A fraud model performs well on training data but poorly on unseen transactions. What problem is likely?

1. Guaranteed GPU hardware failure
2. A required absence of labeled data
3. Overfitting or unrepresentative evaluation
4. Automatic improvement from removing the test set