A model scores very well on its training examples but performs poorly on new, representative examples. What is the most likely issue?

1. An absence of model capacity
2. Guaranteed generalization
3. Overfitting
4. Successful prevention of data leakage