Tutorials › Exam ASTAM - Advanced Short-Term Actuarial Mathematics › Exposure Units and the Credibility Constant

Home  /  Tutorials  /  Exam ASTAM / Buhlmann and Buhlmann-Straub

Exam ASTAM · Buhlmann and Buhlmann-Straub · Tutorial 529 of 979

Exposure Units and the Credibility Constant

Learn what the credibility constant measures, how it scales with the chosen exposure unit, and why a consistent change of convention preserves the premium.

Intermediate 10 min read

What You'll Learn

  • Define the credibility constant from EPV and VHM under a unit-exposure model
  • Explain why its numerical value depends on the exposure convention
  • Convert exposure weights and EPV consistently when changing units
  • Verify that the credibility factor and premium are invariant to a consistent conversion
  • Distinguish exposure-unit changes from changes in the monetary unit
  • Recognize when proportional exposure variance assumptions do not apply

Why the Credibility Constant Has Units

In Bühlmann-Straub credibility, experience is weighted by exposure: a risk observed over more exposure generally provides more information than one observed over less. The credibility constant \(K\) compares two sources of variation, but it is not a unit-free number. Its numerical value depends on what counts as one unit of exposure and on how the model defines the experience measurement.

That dependence is not a defect. The credibility factor combines total exposure and \(K\), so both must use the same exposure convention. If the convention changes consistently, the numerical value of \(K\) changes but the credibility factor and premium do not. This distinction is important whenever exposure weights are reported in different units or are rescaled for computation.

Definition: Let \(\Theta\) denote a risk's latent risk parameter, and let \(X_{ij}\) be the experience measurement for risk \(i\) in observation period \(j\). Under the Bühlmann-Straub model, conditional on \(\Theta_i\), the observations for a risk have common conditional mean \(m(\Theta_i)\), are conditionally uncorrelated, and have conditional variance \(\operatorname{Var}(X_{ij}\mid\Theta_i)=\sigma^2(\Theta_i)/w_{ij}\), where \(w_{ij}>0\) is exposure measured in a stated unit. Define the expected process variance per unit exposure as \(\mathrm{EPV}=E[\sigma^2(\Theta)]\), the variance of hypothetical means as \(\mathrm{VHM}=\operatorname{Var}(m(\Theta))\), and, when \(\mathrm{VHM}>0\), the credibility constant as \(K=\mathrm{EPV}/\mathrm{VHM}\). Thus \(K\) is measured in exposure units.

The exposure-weighted experience mean and total exposure are

$$ \bar{X}_w=\frac{\sum_j w_jX_j}{w}, \qquad w=\sum_j w_j. $$

For one risk, the Bühlmann-Straub premium for the next period, under the same experience scale and unit-exposure convention, is

$$ P=Z\bar{X}_w+(1-Z)\mu,\qquad Z=\frac{w}{w+K},\qquad \mu=E[m(\Theta)]. $$

Here \(\mu\) is the collective mean. The formula assumes that the future quantity being priced is on the same per-unit basis as the observed experience. Exposure describes the amount of observation; it is not itself a multiplier to apply to the resulting per-unit premium.

Why \(K\) Is EPV Divided by VHM

The ratio has a direct interpretation. EPV measures average random variation within a risk, while VHM measures variation in underlying expected experience between risks. A large EPV relative to VHM means that observed experience is noisy compared with the differences credibility is trying to detect, so more exposure is needed before experience receives substantial weight.

The standard Bühlmann-Straub factor can also be justified by comparing the variance of the weighted experience mean with the between-risk variation. Conditional on \(\Theta\), its variance is \(\sigma^2(\Theta)/w\). Taking expectations gives \(\operatorname{Var}(\bar{X}_w\mid\Theta)\) averaged over risks equal to \(\mathrm{EPV}/w\). The weighted mean also varies between risks through \(m(\Theta)\), with variance \(\mathrm{VHM}\). The linear credibility weight on experience is therefore the between-risk component divided by the total variance relevant to the linear prediction:

$$ Z=\frac{\mathrm{VHM}}{\mathrm{VHM}+\mathrm{EPV}/w} =\frac{w\,\mathrm{VHM}}{w\,\mathrm{VHM}+\mathrm{EPV}} =\frac{w}{w+\mathrm{EPV}/\mathrm{VHM}} =\frac{w}{w+K}. $$

This also shows why \(K\) must have exposure units: it is added to \(w\) in the denominator. The ratio \(\mathrm{EPV}/\mathrm{VHM}\) is not a stand-alone measure of a risk's premium or of the amount of risk in dollars; it tells how much exposure, under the chosen convention, is associated with the balance between within-risk and between-risk variation.

Worked Example: Finding \(K\) and the Premium

A hypothetical portfolio uses one annual policy as one exposure unit. Suppose \(\mathrm{EPV}=\$2{,}400^2\) and \(\mathrm{VHM}=\$600^2\) on the model's per-policy annual experience scale. A risk has two annual observations: \(X_1=\$900\) with exposure \(w_1=1\), and \(X_2=\$1{,}100\) with exposure \(w_2=3\). The collective mean is \(\mu=\$1{,}000\). Calculate the credibility constant and premium.

Method and model check. Use the stated Bühlmann-Straub convention: conditional variance is inversely proportional to exposure, and the two observations share the risk-specific conditional mean. Both variance components use squared dollars, so their ratio is exposure units.

Step 1: Calculate \(K\).

$$ K=\frac{\mathrm{EPV}}{\mathrm{VHM}} =\frac{2{,}400}{600}=4\text{ annual policy exposure units}. $$

Step 2: Calculate total exposure and the weighted mean.

$$ w=1+3=4,\qquad \bar{X}_w=\frac{1(900)+3(1{,}100)}{4} =\frac{900+3{,}300}{4} =\$1{,}050. $$

Step 3: Calculate the credibility factor.

$$ Z=\frac{4}{4+4}=0.5. $$

Step 4: Calculate the premium.

$$ P=0.5(1{,}050)+0.5(1{,}000) =525+500 =\$1{,}025. $$

Check. The experience mean is halfway between the collective mean and the premium, as required by \(Z=0.5\). The common mistake this method avoids is treating \(K=4\) as a dollar amount; it is four annual policy exposure units.

Changing the Exposure Unit

Suppose a new convention counts \(c\) new exposure units for each old exposure unit, where \(c>0\). The numerical exposure weights then become \(w'_j=cw_j\). To preserve the same conditional variance for the unchanged experience measurement, the per-unit EPV must also be converted:

$$ \frac{\mathrm{EPV}}{w_j} =\frac{\mathrm{EPV}'}{w'_j} \quad\Longrightarrow\quad \mathrm{EPV}'=c\,\mathrm{EPV}. $$

The risk-specific conditional means have not changed, so VHM remains the same. Consequently \(K'=cK\). The weighted mean is unchanged because multiplying every weight by the same positive constant cancels between its numerator and denominator. The credibility factor is unchanged as well:

$$ Z'=\frac{w'}{w'+K'} =\frac{cw}{cw+cK} =\frac{w}{w+K}=Z. $$

The premium is therefore unchanged, since both its experience component and collective component are unchanged. This is a re-expression of the same model, not a change in the information in the observations.

1
State the old convention.
Record the unit represented by each exposure weight and confirm that the model uses conditional variance \(\sigma^2(\Theta)/w\).
2
Convert every exposure weight.
If each old unit becomes \(c\) new units, set \(w'_j=cw_j\), so total exposure is \(w'=cw\).
3
Convert EPV consistently.
For the same experience measurements and conditional variances, set \(\mathrm{EPV}'=c\,\mathrm{EPV}\). Keep VHM unchanged.
4
Recalculate and check.
Then \(K'=cK\), \(Z'=Z\), and the premium is unchanged. If these checks fail, the exposure convention or another model quantity has likely been changed inconsistently.

Worked Example: Converting Annual Exposure to Monthly Exposure

A hypothetical rating analysis uses annual exposure units, with \(\mathrm{EPV}=300\) and \(\mathrm{VHM}=100\) in the corresponding squared experience units. A risk has observations \(X_1=1{,}200\) and \(X_2=900\), with annual exposures 2 and 1. The collective mean is \(\mu=1{,}000\). Re-express the model using monthly exposure units, where one year is 12 units, and compare the premiums.

Method and common mistake. Keep the observed annual experience values on their original scale. Since each old exposure unit is represented by 12 new units, multiply both the exposure weights and EPV by 12. Scaling only the weights would change the modeled conditional variances and would not be a mere unit conversion.

Step 1: Calculate the original quantities.

$$ K=\frac{300}{100}=3,\qquad w=2+1=3, $$
$$ \bar{X}_w=\frac{2(1{,}200)+1(900)}{3} =\frac{2{,}400+900}{3} =1{,}100. $$

Thus \(Z=3/(3+3)=0.5\), and the annual-scale premium is

$$ P=0.5(1{,}100)+0.5(1{,}000)=\$1{,}050. $$

Step 2: Convert the convention. The new total exposure is \(w'=12(3)=36\), and the new EPV is \(\mathrm{EPV}'=12(300)=3{,}600\). VHM remains 100. Therefore

$$ K'=\frac{3{,}600}{100}=36,\qquad \bar{X}'_w=\frac{24(1{,}200)+12(900)}{36} =\frac{28{,}800+10{,}800}{36} =1{,}100. $$

Step 3: Check the new premium.

$$ Z'=\frac{36}{36+36}=0.5,\qquad P'=0.5(1{,}100)+0.5(1{,}000)=\$1{,}050. $$

Check. Both \(K\) and total exposure increased by a factor of 12, so \(Z\) and the premium stayed the same. The incorrect conversion that multiplies exposure by 12 but leaves EPV at 300 would give \(K'=3\), an erroneous \(Z'=36/39\), and a premium of about \(\$1{,}092.31\). That result describes a different variance model, not the same model in monthly units.

Exposure Units Are Not Monetary Units

A change in exposure unit is different from converting dollars to thousands of dollars. If every experience measurement and collective mean is divided by 1,000, every variance is divided by \(1{,}000^2\). EPV and VHM both scale by the same factor, so their ratio \(K\) does not change. Exposure weights do not change in this conversion. The premium is divided by 1,000 and is then interpreted in thousands of dollars.

This contrast helps diagnose calculations. A change to the exposure unit scales \(w\), EPV, and \(K\) together while leaving experience measurements unchanged. A change to the monetary unit scales experience measurements and both variance components, but not exposure. Mixing these two conversions can make a dimensionally plausible formula produce the wrong credibility factor or a premium in the wrong unit.

Worked Example: Converting the Experience Scale to Thousands

A hypothetical model has \(\mathrm{EPV}=40{,}000\) and \(\mathrm{VHM}=10{,}000\), both in squared dollars per the stated unit-exposure convention. A risk has exposure 1 at an experience value of \(\$900\), and exposure 3 at an experience value of \(\$1{,}100\). The collective mean is \(\$1{,}000\). Recalculate the model in thousands of dollars and verify the premium.

Method and common mistake. This is a monetary conversion, not an exposure conversion. Divide experience amounts by 1,000 and variance components by \(1{,}000^2\). Do not multiply exposure or \(K\) by 1,000.

Step 1: Find \(K\) and the dollar-scale premium.

$$ K=\frac{40{,}000}{10{,}000}=4,\qquad w=1+3=4, $$
$$ \bar{X}_w=\frac{1(900)+3(1{,}100)}{4} =\frac{4{,}200}{4} =\$1{,}050,\qquad Z=\frac{4}{4+4}=0.5. $$
$$ P=0.5(1{,}050)+0.5(1{,}000)=\$1{,}025. $$

Step 2: Convert to thousands. The variance components become \(\mathrm{EPV}'=40{,}000/1{,}000{,}000=0.04\) and \(\mathrm{VHM}'=10{,}000/1{,}000{,}000=0.01\), in squared thousands of dollars. Thus

$$ K'=\frac{0.04}{0.01}=4,\qquad \bar{X}'_w=\frac{1(0.9)+3(1.1)}{4}=1.05,\qquad \mu'=1.00. $$

The exposure and credibility factor are unchanged, so

$$ P'=0.5(1.05)+0.5(1.00)=1.025 $$

thousands of dollars, or \(\$1{,}025\). Check. The numerical value of \(K\) stayed at 4 because EPV and VHM were scaled equally; only the monetary representation of the premium changed.

When the Conversion Rule Applies

The simple rule \(w'=cw\), \(\mathrm{EPV}'=c\,\mathrm{EPV}\), and \(K'=cK\) requires the same observations, the same risk-specific expected values, and the same conditional variances after conversion. It relies on the model's inverse-exposure variance relationship. It does not automatically apply if the new weights represent a genuine change in exposure, if the experience variable is also changed, or if conditional variance is not proportional to the reciprocal of exposure.

For example, if doubling a policy's exposure also changes the expected experience because the coverage or time period differs, merely doubling the weight is not necessarily a unit conversion. The model must describe how both the expected value and variance respond to that changed exposure. Likewise, if the variance relationship is not \(\sigma^2(\Theta)/w\), the standard \(K=\mathrm{EPV}/\mathrm{VHM}\) formula may not yield the same credibility factor.

  • Check the denominator. The exposure \(w\) and \(K\) in \(w/(w+K)\) must be expressed in the same units.
  • Check both variance components. EPV and VHM must refer to the same experience scale; EPV must also match the exposure convention.
  • Check the weighted mean. Multiplying every exposure by the same positive constant must not change \(\bar{X}_w\).
  • Check invariance. Under a consistent exposure conversion, the computed premium should be identical before and after conversion.
Key takeaway: \(K=\mathrm{EPV}/\mathrm{VHM}\) is an exposure-unit quantity, not an intrinsic count that can be compared across conventions without conversion. Rescaling exposure by \(c\) requires rescaling EPV by \(c\); then \(K\) also scales by \(c\), while \(Z\) and the premium remain unchanged.

Check Your Understanding

Use a consistent exposure convention in each calculation. Monetary quantities are hypothetical.

  1. A model has \(\mathrm{EPV}=240\), \(\mathrm{VHM}=60\), and total exposure \(w=6\). Find \(K\) and \(Z\). (Answer: \(K=4\) exposure units; \(Z=0.6000\).)
  2. In the preceding setting, exposure is re-expressed so that each old unit is 5 new units. What are the converted EPV, \(K\), and total exposure, and what is the new \(Z\)? (Answer: \(\mathrm{EPV}'=1{,}200\), \(K'=20\), \(w'=30\), and \(Z'=0.6000\).)
  3. A risk has experience values 800 and 1,200 with exposures 1 and 3. The collective mean is 1,000, EPV is 400, and VHM is 100. Find the weighted mean, \(K\), \(Z\), and premium. (Answer: weighted mean \(=1{,}100\), \(K=4\), \(Z=0.5000\), premium \(=1{,}050\).)
  4. For a model with \(\mathrm{EPV}=500\), \(\mathrm{VHM}=100\), and \(w=4\), an analyst multiplies exposure by 10 but leaves EPV unchanged. Explain the error and give the consistent converted values and credibility factor. (Answer: Leaving EPV unchanged changes the modeled conditional variances; consistently, \(\mathrm{EPV}'=5{,}000\), \(K'=50\), \(w'=40\), and \(Z'=0.4444\), equal to the original \(4/(4+5)=0.4444\).)
  5. All experience values are converted from dollars to thousands of dollars. What happens to EPV, VHM, \(K\), and \(Z\), and how is the premium reported? (Answer: EPV and VHM are each divided by \(1{,}000{,}000\); \(K\) and \(Z\) are unchanged; the premium is divided by 1,000 and reported in thousands of dollars.)

Previous

Next

Back to Tutorials
Clinical Biostats