Propagating Credibility Parameter Uncertainty Into Reserves

Why Parameter Uncertainty Changes a Reserve

A credibility-weighted ultimate estimate depends on more than the observed experience. It also depends on the hypothetical mean, the variance of hypothetical means, the process variance of the experience, and the credibility weight determined from those variances. If the variance assumptions or the weight are themselves uncertain, treating them as fixed can understate uncertainty in the ultimate estimate and its unpaid reserve.

The key is to keep two sources of variation distinct. Process variation is the random variation in experience conditional on a set of model parameters. Parameter uncertainty is variation across plausible values of those parameters, including the variance of hypothetical means and the process variance. A credibility weight can change across parameter scenarios, so it must be recalculated within each scenario before the results are combined.

Definition: Let $M$ be a hypothetical mean, $a$ the variance of hypothetical means, and $p$ the process variance of an experience estimate $X$. Under the Bühlmann credibility form, the credibility weight is $Z=a/(a+p)$, and the credibility-weighted estimate is $\hat U=M+Z(X-M)$. If paid loss $P$ is known, the corresponding unpaid reserve estimate is $\hat R=\hat U-P$. When $M$, $a$, or $p$ are uncertain, calculate these quantities for each parameter scenario rather than assuming one fixed value.

In applications, $p$ may already incorporate the volume of experience. For example, if individual observations have process variance $\sigma^2$ and $X$ is their mean over $n$ independent observations, then $p=\sigma^2/n$. Use the process variance of the actual experience statistic in the credibility formula; do not use $\sigma^2$ in place of $p$ unless $X$ is a single observation.

Decomposing the Uncertainty

Suppose parameter uncertainty is represented by a scenario variable $\Theta$. In a given scenario, let the conditional mean of $X$ be $\mu_\Theta$, and let its conditional process variance be $p_\Theta$. The credibility weight and conditional mean of the estimate are then

Z_\Theta=\frac{a_\Theta}{a_\Theta+p_\Theta}, \qquad h_\Theta=M_\Theta+Z_\Theta(\mu_\Theta-M_\Theta).

Because $\hat U=M_\Theta+Z_\Theta(X-M_\Theta)$, its conditional process variance in that scenario is $Z_\Theta^2p_\Theta$. The factor is $Z_\Theta^2$, not $Z_\Theta$, because variance scales by the square of a multiplier.

Theorem: For a random estimate $\hat U$ with conditional mean $h_\Theta=E[\hat U\mid\Theta]$ and conditional variance $q_\Theta=\operatorname{Var}(\hat U\mid\Theta)$, $$ \operatorname{Var}(\hat U) =E[q_\Theta]+\operatorname{Var}(h_\Theta). $$ For credibility-weighted experience with conditional mean $\mu_\Theta$ and process variance $p_\Theta$, $q_\Theta=Z_\Theta^2p_\Theta$. Thus total variance includes both average process variance and variation in the scenario-specific conditional means.

This decomposition also clarifies how to handle the variance of hypothetical means. If the scenario values of $\mu_\Theta$ represent variation in the underlying hypothetical means, their effect appears in $\operatorname{Var}(h_\Theta)$. Do not add the same variation again as a separate variance term. The scenario design must reflect the intended uncertainty model.

Why the Variance Decomposition Holds

The conditional-mean component measures how far each scenario’s expected estimate lies from the overall expected estimate. Within each scenario, the estimate also varies around its own conditional mean. Write $\hat U-E[\hat U]=(\hat U-h_\Theta)+(h_\Theta-E[\hat U])$, square, and take expectations. The cross-product has expectation zero because $E[\hat U-h_\Theta\mid\Theta]=0$.

\begin{aligned} \operatorname{Var}(\hat U) &=E[(\hat U-E[\hat U])^2]\\ &=E[(\hat U-h_\Theta)^2] +E[(h_\Theta-E[\hat U])^2]\\ &\quad+2E[(\hat U-h_\Theta)(h_\Theta-E[\hat U])]\\ &=E[\operatorname{Var}(\hat U\mid\Theta)] +\operatorname{Var}(E[\hat U\mid\Theta]). \end{aligned}

The last equality follows because $h_\Theta-E[\hat U]$ is fixed conditional on $\Theta$, while $E[\hat U-h_\Theta\mid\Theta]=0$. This is the law of total variance. It applies whether the parameter scenarios are a probability model or a weighted set of scenarios used to represent parameter uncertainty.

Worked Calculations

Worked Example: Combining Process and Scenario Uncertainty

A hypothetical portfolio has fixed hypothetical mean $M=\$1{,}000$. Its experience statistic $X$ is based on $n=4$ observations. Model uncertainty is represented by three scenarios:

Scenario probability	$a$	$p=\sigma^2/n$	$\mu_\Theta=E[X\mid\Theta]$
0.25	100	100	\$900
0.25	300	100	\$1,100
0.50	200	200	\$1,000

Calculate the mean and variance of the credibility-weighted ultimate estimate, then calculate the mean and variance of the unpaid reserve if known paid loss is $P=\$700$. Assume the conditional process error in $X$ has variance $p$ in each scenario.

Method and weights. In each scenario, calculate $Z=a/(a+p)$, the conditional mean $h=M+Z(\mu_\Theta-M)$, and conditional process variance $q=Z^2p$.

Scenario 1: $Z=100/(100+100)=0.5$, $h=1{,}000+0.5(900-1{,}000)=\$950$, and $q=0.5^2(100)=25$.
Scenario 2: $Z=300/(300+100)=0.75$, $h=1{,}000+0.75(1{,}100-1{,}000)=\$1{,}075$, and $q=0.75^2(100)=56.25$.
Scenario 3: $Z=200/(200+200)=0.5$, $h=1{,}000+0.5(1{,}000-1{,}000)=\$1{,}000$, and $q=0.5^2(200)=50$.

Overall mean. Weight the conditional means by the scenario probabilities: $E[\hat U]=0.25(950)+0.25(1{,}075)+0.50(1{,}000)=\$1{,}006.25$.

Between-scenario variance. The deviations from the mean are $-56.25$, $68.75$, and $-6.25$. Therefore, $0.25(56.25^2)+0.25(68.75^2)+0.50(6.25^2)=1{,}992.1875$.

Average process variance. This is $0.25(25)+0.25(56.25)+0.50(50)=45.3125$. The total variance is $1{,}992.1875+45.3125=2{,}037.5$ dollars squared.

The reserve mean is $\$1{,}006.25-\$700=\$306.25$. Because paid loss is fixed, subtracting it does not change variance, so reserve variance is also $2{,}037.5$ dollars squared. Check: the total variance exceeds the average conditional process variance because the scenario means differ. The method avoids omitting parameter-scenario variation.

Worked Example: Uncertain Parameters With Observed Experience

A hypothetical insurer has observed experience $X=\$1{,}500$, fixed hypothetical mean $M=\$1{,}200$, and known paid loss $P=\$900$. There are two equally plausible parameter scenarios: scenario A has $a=100$ and $p=100$; scenario B has $a=400$ and $p=200$. Find the probability-weighted mean reserve estimate and the variance arising from parameter uncertainty.

Method. Since $X$ is already observed, calculate a reserve estimate in each parameter scenario. Do not add process variance to the dispersion of these scenario-specific point estimates: the observed experience is being held fixed in this calculation.

In scenario A, $Z=100/(100+100)=0.5$, so $\hat U=1{,}200+0.5(1{,}500-1{,}200)=\$1{,}350$, and $\hat R=1{,}350-900=\$450$.

In scenario B, $Z=400/(400+200)=2/3$, so $\hat U=1{,}200+(2/3)(300)=\$1{,}400$, and $\hat R=1{,}400-900=\$500$.

The mean reserve is $0.5(450)+0.5(500)=\$475$. Its parameter-scenario variance is $0.5(450-475)^2+0.5(500-475)^2=625$ dollars squared.

Common mistake. Averaging the parameters first gives $a=250$ and $p=150$, hence $Z=250/400=0.625$. This produces $\hat R=1{,}200+0.625(300)-900=\$487.50$, not the scenario-weighted mean of $\$475$. The weight is a nonlinear function of $a$ and $p$, so in general the weight calculated from averaged parameters is not the average scenario weight.

Worked Example: Two Reserve Components With a Shared Weight

A hypothetical book has two ultimate-loss components. Their base hypothetical means are $M_1=\$1{,}000$ and $M_2=\$600$, while their experience values are $X_1=\$1{,}300$ and $X_2=\$800$. A shared uncertain credibility weight is 0.4 or 0.7, each with probability 0.5. Conditional on the scenario, the two experience errors are independent, with process variances 100 and 64. Known paid loss across the book is $\$1{,}100$. Find the mean and variance of the aggregate unpaid reserve estimate.

Scenario conditional means. At $Z=0.4$, the component ultimate estimates are $1{,}000+0.4(300)=\$1{,}120$ and $600+0.4(200)=\$680$, for total $\$1{,}800$. At $Z=0.7$, they are $\$1{,}210$ and $\$740$, for total $\$1{,}950$. The aggregate ultimate mean is $0.5(1{,}800)+0.5(1{,}950)=\$1{,}875$, so the mean reserve is $\$1{,}875-\$1{,}100=\$775$.

The between-scenario variance of aggregate ultimate means is $0.5(1{,}800-1{,}875)^2+0.5(1{,}950-1{,}875)^2=5{,}625$ dollars squared.

Conditional process variance. Independence makes the conditional variance of the aggregate equal to the sum of component variances. In the first scenario it is $0.4^2(100+64)=26.24$; in the second it is $0.7^2(100+64)=80.36$. The average is $0.5(26.24)+0.5(80.36)=53.3$.

Thus aggregate reserve variance is $5{,}625+53.3=5{,}678.3$ dollars squared. Fixed paid loss changes the reserve mean but not its variance. Check: the shared weight moves both component means together; the between-scenario variance of their sum includes their positive covariance. Adding only their separate scenario variances would miss that dependence.

Practical Checks and Limits

Before reporting a reserve uncertainty measure, identify what is random in the calculation. If the observed experience is fixed, scenario dispersion in the credibility estimates represents parameter uncertainty, not new process variation in that observed experience. If the experience statistic is modeled as random within each scenario, its conditional process variance can be included through $E[q_\Theta]$. Do not count the same source twice.

Use the correct process variance. If credibility is applied to an average of independent observations, use the variance of that average, not the variance of one observation.
Recalculate the weight by scenario. Calculate $Z_\Theta$ from the scenario’s $a_\Theta$ and $p_\Theta$ before estimating the ultimate.
Keep the units consistent. If losses are in thousands, variances are in thousands squared. Convert before combining amounts and variances expressed on different scales.
Preserve dependence. When components share uncertain parameters or weights, calculate the variance of their total under the joint scenarios; do not assume their parameter-driven variations are independent.
Check the reserve transformation. Subtracting a fixed paid amount changes the mean by that amount and leaves variance unchanged. If paid loss is also uncertain, its variance and covariance with ultimate loss must be handled separately.

This approach propagates specified uncertainty through a credibility-weighted ultimate and its unpaid reserve. It does not by itself specify a full predictive distribution for future reserve outcomes; the result depends on the parameter scenarios and conditional process model supplied.

Check Your Understanding

For each problem, keep scenario probabilities, process variation, and fixed paid amounts distinct.

A hypothetical portfolio has $M=\$800$, observed $X=\$1{,}000$, $a=120$, $p=80$, and paid loss $P=\$500$. Calculate $Z$, the ultimate estimate, and the reserve estimate. (Answer: $Z=0.6$, ultimate $\$920$, reserve $\$420$)
With fixed $M=\$900$, $X=\$1{,}200$, and $P=\$700$, a parameter scenario gives $Z=0.4$ with probability 0.5 and $Z=0.8$ with probability 0.5. Find the mean reserve and its parameter-scenario variance. (Answer: mean $\$380$; variance $3{,}600$ dollars squared)
Let $M=\$1{,}000$, $X=\$1{,}300$, and $P=\$800$. Scenario A has probability 0.25, $a=100$, $p=100$; scenario B has probability 0.75, $a=300$, $p=100$. Find the mean reserve and parameter-scenario variance. (Answer: mean $\$406.25$; variance $1{,}054.6875$ dollars squared)
An estimate has two equally likely scenarios. Their conditional means are $\$1{,}100$ and $\$900$, and their conditional process variances are 16 and 56.25. Calculate the overall mean and variance. (Answer: mean $\$1{,}000$; variance $10{,}036.125$ dollars squared)
A two-component book has aggregate conditional mean $\$2{,}000+400Z$, where shared $Z$ is 0.3 or 0.6 with equal probability. Conditional process variances are 100 and 144, with independent errors. Paid loss is fixed at $\$1{,}300$. Calculate the mean reserve and total variance. (Answer: mean $\$880$; variance $3{,}654.9$ dollars squared)

What You'll Learn