Credibility-Weighted Classification Relativities

Why Credible Loss Costs Need a Common Base

A classification plan compares expected costs among groups of insureds. A class with a high observed loss cost may genuinely be higher risk, but its experience may also be based on limited data. Credibility addresses that uncertainty by blending the class observation with a portfolio complement. The resulting credible loss costs can then be expressed as relativities, making the differences among classes easier to use and interpret.

The two steps must be kept distinct. First, credibility determines a cost estimate for each class. Second, a common base converts those estimates into relative factors. The base is a stated dollar loss cost, not a separate adjustment to each class. If the objective is for relativities to average to 1.00 across a specified prospective mix, the base must be the weighted average credible loss cost for that mix.

Definition: For class $j$, let $L_j$ be its observed loss cost, calculated as losses divided by the corresponding exposure on a consistent basis. Let $L_0$ be the collective complement loss cost and $Z_j$, between 0 and 1, be the class's credibility weight. The credibility-weighted loss cost is $$ \widehat{L}_j=Z_jL_j+(1-Z_j)L_0. $$ For a positive common base $B$, the class relativity is $$ R_j=\frac{\widehat{L}_j}{B}. $$

The observed class loss costs, complement, and base must use the same currency, coverage, valuation period, and exposure definition. For example, a class's losses per vehicle-year should not be divided by a complement stated per policy unless the exposure measures have first been made consistent. Credibility weights should come from the applicable credibility model or be given; they are not selected simply to make factors look reasonable.

Credibility Weights and the Blend

Under the Bühlmann model, a risk has an unobserved parameter $\Theta$. Conditional on $\Theta$, the equal-exposure observations $X_1,\ldots,X_n$ are independent and identically distributed, with conditional mean $m(\Theta)=E[X_k\mid\Theta]$ and conditional variance $v(\Theta)=\operatorname{Var}(X_k\mid\Theta)$. Define the expected process variance and variance of hypothetical means by $$ \mathrm{EPV}=E[v(\Theta)],\qquad \mathrm{VHM}=\operatorname{Var}(m(\Theta)). $$ When both are finite and $\mathrm{VHM}>0$, the Bühlmann credibility factor is $Z=n/(n+K)$, where $K=\mathrm{EPV}/\mathrm{VHM}$. The Bühlmann linear credibility estimate of the next observation's expected value is $Z\bar X+(1-Z)\mu$, where $\bar X$ is the observed mean and $\mu=E[X_k]$ is the collective mean.

Thus, in a class application with the class observation represented by $L_j$ and the collective estimate by $L_0$, the credibility blend is $Z_jL_j+(1-Z_j)L_0$. The blend assigns more weight to class experience as its credibility increases, while retaining the complement's weight $1-Z_j$. A particular parametric distribution, such as Poisson, is not required for this Bühlmann formulation; it relies on the stated conditional moment assumptions. The simple $n/(n+K)$ expression assumes equal-exposure observations. Do not apply it mechanically to unequal exposures.

Worked Example: Blending Two Class Loss Costs

A hypothetical insurer reviews two vehicle classes. The collective complement is $L_0=\$800$ per vehicle-year. Class A has observed loss cost $L_A=\$1{,}000$ and credibility $Z_A=0.60$; Class B has $L_B=\$700$ and $Z_B=0.40$. For this illustration, assume the equal-exposure Bühlmann conditions described above and regard the credibility weights as already estimated. Equivalently, the weight ratios could arise from $n=3,K=2$ for A and $n=2,K=3$ for B. No interest or mortality assumption is involved, and no parametric loss distribution is assumed. Calculate credible costs and relativities using the complement as the common base.

Method. Apply the credibility blend to each class, then divide each result by the same base, $B=\$800$.

Class A. Its complement weight is $1-0.60=0.40$. Therefore, $$ \widehat{L}_A=0.60(\$1{,}000)+0.40(\$800)=\$600+\$320=\$920. $$ The relativity is $R_A=\$920/\$800=1.15$.

Class B. Its complement weight is $1-0.40=0.60$. Therefore, $$ \widehat{L}_B=0.40(\$700)+0.60(\$800)=\$280+\$480=\$760. $$ The relativity is $R_B=\$760/\$800=0.95$.

Check and trap avoided. Both credible costs lie between their class observation and the complement, as they should. Class A is 15% above the complement base, and Class B is 5% below it. The common mistake is to use the observed costs directly, producing $1{,}000/800=1.25$ and $700/800=0.875$, and thereby omit credibility. No rounding is needed. These factors use the complement as base; they are not guaranteed to average to 1 across the classes.

Choosing the Common Base

There is no single denominator that is correct for every purpose. If the intended comparison is with the collective portfolio complement, set $B=L_0$. This answers, “How does this class's credible cost compare with the collective cost?” If factors must average to 1.00 over a specified prospective class mix, use that mix's weighted average credible cost as the base.

Let $w_j$ be the prospective share of exposure in class $j$, with $w_j\geq 0$ and $\sum_j w_j=1$. Then define $$ B_{\mathrm{mix}}=\sum_j w_j\widehat{L}_j,\qquad R_j=\frac{\widehat{L}_j}{B_{\mathrm{mix}}}. $$ These weights describe the mix for which the factors are to be normalized. Historical exposures can be a starting point, but they need not represent the intended prospective mix.

Establish comparable inputs.
Calculate each class's observed loss cost using the same coverage, valuation, and exposure basis as the collective complement.

Apply the class credibility weight.
Calculate $Z_jL_j+(1-Z_j)L_0$ separately for every class.

State the normalization objective.
Use the collective complement as base for comparison to the portfolio, or calculate the weighted mean credible cost for a specified prospective mix.

Divide and check.
Divide every credible cost by the same base. For mix-normalized factors, verify that $\sum_j w_jR_j=1$.

The normalization check follows directly from the definition. If $B_{\mathrm{mix}}=\sum_jw_j\widehat{L}_j$, then $$ \sum_jw_jR_j =\sum_jw_j\frac{\widehat{L}_j}{B_{\mathrm{mix}}} =\frac{\sum_jw_j\widehat{L}_j}{B_{\mathrm{mix}}} =\frac{B_{\mathrm{mix}}}{B_{\mathrm{mix}}} =1. $$ The result holds because every class uses the same denominator and the weights sum to 1. If the denominator is instead $L_0$, the weighted average of the resulting factors is $\sum_jw_j\widehat{L}_j/L_0$, which need not equal 1.

Worked Example: Calculating Loss Costs Before Relativities

A hypothetical insurer has three classes. Class A incurred $\$84{,}000$ over 120 exposure units, Class B incurred $\$45{,}000$ over 90 units, and Class C incurred $\$36{,}000$ over 60 units. The collective complement is $\$500$ per exposure unit. The credibility weights are $0.50, 0.60,$ and $0.25$, respectively. Assume each class has an equal-exposure Bühlmann model with independent conditional observations and finite conditional moments; these weights correspond, for example, to $n=2,K=2$, $n=3,K=2$, and $n=1,K=3$. The distribution is not specified parametrically; no interest or mortality assumption is involved. Calculate credible costs and relativities normalized to the observed exposure mix. Round relativities to four decimal places.

Step 1: Calculate observed loss costs. For A, $\$84{,}000/120=\$700$. For B, $\$45{,}000/90=\$500$. For C, $\$36{,}000/60=\$600$.

Step 2: Blend each observation with the complement. For A, $\widehat{L}_A=0.50(\$700)+0.50(\$500)=\$350+\$250=\$600$. For B, $\widehat{L}_B=0.60(\$500)+0.40(\$500)=\$300+\$200=\$500$. For C, $\widehat{L}_C=0.25(\$600)+0.75(\$500)=\$150+\$375=\$525$.

Step 3: Determine the exposure weights and common base. Total exposure is $120+90+60=270$. The weights are $w_A=120/270=4/9$, $w_B=90/270=1/3$, and $w_C=60/270=2/9$. Thus, $$ B_{\mathrm{mix}} =\frac{120(\$600)+90(\$500)+60(\$525)}{270} =\frac{\$72{,}000+\$45{,}000+\$31{,}500}{270} =\frac{\$148{,}500}{270} =\$550. $$

Step 4: Calculate the relativities. $R_A=\$600/\$550=1.0909$, $R_B=\$500/\$550=0.9091$, and $R_C=\$525/\$550=0.9545$, rounded to four decimal places.

Check and trap avoided. The exposure-weighted mean credible cost is $\$550$, the chosen base. Before rounding, the weighted relativity check is $\{120(600/550)+90(500/550)+60(525/550)\}/270=1$. A common error is to average the three class costs equally even though the stated mix has unequal exposures. This would produce a different base and would not normalize factors to the specified exposure mix.

Normalization to a Prospective Mix

The mix used in normalization matters even when the credible costs do not change. A rating plan may target a prospective distribution that differs from the observed distribution, for example because its expected class shares have changed. In that case, calculate $B_{\mathrm{mix}}$ using the prospective weights. Do not silently retain historical exposure weights and describe the result as normalized to the future mix.

A normalization base is a scaling choice, not a new credibility complement. Changing the base changes all reported relativities by the same multiplicative factor but does not change their ordering or the ratios between classes. For any two classes $j$ and $k$, $$ \frac{R_j}{R_k} =\frac{\widehat{L}_j/B}{\widehat{L}_k/B} =\frac{\widehat{L}_j}{\widehat{L}_k}. $$ This also provides a useful check: a common base must not alter a class-to-class cost ratio.

Worked Example: Credible Costs With a New Prospective Mix

A hypothetical property program has three classes with observed loss costs of $\$1{,}300$, $\$900$, and $\$700$. Their credibility weights are $0.75$, $0.50$, and $0.25$, and the portfolio complement is $\$1{,}000$. Assume the equal-exposure Bühlmann conditions, with finite conditional moments and independent observations; the weights correspond to $n=3,K=1$, $n=2,K=2$, and $n=1,K=3$. No parametric distribution, interest rate, or mortality assumption is required. The prospective exposure shares are $0.20,0.50,$ and $0.30$. Calculate mix-normalized relativities. Round to four decimal places.

Step 1: Calculate credible loss costs. For Class A, $0.75(\$1{,}300)+0.25(\$1{,}000)=\$975+\$250=\$1{,}225$. For Class B, $0.50(\$900)+0.50(\$1{,}000)=\$450+\$500=\$950$. For Class C, $0.25(\$700)+0.75(\$1{,}000)=\$175+\$750=\$925$.

Step 2: Calculate the prospective-mix base. $$ B_{\mathrm{mix}} =0.20(\$1{,}225)+0.50(\$950)+0.30(\$925) =\$245+\$475+\$277.50 =\$997.50. $$

Step 3: Divide by the one common base. $R_A=\$1{,}225/\$997.50=1.2281$, $R_B=\$950/\$997.50=0.9524$, and $R_C=\$925/\$997.50=0.9273$, rounded to four decimal places.

Check and trap avoided. Before rounding, $0.20R_A+0.50R_B+0.30R_C=1$ by the weighted-base identity. If the complement were used as denominator instead, the factors would be $1.225,0.950,$ and $0.925$; their prospective weighted average is $0.20(1.225)+0.50(0.950)+0.30(0.925)=0.9975$, not 1.00. The complement is a valid comparison base, but it is not the mix-normalized base in this example.

Interpretation and Common Exam Traps

A relativity of $1.10$ means the credible loss cost is 10% above the stated common base. It does not, by itself, mean that the class's observed loss cost is 10% above the base; credibility may have moved the observed cost toward the complement. Likewise, a factor below 1.00 can still correspond to a credible cost above the complement if the chosen common base is lower than the complement.

Using raw losses in place of loss costs. Classes with different exposures cannot be compared by total losses alone. Calculate losses per comparable exposure unit first.
Applying credibility in the wrong direction. The class observation receives weight $Z_j$; the complement receives $1-Z_j$. The weights must sum to 1.
Using different denominators by class. A set of classification relativities requires one explicitly stated common base.
Assuming complement-based factors average to 1. This is only guaranteed if the chosen base equals the weighted average credible cost for the same mix.
Using the wrong mix for normalization. The weights in the base calculation must describe the exposure mix named in the objective.
Applying an equal-exposure credibility expression to unequal observations. The formula $n/(n+K)$ has equal-exposure assumptions; use an appropriate credibility method when exposures differ.

Key takeaway: First calculate each class's credible loss cost by blending its observed cost with the collective complement. Then divide every credible cost by the same stated base. Use the complement for comparison to the collective level; use the weighted average credible cost for a specified mix when the factors must average to 1.00.

Check Your Understanding

Solve each problem using the stated base or mix. Monetary amounts are hypothetical.

A class has observed loss cost $\$900$, credibility $0.30$, and complement $\$800$. Calculate its credible cost and its relativity using the complement as base. (Answer: credible cost $\$830$; relativity $1.0375$)
A class has $\$96{,}000$ of losses over 160 exposure units, credibility $0.50$, and complement $\$520$. Calculate its observed loss cost, credible loss cost, and relativity using a $\$500$ base. (Answer: observed cost $\$600$; credible cost $\$560$; relativity $1.12$)
Two classes have observed loss costs $\$1{,}000$ and $\$500$, with credibility weights $0.40$ and $0.60$, respectively. The complement is $\$700$, and prospective exposure shares are 0.25 and 0.75. Find both credible costs, the mix base, and the normalized relativities. (Answer: credible costs $\$820$ and $\$580$; base $\$640$; relativities $1.2813$ and $0.9063$)
Three classes have observed loss costs $\$900,\$600,\$1{,}200$, credibility weights $0.50,0.25,0.75$, and a common complement of $\$800$. Their prospective shares are $0.20,0.50,0.30$. Calculate credible costs and relativities normalized to that mix. (Answer: credible costs $\$850,\$750,\$1{,}100$; base $\$875$; relativities $0.9714,0.8571,1.2571$)
For the classes in Question 4, calculate the relativities using the complement as denominator instead. What is their weighted average under the stated mix, and why is it not 1.00? (Answer: relativities $1.0625,0.9375,1.3750$; weighted average $1.09375$, because the complement $\$800$ is not the mix-weighted credible cost of $\$875$)

What You'll Learn