Feature distributions are stable, but the relationship between inputs and outcomes changes and accuracy falls. Which issue should be investigated?

1. Only the model's name
2. Only storage capacity
3. A guarantee that no retraining is ever needed
4. Concept drift or changed predictive relationships