Machine Learning for Money Laundering Analysis (International Journal of Network Dynamics and Intelligence)

Summary

• Six supervised models on the Elliptic Bitcoin dataset were compared; Random Forest achieved 98.87% accuracy on the holdout set.p.1
• The IMF estimates money laundering at 2% to 5% of global GDP, or $590 billion to $1.5 trillion annually.p.2
• Weber et al. used over 200,000 Bitcoin transactions, of which illicit ones were only 2%.p.5
• Each transaction has 166 features covering transaction flow, privacy and neighborhood information.p.10
• Before SMOTE the training set held 2727 illicit and 25,211 licit transactions; after SMOTE both classes held 25,211.p.13
• Under temporally ordered evaluation, Random Forest macro-F1 drops from 0.965 to 0.826.p.26

Key insights

1. Logistic regression detects 92% of illicit transactions but its precision is only 49% on the test set and 46% on the holdout set.p.18
2. The SVM reaches accuracy of 96.18% on the test set and 96.72% on the holdout set.p.19
3. K-Nearest Neighbors reaches accuracy of 92.56% on the test set and 92.79% on the holdout set.p.19
4. The SGD classifier reaches accuracy of 90.36% on the test set and 89.63% on the holdout set.p.21
5. The Decision Tree reaches accuracy of 96.77% on the test set and 97.22% on the holdout set.p.21
6. Random Forest reaches accuracy of 98.43% on the test set and 98.87% on the holdout set.p.22
7. Random Forest ROC-AUC is 0.996 under random-split evaluation, but performance degrades substantially under temporal evaluation.p.26