Temporal Validation

Temporal validation tests a model on a period after its training window, so it is graded the way production will use it. It exposes regime change that a random split conceals.

A random train and test split lets a model be graded by its contemporaries, matters from the same period, the same judicial climate and sometimes the same wave of litigation. That flatters the result. A chronological split trains on the past and tests on a later period the model never saw.

If performance holds across the cut, it is evidence. If it does not, the model does not ship. The same principle applies to duration estimates, which are validated on later periods before a duration claim is made.

Temporal validation is necessary and not sufficient. It sits beside calibration review, sufficiency checks and leakage tests, and does not replace them.