Engineering, production, governance, cost, communication · insight-0265

Exclude immature labels when building training data for slow outcomes

Situation
A default model labels loans as 'not defaulted' if no default is recorded yet, including loans issued last month.
Action
Define the outcome window (for example default within 12 months), include only rows whose window has fully elapsed by the training cutoff, and document the label maturity date in the dataset.
Reason
A label is only final once its observation window has closed; including open windows mixes 'no event yet' with 'no event'.
Boundary
Credit default, churn, chargebacks, insurance claims, conversions with long lags.

Symptom or risk

Recent rows are labelled negative only because the outcome has not had time to occur, so the model learns that recent applicants are safer.

Survival or time-to-event models can use censored rows correctly by design.

Waiting for maturity makes training data older; survival models recover recency at the cost of complexity.

Validating on a recent period whose labels are also immature.