Choosing Exposure Units For Credibility Weights

Why the Exposure Unit Matters

A Bühlmann-Straub credibility weight is not just a count of observations. It measures how much risk exposure those observations represent. Ten policy-years and ten vehicle-years may describe very different amounts of insured risk, so the exposure unit must match the quantity whose rate is being modeled.

A consistent choice matters both for the credibility factor and for interpreting the rate being estimated. Changing from policy-years to policy-months is legitimate, but the rate, exposure, and model variance parameters must all be converted together. Changing only the exposure number can produce a different—and invalid—credibility estimate.

Definition: An exposure unit is a specified quantity of insured risk used to measure the volume underlying a modeled rate. Examples include one policy-year, one vehicle-year, or a stated amount of payroll. The modeled rate must be expressed per that same unit, and the exposure weight must count multiples or fractions of it.

The Bühlmann-Straub Quantities and Their Units

For risk $i$ and period $j$, let $Y_{ij}$ be the observed rate per chosen exposure unit and $w_{ij}$ its exposure weight. Under the Bühlmann-Straub model, conditional on the risk parameter $\Theta_i$, observations have mean $\mu(\Theta_i)$, variance $\sigma^2(\Theta_i)/w_{ij}$, and are conditionally independent across periods. The model assumes the same conditional mean structure across periods and that exposure is proportional to the amount of information: multiplying exposure by $c$ divides the conditional variance of the rate by $c$.

Write $a=\operatorname{Var}(\mu(\Theta_i))$ for the between-risk variance and $v=E[\sigma^2(\Theta_i)]$ for the expected process-variance parameter. If the modeled rate is dollars per exposure unit, then $a$ has units dollars squared per (exposure unit) squared, while $v$ has units dollars squared per exposure unit. Consequently, $K=v/a$ has units of exposure, as does the total weight $W_i=\sum_j w_{ij}$.

Key result: With $a>0$, $v>0$, and $W_i$ measured in the same exposure unit as $K=v/a$, the Bühlmann-Straub credibility factor and estimate are

K=\frac{v}{a}, \qquad Z_i=\frac{W_i}{W_i+K}, \qquad \widehat{\mu}_i=Z_i\bar{Y}_i+(1-Z_i)m.

Here $\bar{Y}_i=\sum_j w_{ij}Y_{ij}/W_i$ is the risk's exposure-weighted mean rate, and $m=E[\mu(\Theta_i)]$ is the collective mean rate. The estimate is a rate in the same units as $Y_{ij}$. If $a=0$, there is no between-risk variation and the credibility factor is zero; the displayed ratio $v/a$ is not used.

This is a moment-based credibility result; it does not require a particular parametric distribution for the observations. No interest rate or mortality assumption is involved in the examples below. The numerical calculations use the stated model parameters and assume the exposure weights satisfy the model's variance relationship.

Why Consistent Rescaling Preserves Credibility

Suppose one new exposure unit equals $c$ old exposure units. For example, one policy-year equals 12 policy-months, so $c=12$. A rate per new unit is $1/c$ times the rate per old unit. Its between-risk variance is therefore divided by $c^2$. The new numerical exposure weight is $c$ times the old weight. To preserve the model's conditional variance, the process-variance parameter is divided by $c$.

The structural constant changes in the same way as an exposure count: it is multiplied by $c$. Therefore the credibility factor is unchanged. This gives a useful check on any conversion.

Y'=\frac{Y}{c}, \qquad a'=\frac{a}{c^2}, \qquad v'=\frac{v}{c}, \qquad W'=cW, \qquad K'=\frac{v'}{a'}=cK.

Justification. The rate variance between risks transforms as $\operatorname{Var}(\mu/c)=\operatorname{Var}(\mu)/c^2$, giving $a'=a/c^2$. For a period with old weight $w$, the new weight is $cw$, and the new rate must have conditional variance $\operatorname{Var}(Y/c\mid\Theta)=\sigma^2(\Theta)/(c^2w)$. Under the new model this variance is $\sigma'^2(\Theta)/(cw)$, so $\sigma'^2(\Theta)=\sigma^2(\Theta)/c$, and hence $v'=v/c$. It follows that $K'=v'/a'=c(v/a)=cK$. Thus

Z'=\frac{W'}{W'+K'}=\frac{cW}{cW+cK} =\frac{W}{W+K}=Z.

Although the credibility factor is unchanged, the estimated rate is expressed in the new rate units and must be converted back if a different reporting unit is wanted.

Worked Example: Converting Policy-Years to Policy-Months

A hypothetical risk has 6 policy-years of experience, a weighted mean loss cost of $110 per policy-year, and a collective mean of $100 per policy-year. Suppose $a=900$ dollars squared per (policy-year) squared and $v=3{,}600$ dollars squared per policy-year. Calculate its estimate, then repeat the calculation in policy-month units. No discounting, mortality model, or parametric distribution is used.

Method. First calculate the structural constant in policy-years. Then convert every rate, variance parameter, and exposure weight to policy-month units and check that the credibility factor agrees.

In policy-years,

K=\frac{3{,}600}{900}=4\text{ policy-years}, \qquad Z=\frac{6}{6+4}=0.6.

The estimated annual rate is

\widehat{\mu}=0.6(110)+0.4(100)=66+40 =\$106\text{ per policy-year}.

One policy-year is 12 policy-months. Thus the risk has $W'=12(6)=72$ policy-months. The rate values are divided by 12: the risk mean is $110/12=\$9.1667$ per policy-month, and the collective mean is $100/12=\$8.3333$ per policy-month. The parameters convert as

a'=\frac{900}{12^2}=\frac{900}{144}=6.25 \text{ dollars squared per (policy-month) squared},

and

v'=\frac{3{,}600}{12}=300 \text{ dollars squared per policy-month}.

Therefore $K'=300/6.25=48$ policy-months and $Z'=72/(72+48)=0.6$. The estimated monthly rate is $0.6(9.1667)+0.4(8.3333)=5.5+3.3333=\$8.8333$ per policy-month, which is $12(8.8333)=\$106.00$ per policy-year before rounding effects. The two estimates agree. Rates are shown to four decimal places where needed, with money rounded to the cent at the final step.

Common mistake avoided: Dividing the rate by 12 but leaving $a$ and $v$ unchanged would change $K$ inconsistently and give a false credibility factor. The conversion requires $a/144$, $v/12$, and 12 times the exposure.

Choose the Exposure That Represents the Risk

The appropriate exposure base depends on what generates the modeled loss or claim opportunity. For a liability rate that is intended to reflect insured vehicles, vehicle-years may be more meaningful than policy-years when policies insure different numbers of vehicles. A policy count can still be an appropriate weight when policies have comparable risk volume and the model is explicitly defined per policy-year. The choice is a modeling decision, not a device for obtaining a preferred credibility factor.

The observed rate and weighted mean must also use the chosen base. If $L_{ij}$ is loss for a period and $w_{ij}$ is vehicle-years, the rate is $Y_{ij}=L_{ij}/w_{ij}$, not loss divided by the number of policies. The exposure-weighted mean then recovers total loss divided by total exposure.

Worked Example: A Fleet With Different Vehicle Counts

A hypothetical fleet has three policy-years containing 2, 3, and 5 insured vehicles. The corresponding losses are $180, $240, and $380. Model the rate per vehicle-year. Let $a=400$ dollars squared per (vehicle-year) squared, $v=2{,}000$ dollars squared per vehicle-year, and the collective mean be $100 per vehicle-year. Find the credibility estimate. No interest, mortality, or parametric distribution assumption is needed.

Method. Convert each period's loss to a rate per vehicle-year and use vehicle-years—not policy-years—as the weights. The exposures are 2, 3, and 5 vehicle-years, so $W=2+3+5=10$ vehicle-years.

The three rates are $180/2=\$90$, $240/3=\$80$, and $380/5=\$76$ per vehicle-year. The exposure-weighted mean is

\bar{Y}=\frac{2(90)+3(80)+5(76)}{10} =\frac{180+240+380}{10} =\frac{800}{10}=\$80 \text{ per vehicle-year}.

The structural constant and credibility factor are

K=\frac{2{,}000}{400}=5\text{ vehicle-years}, \qquad Z=\frac{10}{10+5}=\frac{2}{3}.

The credibility estimate is

\widehat{\mu}=\frac{2}{3}(80)+\frac{1}{3}(100) =\frac{160+100}{3} =\$86.67\text{ per vehicle-year}.

As a check, the risk experience is below the collective mean, so the credibility estimate must lie between $80 and $100; $86.67 does. Rounding is to the nearest cent only at the final step.

Common mistake avoided: Substituting the 3 policy-years for $W=10$ vehicle-years while retaining $K=5$ vehicle-years would give $Z=3/(3+5)=0.375$, and an apparent estimate of $0.375(80)+0.625(100)=\$92.50$. This calculation mixes incompatible exposure units. It also treats periods with different vehicle counts as though each represented equal risk volume.

Partial Exposure and a Second Unit Check

Exposure need not be an integer. A policy active for half a year contributes 0.5 policy-years, and a set of such periods can be added just like whole-year exposures. When data are stored in a different unit, convert the accumulated exposure and the parameters together. A useful exam habit is to write the unit next to every value of $W$, $K$, $a$, and $v$ before calculating.

Worked Example: Fractional Exposure Reported in Monthly Units

A hypothetical account has a total of 2.5 policy-years of experience and a weighted mean of $2,400 per policy-year. The collective mean is $2,000 per policy-year. Assume $a=144$ dollars squared per (policy-year) squared and $v=360$ dollars squared per policy-year. Calculate the estimate in annual units and verify it using monthly units. No discounting, mortality assumption, or parametric distribution is required.

In annual units, the structural constant is $K=360/144=2.5$ policy-years. Thus $Z=2.5/(2.5+2.5)=0.5$, and

\widehat{\mu}=0.5(2{,}400)+0.5(2{,}000) =\$2{,}200\text{ per policy-year}.

In monthly units, exposure is $W'=12(2.5)=30$ policy-months. The risk mean becomes $2{,}400/12=\$200$ per policy-month, and the collective mean becomes $2{,}000/12=\$166.6667$ per policy-month. The converted parameters are

a'=\frac{144}{144}=1 \text{ dollar squared per (policy-month) squared}, \qquad v'=\frac{360}{12}=30 \text{ dollars squared per policy-month}.

Therefore $K'=30/1=30$ policy-months and $Z'=30/(30+30)=0.5$. The monthly estimate is $0.5(200)+0.5(166.6667)=\$183.3333$ per policy-month, or $12(183.3333)=\$2{,}200.00$ per policy-year. Rates are rounded to four decimal places where needed; money is rounded to the cent at the final step.

Common mistake avoided: Using $W'=30$ with the original $K=2.5$ would yield $Z=30/(30+2.5)=0.9231$. That treats a structural constant measured in policy-years as if it were measured in policy-months. The resulting estimate would be pulled far too close to the risk's experience and would not represent the same model.

Practical Checks Before Applying Credibility

A consistent unit conversion should leave $Z$ unchanged, and the final estimate converted back to the original rate unit should also be unchanged. If either check fails, inspect the exposure base and the units of $a$ and $v$ before interpreting the result.

Name the modeled rate. State whether it is loss per policy-year, per vehicle-year, or another defined unit.
Check the exposure source. The weight should count the volume represented by that rate, including fractional exposures where appropriate.
Check dimensions. For a rate in dollars per unit, $a$ is rate-squared and $v$ is rate-squared times exposure; $v/a$ is an exposure amount.
Check conversions. If the new unit is $c$ old units, multiply exposure and $K$ by $c$, divide the rate and $v$ by $c$, and divide $a$ by $c^2$.
Check model fit. A simple rescaling is appropriate only if the alternative exposure unit measures the same underlying risk volume and the conditional variance remains inversely proportional to exposure.

Changing from policy-years to vehicle-years is not generally a mere unit conversion: it may change what risk volume is being measured. If vehicle counts materially vary, a vehicle-year model requires rates and parameters defined on the vehicle-year basis; policy-year data cannot be relabeled without rebuilding those quantities. This distinction separates harmless changes in the size of a unit from changes in the underlying exposure definition.

Key takeaway: A credibility weight is meaningful only relative to its exposure definition. Choose a base that represents the modeled risk, express the rate and model parameters on that same base, and verify that a consistent rescaling leaves the credibility factor unchanged.

Check Your Understanding

For each problem, keep the modeled rate and exposure units consistent. Answers are given at the end of each question.

A risk has $W=8$ exposure units, $a=225$ dollars squared per exposure unit squared, and $v=900$ dollars squared per exposure unit. Find $K$ and $Z$. (Answer: $K=4$ exposure units; $Z=2/3$.)
A portfolio has $W=15$ policy-years and $K=3$ policy-years. Express both in policy-months and calculate the credibility factor. (Answer: $W'=180$ policy-months, $K'=36$ policy-months, $Z=180/216=5/6$.)
Three periods have exposures 1, 2, and 3 units and losses of $70, $120, and $210. Given $K=3$ units and collective mean $80 per unit, calculate the credibility estimate. (Answer: $\bar{Y}=400/6=\$66.6667$, $Z=6/9=2/3$, and the estimate is $(2/3)(66.6667)+(1/3)(80)=\$71.11$ per unit.)
A fleet has four policy-years, each covering four vehicles. Its rate is $90 per vehicle-year, the collective mean is $105 per vehicle-year, and $K=8$ vehicle-years. Find the credibility estimate using the appropriate exposure. (Answer: $W=4(4)=16$ vehicle-years, $Z=16/(16+8)=2/3$, and the estimate is $(2/3)(90)+(1/3)(105)=\$95$ per vehicle-year.)
For a model with $W=4$ policy-years, $a=64$ dollars squared per (policy-year) squared, $v=256$ dollars squared per policy-year, risk mean $70 per policy-year, and collective mean $90 per policy-year, convert to policy-month units and calculate the estimate in both units. (Answer: $K=4$ years and $Z=1/2$, so the annual estimate is $80 per policy-year. In months, $W'=48$, $a'=64/144=4/9$ dollars squared per (policy-month) squared, $v'=256/12=64/3$ dollars squared per policy-month, $K'=48$ months, and $Z'=1/2$. The monthly estimate is $6.6667 per policy-month, or $80 per policy-year.)

What You'll Learn