Blockchain network complexity and illicit transaction detection: Machine-learning evidence from the Elliptic Bitcoin benchmark

Abstract

Type of the article: Research ArticleAbstractThe blockchain financial system allows users to send money fast without any border restrictions. However, the same structure of the blockchain may be used as a means of laundering money. This paper assesses the relationship between the complexity of transaction networks and the likelihood of their illicit nature within the public Elliptic Bitcoin benchmark and examines whether anomaly detection using machine learning helps to interpret risks from an AML/CFT perspective. This empirical analysis assumes that Elliptic provides an anonymized transaction network in which nodes correspond to Bitcoin transactions, edges reflect directed transactions, and anonymized features facilitate licit/illicit classification of transactions. Furthermore, the dataset is not considered evidence of sender wallet addresses, receiver wallet addresses, transaction amount, timestamp, ownership of exchanges, user geography, and national AML/CFT effectiveness. Based on the labelled analytical dataset presented in the uploaded workbook (46,564 observations, including 42,019 licit transactions and 4,545 illicit transactions), a logit model found a significant positive correlation between illicit transactions and degree centrality (beta = 1.870, p < 0.001), clustering coefficient (beta = 0.940, p < 0.001), and flow entropy (beta = 0.680, p < 0.001). Isolation Forest and Autoencoder reached AUCs of 0.866 and 0.841, respectively. In turn, the coefficient measuring a country’s regulatory capacity and its interaction term are not included in the estimation because there is no country-window marginal effect. Therefore, this paper does not test for the impact of regulatory capacity of the USA, Singapore, and UAE on transaction classification.