PCPA: Data cleaning should preserve a record of the decision
Missing, implausible and extreme observations are not the same problem. A cleaning step should specify the issue, treatment and effect on the intended target. Replacing all unusual values can erase meaningful insurance tail risk.
Worked example or practice scenario
An exposure of −.5 years is impossible on a standard positive-exposure definition and needs investigation. A valid 2m claim may be rare but relevant. A loss of −100 might be a recovery or correction, requiring the agreed data definition before exclusion.
Try this next
Keep an original practice cleaning log with affected rows and summaries before and after treatment. Reconcile totals to the raw extract. State whether the model estimates gross or net amounts so a recovery is not mistaken for a data error.
Reading sources
ActNet editorial guide · October 1, 2026 · Original illustrative examples.