The full abstract on arXiv](https://arxiv.org/abs/2607.13916)

[Read the original paper](https://arxiv.org/abs/2607.13916)

---

Canonical: https://crypto.world/paper/detecting-unusual-trading-patterns-on-cryptocurrency-exchanges-by-means-2607.13916
