99% Accuracy Is Not Always Good: What Fraud Detection Taught Me About ML

I built a fraud detection model that hit 99% accuracy. I was proud of it for about twelve minutes. Then I looked closer and realized the model was basically doing nothing useful. This is the story of