Stabilizing Sparse Triangle Estimates With External Experience

Why Sparse Accident-Period Estimates Need Stabilizing

An accident period with little reported or paid experience can produce a volatile ultimate-loss estimate. A projection based mainly on that period's own data may be far above or below a reasonable benchmark simply because the internal information is sparse. External development experience or an expected-loss estimate can provide a useful reference point, provided it is relevant to the same coverage, time period, and measure of loss.

Credibility formalizes the balance between that reference point and the internal estimate. The central task is not to choose a blend by intuition; it is to specify how uncertain each source is. Greater internal uncertainty gives more weight to the external benchmark. As internal information accumulates and its estimate becomes more precise, the internal weight increases.

Definition: Let $U$ be an accident period's ultimate loss, $M$ an external benchmark for that ultimate loss, and $\widehat{U}$ an internal estimate of the same ultimate loss. A credibility estimate is a weighted combination $Z\widehat{U}+(1-Z)M$, where $0\leq Z\leq1$. The benchmark and internal estimate must be expressed on a consistent basis before they are combined.

The benchmark $M$ can come from external expected-loss information, such as exposure multiplied by an expected loss per exposure unit, or from external development information, such as a factor that projects reported or paid losses to ultimate. These are ways to construct the benchmark; neither makes it automatically appropriate. It must reflect the accident period's coverage, limits, trend level, and other material characteristics.

A Variance-Based Credibility Model

A standard justification for the blend uses a normal prior and a normally distributed internal estimate. Assume that the unknown ultimate loss has prior distribution $U\sim N(M,\tau^2)$. Conditional on $U$, assume the internal estimate is unbiased and satisfies $\widehat{U}\mid U\sim N(U,v)$. Here $\tau^2$ represents uncertainty about the ultimate loss around the external benchmark, while $v$ represents the sampling or estimation variance of the internal estimate. The model treats $M$, $\tau^2$, and $v$ as specified, and assumes the internal estimate's error is conditionally independent of the prior information.

Under these assumptions, the conditional mean of $U$ given $\widehat{U}$ is the credibility estimate. Its weight on the internal estimate is determined by the relative precisions, where precision is the reciprocal of variance.

Write the prior and estimate distributions.
The prior density contributes a factor proportional to $\exp[-(U-M)^2/(2\tau^2)]$, and the internal estimate likelihood contributes a factor proportional to $\exp[-(\widehat{U}-U)^2/(2v)]$.

Combine the quadratic terms in $U$.
The coefficient of $U^2$ in the negative log-density is $\frac{1}{2}(1/\tau^2+1/v)$. Thus the conditional variance is $(1/\tau^2+1/v)^{-1}=\tau^2v/(\tau^2+v)$.

Find the conditional mean.
The coefficient-weighted mean is $\frac{M/\tau^2+\widehat{U}/v}{1/\tau^2+1/v}$. Multiplying numerator and denominator by $\tau^2v$ gives $\frac{v}{\tau^2+v}M+\frac{\tau^2}{\tau^2+v}\widehat{U}$.

Z=\frac{\tau^2}{\tau^2+v}, \qquad E[U\mid\widehat{U}]=Z\widehat{U}+(1-Z)M, \qquad \operatorname{Var}(U\mid\widehat{U})=\frac{\tau^2v}{\tau^2+v}.

The variance comparison explains the direction of the weights. If $v$ is large relative to $\tau^2$, the internal estimate is noisy and $Z$ is small. If $v$ decreases, $Z$ increases. The formula is appropriate only under the stated model and when the variances reasonably represent the two uncertainty sources. It is not a substitute for checking whether the external benchmark is relevant or whether the internal estimate is biased.

Worked Example: Blending a Sparse Estimate With Expected-Loss Information

A hypothetical accident period has an external expected-ultimate benchmark of $900,000. A sparse internal projection is $1,080,000. In thousands of dollars, assume $U\sim N(900,200^2)$ and $\widehat{U}\mid U\sim N(U,300^2)$. Find the credibility estimate and posterior standard deviation. Use the normal-model method, and avoid treating the external benchmark as certain.

Step 1: Identify the variances. In thousands of dollars, the prior variance is $\tau^2=200^2=40{,}000$, and the internal estimate variance is $v=300^2=90{,}000$.

Step 2: Calculate the internal credibility weight.

Z=\frac{40{,}000}{40{,}000+90{,}000} =\frac{40}{130} =\frac{4}{13} \approx 0.3077.

Step 3: Blend the estimates. The external weight is $1-Z=9/13$. Therefore,

E[U\mid\widehat{U}] =\frac{4}{13}(1{,}080)+\frac{9}{13}(900) =\frac{4{,}320+8{,}100}{13} =\frac{12{,}420}{13} =955.3846

The credibility estimate is $955,384.62, or $955,385 to the nearest dollar. The posterior variance is $40{,}000(90{,}000)/130{,}000=27{,}692.31$ in squared thousands of dollars, so the posterior standard deviation is $\sqrt{27{,}692.31}=166.4$ thousand dollars. As a check, the result lies between $900,000 and $1,080,000 and is closer to the benchmark, consistent with the internal weight being below one-half. The method avoids the mistake of giving the sparse internal projection full weight merely because it is an accident-period-specific estimate.

External Development Information as a Benchmark

An external development factor can be used to construct an external ultimate estimate before credibility is applied. For example, if cumulative paid losses are $360,000 and external experience suggests that paid losses at the observed maturity represent 40% of ultimate, the external development estimate is $360{,}000/0.40=$900{,}000$. An internal projection can be calculated separately using internal development information. If both figures estimate ultimate losses on the same basis, the variance-based credibility model can combine them.

The fact that both estimates begin with the same observed paid amount does not by itself make the blend invalid. The relevant question is whether the estimates and variances are defined consistently, and whether the uncertainty model accounts for dependence where necessary. If one source already incorporates the other source's information, treating them as independent evidence can overstate the amount of information available. The calculation below assumes the stated prior and internal-estimate distributions are valid for the intended blend.

Worked Example: Stabilizing an Internal Development Projection

A hypothetical portfolio has $360,000 of cumulative paid loss at a given maturity. External development experience indicates that this amount is 40% of ultimate. An internal development projection is $1,080,000. In thousands of dollars, model the external ultimate benchmark as $U\sim N(900,200^2)$ and the internal projection as $\widehat{U}\mid U\sim N(U,400^2)$. Calculate the credibility estimate. Use the variance-based method, and avoid blending a reported proportion directly with a dollar ultimate.

Step 1: Convert the external development information to an ultimate benchmark.

\text{External ultimate} =\frac{360}{0.40} =900 \quad\text{thousand dollars}.

The internal ultimate projection is 1,080 thousand dollars. The prior variance is $200^2=40{,}000$, and the internal estimate variance is $400^2=160{,}000$, both in squared thousands of dollars.

Step 2: Calculate credibility and the blended estimate.

Z=\frac{40{,}000}{40{,}000+160{,}000}=0.20, \qquad \widehat{U}_{\text{cred}} =0.20(1{,}080)+0.80(900) =216+720 =936.

The credibility estimate is $936,000. It is between the $900,000 external estimate and the $1,080,000 internal projection, and it is closer to the external estimate because the internal estimate has the larger variance. The method avoids the dimensional error of averaging the 40% reported proportion with an ultimate-loss dollar amount.

How Credibility Changes as Internal Experience Grows

A useful accumulation model makes the relationship between internal volume and credibility explicit. Suppose the internal data consist of $n$ independent, unbiased signals of the same ultimate, each with conditional variance $\sigma^2$. Their average $\bar{X}$ has conditional variance $v=\sigma^2/n$. Substitution into the normal-model credibility formula gives:

Z_n =\frac{\tau^2}{\tau^2+\sigma^2/n} =\frac{n\tau^2}{n\tau^2+\sigma^2}.

For $\tau^2>0$ and $\sigma^2>0$, this weight increases with $n$, since

\frac{dZ_n}{dn} =\frac{\tau^2\sigma^2}{(n\tau^2+\sigma^2)^2}>0.

As $n$ becomes large, $Z_n$ approaches 1; when $n$ is small, the external benchmark has greater influence. In a reserving application, $n$ should represent information consistent with the stated variance model, not automatically the number of claims, policies, or exposure units. Different data structures can carry different amounts of information. This model also assumes independent signals with common variance; it does not apply unchanged when signals are materially dependent or their variances differ.

Worked Example: Increasing Credibility With More Internal Signals

A hypothetical accident-period estimate is based on $n$ independent internal signals. In thousands of dollars, each signal has conditional standard deviation $\sigma=400$, and the external benchmark has prior mean $M=1{,}000$ and standard deviation $\tau=200$. For $n=1,4,$ and $16$, suppose the observed signal average is 1,100. Calculate the credibility weight and estimate at each volume. Use the independent-signal model, and avoid holding the credibility weight fixed as the information count changes.

The prior variance is $\tau^2=40{,}000$, and each signal has variance $\sigma^2=160{,}000$. For each $n$, the average has variance $160{,}000/n$.

For $n=1$, $v=160{,}000$, so $Z_1=40{,}000/(40{,}000+160{,}000)=0.20$. The estimate is $0.20(1{,}100)+0.80(1{,}000)=1{,}020$.

For $n=4$, $v=40{,}000$, so $Z_4=40{,}000/(40{,}000+40{,}000)=0.50$. The estimate is $0.50(1{,}100)+0.50(1{,}000)=1{,}050$.

For $n=16$, $v=10{,}000$, so $Z_{16}=40{,}000/(40{,}000+10{,}000)=0.80$. The estimate is $0.80(1{,}100)+0.20(1{,}000)=1{,}080$.

The estimates are $1,020,000, $1,050,000, and $1,080,000, respectively. Each lies between the $1,000,000 benchmark and the $1,100,000 internal average, and the estimate moves toward the internal average as $n$ grows. The method avoids the mistake of interpreting a larger internal data volume as more credibility without recalculating the internal variance.

Exam Traps and Practical Checks

The most important check is comparability. A benchmark for ultimate losses should not be blended directly with paid-to-date losses, a loss ratio on a different premium basis, or an estimate using a different coverage definition. Convert each source to a common measure first. If external development is used to obtain ultimate losses, make clear which observed amount and maturity the factor applies to.

A second trap is reversing the variance weights. In this model, $Z$ is $\tau^2/(\tau^2+v)$, not $v/(\tau^2+v)$. The internal weight increases when internal variance falls; if a proposed calculation does the opposite, the weights have likely been reversed. For a numerical check, the credibility estimate must lie between $M$ and $\widehat{U}$ whenever $0\leq Z\leq1$.

Finally, external information is not automatically independent of internal experience or free of bias. If the benchmark was developed using the same data that produced the internal projection, a simple independent-prior interpretation may overstate its contribution. Also, uncertainty in the benchmark itself is represented by $\tau^2$; treating an uncertain external estimate as a fixed truth can give it too much influence. The normal credibility formula is a model-based tool, not a guarantee that the inputs are appropriate.

Key takeaway: Stabilize a sparse accident-period estimate by comparing uncertainty on a common ultimate-loss basis. In the normal model, the internal weight is $\tau^2/(\tau^2+v)$: it rises as the internal estimate becomes more precise, while the external benchmark retains more influence when internal information is sparse.

Check Your Understanding

For each problem, work in the stated units and use the variance-based credibility model when its assumptions are given.

A hypothetical internal ultimate estimate is $820,000. Its external benchmark is $700,000, with prior standard deviation $100,000; the internal estimate has standard deviation $200,000. Find the credibility estimate. (Answer: $724,000)
An external expected-ultimate estimate is $1,200,000 with prior standard deviation $30,000. The internal estimate is $1,500,000 with standard deviation $60,000. Find the blended estimate. (Answer: $1,260,000)
An external benchmark is $800,000 with $\tau=$150,000$. The average of $n=9$ independent internal signals is $920,000, and each signal has conditional standard deviation $450,000. Find the credibility estimate. (Answer: $860,000)
Paid loss is $500,000. An external development factor is 2.0 and an internal development factor is 2.4. Treat the resulting estimates as an external benchmark with standard deviation $100,000 and an internal estimate with standard deviation $200,000. Find the blended ultimate estimate. (Answer: $1,040,000)
An internal estimate is $1,040,000, the benchmark is $1,000,000, and the prior variance and initial internal-estimate variance are each 25,000 in squared thousands of dollars. If additional independent information halves the internal-estimate variance, find the revised credibility estimate. (Answer: $1,026,666.67)

What You'll Learn