ZSH: Rank-Weighted Profiling of Bitcoin Transactions and a Pre-Specified Evaluation of What the Profiles Capture

Abstract

Bitcoin’s ledger records what each transaction moves but not what it does; unsupervised profiles describe it without labels, but are rarely tested against information the clustering never saw. We evaluate ZSH, which weights twelve features by the rank of their mutual information with a proxy partition, initialises K-means hierarchically and caps cluster size; plan and code were frozen before the reported runs. Fitted on 3.3 million transactions from 2022 to 2023, it was tested on 2.6 million from 2024 and 456,292 collected to August 2026. Its 31 profiles reached 78–84% of the ceiling for the four annotations they capture best, under 11% for the five least frequent. Neither oracle weights nor six other families changed this, though for P2SH inputs supervision on the same features reached 95% against the profiles’ 24%. Refits agreed at an adjusted Rand index of 0.83, yet by 2026 three profiles held three quarters of the sample and none survived a free refit; constraining it to the previous centroids followed 24 of 31 in 2024 and 14 in 2026. For most annotations, the limit, among the methods tested, appears to be the clustering objective, not the features; for exchange tags richer features raise what both supervision and the profiles extract—the profiles from 3.0% to 9.4% of the ceiling and supervision to 59.6%. These profiles are a readable map of activity, not a detector.