Can AI Detect Wash Trading? Evidence from NFTs
Brett Hemenway Falk, Gerry Tsoukalas, Niuniu Zhang
TL;DR
This paper provides direct, on-chain evidence of NFT wash trading across three major marketplaces, estimating that roughly 38% of trades and 60% of value involve manipulation with marked cross-exchange variation. It critically evaluates indirect detection methods, showing that the traditional 1\% roundedness threshold is often inadequate in NFTs, and introduces a platform-tuned $\tau$-regression that reduces error. To further enhance accuracy, the authors develop an AI-based estimator that integrates optimized regressions with rich trade features, achieving aggregate errors below 2.15% and high ROC-AUC on NFT data and robust generalization to fungible assets like Bitcoin Mt. Gox. The work has regulatory relevance, offering a scalable detection framework that can inform live surveillance while acknowledging exchange-specific dynamics and the limits of purely distributional tests.
Abstract
Existing studies on crypto wash trading often use indirect statistical methods or leaked private data, both with inherent limitations. This paper leverages public on-chain NFT data for a more direct and granular estimation. Analyzing three major exchanges, we find that ~38% (30-40%) of trades and ~60% (25-95%) of traded value likely involve manipulation, with significant variation across exchanges. This direct evidence enables a critical reassessment of existing indirect methods, identifying roundedness-based regressions à la Cong et al. (2023) as most promising, though still error-prone in the NFT setting. To address this, we develop an AI-based estimator that integrates these regressions in a machine learning framework, significantly reducing both exchange- and trade-level estimation errors in NFT markets (and beyond).