PCPA: Entity and time leakage require deliberate split design
Records from one insured or one event can share information. Placing related records on both sides of a random split may overstate performance for genuinely new insureds or future events. Match the split to the prediction task.
Worked example or practice scenario
An original practice dataset contains five years for each business. A random row split allows the same business into training and validation. A group-held-out split instead tests new-business generalisation; a later-year split tests temporal generalisation. Neither automatically answers both questions.
Try this next
Describe the deployment population, then justify the split. Fit transformations within the training procedure. Compare segment and period results and identify what remains uncertain, rather than promising that one split proves universal robustness.
Reading sources
ActNet editorial guide · October 1, 2026 · Original illustrative examples.