Shadin A Wallet risk
Machine learning study
78% accuracy
0 high-risk wallets caught
Reported instead: recall and ROC-AUC on the high-risk class.
- What ran
- A classifier for high-risk crypto wallets on a synthetic, heavily imbalanced dataset, in Python and scikit-learn.
- What worked
- The model scored 78% accuracy and caught none of the high-risk wallets. The study reports recall and ROC-AUC instead, because accuracy said it worked when it hadn't.