Data, modelling and actuarial judgement in general insurance
This domain tests the quality and use of data, the choice and validation of models, and the disciplined exercise of judgement when data are limited.
Core concepts
- Data quality: completeness, accuracy, granularity, consistency and relevance. Exposure data are as important as claims data.
- Credibility: weighting an insured's own experience against a broader benchmark. Bühlmann credibility is Z = n/(n + k), where k reflects between-risk variance.
- Models: GLMs and GAMs for frequency-severity; machine learning for segmentation; catastrophe models for tail risk. Validation includes backtesting, out-of-sample testing, sensitivity and scenario analysis.
- Judgement: documented, reasoned and challenged. Model risk must be governed, not hidden.
Worked example
A motor portfolio has expected frequency 5%. An insured has 400 exposure units and observed frequency 8%. With k = 1,500, Z = 400/(400 + 1,500) = 0.21. The credibility-weighted estimate is 0.21 × 8% + 0.79 × 5% = 5.63%. The insured's experience is credible enough to move the estimate, but not enough to replace the benchmark.
Traps that cost marks
Fitting a model without an exposure offset; overfitting a flexible model to sparse data; ignoring data leakage in machine-learning validation; and presenting a model output without sensitivity or limitations.
Exam tip
Link the model to the decision. A pricing model needs predictive stability; a reserving model needs explainable development; a capital model needs tail credibility.