A feature contains information that would not exist at prediction time. What is this?
Why B is the answer
Data leakage
Leakage makes validation scores look better than real performance.
In detail
- Information unavailable at prediction time is data leakage.
- Ask if each feature exists at prediction time.
Use case
An illustrative scenario, not a real incident.
A churn model uses a "cancellation date" feature and scores near perfectly, but the date does not exist when predicting.
See it in different ways
flowchart LR A["All data"] --> B["Training: fit the model"] A --> C["Validation: tune the model"] A --> D["Test: final check"]
mindmap
root((Data preparation))
Missing values
Impute
Categories
One-hot encoding
Ranges
Scaling
Imbalance
Resampling
Class weights
Leakage
Remove future dataflowchart TD
Q{"Does it exist at prediction time?"}
Q -->|"Yes"| A0["Keep it"]
Q -->|"No"| A1["Remove it, this is leakage"]Key point: Leakage inflates validation scores.
Why the other options are wrong
- A. Regularization
- Regularization reduces overfitting.
- C. Underfitting
- Underfitting is a model that is too simple.
- D. Concept drift in production
- Drift happens after deployment.