Why Relativity Calculations Need Diagnostic Checks
A class relativity is useful only when the data behind it are comparable. A numerical difference between two classes can reflect different underlying risk costs, but it can also arise because the classes use different exposure units, the losses and premiums cover different periods, or the calculation gives the wrong weights to the groups. Normalizing a set of factors does not repair those problems.
This tutorial focuses on diagnosing those errors. The calculations assume that losses have already been put on a consistent coverage and valuation basis, and that any needed development, trend, or premium on-level adjustments have been completed. The central question here is whether the data and calculations are internally consistent before a relativity is selected or applied.
A simple multiplicative model is \(q_j=q_b r_j\): the loss cost for class \(j\) is the reference class loss cost multiplied by a class factor. This model is a way to express relative costs, not proof that the data support a stable difference. In particular, an observed factor can be distorted by inconsistent inputs or by a change in the population being summarized.
Check the Exposure Basis Before Comparing Loss Costs
The denominator \(E_j\) must measure the same kind of exposure for every class. A policy-year, vehicle-year, payroll unit, and insured-value-year are not interchangeable merely because each is called an exposure. If the intended rating basis is vehicle-years, for example, loss costs calculated per policy-year cannot be compared directly across classes when the average number of vehicles per policy differs.
When the exposure definitions differ, convert the data to a common basis if the necessary information is available. If it is not, the resulting ratios should not be presented as comparable class relativities. A large or small factor may simply reflect the units chosen for the denominator.
Worked Example: Correcting Inconsistent Exposure Units
A hypothetical insurer is examining two vehicle classes and wants loss costs per vehicle-year. Class A reports 1,000 policy-years, with exactly two vehicles on each policy, and has losses of $120,000. Class B reports 1,500 vehicle-years and has losses of $135,000. Calculate the B-to-A relativity, and show the error from comparing the reported exposure counts without adjustment.
Method. Convert both classes to vehicle-years before calculating loss costs. The losses are assumed to cover comparable periods and coverage.
For class A, the exposure is \(1{,}000\times2=2{,}000\) vehicle-years. Its loss cost is \(\$120{,}000/2{,}000=\$60\) per vehicle-year. Class B's loss cost is \(\$135{,}000/1{,}500=\$90\) per vehicle-year.
The B-to-A relativity is \(\$90/\$60=1.50\). On this common exposure basis, class B has 1.50 times the observed loss cost of class A.
Common mistake. Dividing class A losses by its 1,000 policy-years gives $120 per reported exposure; dividing class B losses by its 1,500 vehicle-years gives $90. The resulting ratio, \(90/120=0.75\), compares different units and reverses the conclusion. Converting class A to 2,000 vehicle-years corrects the denominator. As a check, \(1.50\times\$60=\$90\), the calculated class B loss cost.
Align Loss and Premium Periods
A comparison of loss ratios must pair losses with premiums earned for the same period and on a compatible basis. If losses from one year are divided by premiums from another year, changes in rates, exposure, or business composition can create an apparent class difference. Even if the same calendar-year label appears in both columns, verify that the measures are earned over the same period and that the premium basis is appropriate for the comparison.
The same principle applies when comparing losses per exposure: the losses and exposures must represent the same experience period, subject to the intended adjustments. A ratio is not made valid simply by dividing two totals that happen to be available.
Worked Example: Finding a Period Mismatch in Class Loss Ratios
A hypothetical insurer compares classes A and B using 2024 losses. Those losses are $144,000 for A and $150,000 for B. Earned premiums for 2024 are $240,000 for A and $250,000 for B; 2023 earned premiums are $200,000 for each class. Assume the losses and premiums are otherwise on a consistent basis. Calculate the 2024 loss ratios and show how using the wrong premium period affects the comparison.
Method. A loss ratio is losses divided by earned premium, so use 2024 losses over 2024 earned premiums.
For A, the 2024 loss ratio is \(\$144{,}000/\$240{,}000=0.60\). For B, it is \(\$150{,}000/\$250{,}000=0.60\). The B-to-A ratio of these loss ratios is \(0.60/0.60=1.00\): the supplied, aligned experience indicates equal loss ratios.
Common mistake. Using 2023 premiums gives A a ratio of \(\$144{,}000/\$200{,}000=0.72\) and B a ratio of \(\$150{,}000/\$200{,}000=0.75\). The apparent B-to-A ratio is \(0.75/0.72=1.0417\), or about 1.04. That difference comes from pairing 2024 losses with 2023 premiums, not from the aligned 2024 comparison. The check is to match each numerator and denominator by period before interpreting the ratio.
Choose Weights That Match the Purpose
Weighting is not a clerical detail: it determines which population or rating-plan constraint the resulting average represents. For example, exposure weights describe an average across exposures when each exposure is the intended unit. Current-premium weights are appropriate when the aim is revenue neutrality under current premiums. These targets need not produce the same normalized factors.
Suppose raw factors are \(r_j\), with nonnegative weights \(w_j\). To set their weighted average to 1, calculate the weighted mean and divide each factor by it. This is valid only when the weights reflect the stated objective and are applied consistently.
The formula works because the weighted average after normalization is \(\sum_j w_j r'_j/\sum_j w_j =\sum_j w_j(r_j/\bar{r}_w)/\sum_j w_j =\bar{r}_w/\bar{r}_w=1\). If the weights are current class premiums \(P_j\), this condition says that the premium-weighted average of the revised factors is 1. Under the simplifying assumption that each current class premium is multiplied directly by its revised factor, total premium is unchanged. A different objective requires different weights or a different constraint.
Worked Example: Normalizing Factors to Current Premium
A hypothetical rating plan has raw class factors 0.80, 1.20, and 1.50. The respective current class premiums are $500,000, $300,000, and $200,000. Normalize the factors to keep total premium unchanged under the direct multiplier assumption.
Method. Use current premiums as weights because the stated goal is revenue neutrality under the current premium distribution.
The weighted factor total is \(500{,}000(0.80)+300{,}000(1.20)+200{,}000(1.50) =400{,}000+360{,}000+300{,}000=\$1{,}060{,}000\). Total current premium is $1,000,000, so the weighted mean factor is \(1{,}060{,}000/1{,}000{,}000=1.06\).
Divide each raw factor by 1.06: class 1 has \(0.80/1.06=0.7547\), class 2 has \(1.20/1.06=1.1321\), and class 3 has \(1.50/1.06=1.4151\), rounded to four decimals. The premium-weighted check is \(500{,}000(0.7547)+300{,}000(1.1321)+200{,}000(1.4151)\), approximately \(\$377{,}350+\$339{,}630+\$283{,}020=\$1{,}000{,}000\). The slight displayed difference that can arise from rounding is avoided by using unrounded factors in the calculation.
Common mistake. The unweighted mean of the raw factors is \((0.80+1.20+1.50)/3=1.1667\). Dividing by that mean makes the simple average 1, but does not preserve current premium: the resulting factors are approximately 0.6857, 1.0286, and 1.2857, whose current-premium-weighted average is about 0.9086. The correction is to normalize using the weights that match the stated revenue-neutrality goal.
Do Not Normalize Twice Without a New Objective
Normalization imposes a constraint on the factor scale. Once factors have been normalized to have a current-premium-weighted mean of 1, applying a second normalization using equal weights generally changes that mean and removes the original revenue-neutrality condition. This is not a harmless formatting step.
A second rescaling is appropriate only if a clearly stated new constraint is intended and its consequences are understood. For example, anchoring one reference class at 1 is a different constraint from setting a weighted average to 1. It may be necessary to estimate or select factors under one constraint and then rescale for another, but the analyst must identify which constraint the final factors satisfy. Do not claim that a vector simultaneously meets incompatible constraints unless the calculations show that it does.
Separate Class Experience From Class Mix
An aggregate loss cost depends both on the loss costs within the classes and on the proportion of exposure in each class. Consequently, an aggregate change does not by itself show that the underlying class relativities changed. If the distribution of exposures shifts toward a higher-cost class, the overall loss cost can rise even when every class's own loss cost remains constant.
For class loss costs \(q_j\) and exposures \(E_j\), the aggregate loss cost is \(\sum_j E_j q_j/\sum_j E_j\). The exposure shares \(E_j/\sum_j E_j\) are part of that result. Comparing aggregate experience without checking those shares can turn a change in mix into a misleading conclusion about class factors.
Worked Example: Detecting a Class-Mix Effect
A hypothetical insurer has two classes with unchanged loss costs: $100 per exposure for the lower-cost class and $300 per exposure for the higher-cost class. In period 1 there are 90 lower-cost and 10 higher-cost exposures. In period 2 there are 50 of each. Calculate the aggregate loss cost in each period and determine whether the class relativity changed.
Method. Calculate each period's total losses and divide by its total exposure; also calculate the within-class ratio.
In period 1, losses are \(90(\$100)+10(\$300)=\$9{,}000+\$3{,}000=\$12{,}000\). Total exposure is 100, so aggregate loss cost is \(\$12{,}000/100=\$120\).
In period 2, losses are \(50(\$100)+50(\$300)=\$5{,}000+\$15{,}000=\$20{,}000\). Total exposure is also 100, so aggregate loss cost is \(\$20{,}000/100=\$200\). The aggregate has increased by \(\$200/\$120=1.6667\), or about 66.67%.
But the higher-to-lower class relativity is \(\$300/\$100=3.00\) in both periods. The aggregate increase comes from the higher-cost class's exposure share rising from 10% to 50%, not from a change in either class loss cost. The check is to compare class-level costs and class mix separately.
A Practical Diagnostic Sequence
Before interpreting a class factor, apply the checks in order. Each step addresses a different source of distortion; passing one check does not guarantee that the others are satisfied.
State the target: a loss-cost relativity, a loss-ratio comparison, or a normalized rating factor. These are not interchangeable measures.
Check coverage, valuation, experience periods, earned status, and exposure units. Confirm that each numerator corresponds to its denominator.
Calculate the class loss costs or loss ratios directly from their underlying totals. Investigate any result that cannot be reproduced.
Name the weights and explain the target they represent, such as exposure average or current-premium neutrality.
Verify the final weighted average under the chosen constraint, and separate within-class changes from changes in class shares.
The main exam trap is to treat a plausible-looking factor as self-validating. Ratios can be calculated correctly from inconsistent inputs, and normalized factors can satisfy a mathematical average while representing the wrong weighting objective. Always show the denominator, period, and weights that produced the result.
Check Your Understanding
For each problem, identify the relevant diagnostic and calculate the requested result.
- A hypothetical class has $84,000 of losses over 700 policy-years, with three units per policy. A comparison class has $108,000 of losses over 900 unit-years. Find the second class's relativity to the first per unit-year. (Answer: \(120/40=3.00\).)
- Class X has 2025 losses of $96,000 and 2025 earned premium of $160,000. Class Y has 2025 losses of $108,000 and 2025 earned premium of $180,000. If someone instead uses 2024 premiums of $150,000 for X and $180,000 for Y, what are the correct 2025 loss ratios and the incorrect Y-to-X ratio from the mismatched premiums? (Answer: Correct ratios are 0.6000 for both; the mismatched Y-to-X ratio is \(0.6000/0.6400=0.9375\).)
- Raw factors for two classes are 0.90 and 1.30; current premiums are $400,000 and $600,000. Find the current-premium-neutral normalized factors, rounded to four decimals. (Answer: Weighted mean \(=(360{,}000+780{,}000)/1{,}000{,}000=1.14\); factors \(0.7895\) and \(1.1404\).)
- Two class loss costs remain $80 and $240. Exposure counts change from 80 and 20 to 40 and 60. Calculate the aggregate loss cost in each mix and state whether the within-class relativity changed. (Answer: First \(=(80(80)+20(240))/100=\$112\); second \(=(40(80)+60(240))/100=\$176\); relativity remains \(240/80=3.00\).)
- Factors 0.80 and 1.40 have current premiums of $250,000 and $750,000. They are first normalized to a current-premium-weighted average of 1. If they are then normalized again to an equal-weighted average of 1, what happens to the current-premium-weighted average? (Answer: First weighted mean \(=1.25\), giving factors 0.64 and 1.12; their equal mean is 0.88, so second factors are approximately 0.7273 and 1.2727, and the current-premium-weighted average becomes approximately 1.1364, no longer 1.)