On-Chain Behavioral Signatures for Classifying Centralized Exchange Addresses on Ethereum Using Proof-of-Reserves
In briefClassifying Exchange Addresses on Ethereum Using On-Chain Behavior and Proof-of-Reserves
SummaryThis paper builds a classifier for identifying Ethereum addresses belonging to centralized exchanges, using Proof-of-Reserves disclosures to label roughly 3.03 million addresses from 15 exchanges and five non-exchange categories. It derives 58 on-chain behavioral features and tests LightGBM against logistic regression under a leave-one-exchange-out protocol. LightGBM outperformed logistic regression across all five lookback windows, and SHAP analysis highlights gas behavior, ERC-20 token activity, and internal transactions as key signals.