Your chain ladder spits out a number. Before you book it, you owe two questions: is the data that fed the model clean, and does the answer make sense? Vetting answers both.
Vetting is not one step. First you scrub the input data. Then you interrogate the output. A clean model on dirty data gives a precise wrong answer. A clean model on clean data can still give an unreasonable answer if the method is wrong for the environment.
KEY: Every anomaly you find resolves into one of three explanations: a data error, a genuine operational change (faster settlement, stronger case reserves), or an inappropriate method. Your job is to decide which.
Friedland frames data verification as four reviews. Run all four before you trust the triangle.
Consistency with financial statement data. Reconcile the loss triangles to the company's financial records. Total paid losses in the triangle should tie to paid losses on the Annual Statement. If they differ, something is missing or double counted.
Common mistakes
- Averaging away an outlier factor. A single 2.500 among 1.5s is not noise to smooth; it is a cell to investigate. Averaging it in overstates every green year.
- Treating a sub-1.0 reported factor as impossible. A 0.660 reported age-to-age factor is a flag, not a verdict. Reported losses fall from claims closing under case reserves, case takedowns, or salvage and subrogation. Rule those out before you book it as a data error.
- Booking without reconciling to the financials. If triangle paid does not tie to Annual Statement paid, the ultimate is built on an incomplete or double-counted base.
Bottom line
- Vetting has two layers: validate the input data, then test whether the results pass a reasonableness check.
- Verify data four ways: reconcile to the financial statements, compare against the prior study, test reasonableness, and confirm each field's definition.
- Cumulative paid cannot decrease and paid cannot exceed reported, but cumulative reported can fall as normal favorable development.
- Diagnostic ratios flag distortions: paid-to-reported, average case outstanding, average paid, and closed-to-reported claim counts.
Exam shortcut
Run data checks in a fixed order: reconcile to the financials, compare against the prior study, confirm each field's definition, then confirm paid is non-decreasing and never above reported, and confirm closed counts never exceed reported counts. Most planted errors trip one of these. For any reported development factor below 1.0, stop and investigate.
The full lesson (about 2,338 words, 16 min read) adds 2 worked examples, all 7 common mistakes, a self-check, free in the app.
Learning objectives
- B2
Browse all free Exam 5 lessons or jump into free Exam 5 practice questions.