Principles of Actuarial Modelling
This domain tests how you turn a messy real-world problem into a formal, testable representation, and how you judge whether the resulting model can be trusted. CM2 expects you to justify modelling choices, separate deterministic from stochastic approaches, and describe the control cycle linking data, assumptions, projection and monitoring.
Core rules
- A model is a deliberate simplification. Parsimony and relevance matter more than realism; anything you simplify away becomes model risk that must be disclosed.
- Deterministic models produce a single path (e.g. fixed-interest accumulation). Stochastic models produce a distribution, so you can quote percentiles, ruin probabilities and confidence intervals.
- The actuarial control cycle: define the problem, build the model, calibrate to data, validate, communicate, then monitor and revise.
- Separate the sources of uncertainty: process risk (inherent randomness), parameter/estimation risk, and model error.
- The Markov property is an assumption you impose, not a fact of nature.
- Validation: back-testing, out-of-sample checks, sensitivity and scenario testing, and stochastic simulation with an economic scenario generator.
Worked example
A fund of 10,000 is projected for 10 years. At 4%, 10,000 x 1.04^10 = 14,802. At 3%, 10,000 x 1.03^10 = 13,439. A 1 percentage point change in the assumption cuts the projected fund by 1,363, i.e. 9.2%. That one-line sensitivity test quantifies assumption risk; the same rate change moves a matching liability in the opposite direction, so the sign of the effect matters as much as its size.
Traps that cost marks
- Presenting a single deterministic number as though it were a distribution.
- Confusing precision (many decimal places) with accuracy.
- Applying a model outside the range over which it was calibrated.
- Double counting uncertainty by adding model risk and parameter risk that overlap.
- Forgetting to state assumptions as part of the answer.