The detector that went blind: crypto assets, money laundering, and the Lucas critique

Abstract

This paper trains a detector of illicit bitcoin transactions on the public Elliptic data set (203,769 transactions over 49 time steps of about two weeks each) and shows that its out-of-sample performance collapses after the shutdown of a major dark web marketplace: the detection rate falls from 86% to 2% from one time step to the next, and retraining on the few post-shock observations does not restore it. The episode, already noted by the authors of the data set, is read through the lens of the Lucas critique (1976): a detector estimated on past data summarizes how agents behaved under a given enforcement regime and breaks down when enforcement itself changes that behavior. The paper discusses the implications for Brazilian anti-money laundering regulation, recently extended to virtual asset service providers: continuous validation of monitoring systems, period-by-period rather than aggregate evaluation, the complementarity between algorithms and investigation, and the cost of false alarms.