Accuracy
The fraction of predictions that match the true label.
- Simple and intuitive, but misleading on imbalanced data.
- A "predict majority" model can score 99% and be useless.
- Always pair with precision, recall, or F1.
When to use: Balanced classification where both error types cost roughly the same.
Example: Fraud detection where 0.5% of transactions are fraud: a model that always predicts "not fraud" scores 99.5% accuracy and has zero value.