Model Evaluation: Beyond Just Accuracy

Early on, I would just look at accuracy to see if my model was good. Then I worked on a fraud detection project. If 99% of transactions are normal, a model that just predicts 'normal' every time has 99% accuracy—but it’s completely useless. That’s when I learned about precision, recall, F1 scores, and confusion matrices. Accuracy can be a liar. You have to think about the cost of false positives versus false negatives. Understanding the business problem is key to choosing the right metric.