Representing Parameter and Model Uncertainty

Why the Source of Uncertainty Matters

A capital estimate can be uncertain for different reasons. Even if a model and its parameters were known exactly, future losses would still vary. Separately, the parameter values may have been estimated from limited data, and the chosen model structure may be only one of several defensible representations. These uncertainties are related, but they are not interchangeable.

For example, claim counts can vary from year to year under a fixed frequency model. That is uncertainty in the loss outcome. The underlying frequency rate may also be uncertain because the available experience is sparse. That is parameter uncertainty. Finally, a Poisson count model may not capture clustering, so a more dispersed count distribution may be a plausible alternative. That is structural model uncertainty. Treating all three as one undifferentiated “margin” can obscure why a capital result changes and what action might address it.

Definition: Outcome uncertainty is the variation in loss $X$ conditional on a specified model and its parameters. Parameter uncertainty is uncertainty about the values of parameters within a specified model structure. Structural model uncertainty is uncertainty about the form, assumptions, or relationships specified by the model itself. These are distinct from the realized loss outcome, which is one value of $X$.

A useful notation makes the distinction explicit. Let $M$ identify a model structure, let $\theta$ denote its parameters, and let $X$ denote the loss outcome. The model specifies a conditional distribution $P(X\mid \theta,M)$. A fixed-parameter capital calculation uses one chosen $M$ and one chosen $\theta$. An uncertainty analysis instead asks how results vary across plausible values of $\theta$, plausible structures $M$, or both.

A distribution assigned to $\theta$ is a representation of uncertainty about the parameter; it does not mean that the parameter physically changes from one claim to another. Likewise, a weight assigned to alternative models is a way to represent their relative plausibility for an analysis, not an observed frequency of model outcomes. Those interpretations should be stated rather than implied.

Representing Parameter Uncertainty

Parameter uncertainty can be represented by a probability distribution for a parameter, a set of plausible parameter values, or a collection of parameter scenarios. A probability distribution is useful when the analysis supports a coherent probabilistic representation, such as a prior distribution updated with data. A scenario set may be more appropriate when probabilities cannot be justified. In either case, report the source and meaning of the parameter values.

When parameter uncertainty is represented probabilistically, the predictive distribution averages the outcome distribution over the parameter distribution:

P(X\leq x)=\int P(X\leq x\mid\theta)\,dG(\theta),

where $G$ is the stated distribution for $\theta$. This is not the same calculation as substituting the mean of $\theta$ into the model. The latter produces a conditional distribution at one parameter value; averaging conditional distributions can produce a different predictive distribution and different tail measures.

Worked Example: Updating a Claim-Frequency Parameter

A hypothetical portfolio has annual claim counts that, conditional on its frequency rate $\lambda$, follow a Poisson distribution. The rate is measured in claims per exposure-year. Before observing experience, use a Gamma prior with shape $\alpha=2$ and rate $\beta=1$. The portfolio then records 3 claims over 2 exposure-years. Estimate the predictive mean and variance of next year's claim count, and the corresponding mean and variance of aggregate loss if each claim costs exactly $100,000.

Assumptions and method. Conditional counts follow $N\mid\lambda\sim\operatorname{Poisson}(\lambda)$ for one exposure-year. The prior is $\lambda\sim\operatorname{Gamma}(2,1)$, using the shape-rate convention, so $E[\lambda]=\alpha/\beta$ and $\operatorname{Var}(\lambda)=\alpha/\beta^2$. Claim severities are fixed at $100,000. No interest or mortality assumption applies.

Step 1: Update the parameter distribution. For $n=3$ claims over exposure $T=2$, the Poisson likelihood as a function of $\lambda$ is proportional to $\lambda^3e^{-2\lambda}$. Multiplying by the Gamma prior gives a Gamma posterior with shape $2+3=5$ and rate $1+2=3$. Thus $\lambda\mid\text{data}\sim\operatorname{Gamma}(5,3)$.

Step 2: Find the predictive count mean. Since $E[N\mid\lambda]=\lambda$, averaging over the posterior gives $E[N\mid\text{data}]=E[\lambda\mid\text{data}]=5/3=1.6667$ claims.

Step 3: Find the predictive count variance. The conditional Poisson variance is $\operatorname{Var}(N\mid\lambda)=\lambda$. The posterior parameter variance is $\operatorname{Var}(\lambda\mid\text{data})=5/3^2=5/9$. Therefore the predictive variance is $5/3+5/9=20/9=2.2222$ claims squared.

Step 4: Convert to aggregate loss. If $X=100{,}000N$, then $E[X]=100{,}000(5/3)=\$166{,}666.67$, and $\operatorname{Var}(X)=100{,}000^2(20/9)=\$22{,}222{,}222{,}222.22$ dollars squared. The mean is consistent with the predictive mean count multiplied by the fixed cost per claim. The variance includes both count variation and uncertainty about $\lambda$.

Common mistake avoided: using only the conditional Poisson variance at the posterior mean rate would give $5/3$, omitting the $5/9$ contribution from uncertainty about $\lambda$. The predictive variance is $20/9$, not $5/3$.

Separating the Components of Predictive Variation

The distinction in the example follows from the law of total variance. For a fixed model structure $M$, and a parameter $\theta$ with a specified uncertainty distribution, the result is:

\operatorname{Var}(X\mid M) = E\!\left[\operatorname{Var}(X\mid\theta,M)\mid M\right] + \operatorname{Var}\!\left(E[X\mid\theta,M]\mid M\right).

The first term is the average conditional outcome variance: it measures variation in $X$ when the parameter is held fixed, averaged over plausible parameter values. The second term is variation in the conditional mean as the parameter varies. Together they give predictive variance under the specified parameter distribution and model structure.

To see why the formula holds, write $X-E[X]=(X-E[X\mid\theta])+(E[X\mid\theta]-E[X])$. Squaring and taking expectations gives the sum of the two squared terms plus twice their cross-product. The cross-product has expectation zero, because conditional on $\theta$, $E[X-E[X\mid\theta]\mid\theta]=0$. The first squared term therefore has expectation $E[\operatorname{Var}(X\mid\theta)]$, and the second has expectation $\operatorname{Var}(E[X\mid\theta])$. This yields the stated decomposition.

Worked Example: Quantifying Parameter Uncertainty in a Poisson Model

For a hypothetical annual portfolio, the model structure is Poisson frequency, but uncertainty analysis assigns equal probability to $\lambda=1$ and $\lambda=3$ claims per year. Each claim has fixed severity of $100,000. Calculate the predictive mean and variance of aggregate loss, and compare the variance with the result from simply using the average rate. No interest or mortality assumption applies.

Step 1: Calculate the average rate and its variance. The mean rate is $E[\lambda]=(1+3)/2=2$. The rate variance is $((1-2)^2+(3-2)^2)/2=1$.

Step 2: Calculate the predictive count mean and variance. Conditional on either rate, $N\mid\lambda\sim\operatorname{Poisson}(\lambda)$. Thus $E[N]=E[\lambda]=2$. By total variance, $\operatorname{Var}(N)=E[\lambda]+\operatorname{Var}(\lambda)=2+1=3$ claims squared.

Step 3: Convert to aggregate loss. With $X=100{,}000N$, $E[X]=100{,}000(2)=\$200{,}000.00$, and $\operatorname{Var}(X)=100{,}000^2(3)=\$30{,}000{,}000{,}000.00$ dollars squared.

Step 4: Check the plug-in calculation. A Poisson model fixed at the mean rate $\lambda=2$ has count variance 2, giving aggregate loss variance of $2(100{,}000)^2=\$20{,}000{,}000{,}000.00$. This is smaller than the predictive variance by $1(100{,}000)^2=\$10{,}000{,}000{,}000.00$, exactly the contribution from rate uncertainty. The mean agrees in both calculations, but the variance does not.

Common mistake avoided: the average rate does not preserve all predictive variation. Substituting $\lambda=2$ into a Poisson model removes the between-rate variance, so it understates the variance under the stated uncertainty representation.

Representing Structural Model Uncertainty

Structural uncertainty concerns which model form is appropriate, not merely which parameter value belongs in one form. Examples include choosing between count distributions with different dispersion, alternative dependence structures, or competing ways to represent a loss process. A model comparison should state what changes and what is held fixed; otherwise a difference in results cannot be attributed clearly.

One approach is to report results separately under each plausible model. Another is model averaging: assign explicit weights to models, form a mixture distribution, and calculate the desired measure from that mixture. A third is to report a range across models without assigning probabilities. These approaches answer different questions. A weighted mixture requires defensible weights, while a range reports model sensitivity without claiming that the alternatives have known probabilities.

Definition: For alternative models $M_1,\ldots,M_k$ with stated weights $w_j\geq0$ and $\sum_j w_j=1$, the model-averaged distribution is $F_X(x)=\sum_{j=1}^k w_jF_{X\mid M_j}(x)$. A quantile of this mixture is calculated from the combined distribution; in general, it is not the weighted average of the separate model quantiles.

Worked Example: Comparing Alternative Count Structures

A hypothetical annual loss is $250,000 for each event. Two equally weighted models are considered. Model A specifies $N\sim\operatorname{Poisson}(1)$. Model B specifies a geometric count with $P(N=n)=0.5^{n+1}$, for $n=0,1,\ldots$; this is a geometric distribution with mean 1. Calculate the model-averaged 95th percentile of aggregate loss. No interest or mortality assumption applies.

Step 1: State the quantile convention. For a loss distribution $F_X$, the 95th percentile is the smallest loss $x$ for which $F_X(x)\geq0.95$. Since each event costs $250,000, it is sufficient to find the corresponding count quantile.

Step 2: Calculate each model's cumulative probabilities. Under Model A, $P(N\leq2)=e^{-1}(1+1+1/2)=2.5/e=0.9197$, and $P(N\leq3)=e^{-1}(1+1+1/2+1/6)= (8/3)/e=0.9810$. Under Model B, $P(N\leq k)=1-0.5^{k+1}$, so $P(N\leq2)=0.8750$ and $P(N\leq3)=0.9375$.

Step 3: Average the cumulative probabilities. The equally weighted mixture has $P(N\leq2)=(0.9197+0.8750)/2=0.8974$, below 0.95. It has $P(N\leq3)=(0.9810+0.9375)/2=0.9593$, at least 0.95. Therefore the mixture's 95th percentile count is 3.

Step 4: Convert to loss and check. The model-averaged 95th percentile of aggregate loss is $3(250{,}000)=\$750{,}000$. The mixture cumulative probability is below the target at 2 events and above it at 3 events, confirming the quantile.

Common mistake avoided: Model A's 95th percentile is 3 events and Model B's is 4 events. Averaging those quantiles gives 3.5 events, or $875,000, but that is not the 95th percentile of the stated mixture. The correct method averages the model cumulative distributions first and then finds the quantile.

Communicating Results Without Conflating Them

A useful uncertainty statement identifies the modeled outcome, the parameters treated as uncertain, the alternative structures considered, and the method used to combine or compare them. It should also distinguish an estimate conditional on one selected model from a predictive result that incorporates stated uncertainty. When weights or distributions reflect judgment, say so; do not present them as if they were directly observed facts.

The distinction matters for decision-making. Outcome variation describes losses that can occur under the assumed model. Parameter uncertainty can indicate that more or better data might narrow an estimate. Structural uncertainty can indicate that a result depends on assumptions about how the loss process behaves. A single capital number may be required for a particular use, but its interpretation should not conceal these sources of uncertainty.

Common reporting errors include calling the modeled parameter “random” without explaining that its distribution represents uncertainty, presenting a model-weighted result without disclosing the weights, and adding a generic uncertainty margin to a capital measure without showing which uncertainty it addresses. A disciplined report names the conditional model result, the uncertainty representation, and the resulting predictive or comparative result separately. This treatment is aligned with the CAS Exam 9 (Risk Management for Actuaries) syllabus.

Key takeaway: Keep three questions separate: how can the loss outcome vary under a specified model; how uncertain are the model's parameter values; and how much do results depend on the model structure? State the representation used for each before interpreting a capital result.

Check Your Understanding

For each problem, identify the uncertainty representation before calculating. Monetary answers are in dollars unless otherwise stated.

A hypothetical portfolio has $N\mid\lambda\sim\operatorname{Poisson}(2.5)$ and fixed claim severity of $40,000. Calculate the mean and variance of annual aggregate loss. No interest or mortality assumption applies. (Answer: mean $100,000.00; variance $4,000,000,000.00 dollars squared.)
Use a Gamma prior with shape 3 and rate 2 for a hypothetical Poisson claim rate. The experience is 4 claims over 2 exposure-years. Find the posterior distribution and the next-year predictive count mean and variance. (Answer: posterior $\operatorname{Gamma}(7,4)$; predictive mean 1.75 and variance 2.1875.)
A hypothetical count model is Poisson conditional on $\lambda$, with equal uncertainty weights on $\lambda=0.5$ and $\lambda=2.5$. Each event costs $80,000. Find the predictive count mean and variance, then the aggregate loss mean and variance. (Answer: count mean 1.5 and variance 2.5; loss mean $120,000.00 and variance $16,000,000,000.00 dollars squared.)
Two equally weighted structural models describe a hypothetical event count. Model A is Poisson with mean 0.8; Model B is geometric with mean 0.8 and support $0,1,\ldots$. Each event costs $300,000. Find the model-averaged 90th-percentile loss. (Answer: 2 events, or $600,000.00.)
A model has a fixed Poisson structure and assigns equal uncertainty weights to rates 1 and 3. Explain how the predictive count variance compares with the variance obtained by substituting the mean rate into a fixed Poisson model. (Answer: predictive variance is 3; plug-in variance is 2, because the predictive result includes 1 unit squared of between-rate variance.)

What You'll Learn