Choosing metrics for an imbalanced fraud model
Interview insights are lockedSee how often this is asked and which companies asked it with Buncho Pro.A bank has 1,000,000 card transactions per month, of which 0.2% (2,000) are fraudulent. Your model flags 3,000 transactions and 1,500 of them are truly fraud. Its accuracy is 99.8%.
- Is 99.8% accuracy good? What accuracy would a model that always predicts "not fraud" get?
- Compute precision, recall and F1 for this model.
- Which metrics and which decision threshold would you choose, and how would you evaluate before launch?