Avoiding Double Counting Between Priors And Credibility Weights

Why the Same Benchmark Can Be Counted Twice

A reserve estimate may use both recent claim experience and an external expected-loss benchmark. A credibility blend is appropriate when those are the two inputs: it gives more weight to the experience when that experience is sufficiently informative, and more weight to the benchmark when it is not. The difficulty arises when an estimate already contains the benchmark and is then blended with that same benchmark again. The calculation remains an average, but its effective weights may no longer express the intended judgment about how much evidence the experience provides.

The key diagnostic is to trace the ingredients of each number before applying a credibility formula. A figure labelled “experience estimate” may be a raw estimate based on observed claims, or it may already be adjusted toward an external expectation. Those are not interchangeable inputs.

Definition: Let $U$ be an experience-based estimate, $B$ an external expected-loss benchmark, and $Z$, with $0\leq Z\leq 1$, the credibility assigned to $U$. A direct credibility blend is $Q=ZU+(1-Z)B$. It uses $U$ and $B$ as its two components. If $Q$ already contains $B$, blending $Q$ toward the same $B$ again changes the effective weights and must be justified as a new, intentional adjustment.

The benchmark $B$ might be an expected reserve from an external model, a portfolio-level expectation, or an expected loss estimate used as a prior center. “External” describes its role relative to the experience being assessed; it does not mean the information is necessarily independent of every other input. Before blending, identify the data and assumptions behind each component and check whether any information is shared.

What a Credibility Weight Means

In a direct blend, $Z$ is the weight on the experience estimate, not a generic measure of how credible the final number is. The complementary weight $1-Z$ is assigned to the benchmark. The formula has this interpretation only when the two stated inputs are the components being blended.

$$ Q=ZU+(1-Z)B $$

A useful special case is Bayesian credibility under a normal prior and normal likelihood. Suppose the unknown quantity $\theta$ has prior distribution $\theta\sim N(B,\tau^2)$. Suppose also that the sample estimate $U$, conditional on $\theta$, has distribution $U\mid\theta\sim N(\theta,\sigma^2/n)$, with known $\tau^2$, $\sigma^2$, and $n$. Then the posterior mean is a credibility blend of $U$ and the prior mean $B$. The conditional normal likelihood is essential: prior and sampling means and variances alone do not establish this posterior result.

Theorem: Under the normal prior and conditional normal sampling assumptions just stated, the posterior mean is $ZU+(1-Z)B$, where $Z=\tau^2/(\tau^2+\sigma^2/n)$. Here $Z$ is determined by the prior variance and the sampling variance of $U$; it is not an additional weight to apply again to the posterior mean using the same prior mean.

Why the Normal Credibility Formula Holds

The posterior density is proportional to the product of the prior density and the likelihood. Keeping the terms involving $\theta$, its exponent is minus one-half times $\theta^2/\tau^2+(U-\theta)^2/(\sigma^2/n)$. Completing the square gives a normal posterior. Its precision (the reciprocal of its variance) is the sum of the prior precision and the likelihood precision.

\begin{aligned} \operatorname{Var}(\theta\mid U) &=\left(\frac{1}{\tau^2}+\frac{n}{\sigma^2}\right)^{-1},\\ E[\theta\mid U] &=\frac{B/\tau^2+Un/\sigma^2}{1/\tau^2+n/\sigma^2}\\ &=\frac{\sigma^2/n}{\tau^2+\sigma^2/n}B +\frac{\tau^2}{\tau^2+\sigma^2/n}U\\ &=(1-Z)B+ZU. \end{aligned}

The final line shows why the blend has the stated weights: the data estimate receives weight proportional to its precision, while the prior mean receives weight proportional to prior precision. Applying another credibility blend toward $B$ would introduce an extra operation, not a second step implied by the posterior formula.

Worked Example: A Normal Prior for an Unpaid-Claim Estimate

At a valuation date, an actuary models an unknown expected unpaid-loss amount $\theta$. The external prior is $\theta\sim N(\$1{,}000{,}000,(\$200{,}000)^2)$. A hypothetical portfolio provides $n=9$ conditionally independent observations, each normally distributed with mean $\theta$ and standard deviation $\$300{,}000$. Their average is $U=\$1{,}050{,}000$. All amounts are at the same valuation date; no discounting or interest adjustment is being made. Find the posterior credibility weight and posterior mean.

Method and assumptions. Use the normal-prior, normal-likelihood formula. The prior benchmark is $B=\$1{,}000{,}000$, the prior variance is $\tau^2=(\$200{,}000)^2=40{,}000{,}000{,}000$, and the variance of the sample average is $\sigma^2/n=(\$300{,}000)^2/9=10{,}000{,}000{,}000$.

Calculate the weight. Thus $Z=40{,}000{,}000{,}000/(40{,}000{,}000{,}000+10{,}000{,}000{,}000)=0.80$.

Calculate the posterior mean. The prior receives weight $1-Z=0.20$, so $E[\theta\mid U]=0.80(\$1{,}050{,}000)+0.20(\$1{,}000{,}000)=\$840{,}000+\$200{,}000=\$1{,}040{,}000$.

Check. The posterior mean is between the prior mean of $\$1{,}000{,}000$ and the observed average of $\$1{,}050{,}000$, and it is closer to the observation because $Z=0.80$. The common mistake this avoids is calculating the variance of the sample average as though it were the variance of one observation.

Tracing Effective Weights Through a Repeated Blend

Suppose a preliminary estimate $Q$ has already been formed from an experience estimate $U$ and a benchmark $B$. If an analyst then blends $Q$ toward that same benchmark, substitute the expression for $Q$ into the second blend. This reveals the actual, or effective, weights.

\begin{aligned} Q&=Z_1U+(1-Z_1)B,\\ R&=Z_2Q+(1-Z_2)B\\ &=Z_2Z_1U+\{Z_2(1-Z_1)+(1-Z_2)\}B\\ &=(Z_1Z_2)U+(1-Z_1Z_2)B. \end{aligned}

The coefficient on experience is the product $Z_1Z_2$. Since both weights lie between zero and one, $Z_1Z_2\leq Z_1$: the second blend moves the result further toward the benchmark. The coefficients still sum to one, so the result is a valid weighted average. The issue is not invalid arithmetic; it is whether the extra movement toward $B$ was intended.

Worked Example: Find the Effective Weights Before Interpreting the Result

A hypothetical reserve estimate starts with raw experience $U=\$1{,}500{,}000$ and an external benchmark $B=\$1{,}100{,}000$. The first calculation assigns 70% credibility to experience, giving $Q=0.70U+0.30B$. A later step assigns 80% credibility to $Q$ and 20% to the same benchmark: $R=0.80Q+0.20B$. Find $Q$, $R$, and the effective weights in $R$.

First blend. $Q=0.70(\$1{,}500{,}000)+0.30(\$1{,}100{,}000)=\$1{,}050{,}000+\$330{,}000=\$1{,}380{,}000$.

Second blend. $R=0.80(\$1{,}380{,}000)+0.20(\$1{,}100{,}000)=\$1{,}104{,}000+\$220{,}000=\$1{,}324{,}000$.

Effective weights. The weight on raw experience is $0.80(0.70)=0.56$. The weight on the benchmark is $0.80(0.30)+0.20=0.44$. Checking directly, $0.56(\$1{,}500{,}000)+0.44(\$1{,}100{,}000)=\$840{,}000+\$484{,}000=\$1{,}324{,}000$, as required.

Interpretation and trap. The two operations have reduced experience’s effective weight from 70% to 56%. A common error is to preserve the first 70% as the final experience weight, or to miscalculate the component amounts. If the intended final weight on raw experience was 70%, the second blend should not be applied to $Q$ with the same benchmark.

Correcting the Blend Without Losing the Intended Credibility

When a preliminary estimate already contains the benchmark, first decide what the requested weight is supposed to apply to. If the intended final estimate gives raw experience weight $Z$ and benchmark weight $1-Z$, calculate that blend directly from raw $U$ and $B$. Do not treat a benchmark-adjusted estimate as if it were raw experience.

An alternative is to retain the nested calculation but calibrate the second weight to achieve the desired effective weight. If the first estimate has experience weight $Z_1>0$, and the target final experience weight is $Z^*$, then the second-stage weight must satisfy $Z_1Z_2=Z^*$. Thus $Z_2=Z^*/Z_1$. This is possible as a credibility weight only if $0\leq Z^*/Z_1\leq1$. Usually the clearest approach is to show the components and apply the target weight once.

Label the raw inputs.
Identify the experience estimate, external benchmark, and any other information used to construct each estimate.

Expand the existing estimate.
Write it as a weighted combination of its underlying components where possible, rather than relying only on its label.

Combine coefficients.
Multiply weights along each path and add the contributions to a shared input, such as the same external benchmark.

Compare with the intended weights.
If the effective weight on experience differs from the target, return to the raw inputs or recalibrate the second-stage weight.

Check the arithmetic and interpretation.
Confirm the weights sum to one and that the result lies between the component estimates when all weights are nonnegative.

Worked Example: Correct a Benchmark Already Included in a Reserve Estimate

A hypothetical actuary has a raw development-based reserve estimate $U=\$900{,}000$ and an external expected reserve $B=\$600{,}000$. A preliminary calculation already blends them equally, $Q=0.50U+0.50B$. A review requires the final estimate to assign 70% to raw experience and 30% to the benchmark. Determine the preliminary estimate, the incorrectly repeated blend if someone applies 70% to $Q$, and the corrected final estimate. All amounts are at the same valuation date, with no discounting.

Preliminary estimate. $Q=0.50(\$900{,}000)+0.50(\$600{,}000)=\$450{,}000+\$300{,}000=\$750{,}000$.

Repeated blend. Applying 70% to $Q$ and 30% to $B$ gives $R=0.70(\$750{,}000)+0.30(\$600{,}000)=\$525{,}000+\$180{,}000=\$705{,}000$. Its effective experience weight is $0.70(0.50)=0.35$, and its benchmark weight is $0.70(0.50)+0.30=0.65$. Direct verification gives $0.35(\$900{,}000)+0.65(\$600{,}000)=\$315{,}000+\$390{,}000=\$705{,}000$.

Corrected estimate. The required final weights apply to raw experience and the benchmark, so the correct calculation is $0.70(\$900{,}000)+0.30(\$600{,}000)=\$630{,}000+\$180{,}000=\$810{,}000$.

Check and method. The corrected estimate is between $\$600{,}000$ and $\$900{,}000$, and its experience weight is exactly 70%. The repeated blend also lies between the two inputs, but it has only 35% effective weight on experience. The mistake is treating $Q$, which already contains the benchmark, as raw experience.

Common Traps and Practical Checks

Taking an estimate’s label at face value. “Experience estimate” does not guarantee that the estimate is unadjusted. Trace whether a prior, benchmark, or expected-loss complement has already entered it.
Applying the same benchmark twice without recalibration. A second blend changes the effective experience weight to the product of the successive weights.
Confusing a Bayesian update with an extra credibility step. Under the stated normal model, the posterior mean already combines the likelihood information with the prior. A second blend toward the same prior mean is a separate modeling choice.
Using only moments to claim a normal posterior. The normal-prior credibility result requires a normal likelihood (or an equivalent specified normal sampling model), not merely a conditional mean and variance.
Forgetting to verify dollar arithmetic. Calculate each weighted component separately, add them, and then check the same result using the effective weights.

Repeated use of an external benchmark is not automatically prohibited. It can represent a deliberate additional decision, such as a governance adjustment or an explicitly calibrated shrinkage step. In that case, state the purpose of the second step and report the resulting effective weights. The problem is presenting the final estimate as though it retained the original weight on raw experience when algebra shows otherwise.

Key takeaway: Before using a credibility weight, identify exactly what estimate it weights. Expand any estimate that already incorporates external information, combine the effective coefficients, and compare them with the intended weights. This makes the use of a prior or benchmark transparent and prevents accidental over-weighting.

Check Your Understanding

For each problem, identify the components before calculating; assume all stated amounts are at the same valuation date and no discounting is required.

A reserve estimate uses raw experience $U=\$820{,}000$ and benchmark $B=\$700{,}000$, with $Q=0.60U+0.40B$. Find $Q$ and state its effective weights on the two raw inputs. (Answer: $Q=\$772{,}000$; weights 0.60 on $U$ and 0.40 on $B$)
An estimate is $Q=0.80U+0.20B$, then a second step is $R=0.75Q+0.25B$. Find the effective weights on $U$ and $B$. (Answer: 0.60 on $U$ and 0.40 on $B$)
A hypothetical normal model has prior $\theta\sim N(\$500{,}000,(\$100{,}000)^2)$. The sample average is $\$560{,}000$ from four conditionally independent normal observations, each with standard deviation $\$100{,}000$. Find the sample-average variance, $Z$, and posterior mean. (Answer: sample-average variance $\$2{,}500{,}000{,}000$; $Z=0.80$; posterior mean $\$548{,}000$)
A preliminary estimate is $Q=0.40U+0.60B$, but the required final weights are 70% on raw experience and 30% on the benchmark. If $U=\$1{,}000{,}000$ and $B=\$800{,}000$, calculate the correct final estimate and the second-stage weight on $Q$ that would produce the required effective experience weight. (Answer: correct estimate $\$940{,}000$; required second-stage weight $Z_2=1.75$, so it cannot be achieved by a credibility weight between 0 and 1 applied to $Q$; recompute from raw inputs)
A raw estimate is $U=\$1{,}400{,}000$, a benchmark is $B=\$1{,}000{,}000$, and the first blend is $Q=0.75U+0.25B$. An analyst then calculates $R=0.80Q+0.20B$. Find $Q$, $R$, and explain why $R$ does not have 80% weight on raw experience. (Answer: $Q=\$1{,}300{,}000$; $R=\$1{,}240{,}000$; effective weights are 0.60 on $U$ and 0.40 on $B$)

What You'll Learn