Pooling The Collective Mean Across Risks

Why Pooling Weights Matter

In Bühlmann-Straub credibility, an individual risk’s observed loss cost is blended with a collective complement. When risks have different total exposures, the complement depends on how their experience is pooled. Giving each risk equal weight answers a different question from pooling all exposure units together. That distinction can materially change a credibility indication.

This tutorial focuses on estimating the collective complement from multiple risks. It assumes the Bühlmann-Straub framework and its exposure-weighted risk-period means. The central calculation is straightforward, but its interpretation depends on the target: the usual pooled portfolio loss cost gives greater weight to risks with more exposure.

Definition: Suppose risk $i$ has observations $X_{ij}$ in periods $j$, with positive exposures $w_{ij}$. Its total exposure is $w_i=\sum_j w_{ij}$, and its exposure-weighted mean is $\bar X_i=\sum_j w_{ij}X_{ij}/w_i$. The exposure-weighted pooled collective mean is $$ \hat m_{\mathrm{pool}}=\frac{\sum_i w_i\bar X_i}{\sum_i w_i} =\frac{\sum_i\sum_j w_{ij}X_{ij}}{\sum_i\sum_j w_{ij}}. $$ It is the total observed loss divided by total exposure, when each $X_{ij}$ is a loss cost per exposure unit.

Model and Meaning of the Pooled Complement

In the Bühlmann-Straub model, let $\Theta_i$ be the underlying expected loss cost for risk $i$, and let $m=E[\Theta_i]$ be the collective mean. The standard assumptions are that risks are drawn from the same risk population, exposures are fixed or non-informative about $\Theta_i$, and, conditional on $\Theta_i$, period observations have mean $\Theta_i$, variance proportional to the inverse of exposure, and are independent across periods:

E[X_{ij}\mid\Theta_i]=\Theta_i,\qquad \operatorname{Var}(X_{ij}\mid\Theta_i)=\frac{\phi}{w_{ij}},\qquad E[\Theta_i]=m,\qquad \operatorname{Var}(\Theta_i)=\tau^2.

Here $\phi$ is the process-variance parameter and $\tau^2$ is the between-risk variance. The pooled mean is an estimate of the exposure-weighted portfolio complement: it treats each unit of exposure as one unit of portfolio experience. If risks have unequal total exposures, a risk with twice the exposure contributes twice as much to the pooled numerator and denominator.

That target is not the same as the arithmetic average of the risk means, $\sum_i\bar X_i/n$, which gives every risk equal weight regardless of exposure. Neither weighting rule is universally correct without specifying the target. For the usual exposure-pooled collective complement, use total exposure weights. If the intended target is instead the average underlying mean across risks, equal risk weights may be relevant; do not silently substitute one estimand for the other.

Why the Pooling Formula Holds

The formula follows by converting each risk-period loss cost back to its aggregate loss. For period $j$ of risk $i$, the observed aggregate loss is $w_{ij}X_{ij}$. Summing those amounts gives total loss across the portfolio. Dividing by total exposure returns loss cost per exposure unit:

\frac{\text{total observed loss}}{\text{total exposure}} = \frac{\sum_i\sum_j w_{ij}X_{ij}}{\sum_i\sum_j w_{ij}}.

Within each risk, $\sum_j w_{ij}X_{ij}=w_i\bar X_i$, so regrouping the numerator by risk produces $\sum_i w_i\bar X_i$. This proves the equivalent risk-mean formula. Under the stated assumptions, each period observation has conditional mean $\Theta_i$, so its aggregate contribution has expected value $w_{ij}\Theta_i$. For fixed exposures, the expected pooled numerator is $m\sum_i\sum_j w_{ij}$; dividing by total exposure gives $m$. Thus the pooled estimate is unbiased for the exposure-weighted collective mean under this model.

This justification establishes the target and the natural pooling calculation; it does not assert that the exposure-weighted estimator is always the minimum-variance estimator among every possible estimator of $m$. In routine Bühlmann-Straub calculations, it is the pooled exposure experience used as the collective complement.

Worked Example: Pooling Three Risks With Unequal Exposure

A hypothetical portfolio has three risks. Their total exposures and observed loss costs are: Risk A, 40 units at $100 per unit; Risk B, 30 units at $140 per unit; and Risk C, 30 units at $160 per unit. Calculate the exposure-weighted collective mean and compare it with the equal-risk average.

Method. Multiply each loss cost by its exposure to recover aggregate loss, then divide total loss by total exposure.

Risk A aggregate loss is $40(100)=\$4{,}000$. Risk B aggregate loss is $30(140)=\$4{,}200$. Risk C aggregate loss is $30(160)=\$4{,}800$. Total observed loss is therefore $\$4{,}000+\$4{,}200+\$4{,}800=\$13{,}000$, and total exposure is $40+30+30=100$.

\hat m_{\mathrm{pool}}=\frac{13{,}000}{100}=\$130.00\text{ per exposure unit}.

The equal-risk average would instead be $(100+140+160)/3=\$133.33$ per unit, rounded to cents. That is a different weighting rule, not an alternative arithmetic route to the exposure-pooled result. Check: the pooled estimate lies between the smallest and largest observed costs, and multiplying it by 100 exposure units reproduces the total loss of $\$13{,}000$. The common mistake avoided is averaging the three risk costs equally when the requested complement is exposure-weighted.

Using the Collective Mean in Credibility

Once the collective complement is set, a Bühlmann-Straub credibility estimate for risk $i$ has the form

\hat\Theta_i=Z_i\bar X_i+(1-Z_i)\hat m_{\mathrm{pool}}, \qquad Z_i=\frac{w_i}{w_i+K}, \qquad K=\frac{\phi}{\tau^2}.

Here $w_i$ is the risk’s total exposure, $\bar X_i$ is its exposure-weighted observed mean, and $K$ is the structural constant. The factor $Z_i$ determines how much weight the risk’s own experience receives; $1-Z_i$ is the weight on the collective complement. This tutorial uses that credibility formula to show how a pooled complement enters an indication; it does not re-estimate $\phi$ or $\tau^2$.

Changing the pooling weights changes $\hat m_{\mathrm{pool}}$. For any risk with $Z_i<1$, changing the complement by an amount $\Delta$ changes its indication by $(1-Z_i)\Delta$. The complement has the greatest effect on risks with lower credibility, while a risk with credibility approaching one is less affected.

Worked Example: How Pooling Changes a Credibility Indication

Use the three-risk data from the previous example. For Risk A, the total exposure is 40 and its observed loss cost is $100. Suppose the structural constant is $K=60$. Calculate its indication using the exposure-weighted complement, then show what happens if the equal-risk average is mistakenly used.

Method. First calculate the credibility factor, then blend Risk A’s experience with the selected complement.

Risk A’s credibility is $Z_A=40/(40+60)=0.4$, so the complement weight is $1-Z_A=0.6$. With the correct exposure-weighted complement of $\$130.00$, the indication is

\hat\Theta_A=0.4(100)+0.6(130)=40+78=\$118.00.

If the equal-risk average of $\$133.33$ is substituted, the indication is $0.4(100)+0.6(133.33)=40+79.998=\$120.00$, to the nearest cent. The incorrect complement is $\$3.33$ higher, and the resulting indication is approximately $0.6(3.33)=\$2.00$ higher. Check: both indications lie between Risk A’s experience and the complement used. The mistake is not in calculating $Z_A$; it is using a complement that does not match the stated exposure-weighted target.

Worked Example: Pooling Multiple Periods Before Credibility

A hypothetical insurer records the following loss costs and exposures. Risk X has 20 exposure units at $90 and 30 at $150. Risk Y has 80 units at $130 and 20 at $170. Risk Z has 50 units at $110. Calculate the pooled complement and the credibility indications for X, Y, and Z when $K=40$.

Method. Aggregate each risk’s periods using exposure weights, pool all aggregate losses, and then apply the risk-specific credibility factor using each risk’s total exposure.

For X, total exposure is $20+30=50$, and total loss is $20(90)+30(150)=1{,}800+4{,}500=\$6{,}300$. Thus $\bar X_X=6{,}300/50=\$126.00$. For Y, total exposure is $80+20=100$, and total loss is $80(130)+20(170)=10{,}400+3{,}400=\$13{,}800$, so $\bar X_Y=\$138.00$. For Z, total exposure is 50 and total loss is $50(110)=\$5{,}500$, so $\bar X_Z=\$110.00$.

Across all risks, total loss is $6{,}300+13{,}800+5{,}500=\$25{,}600$, and total exposure is $50+100+50=200$. Therefore, $\hat m_{\mathrm{pool}}=25{,}600/200=\$128.00$.

For X, $Z_X=50/(50+40)=5/9$, giving $\hat\Theta_X=(5/9)(126)+(4/9)(128)=(630+512)/9=\$126.89$, rounded to cents.

For Y, $Z_Y=100/(100+40)=5/7$, giving $\hat\Theta_Y=(5/7)(138)+(2/7)(128)=(690+256)/7=\$135.14$, rounded to cents.

For Z, $Z_Z=50/(50+40)=5/9$, giving $\hat\Theta_Z=(5/9)(110)+(4/9)(128)=(550+512)/9=\$118.00$.

Check: the three indicated costs lie between each risk’s own observed mean and the complement of $\$128.00$. Risk Y has the largest total exposure and therefore the largest credibility factor. The common mistake avoided is using a simple average of the five period observations: that would give each period equal weight despite their different exposures.

Common Pooling Traps

Averaging risk means equally without checking the target. The arithmetic mean of risk means is an equal-risk average. The exposure-pooled complement is total loss divided by total exposure.
Weighting a risk by its number of periods. The pooling weight is its total exposure, not its count of observed periods. Periods can have different exposures.
Using only one exposure weight. For a multi-period risk, first calculate its total exposure and exposure-weighted mean; do not weight its periods equally unless their exposures are equal.
Confusing pooling weights with credibility factors. Total exposure weights form the pooled complement. The factor $Z_i=w_i/(w_i+K)$ then determines how that complement is blended with a particular risk’s experience.
Ignoring the intended portfolio. If exposures are informative about risk quality or the target is an equal-risk mean, the usual exposure-pooled complement may not represent the intended quantity. State the target and check the assumptions before pooling.

Key takeaway: For an exposure-weighted collective complement, sum aggregate losses across risks and divide by total exposure. Use the same complement consistently in each credibility blend, and distinguish an exposure-weighted portfolio mean from an equal-risk average.

Check Your Understanding

Work through each problem, keeping the target and the pooling weights explicit.

A hypothetical portfolio has two risks: 40 exposure units at $75 per unit and 60 units at $105 per unit. What is the exposure-weighted pooled mean? (Answer: $93.00 per exposure unit)
Three risks have total exposures 20, 30, and 50, and respective loss costs of $90, $120, and $150. Calculate the exposure-weighted pooled mean. (Answer: $129.00 per exposure unit)
A risk has exposure 40 and observed loss cost $180. The pooled complement is $129 and $K=60$. Calculate its Bühlmann-Straub indication. (Answer: $149.40 per exposure unit)
Three risks have exposures 10, 30, and 60 and loss costs $200, $100, and $80. Calculate both the exposure-weighted pooled mean and the equal-risk average. Which is the exposure-pooled complement? (Answer: Exposure-weighted pooled mean = $98.00; equal-risk average = $126.67; the exposure-pooled complement is $98.00 per exposure unit.)

What You'll Learn