Risk Modelling and Survival Analysis · Mortality projection

Parameter Uncertainty and Stochastic Projections in Mortality

Updated 11 October 2026 · Fact-checked

Mortality projections are uncertain for three reasons: the model may be wrong (model risk), its fitted parameters are estimated with error (parameter risk), and future deaths are random (stochastic risk). You measure this by simulating many future paths, then reading percentiles of the outputs such as annuity values.

Understand Parameter Uncertainty and Stochastic Projections

A mortality projection is a forecast of future mortality rates. It is never a single certain answer. You must be able to say where the doubt comes from and how to measure it.

There are three main sources of uncertainty. Model risk means the chosen model (for example Lee-Carter or CBD) may have the wrong structure. Parameter risk means the model is right in form, but its parameters are estimated from limited data, so they carry error. Stochastic risk (process risk) means that even with a perfect model and perfect parameters, actual deaths in a finite group are random.

Longevity risk is the risk that people live longer than assumed. It has two parts. Trend risk is the risk that the whole population improves faster than projected. This cannot be diversified by writing more policies. Idiosyncratic risk is the random variation of individual lives. This falls as the portfolio grows.

A deterministic projection gives one set of rates. A stochastic projection gives many. You simulate the model many times, each time drawing random errors for the period effect, and often a fresh set of parameters. Each simulation produces a full set of future mortality rates. You then compute a quantity of interest, such as a pension annuity value, for each path. The spread of results shows the uncertainty.

A scenario is a chosen what-if, such as a sudden medical breakthrough or a stalled improvement. Scenarios are easy to explain and can test extreme but plausible cases. Stochastic models give probabilities, but only inside the model. Use both. In pension valuation, these results help set reserves, test funding, and price longevity swaps or buy-outs.

Key rules to remember

Lee-Carter model: ln m(x,t) = a(x) + b(x) × k(t) + ε(x,t); Usual constraints: Σ b(x) = 1 and Σ k(t) = 0. k(t) is the period index, often projected as a random walk with drift.
Random walk with drift for k(t): k(t+1) = k(t) + μ + σ × Z(t+1), with Z ~ N(0,1) independent; μ is the drift, σ the volatility of the period index. Parameter risk arises because μ and σ are estimated.
Projected k(t) after s years: k(T+s) = k(T) + s × μ + σ × √s × Z; Mean is k(T) + sμ and variance is sσ². Uncertainty grows with the square root of the horizon.
Total variance decomposition: Var(Y) = E[Var(Y | θ)] + Var(E[Y | θ]); Here θ is the parameters. The first term is process-type variation, the second is parameter uncertainty.
Simulated estimate: Estimate = (1/N) × Σ Y(i), for i = 1 to N; Standard error of the mean is s ÷ √N, where s is the sample standard deviation of the simulated values.
Percentile from simulations: Sort the N values; the p-th percentile is about the (p × N)-th smallest value; Used to quote a 95th or 99.5th percentile annuity value.

How to solve Parameter Uncertainty and Stochastic Projections questions

Use this method for most exam questions on uncertainty and stochastic projection.

1Identify what is being projected and what output matters, such as k(t), mortality rates, life expectancy or an annuity value.
2Name the sources of uncertainty that apply: model, parameter and stochastic (process) risk. Say which the question is about.
3State the model and its assumptions, for example k(t) follows a random walk with drift and normal errors.
4Write down the formula for the projected quantity, using standard notation, and insert the given values.
5If simulating, describe the loop: draw random errors, build k(t) paths, convert to rates, compute the output, repeat N times.
6Summarise results with a mean, standard deviation and the percentile asked for.
7Comment on limits: the result holds only if the model is right, and parameter risk may be understated if parameters are treated as fixed.
8Link to the use, such as reserves, capital or hedging, if asked.

Quickest way: Three-source check plus random walk arithmetic

When to use it: Use for short MCQs and for the first marks of written questions.

Ask: is the doubt about the structure (model), the estimated numbers (parameter) or chance deaths (stochastic)?
For k(t) questions, use mean = k(T) + sμ and standard deviation = σ√s.
For a percentile, use mean ± z × standard deviation, with z = 1.645 for 95% one-sided and 1.96 for 95% two-sided.
For simulation accuracy, remember standard error = s ÷ √N. To halve it, you need four times as many runs.
Finish with one sentence on what the result is conditional on.

Common mistakes in Parameter Uncertainty and Stochastic Projections

Confusing parameter risk with stochastic risk.
Both involve randomness and both appear in simulations.
Fix: Parameter risk is about estimation error in fitted values. Stochastic risk is chance variation in outcomes even when the parameters are known exactly.
Saying that a larger portfolio removes longevity risk.
Students apply the law of large numbers to everything.
Fix: A larger portfolio reduces idiosyncratic risk only. Trend risk affects all lives together and stays.
Using σ × s instead of σ × √s for uncertainty in k(t) over s years.
The drift grows linearly, so students scale the volatility the same way.
Fix: The drift term is sμ. The variance is sσ², so the standard deviation is σ√s.
Treating a simulation percentile as a true probability of the real world.
The output looks precise.
Fix: State that it holds only under the chosen model and parameters. Model risk is not captured.
Ignoring parameter uncertainty in the simulation.
It is simpler to fix μ and σ at their estimates.
Fix: Resample or draw parameters from their estimated distribution for each run, or say that the spread is understated if you do not.
Giving only a stochastic answer and no scenarios.
Students think a model is always better than judgement.
Fix: Mention scenarios as a check on extreme cases and as an easy way to communicate risk to non-technical users.

Worked examples

Example 1

A Lee-Carter period index has last fitted value k(T) = −20. It is projected as a random walk with drift μ = −1.5 per year and σ = 2. Find the mean and standard deviation of k(T+9), and the value that k(T+9) falls below with 95% probability (one-sided, z = 1.645).

Show the solution

Mean = k(T) + sμ = −20 + 9 × (−1.5) = −20 − 13.5 = −33.5.
Variance = sσ² = 9 × 4 = 36, so the standard deviation = √36 = 6.
95% one-sided upper bound = mean + 1.645 × 6 = −33.5 + 9.87 = −23.63.

Answer: Mean −33.5, standard deviation 6, and k(T+9) is below about −23.63 with 95% probability. Higher k means higher mortality, so −23.63 is the pessimistic-for-mortality bound, a favourable-to-mortality view is the lower tail.

Example 2

An insurer simulates N = 400 values of the present value of a pension annuity. The sample mean is ₹12,00,000 and the sample standard deviation is ₹90,000. (a) Find the standard error of the mean. (b) Give an approximate 95% confidence interval for the true model mean. (c) Explain why this interval understates the total uncertainty.

Show the solution

(a) Standard error = 90,000 ÷ √400 = 90,000 ÷ 20 = ₹4,500.
(b) Interval = 12,00,000 ± 1.96 × 4,500 = 12,00,000 ± 8,820, which is ₹11,91,180 to ₹12,08,820.
(c) The interval only reflects simulation error, which falls as N rises. It does not include model risk. It also ignores parameter risk if the parameters were held fixed in the runs.

Answer: Standard error ₹4,500. The 95% interval for the model mean is ₹11,91,180 to ₹12,08,820. It covers only sampling error from simulation, not model or parameter uncertainty.

Exam tips

Always name the three sources of uncertainty and say which one the question tests.
Show the random walk arithmetic: mean uses s, standard deviation uses √s. Examiners look for this.
When asked to comment, give both a strength and a limit of stochastic models, and mention scenarios.
For pension questions, link results to reserves, funding and longevity hedging, and note that trend risk cannot be diversified.
In computer-based questions, describe the simulation loop clearly and state the number of runs and the seed if used.

Practice questions from Mortality projection

Parameter Uncertainty and Stochastic Projections: frequently asked questions

What is the difference between parameter risk and model risk?

Parameter risk is error in the estimated values of a model whose form is assumed correct. Model risk is the chance that the form itself is wrong. Comparing several models helps assess model risk.

How do you simulate a stochastic mortality projection?

Fit the model, then project the period index with a time series such as a random walk with drift. For each run, draw random errors, build future rates, and compute the output such as an annuity value. Repeat many times and study the spread.

Why does longevity risk matter for pension valuation?

Pension liabilities last for decades, so small changes in assumed improvement change the value a lot. Trend risk affects all members together, so it cannot be removed by pooling.

When should I use scenarios instead of stochastic models?

Use scenarios to test specific extreme events and to explain risk simply. Use stochastic models when you need a distribution and percentiles. Good practice uses both.