Out-Of-Sample Testing The Co-Spend Heuristic

Abstract

We conduct out-of-sample analysis of the "co-spend heuristic" in blockchain tracing and conclude it is prone to hallucinating structure in data where none is present. We employ a novel approach that allows us to conduct out-of-sample testing using real-world transaction data taken from real obfuscation transactions conducted on real services with real money. By using a simple isomorphism to convert obfuscation transactions from a zero-knowledge mixing service on Ethereum into a Coinjoin-based service on Bitcoin we are able to measure the out-of-sample accuracy of the co-spend heuristic under standard scientific principles. We find this heuristic is far less reliable than claimed, using only in-sample testing, by many proponents of the technique. Our results also strongly suggest this heuristic, as presented in the public literature, has severe overfitting problems.