Actuarial Modeling, Data Analysis and Communication
The final domain tests whether you can build, validate and explain a model rather than only run one. The exam combines model governance, data quality, statistical technique and - often underweighted by candidates - the ability to communicate uncertainty to a board or a regulator in plain terms.
Model lifecycle and governance
- Maintain a model inventory; each model needs documented purpose, scope, owner, assumptions, limitations and change-control history.
- Independent validation: a party separate from the developer replicates results, challenges assumptions and tests sensitivities. Validation is not the same as back-testing.
- Data controls: completeness, accuracy, consistency and timeliness, with reconciliation between policy administration, claims and reinsurance systems before any assumption is set.
- Assumption setting uses experience analysis (mortality, morbidity, lapse, expense), credibility weighting, trend analysis and allowance for trend tails.
Analytical technique
- GLMs remain the pricing workhorse: Tweedie for pure premium, Poisson frequency × Gamma severity, a log exposure offset, and checks on coefficient signs and significance.
- Machine learning (gradient boosting, random forests, neural nets) improves motor pricing and fraud detection but requires out-of-sample validation, interpretability tools and monitoring for drift.
- Credibility theory blends collateral data with own experience.
Communication
- Lead with the conclusion, then the range of uncertainty, then the key assumptions and their sensitivities (for example interest ±50bp, mortality ±10%).
- Every figure quoted externally must reconcile to the regulatory filing; reconciling items must be explained, not buried.
Worked example
A group has 12 years of experience. Expected process variance is 400 and the variance of hypothetical means is 100, so K = 400/100 = 4 and Z = n/(n + K) = 12/16 = 0.75. The observed loss ratio is 0.30 and the manual rate 0.20, giving a credibility premium of 0.75 × 0.30 + 0.25 × 0.20 = 0.275. Reporting only 0.30 overstates the signal from a modest sample; the exam rewards stating Z alongside the estimate and explaining that a larger K would pull the answer further toward the manual rate.
Traps that cost marks
- Swapping the expected process variance and the variance of hypothetical means.
- Presenting a point estimate with no sensitivity or uncertainty range.
- Treating in-sample fit as validation.
- Quoting a model output that does not reconcile to the regulatory filing.