Diagnostics for Rating Relativities

Why Relativities Need Diagnostics

A rating relativity compares the expected cost of one class with the cost of a reference class. A large observed relativity can be meaningful, but it can also arise from a small exposure count, an unusual mix of risks, or inconsistent data. Conversely, a modest difference can matter to many policyholders when the class has substantial premium volume. Diagnostics help distinguish these situations before a factor is treated as persuasive evidence.

The calculations below use hypothetical insurance data. Losses and exposures must be measured on a consistent basis: for example, losses for the same coverage and experience period divided by earned exposures for that period. A loss cost is an observed average, not automatically a prospective expected cost. Adjustments for changes in level or coverage must be addressed before comparing experience if they are relevant.

Definition: For class $j$, let $E_j$ be its exposure count and $L_j$ its aggregate losses on a consistent basis. Its observed loss cost is $LC_j=L_j/E_j$. Relative to a reference class $0$, the raw loss-cost relativity is $R_j=LC_j/LC_0$. If $P_j$ is the current premium for class $j$, its premium weight is $w_j=P_j/\sum_k P_k$, so $\sum_j w_j=1$.

These are diagnostic quantities, not a complete rate indication. In particular, a raw relativity does not account by itself for expenses, profit provisions, credibility, or prospective changes. The earlier material on credibility and rate indications supplies methods for those broader decisions; here the emphasis is on checking the data and the implications of class factors.

Three Questions to Ask of a Class Relativity

Is the comparison on the same basis?
Check that class losses, exposure definitions, coverage, and experience periods are comparable. An apparent class difference can be an artifact if one class has a different loss basis or exposure definition.

How much experience supports it?
Review exposure counts and, where available, claim counts and claim-size variation. A sharp factor based on little experience is less stable than the same factor supported by substantial data.

What does it do to the portfolio?
Use current premium weights to measure the aggregate effect of class factors. Premium weights describe portfolio impact; they are not measures of statistical precision.

A useful first table includes exposures, aggregate losses, loss costs, raw relativities, current premiums, and premium weights. Keep the numerators and denominators visible. A class with a high loss cost but only a few exposures deserves investigation, not automatic adoption of its raw factor. Also check whether the observed difference is driven by one unusually large claim or by a broadly higher frequency or severity pattern.

Assessing Uncertainty With a Simple Model

One transparent, though restrictive, model illustrates why exposure counts matter. Suppose each exposure independently produces a claim count with a Poisson distribution, and every claim has the same fixed payment $c$. For $E$ exposures, the total number of claims $K$ is Poisson with mean $E\lambda$, where $\lambda$ is the claim frequency per exposure. Aggregate loss is $cK$, so observed loss cost is $cK/E$.

The Poisson mean and variance are both $E\lambda$. Therefore, conditional on the stated model, the loss-cost expectation and variance are:

E\left[\frac{cK}{E}\right]=c\lambda, \qquad \operatorname{Var}\left(\frac{cK}{E}\right) =\frac{c^2}{E^2}\operatorname{Var}(K) =\frac{c^2\lambda}{E}.

Replacing the unknown frequency $\lambda$ by its estimate $K/E$ gives an estimated standard error $c\sqrt{K}/E$. The denominator shows the key diagnostic: with similar frequency and claim payment, more exposures generally mean less sampling variation in loss cost. But the fixed-payment assumption excludes claim severity variation, and the Poisson assumption may not fit clustered or heterogeneous risks. If individual exposure-level losses are available, their observed dispersion is often more informative than this simplified calculation.

Worked Example: A High Relativity With Sparse Experience

Two hypothetical classes have the same fixed claim payment of $2,000 per claim. Class A has 1,000 exposures and 50 claims; Class B has 100 exposures and 8 claims. Assume independent Poisson claim counts within each class, with a constant frequency within each class. Calculate the loss costs, Class B’s raw relativity to Class A, and an approximate uncertainty check for the difference.

Method. Calculate aggregate losses and loss costs first. Then use the Poisson fixed-payment standard error. This check avoids treating a point estimate from 100 exposures as if it were as stable as one based on 1,000 exposures.

For Class A, aggregate losses are $50\times \$2{,}000=\$100{,}000$. Its loss cost is $\$100{,}000/1{,}000=\$100$ per exposure. Its estimated standard error is:

\frac{\$2{,}000\sqrt{50}}{1{,}000} =\$14.1421.

For Class B, aggregate losses are $8\times \$2{,}000=\$16{,}000$. Its loss cost is $\$16{,}000/100=\$160$ per exposure. Its estimated standard error is:

\frac{\$2{,}000\sqrt{8}}{100} =\$56.5685.

The raw relativity is $\$160/\$100=1.60$, or 60% above Class A. The estimated standard error of the difference between independent class loss costs is:

\sqrt{14.1421^2+56.5685^2} =\sqrt{200+3{,}200} =\$58.3095.

The observed difference is $\$160-\$100=\$60$, about $60/58.3095=1.03$ estimated standard errors. Under this simplified model, the difference is not compelling evidence of a true difference at a conventional two-sided 5% level: the approximate 95% interval for the difference is $\$60\pm1.96(\$58.3095)$, or approximately $(-\$54.29,\$174.29)$. This is a model-based diagnostic, not a definitive test. The check does not include severity variation, and it depends on the assumptions stated above.

Check. The computed aggregate losses divided by exposures return $100 and $160, and their ratio is 1.60. The large estimated error for Class B reflects its much smaller exposure count. The common mistake is to report 1.60 as a stable indicated factor without considering that uncertainty.

Premium Weights and Portfolio Impact

Suppose proposed class factors are $q_j$, expressed on a common scale. Their premium-weighted average factor is $\sum_j w_jq_j$. If the goal is to keep the overall premium level unchanged, the factors must satisfy $\sum_j w_jq_j=1$. When raw factors do not satisfy this condition, dividing each by their premium-weighted average puts them on a revenue-neutral scale. This balancing operation changes the scale, not the relative ratios among classes.

Worked Example: Balancing Factors With Premium Weights

A hypothetical insurer has three classes with current premiums of $600,000, $300,000, and $100,000. Its raw factors are 0.90, 1.20, and 1.80. Find the premium weights, calculate the weighted average raw factor, and balance the factors to an overall average of 1.

Method. Divide each current premium by total premium to obtain its weight. Apply the weights to the factors, then divide every raw factor by the resulting weighted average. This avoids using an unweighted average when the classes contribute different amounts of premium.

Total current premium is $\$600{,}000+\$300{,}000+\$100{,}000=\$1{,}000{,}000$. The weights are $0.60$, $0.30$, and $0.10$, which sum to 1. The weighted average raw factor is:

$$ 0.60(0.90)+0.30(1.20)+0.10(1.80) =0.54+0.36+0.18=1.08. $$

Divide each raw factor by 1.08. The balanced factors are $0.90/1.08=0.8333$, $1.20/1.08=1.1111$, and $1.80/1.08=1.6667$, rounded to four decimal places. Checking the weighted average gives:

0.60(0.8333)+0.30(1.1111)+0.10(1.6667) \approx 0.5000+0.3333+0.1667=1.0000.

The check confirms approximate revenue neutrality at the stated rounding. The relative ratio between the second and first factors remains $1.20/0.90=1.3333$ before balancing and $1.1111/0.8333\approx1.3333$ afterward.

Check. The premiums sum to $1,000,000, and the original weighted average is 1.08. A common mistake is to use the simple mean, $(0.90+1.20+1.80)/3=1.30$, as the balancing factor. That would ignore the premium mix; the correct factor is 1.08 because 60% of current premium is in the first class.

Separate Class Experience From Risk Mix

A class’s aggregate loss cost can be counterintuitive because classes may contain different proportions of lower- and higher-cost risks. When relevant subdivisions are available, compare experience within those subdivisions as well as in aggregate. A standardized comparison applies the same chosen mix to each class. This is a diagnostic: it can reveal composition effects, but it does not by itself establish which factors should be adopted.

Worked Example: A Crude Difference Caused by Exposure Mix

A hypothetical portfolio has two classes, A and B, and two risk groups. The loss costs in the lower-cost group are $100 per exposure in either class; those in the higher-cost group are $300 per exposure in either class. Class A has 900 lower-cost and 100 higher-cost exposures. Class B has 100 lower-cost and 900 higher-cost exposures. Calculate the crude class loss costs and compare them with loss costs standardized to a 50%/50% mix.

Method. First aggregate losses using each class’s actual exposure mix. Then calculate both standardized averages using the same mix. This exposes the trap of attributing a difference between aggregate averages to the class label when the risk-group composition differs.

For Class A, losses are $900(\$100)=\$90{,}000$ in the lower-cost group and $100(\$300)=\$30{,}000$ in the higher-cost group. Total losses are $\$120{,}000$ over 1,000 exposures, so its crude loss cost is $\$120$.

For Class B, losses are $100(\$100)=\$10{,}000$ and $900(\$300)=\$270{,}000$. Total losses are $\$280{,}000$ over 1,000 exposures, so its crude loss cost is $\$280$. The crude relativity of B to A is $\$280/\$120=2.3333$.

At a common 50%/50% mix, each class has standardized loss cost $0.50(\$100)+0.50(\$300)=\$200$. The standardized relativity is $\$200/\$200=1.0000$. In this constructed example, the crude difference is entirely explained by the differing proportions of the two risk groups; neither group has a within-group class difference.

Check. Both classes have 1,000 exposures, but their group distributions differ sharply. Recomputing the standardized average from either class’s group loss costs gives $200. The crude factor alone would conceal the composition effect.

Common Diagnostic Traps

Confusing premium weight with exposure count. Premium weight measures a class’s contribution to current premium; exposure count helps describe the amount of experience. A high premium weight does not automatically mean low uncertainty, and a high exposure count does not automatically mean a large portfolio impact.
Comparing inconsistent data. Different coverage, loss development, exposure definitions, or periods can create apparent differences unrelated to rating class. State the basis before interpreting a ratio.
Treating a raw factor as proof. A loss-cost ratio is an observed statistic. Sampling variation, large claims, and risk mix can all affect it. A simple uncertainty calculation only supports conclusions under its assumptions.
Balancing with equal class weights. If the objective is a premium-weighted overall factor of 1, use current premium shares, not the number of classes or their unweighted mean.
Overinterpreting a lack of statistical evidence. A wide uncertainty range may reflect sparse data; it does not prove that class costs are equal. Conversely, a statistically detectable difference does not alone determine whether a factor is practical or appropriate.

Key takeaway: Diagnose a relativity by checking its data basis, exposure support, and portfolio effect. Use uncertainty calculations only under explicit assumptions, and use premium weights for aggregate impact rather than as a substitute for evidence about stability.

Check Your Understanding

For each hypothetical problem, calculate the requested diagnostic and check that the units and weighting basis match the question.

A class has 250 earned exposures and $45,000 of losses; its reference class has a loss cost of $150 per exposure. Find the first class’s loss cost and raw relativity. (Answer: $180 per exposure; relativity $1.2000$.)
Assume a fixed claim payment of $1,000 and a Poisson claim count. A class has 400 exposures and 18 claims. Find its loss cost and the plug-in standard error under the fixed-payment model. (Answer: $45.00 per exposure; standard error $=\$1{,}000\sqrt{18}/400=\$10.61$.)
Three classes have premium weights 0.50, 0.30, and 0.20, and raw factors 0.80, 1.10, and 1.50. Find the weighted average factor and the balanced factors, rounded to four decimals. (Answer: weighted average $1.0300$; balanced factors $0.7767,\ 1.0680,\ 1.4563$.)
Two classes have identical loss costs by risk group: $80 for Group 1 and $200 for Group 2. Class A has 75% Group 1 exposures and Class B has 25% Group 1 exposures. Find each crude loss cost and the common loss cost standardized to a 50%/50% mix. (Answer: Class A $110; Class B $170; standardized loss cost $140 for each.)
Under the independent Poisson fixed-payment model, two classes have estimated loss costs of $90 and $120, with estimated standard errors of $8 and $15, respectively. Calculate the difference and its estimated standard error, then state whether the difference is more or less than two standard errors. (Answer: difference $30; standard error $\sqrt{8^2+15^2}=17$; the difference is less than two standard errors, since $30<34$.)

What You'll Learn