Executive summary
- An actuarial model produces a distribution, not a prediction; the distribution is the product.
- The same model chain feeds pricing, reserving, capital, reinsurance and risk appetite, which is why inconsistency between these is a governance signal.
- Model error is concentrated in data quality, exposure definition and assumption choice — not usually in the statistical technique.
- Executives should ask what would have to be true for the model to be materially wrong, and what would be observed first if it were.
- The value of the modelling is realised only where an actual decision changes as a result.
The product is a distribution, not a number
A pricing model that returns an expected loss of 4.2 per cent of sum insured has said very little on its own. The same expected loss can arise from a portfolio with modest, frequent claims and from one with rare, severe losses. The first can be written with limited capital; the second cannot. The distinguishing information is in the shape of the distribution, not its mean.
This is why actuarial output should be presented with dispersion and tail measures alongside the central estimate. A reserve best estimate accompanied by a range, with the drivers of that range named, supports a different conversation from a single booked figure. A capital number accompanied by the scenarios that generate it supports a different conversation from a percentile.
Three sources of uncertainty
- Process uncertainty: the inherent randomness of outcomes even if the model and parameters were exactly right.
- Parameter uncertainty: the estimation error in the parameters, which is large when data is thin.
- Model uncertainty: the possibility that the chosen structure is wrong. This is rarely quantified and is often the largest of the three.
Most stochastic reporting captures the first, sometimes the second, and almost never the third. Presenting a confidence interval derived only from process uncertainty conveys a precision the analysis does not possess.
One model chain, five decisions
The same underlying view of frequency, severity, dependence and inflation should inform pricing, reserving, capital, reinsurance purchase and risk appetite. In practice these are often produced by different teams on different cycles with different assumptions, and the inconsistency is discovered only when results diverge.
| Decision | What the model supplies | Typical failure mode |
|---|---|---|
| Pricing | Expected loss, expense and capital loading by segment | Capital cost applied as a flat percentage rather than by segment volatility |
| Reserving | Distribution of ultimate cost and its development | Booked figure disconnected from the modelled range without documented rationale |
| Capital | Aggregated tail across risk types | Diversification assumption unexamined and undisclosed |
| Reinsurance | Retained tail under alternative structures | Structures compared on historical recoveries rather than modelled relief |
| Risk appetite | Probability of breaching stated thresholds | Appetite expressed qualitatively so no model output can test it |
What executives should ask
- What data is this built on, over what period, and what has changed in the portfolio since then?
- Which three assumptions move the answer most, and what is the result under reasonable alternatives?
- What would have to be true for this to be materially wrong, and what would we observe first?
- How does this compare with the previous version, and what explains the movement — experience, model change or assumption change?
- What decision changes if the answer moves by twenty per cent in either direction?
The last question is the most useful. If no decision changes across a wide range of outputs, the modelling effort is disproportionate to its purpose. If a decision flips on a small movement, the position is more finely balanced than anyone has acknowledged.
The place of actuarial judgement
Judgement is unavoidable. Data is incomplete, portfolios change, and emerging exposures have no relevant history. The professional standard is not to eliminate judgement but to make it explicit: state the assumption, state the basis, state the alternative considered and state the effect of choosing differently.
Judgement becomes a governance failure when it is embedded in a spreadsheet cell with no annotation, when it is applied inconsistently between reporting periods without explanation, or when it consistently moves results in the direction that suits the current commercial objective. Each of these is detectable through review; none is detectable through statistical testing alone.
Practical example
Two portfolios each have an expected annual loss of 10 units. Portfolio A has a standard deviation of 2 units and a 99th percentile loss of 16 units. Portfolio B has a standard deviation of 9 units and a 99th percentile loss of 48 units, driven by a small probability of a large accumulation event.
Priced on expected loss plus a uniform 20 per cent loading, both would carry a premium of 12 units. Priced on the capital each consumes — approximately 6 units for A and 38 units for B above the expected loss — the required premium differs substantially, and Portfolio B may be uneconomic at any price the market will accept. The modelling has not predicted anything; it has revealed that a uniform loading was cross-subsidising volatility, which is a decision the board can act on.
Limitations and caveats
- Models are estimated on historical experience that may not represent the exposure being written today.
- Model uncertainty is rarely quantified, so reported ranges generally understate true uncertainty.
- Dependence assumptions between segments and perils are weakly evidenced and strongly influential in the tail.
- Statistical validation cannot detect a well-fitted model applied to the wrong problem; that requires conceptual review.
Conclusion
Actuarial risk modelling earns its cost when it changes decisions: which business to write, at what price, with what reinsurance, against what capital and within what appetite.
The technical quality of the model matters, but the determining factor is whether the output arrives in a form the decision-maker can interrogate — with assumptions visible, ranges stated and limitations declared.
References
- Actuarial Standards and Professional Guidance — Institute and Faculty of Actuaries
- Actuarial Standards of Practice — Society of Actuaries / Actuarial Standards Board
- Solvency II Directive 2009/138/EC — European Commission, 2009
Author bio
Jonas Mohamed Osman Abdelghafour, known as Yonas Osman is an actuary, FRM and financial risk professional specialising in banking, insurance, model risk, capital modelling and quantitative risk management.