On-Chain Behavioral Signatures for Classifying Centralized Exchange Addresses on Ethereum Using Proof-of-Reserves (arXiv (Cornell University))

Summary

• The study constructs a dataset of 3,033,857 Ethereum addresses from 15 CEXs and five Non-CEX categories using Proof-of-Reserves disclosures.p.1
• Using 17 selected features and a 730-day lookback window, LightGBM achieved a macro PR-AUC of 0.9594, outperforming logistic regression at 0.7697.p.1
• Performance increased from 0.6003 with four features and a 30-day lookback to 0.9594 with 17 features and a 730-day lookback.p.2
• Feature selection reduced the initial taxonomy of 58 on-chain behavioral features down to a compact set of 17 features.p.6
• Gas behavior and internal transactions proved to be key signals, with avg_gas_used accounting for 23.8% and internal_tx_count_out for 10.7% of total feature attribution.p.7
• On an external dataset of 656 CEX addresses from third-party labels, the model achieved a 90.34% recall at a 1% false-positive rate threshold.p.7
• Applying exchange-balanced reweighting to CEX training samples raised the macro PR-AUC from 0.9594 to 0.9667.p.8

Key insights

1. Centralized exchange address identification improves monotonically with longer temporal lookback windows and broader feature coverage.p.2
2. CEX addresses exhibit concentration at lower values for average gas usage and outbound internal transaction counts compared to Non-CEX addresses.p.5
3. The leave-one-exchange-out protocol ensures complete exchange disjointness between model training and test evaluation.p.5
4. Model predictions generalize better to hot wallets with 95.82% recall and deposit addresses with 92.82% recall than to cold wallets at 83.70%.p.7
5. Feature-group ablation shows that removing gas usage and fee features causes the largest performance drop of 0.0289 in macro PR-AUC.p.8
6. Reweighting training samples to equalize exchange influence mitigates dominance by larger exchanges like Binance.p.8