Deriving Exposure-Based Credibility For Reserve Experience

Why Exposure-Based Credibility Matters

Accident periods often contribute different amounts of experience. A rate based on a large exposure is generally more informative about the underlying expected loss rate than a rate based on a small exposure. Credibility must therefore reflect both the size of each period's exposure and the variation between periods.

The Buhlmann-Straub model provides a standard way to make this adjustment. It first combines period-specific rates using exposure weights, then blends that experience rate with a collective mean. In a reserve application, the result is a credible ultimate loss rate; multiplying by the appropriate target exposure converts that rate into an ultimate loss estimate.

Definition: In the Buhlmann-Straub model, $R_j$ is the observed loss rate for period $j$, based on exposure $w_j>0$. Conditional on a risk parameter $\Theta$, the rates are independent and satisfy $E[R_j\mid\Theta]=\mu(\Theta)$ and $\operatorname{Var}(R_j\mid\Theta)=\sigma^2(\Theta)/w_j$. The collective mean is $m=E[\mu(\Theta)]$, the expected process variance is $\mathrm{EPV}=E[\sigma^2(\Theta)]$, and the variance of hypothetical means is $\mathrm{VHM}=\operatorname{Var}(\mu(\Theta))$. When $\mathrm{VHM}>0$, define $K=\mathrm{EPV}/\mathrm{VHM}$.

The model assumes the rates are measured on a consistent basis and that exposure is proportional to the amount of information in a period. It also assumes a common underlying collective distribution for the risks being blended. If rates have not been adjusted to a comparable price level or definition of loss, weighting them cannot fix that mismatch. In reserve work, the rates used here must represent ultimate losses per exposure, or otherwise be converted to that basis before this calculation.

The Unequal-Exposure Credibility Factor

Let $W=\sum_j w_j$ be the total experience exposure. The exposure-weighted experience rate is $\bar R_w=\sum_j w_jR_j/W$. Its credibility estimate is a blend of $\bar R_w$ and the collective mean $m$.

\widehat{\mu}=Z\bar R_w+(1-Z)m,\qquad \bar R_w=\frac{\sum_j w_jR_j}{W},\qquad Z=\frac{W}{W+K} =\frac{W\,\mathrm{VHM}}{W\,\mathrm{VHM}+\mathrm{EPV}}.

The factor $Z$ is the weight on the experience rate. It increases with total exposure $W$, and decreases as process variation grows relative to variation between risk means. The parameter $K$ is expressed in exposure units: it is the amount of exposure at which the credibility factor is one-half.

Why the Factor Has This Form

The formula follows by finding the best linear predictor of the individual risk mean $\mu(\Theta)$ from the observed weighted rate. The derivation uses the model assumptions stated above and requires finite variances.

Find the conditional variance of the weighted rate.
Given $\Theta$, conditional independence makes the variance of the weighted sum equal to the sum of its component variances. Thus $\operatorname{Var}(\bar R_w\mid\Theta)=\sum_j w_j^2\sigma^2(\Theta)/w_j\,/W^2=\sigma^2(\Theta)/W$.

Apply total variance and covariance.
The law of total variance gives $\operatorname{Var}(\bar R_w)=\mathrm{VHM}+\mathrm{EPV}/W$, since $E[\bar R_w\mid\Theta]=\mu(\Theta)$. Also, $\operatorname{Cov}(\mu(\Theta),\bar R_w)=\mathrm{VHM}$: conditionally, the weighted rate has mean $\mu(\Theta)$.

Choose the best linear coefficient.
For a predictor $a+b\bar R_w$, the least-squares coefficient on $\bar R_w$ is $\operatorname{Cov}(\mu(\Theta),\bar R_w)/\operatorname{Var}(\bar R_w)$. Substitution gives $b=\mathrm{VHM}/(\mathrm{VHM}+\mathrm{EPV}/W)=W/(W+K)$.

Set the intercept.
The best linear predictor has the same mean as $\mu(\Theta)$. Since $E[\bar R_w]=m$, its intercept is $m(1-b)$. The predictor is therefore $Z\bar R_w+(1-Z)m$, as stated.

This is a best linear predictor under the model, not a claim that the conditional distribution of $\mu(\Theta)$ given the observations is normal. Nor does the formula apply unchanged if the conditional variance does not decrease in inverse proportion to exposure, or if the periods are dependent after conditioning on $\Theta$.

From Credible Rate to Ultimate Loss

Once a credible rate has been calculated, the conversion to an ultimate loss estimate is a separate step. If the rate is dollars of ultimate loss per exposure unit and the target accident period has exposure $E^*$, then the estimate is $E^*\widehat{\mu}$. The exposure used in the credibility factor is the exposure supporting the observed experience; $E^*$ is the exposure to which the estimated rate will be applied. They need not be equal.

Key result: With a consistent ultimate-loss-rate basis, the estimated ultimate loss for target exposure $E^*$ is $E^*\widehat{\mu}$. This scaling does not itself estimate development, trend, or changes in coverage; those must be addressed before applying the credible rate if they affect comparability.

Worked Example: Unequal Accident-Period Exposures

A hypothetical portfolio has two accident-period ultimate loss rates: $800 per exposure unit on exposure 2, and $1,100 per exposure unit on exposure 4. Assume $\mathrm{EPV}=20{,}000$, $\mathrm{VHM}=10{,}000$, and collective mean $m=\$900$ per exposure unit. Estimate the credible rate and the ultimate loss for a target period with exposure 15.

Method and assumptions. Use Buhlmann-Straub credibility. The rates are assumed to be comparable ultimate loss rates, with conditional variance inversely proportional to exposure. First calculate the credibility parameter and the exposure-weighted rate.

The credibility parameter is $K=\mathrm{EPV}/\mathrm{VHM}=20{,}000/10{,}000=2$ exposure units. Total experience exposure is $W=2+4=6$.

The weighted rate is

\bar R_w=\frac{2(800)+4(1{,}100)}{6} =\frac{1{,}600+4{,}400}{6} =\frac{6{,}000}{6} =\$1{,}000.

The credibility factor is $Z=W/(W+K)=6/(6+2)=6/8=0.75$. Thus the credible rate is

\widehat{\mu}=0.75(1{,}000)+0.25(900) =750+225 =\$975.00\text{ per exposure unit}.

Apply this rate to target exposure 15: $15(975)=\$14{,}625.00$. The estimated ultimate loss is therefore $14,625.00. The check is that the rate lies between the experience rate of $1,000 and the collective mean of $900, as a blend with weights 0.75 and 0.25 must.

Common mistake avoided: Giving both periods equal weight would produce an experience rate of $950, despite their unequal exposures. Buhlmann-Straub uses the exposure-weighted rate, not the simple average of the two rates.

Worked Example: Several Periods and a Different Target Exposure

A hypothetical book has three comparable accident-period ultimate loss rates: $720 on exposure 1, $960 on exposure 3, and $840 on exposure 6. The collective mean is $800 per exposure unit, $\mathrm{EPV}=25{,}000$, and $\mathrm{VHM}=5{,}000$. Find the credible rate and estimate ultimate losses for a target exposure of 25.

Method and assumptions. Use the Buhlmann-Straub weighted average and credibility factor. The observations are conditionally independent, share the same collective model, and are on the same ultimate-loss basis.

The total exposure is $W=1+3+6=10$. The credibility parameter is $K=25{,}000/5{,}000=5$. The weighted loss amount represented by the observed rates is

$$ 1(720)+3(960)+6(840)=720+2{,}880+5{,}040=8{,}640. $$

Therefore, $\bar R_w=8{,}640/10=\$864$ per exposure unit, and $Z=10/(10+5)=10/15=2/3$. The credible rate is

\widehat{\mu}=\frac{2}{3}(864)+\frac{1}{3}(800) =576+266.6667 =\$842.6667\text{ per exposure unit}.

Multiplying by target exposure gives $25(842.6667)=\$21{,}066.6667$, or $21,066.67 to the nearest cent. The experience rate is above the collective mean, so the credible rate should be between $800 and $864; $842.6667 satisfies this check.

Common mistake avoided: The target exposure of 25 is not added to the experience exposure when calculating $Z$. It is used only after the credible rate has been obtained.

Worked Example: Deriving the Credibility Parameter from Variances

A hypothetical portfolio has experience rates of $1,300 on exposure 2 and $900 on exposure 5. The collective mean is $1,100 per exposure unit, with $\mathrm{EPV}=16{,}000$ and $\mathrm{VHM}=8{,}000$. Calculate the credible rate and the ultimate estimate for target exposure 12.

Method and assumptions. Calculate $K$ from the variance components, then calculate the weighted experience rate, credibility, and scaled ultimate loss. Treat all rates as comparable ultimate loss rates and exposures as proportional to information.

First, $K=16{,}000/8{,}000=2$. Total experience exposure is $W=2+5=7$, and

\bar R_w=\frac{2(1{,}300)+5(900)}{7} =\frac{2{,}600+4{,}500}{7} =\frac{7{,}100}{7} =\$1{,}014.2857.

The credibility factor is $Z=7/(7+2)=7/9$. The experience component in the blend is $(7/9)(7{,}100/7)=7{,}100/9$. The collective component is $(2/9)(1{,}100)=2{,}200/9$. Therefore,

\widehat{\mu}=\frac{7{,}100+2{,}200}{9} =\frac{9{,}300}{9} =\$1{,}033.3333\text{ per exposure unit}.

The target ultimate loss is $12(1{,}033.3333)=\$12{,}400.00$. The credible rate is between the collective mean of $1,100 and the experience rate of approximately $1,014.29, and the target estimate equals 12 times the unrounded credible rate.

Common mistake avoided: $K$ is the ratio $\mathrm{EPV}/\mathrm{VHM}$, not the reverse. Reversing it would change the credibility assigned to experience and generally produce an incorrect blend.

Checks and Common Traps

A short sequence of checks catches many errors in exposure-based reserve credibility:

Check the weighted average. Its weights must be the period exposures, and they must sum to $W$.
Check the credibility direction. With $K>0$, $0<Z<1$; increasing $W$ increases $Z$.
Check the blend. The credible rate must lie between $\bar R_w$ and $m$.
Check the units. $K$ has exposure units, $\widehat{\mu}$ has loss-per-exposure units, and $E^*\widehat{\mu}$ has loss units.
Keep the bases aligned. Do not blend a paid rate from one period with an ultimate rate from another without first making them comparable.

A frequent exam trap is reversing the weight: $Z$ multiplies the observed exposure-weighted rate, while $1-Z$ multiplies the collective mean. Another is to use an unweighted average simply because only a few period rates are listed. The model gives more influence to periods with more exposure, and the collective mean receives the complementary weight.

Check Your Understanding

For each problem, use the Buhlmann-Straub model and keep the experience exposure separate from the target exposure.

A hypothetical portfolio has rates $600 on exposure 1 and $1,000 on exposure 3. Let $m=\$750$, $K=2$, and target exposure be 8. Find the credible rate and ultimate loss. (Answer: credible rate $850.00 per exposure unit; ultimate loss $6,800.00)
Rates are $1,200 on exposure 2 and $800 on exposure 5. Let $m=\$950$, $\mathrm{EPV}=30{,}000$, $\mathrm{VHM}=10{,}000$, and target exposure be 20. Find $K$, the credible rate, and the ultimate loss. (Answer: $K=3$; credible rate $925.00 per exposure unit; ultimate loss $18,500.00)
A portfolio has rates $700 on exposure 4 and $1,000 on exposure 5. Let $m=\$900$, $\mathrm{EPV}=18{,}000$, $\mathrm{VHM}=6{,}000$, and target exposure be 16. Find the exposure-weighted rate, credibility factor, and ultimate loss. (Answer: weighted rate $866.67; $Z=0.7500$; ultimate loss $14,000.00)
Three comparable rates are $500 on exposure 1, $900 on exposure 2, and $1,000 on exposure 5. Let $m=\$850$, $K=2$, and target exposure be 40. Calculate the credible rate and target ultimate loss. (Answer: credible rate $900.00 per exposure unit; ultimate loss $36,000.00)
A hypothetical book has a weighted experience rate of $1,100 based on total exposure 12. Given $m=\$950$, $\mathrm{EPV}=24{,}000$, $\mathrm{VHM}=6{,}000$, and target exposure 30, calculate the credibility parameter, credible rate, and ultimate loss. (Answer: $K=4$; credible rate $1,062.50 per exposure unit; ultimate loss $31,875.00)

What You'll Learn