Tutorials › CAS Exam 8 - Advanced Ratemaking › Integrated Bühlmann-Straub Calculation Workshop

Home  /  Tutorials  /  CAS Exam 8 / Bühlmann-Straub Credibility

CAS Exam 8 · Bühlmann-Straub Credibility · Tutorial 416 of 989

Integrated Bühlmann-Straub Calculation Workshop

Learn to move from unequal historical exposures to credibility-weighted risk estimates and projected losses, while keeping parameter estimation and exposure weighting consistent.

Intermediate 10 min read

What You'll Learn

  • Estimate the collective mean, process variance, and between-risk variance from unequal-exposure data
  • Calculate Bühlmann-Straub credibility factors from each risk’s total exposure
  • Distinguish exposure-weighted rates from unweighted averages
  • Handle partial histories without treating missing periods as zero losses
  • Project individual and aggregate losses at new exposure levels
  • Identify rounding and weighting errors that can change a final answer

From Unequal Histories to Future Losses

Bühlmann-Straub calculations combine three tasks: summarize each risk’s experience using its exposure, estimate the credibility model’s variance parameters, and use the resulting risk estimates at a future exposure level. The bookkeeping matters. A rate observed over a large exposure is more precise than the same rate observed over a small exposure, and a missing period is not a zero-loss observation.

The model below concerns rates such as loss cost per exposure unit. It does not require a particular probability distribution, interest rate, or mortality assumption: the calculation uses conditional means and variances, with all loss amounts measured on the same valuation basis. The variance structure and independence assumptions must nevertheless be suitable for the data.

Definition: For risk \(i\) in observed period \(j\), let \(X_{ij}\) be the observed rate and \(w_{ij}>0\) its exposure weight. Write \(w_i=\sum_j w_{ij}\) for the risk’s total observed exposure and \(\bar X_i=\sum_j w_{ij}X_{ij}/w_i\) for its exposure-weighted mean rate. In the Bühlmann-Straub model, a risk has an underlying mean \(\mu_i\), with \(E[X_{ij}\mid\mu_i]=\mu_i\) and \(\operatorname{Var}(X_{ij}\mid\mu_i)=\sigma^2/w_{ij}\). Across risks, \(E[\mu_i]=\mu\) and \(\operatorname{Var}(\mu_i)=\tau^2\). Thus \(\sigma^2\) is the process variance per unit of exposure and \(\tau^2\) is the variance of underlying risk means.

The standard model treats the exposures as known weights, assumes conditional independence of observations given the risk means, and assumes that a risk’s mean does not change across its observed periods. These are substantive assumptions, not bookkeeping conventions. If process variance changes by period, periods are dependent, or risk means trend over time, the basic formulas need adjustment. A period with zero exposure contributes no rate observation; a missing period is omitted rather than entered as a zero.

Parameter Estimates and the Credibility Rule

For \(I\) risks, with \(n_i\) observed periods for risk \(i\), the pooled mean weights every rate by its exposure. A standard method-of-moments estimate of process variance uses within-risk deviations. The between-risk estimate then compares variation among risk means with the amount of variation expected from process noise.

$$ \widehat{\mu}= \frac{\sum_{i=1}^{I}\sum_{j=1}^{n_i}w_{ij}X_{ij}} {\sum_{i=1}^{I}w_i}, \qquad \widehat{\sigma}^{\,2}= \frac{\sum_{i=1}^{I}\sum_{j=1}^{n_i}w_{ij}(X_{ij}-\bar X_i)^2} {\sum_{i=1}^{I}(n_i-1)}. $$

Let \(W=\sum_i w_i\) and \(Q=\sum_i w_i(\bar X_i-\widehat{\mu})^2\). The corresponding method-of-moments estimate of the between-risk variance is

$$ \widehat{\tau}^{\,2}= \frac{Q-(I-1)\widehat{\sigma}^{\,2}} {W-\dfrac{\sum_{i=1}^{I}w_i^2}{W}}. $$

The subtraction in the numerator removes the variation in risk means attributable to process noise. In applications, this estimate can be negative in a small or unrepresentative sample; a common constrained procedure replaces a negative estimate by zero. The ordinary credibility formula assumes positive process and between-risk variance estimates. Define \(K=\sigma^2/\tau^2\), or estimate it by \(\widehat K=\widehat{\sigma}^{\,2}/\widehat{\tau}^{\,2}\). Risk \(i\)’s credibility factor is \(Z_i=w_i/(w_i+K)\), and its credibility-weighted estimate is \(Z_i\bar X_i+(1-Z_i)\mu\), with estimated parameters substituted in practice.

Why the credibility factor has this form: Consider a linear estimate \(a\bar X_i+(1-a)\mu\), where \(a\) is the weight assigned to the risk’s experience. Since \(\operatorname{Var}(\bar X_i\mid\mu_i)=\sigma^2/w_i\), the mean squared error for estimating \(\mu_i\) is \((1-a)^2\tau^2+a^2\sigma^2/w_i\). Differentiating with respect to \(a\) and setting the derivative to zero gives \(a\tau^2-(1-a)\sigma^2/w_i=0\). Solving yields \(a=\tau^2/(\tau^2+\sigma^2/w_i)=w_i/(w_i+K)\). The more exposure a risk has, the larger its credibility weight.

Worked Example: Estimate the Model Parameters

Worked Example: Estimate the Model Parameters

A hypothetical portfolio has three risks, each observed in two periods, with one exposure unit in every period. Their rates are A: 8 and 12; B: 10 and 14; C: 14 and 18. Estimate the collective mean, process variance, between-risk variance, and \(K\), then calculate the credibility-weighted estimate for each risk. Assume the Bühlmann-Straub moment structure above and conditional independence; no parametric distribution is specified or needed.

1. Calculate the risk means and pooled mean. The risk means are \(\bar X_A=(8+12)/2=10\), \(\bar X_B=(10+14)/2=12\), and \(\bar X_C=(14+18)/2=16\). Total exposure is \(W=6\), so \(\widehat{\mu}=(10+12+16)/3=38/3=12.6667\).

2. Estimate process variance. For each risk the weighted squared deviations from its own mean sum to \((2)^2+(-2)^2=8\). The total within-risk sum is \(8+8+8=24\), and the denominator is \(\sum_i(n_i-1)=3\). Therefore \(\widehat{\sigma}^{\,2}=24/3=8\).

3. Estimate between-risk variance. Each risk has total exposure 2. The weighted squared deviations of the risk means from \(38/3\) sum to \(Q=2[(-8/3)^2+(-2/3)^2+(10/3)^2]=112/3\). The denominator is \(W-\sum_iw_i^2/W=6-12/6=4\), and the process correction is \((I-1)\widehat{\sigma}^{\,2}=2(8)=16\). Thus \(\widehat{\tau}^{\,2}=[112/3-16]/4=16/3\).

4. Calculate credibility and estimates. Here \(\widehat K=8/(16/3)=1.5\). Each risk has \(w_i=2\), so \(Z_i=2/(2+1.5)=4/7\). For A, \(\widehat{\mu}_A=38/3+(4/7)(10-38/3)=78/7=11.1429\). For B, \(\widehat{\mu}_B=38/3+(4/7)(12-38/3)=86/7=12.2857\). For C, \(\widehat{\mu}_C=38/3+(4/7)(16-38/3)=102/7=14.5714\).

Check. Each estimate lies between its risk’s observed mean and the collective mean, as a credibility-weighted average must. All three have the same exposure and therefore the same credibility factor.

Unbalanced Periods and Future Exposure

The formulas do not require every risk to have the same periods or the same exposure in each period. Calculate \(\bar X_i\) using that risk’s observed exposures, then use the risk’s total exposure \(w_i\) in \(Z_i\). To project expected losses, multiply the estimated rate by the future exposure. If risk \(i\) is expected to have future exposure \(e_i\), the estimated aggregate loss is \(\sum_i e_i\widehat{\mu}_i\). This is a projection of expected loss, not a prediction that actual future losses will equal that amount.

1
Keep observed data aligned.
Record each observed rate with its own exposure. Leave unobserved risk-periods out; do not code them as zero rates.
2
Calculate weighted means and model parameters.
Use exposure-weighted risk means and pooled moments, checking that the estimates correspond to the stated variance assumptions.
3
Credibility-weight each risk.
Use its total observed exposure in \(Z_i=w_i/(w_i+K)\), then combine its rate with the collective mean.
4
Project at future exposure.
Multiply each credibility estimate by its own future exposure and add unrounded amounts before rounding the total.

Worked Example: Unequal Exposures and Partial Histories

A hypothetical rating portfolio has current observed rates A: 10 at exposure 1 and 13 at exposure 2; B: 8 at exposure 0.5 and 12 at exposure 1.5; C: 16 at exposure 2 in one observed period. Assume an independently calibrated model has \(\sigma^2=8\) and \(\tau^2=16/3\), hence \(K=1.5\). Estimate each risk’s rate and project losses at future exposures 2, 1.5, and 3, respectively. Conditional independence and the same constant process-variance structure apply; no particular distribution, interest, or mortality assumption is required.

1. Calculate exposure-weighted means. For A, \(w_A=1+2=3\) and \(\bar X_A=(1(10)+2(13))/3=36/3=12\). For B, \(w_B=0.5+1.5=2\) and \(\bar X_B=(0.5(8)+1.5(12))/2=(4+18)/2=11\). For C, \(w_C=2\) and \(\bar X_C=16\).

2. Calculate the collective mean. Total observed exposure is \(3+2+2=7\). The weighted total is \(36+22+32=90\), so \(\widehat{\mu}=90/7=12.8571\). The single observed period for C is valid experience; it does not imply that C had a zero rate in any unobserved period.

3. Calculate risk credibility factors and estimates. For A, \(Z_A=3/(3+1.5)=2/3\), giving \(\widehat{\mu}_A=90/7+(2/3)(12-90/7)=86/7=12.2857\). For B, \(Z_B=2/(2+1.5)=4/7\), giving \(\widehat{\mu}_B=90/7+(4/7)(11-90/7)=578/49=11.7959\). For C, \(Z_C=2/(2+1.5)=4/7\), giving \(\widehat{\mu}_C=90/7+(4/7)(16-90/7)=718/49=14.6531\).

4. Project aggregate losses. At the stated future exposures, A contributes \(2(86/7)=172/7=24.5714\), B contributes \(1.5(578/49)=867/49=17.6939\), and C contributes \(3(718/49)=2{,}154/49=43.9592\). Their sum is \(4{,}225/49=86.2245\) loss units, or 86.22 loss units to two decimal places.

Check. A has more observed exposure than B or C and consequently has the largest credibility factor. Its estimate is closest to its own observed mean. The future total uses the separate future exposure for each risk.

Integrated Projection and Common Exam Traps

A full calculation should keep precision through the credibility factors, risk estimates, and projected amounts. Rounding a factor early can change an answer at the final cent. The following hypothetical calibration exercise also illustrates why exposure totals and observed rates must be interpreted consistently.

Worked Example: Preserve Precision Through the Aggregate Projection

A hypothetical calibration has produced \(\widehat{\sigma}^{\,2}=52\) and \(\widehat{\tau}^{\,2}=856/6.6\). The collective rate is 108. Risks A and B each have exposure 3 and observed mean rate 100; risk C has exposure 4 and observed mean rate 120. Project at future exposures 2 for A, 1.5 for B, and 3 for C. Assume the calibrated parameters apply to these risks and periods, conditional independence holds, and all rates and projected amounts use the same monetary basis. No parametric distribution, interest rate, or mortality assumption is involved.

1. Find \(K\). The exact ratio is \(K=52/(856/6.6)=343.2/856\), or \(K\approx0.4009345794\). Rounded to four decimal places this is 0.4009.

2. Calculate the credibility factors. For A and B, \(Z_3=3/(3+343.2/856)=1070/1213\approx0.8821\). For C, \(Z_4=4/(4+343.2/856)=4280/4709\approx0.9089\). The exact fractions, rather than the displayed rounded decimals, are retained for the remaining calculations.

3. Calculate the credibility estimates. For A and B, the estimate is \(108+(1070/1213)(100-108)=122444/1213\approx100.943116\), or 100.9431 to four decimal places. For C, it is \(108+(4280/4709)(120-108)=559932/4709\approx118.906774\), or 118.9068 to four decimal places.

4. Project and sum. A contributes \(2(122444/1213)\approx201.886233\); B contributes \(1.5(122444/1213)\approx151.414674\); and C contributes \(3(559932/4709)\approx356.720323\). The unrounded sum is approximately 710.021230, so the projected aggregate is $710.02 to the nearest cent.

Common mistake and correction. Substituting the rounded value \(Z_4=0.9090\) gives \(108+12(0.9090)=118.9080\), which is not the estimate obtained from the stated parameters. The exact factor is \(4280/4709\), giving 118.906774 before rounding. Retaining exact factors also gives the correct projected aggregate; rounding each intermediate result can shift the total.

Other frequent errors are averaging period rates without exposure weights, using a period’s exposure rather than the risk’s total observed exposure in \(Z_i\), and projecting every risk at the same future exposure when the problem states different exposures. A useful final check is that each credibility estimate lies between its own experience mean and the collective mean when \(0\leq Z_i\leq1\). A zero-exposure risk has \(Z_i=0\) and receives the collective estimate; zero exposure does not make its expected rate zero.

Key takeaway: The Bühlmann-Straub workflow is exposure-weighted at every stage: estimate risk and collective rates, calculate each risk’s credibility from its total observed exposure, and multiply its final rate estimate by its own future exposure. This treatment of unbalanced experience is aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.

Check Your Understanding

Work each problem with the stated model inputs and retain unrounded values until the final answer.

  1. A risk has total observed exposure 6, observed weighted mean rate 44, collective mean 50, and \(K=2\). Calculate its credibility factor and credibility estimate. (Answer: \(Z=0.75\); estimate \(45.50\).)
  2. A risk’s rates are 20 at exposure 0.5 and 28 at exposure 1.5. The collective mean is 24 and \(K=1\). Find its weighted experience mean, credibility factor, and credibility estimate. (Answer: weighted mean \(26\); \(Z=2/3\); estimate \(25.3333\).)
  3. Two risks have credibility estimates of 70 and 90. Their future exposures are 1.25 and 2, respectively. Find projected aggregate loss. (Answer: \(1.25(70)+2(90)=267.50\).)
  4. A model has \(\sigma^2=18\) and \(\tau^2=6\). A risk has exposure 4, experience mean 121, and the collective mean is 100. Calculate \(K\), \(Z\), and the credibility estimate. (Answer: \(K=3\); \(Z=4/7\); estimate \(112\).)
  5. A risk has no observed exposure, the collective mean rate is 36, and its future exposure is 2.5. Under the ordinary Bühlmann-Straub rule, what estimate and projected loss apply? (Answer: rate estimate \(36\); projected loss \(90\).)

Previous

Next

Back to Tutorials
Clinical Biostats