Evaluation Targets for Experience Rating

Three Questions, Three Evaluation Targets

An experience rating plan can be evaluated for different purposes. Does a modification help predict a risk’s future losses? Does the premium vary across risks in a way that reflects their expected losses? Does a proposed classification divide risks into groups with meaningfully different expected loss costs? These questions are related, but they are not interchangeable. Each calls for a defined observation period, exposure basis, and loss measure.

A useful evaluation begins by fixing the information available at the rating date, the period being rated, and the outcome to be measured. For example, losses that occur after a modification is calculated can serve as a future outcome; losses already used to calculate that modification cannot independently validate its prospective performance. The loss basis must also match the intended target: paid losses, developed ultimate losses, capped losses, and claim counts measure different things.

Definition: Prospective loss prediction evaluates whether information available at a rating date, including an experience modification, predicts losses in a later period. Premium differentiation evaluates whether premiums or premium relativities distinguish risks in a manner consistent with their expected loss costs. Risk classification evaluates whether a classification system assigns risks to groups that differ in expected loss cost, after making appropriate comparisons of exposure and other relevant characteristics.

The three targets can use some of the same records, but they ask different questions. Prediction concerns a forecast and a later outcome. Premium differentiation concerns how the charged premium varies relative to risk. Classification concerns whether the groups themselves separate risks. None of these definitions, on its own, establishes that the overall premium level is sufficient for all costs.

Specify the Evaluation Design Before Calculating

For an experience rating evaluation, distinguish at least three time periods. The experience period contains losses and exposures used to calculate the rating information. The rating date is when the modification or classification information is determined. The outcome period is the later period whose losses are used to evaluate the target. For a prospective prediction study, the outcome period must follow the rating date; the experience period and outcome period should not overlap in a way that allows the outcome losses to influence the prediction being assessed.

“Later” refers to when losses occur, not necessarily when claims are reported or settled. A mature evaluation may need time for claims from the outcome period to develop. If ultimate losses are the target, compare developed ultimate losses on a consistent valuation basis rather than comparing one risk’s near-final paid losses with another risk’s immature incurred losses. Adjustments for trend, coverage changes, or other differences should be made when needed to make the comparison meaningful.

State the question.
Specify whether the target is a future loss forecast, premium differentiation, or separation of risks into classes.

Fix the information date and periods.
Record what was known at the rating date, which experience informed the rating, and which later or comparison period supplies the outcome.

Choose the loss and exposure basis.
Define whether the outcome is claim count, severity, aggregate loss, or loss per exposure unit, and make the coverage, valuation, and limitation basis consistent.

Interpret only the target measured.
A result about prediction, premium variation, or class separation does not automatically answer either of the other two questions.

Define the Outcome and the Denominator

A loss measure should reflect the purpose of the evaluation. Claim count can be useful when the question concerns frequency; average cost per claim can describe severity; aggregate losses combine frequency and severity; and loss cost per exposure combines aggregate loss with an exposure denominator. For exposure $E$ and total loss $L$, the observed loss cost is $L/E$, provided $L$ and $E$ refer to compatible coverage, time, and units.

\text{Observed loss cost}=\frac{\text{losses on the defined basis}}{\text{exposure on the matching basis}}

For instance, a policy-year exposure denominator should not be paired with losses from a substantially different time period without an adjustment or a clear justification. If an evaluation uses a per-claim limit, the loss outcome must apply that limit consistently; if it uses uncapped losses, the result answers a different question. Paid, incurred, and ultimate losses are also not interchangeable labels. State the selected measure and use it consistently.

It is often helpful to separate a prediction into an expected loss cost per exposure unit and an exposure amount. If the expected loss cost is $r$ and the future exposure is $E$, the predicted aggregate loss is $rE$. A plan can appear to predict total dollars well merely because it has the correct exposure estimate, even if its loss cost differentiation is poor. Conversely, a reasonable loss cost estimate can yield an inaccurate total when the exposure forecast is wrong. Make clear which quantity the evaluation addresses.

Prospective Loss Prediction

For prospective prediction, calculate the indicated prediction using information available at the rating date, then compare it with an outcome observed later on the same defined basis. A useful target might be future ultimate losses per payroll unit, per vehicle-year, or per another relevant exposure unit. The specific outcome should match the intended use of the rating estimate. A prediction of limited losses should not be judged against unlimited losses without an adjustment or an explicit reason.

The comparison requires more than matching a forecast to a total. Record the outcome period, the exposure realized during that period, and the loss development status. If the rating estimate is an expected loss cost, compare it with an outcome loss cost; if it is a predicted aggregate for a stated exposure, compare it with aggregate losses over that period and exposure. This defines the prediction target without confusing it with the question of how the charged premium is set.

Worked Example: Setting a Prospective Prediction Target

A hypothetical service contractor has 100 risk-years in a future evaluation period. At the rating date, its experience rating estimate is a predicted loss cost of $1,200 per risk-year. The later outcome period produces 118 claims, each with an ultimate loss of $1,000. Determine the predicted aggregate loss, the observed aggregate loss, and the observed loss cost. State the comparison being made.

Assumptions. All amounts are undiscounted dollars, with effective annual interest rate $i=0$; no mortality assumption applies. For a stochastic frequency description, claim count is Poisson with mean $1.2$ per risk-year, so the count over 100 risk-years has mean $\lambda=120$. Claim severity is fixed at $1,000, with no limit or trend adjustment. The supplied 118 claims are the observed outcome, not a replacement for the stated expected count.

Method. This is a prospective loss prediction comparison. Multiply the predicted loss cost by the future exposure, then calculate the later observed loss total and divide by its matching exposure.

Predicted aggregate: $100 \times \$1{,}200=\$120{,}000$.

Observed aggregate: $118 \times \$1{,}000=\$118{,}000$.

Observed loss cost: $\$118{,}000/100=\$1{,}180$ per risk-year.

The specified prediction is $1,200 per risk-year and the observed outcome is $1,180 per risk-year, on the stated basis. This example defines a forecast-to-outcome comparison; one observed period alone does not establish general performance. The common mistake is comparing the $120,000 forecast with the $118,000 outcome while omitting exposure. Here the exposures happen to be equal to 100 risk-years; if they differed, aggregate dollars alone would not make a like-for-like loss-cost comparison.

Check: The observed count is 118, which is a possible outcome under a Poisson count with mean 120. Multiplying 118 claims by $1,000 gives $118,000, and dividing by 100 gives $1,180 per risk-year.

Premium Differentiation Is Not a Premium Adequacy Conclusion

Premium differentiation asks whether the premium varies appropriately across risks with different expected loss costs. A common comparison is between a premium relativity and a loss-cost relativity, calculated on consistent exposure and coverage bases. For example, one group may have an expected loss cost 1.5 times another group’s. A premium relativity of 1.5 would reflect that difference in the relative loss-cost component, while a relativity of 1.2 would reflect a smaller premium distinction.

That comparison is not, by itself, an assessment of whether total premiums cover expected losses and other required costs. Premiums may include expenses, risk loads, taxes, and other components. The target here is the relative variation by risk, not whether the overall level is adequate. Keep the base premium and the component being compared explicit.

Worked Example: Comparing Premium and Loss-Cost Relativities

A hypothetical insurer assigns two groups the following prospective estimates and annual premiums per exposure unit. Group M has an expected loss cost of $800 and a premium of $1,000 per unit. Group N has an expected loss cost of $1,200 and a premium of $1,350 per unit. Calculate the loss-cost relativity and premium relativity of N to M, then describe what the comparison does and does not answer.

Assumptions. Amounts are undiscounted, $i=0$, and there is no mortality assumption. For the loss estimate, claim counts per exposure unit are Poisson: $\lambda_M=0.8$ and $\lambda_N=1.2$. Each claim has fixed severity $1,000, so the expected loss costs are $0.8 \times \$1{,}000=\$800$ and $1.2 \times \$1{,}000=\$1{,}200$. No claim limit is applied.

Method. Divide each group’s amount by the corresponding amount for Group M. This compares relativities rather than total dollars, so unequal group exposure volumes do not enter these per-unit calculations.

Loss-cost relativity: $\$1{,}200/\$800=1.50$. The indicated expected loss cost for N is 1.50 times M’s.

Premium relativity: $\$1{,}350/\$1{,}000=1.35$. The charged premium per unit for N is 1.35 times M’s.

The premium varies by less than the stated expected loss-cost relativity: 1.35 versus 1.50. This describes the degree of relative differentiation on the stated bases. It does not establish whether either group’s total premium is adequate, because the comparison does not specify expense provisions, loads, or the overall rate level. The common mistake is to call 1.35 “35% inadequate.” It is a premium relativity, not a measure of total premium adequacy.

Check: Both ratios use N divided by M. The loss-cost ratio is 1.50, and the premium ratio is 1.35; reversing the order would answer a different comparison.

Risk Classification Requires Comparable Groups

Risk classification evaluates whether a classification characteristic separates risks by expected loss cost. A crude difference between class totals may reflect different exposure volumes or different mixes of underlying risk characteristics rather than a genuine class distinction. Comparing loss costs per exposure helps address volume, but it does not automatically control for differences in the kinds of risks assigned to each class.

When an important underlying characteristic has a different distribution across classes, compare classes using a common distribution or another appropriate adjustment. The choice of exposure unit and loss basis remains essential. A class comparison based on loss per payroll unit, for example, answers a different question from one based on loss per vehicle-year. Also distinguish the classification feature being evaluated from other risk factors that may explain the observed losses.

Key distinction: A class can have a higher crude loss cost because its members have a different underlying exposure mix. To assess whether the class label itself differentiates risk, compare class loss costs at a common mix or control for the relevant differences. State the standardization basis; the result applies to that basis.

Worked Example: Standardizing a Class Comparison

A hypothetical carrier compares two proposed classes, A and B, across low- and high-hazard operations. The exposure and ultimate losses are: A has 80 low-hazard units with $8,000 of losses and 20 high-hazard units with $6,000; B has 20 low-hazard units with $3,000 of losses and 80 high-hazard units with $24,000. Calculate each class’s crude loss cost, then its loss cost standardized to a mix of 50% low- and 50% high-hazard exposure.

Assumptions. Amounts are undiscounted dollars, $i=0$, and there is no mortality assumption. For a stochastic description, claim counts in each class-and-hazard cell are Poisson, with expected count equal to exposure multiplied by the cell’s stated frequency. Severity is fixed at $1,000. The observed counts are therefore A-low 8, A-high 6, B-low 3, and B-high 24; no limits or trend adjustments apply.

Method. First calculate losses divided by exposure within each cell. Then apply the same 50/50 exposure weights to both classes. This standardization removes the different low/high exposure mix from the comparison.

Class A crude rate: total losses are $\$8{,}000+\$6{,}000=\$14{,}000$, and total exposure is $80+20=100$. Thus, $\$14{,}000/100=\$140$ per unit.

Class B crude rate: total losses are $\$3{,}000+\$24{,}000=\$27{,}000$, and total exposure is $20+80=100$. Thus, $\$27{,}000/100=\$270$ per unit.

For Class A, the low-hazard rate is $\$8{,}000/80=\$100$, and the high-hazard rate is $\$6{,}000/20=\$300$. Its standardized rate is $0.50(\$100)+0.50(\$300)=\$200$ per unit.

For Class B, the low-hazard rate is $\$3{,}000/20=\$150$, and the high-hazard rate is $\$24{,}000/80=\$300$. Its standardized rate is $0.50(\$150)+0.50(\$300)=\$225$ per unit.

On the stated common mix, B’s standardized loss cost is $\$225-\$200=\$25$ per unit higher than A’s. The crude difference was $\$270-\$140=\$130$ per unit, in part because B had more high-hazard exposure. The standardized comparison is the relevant one for this defined classification question; it does not prove that class assignment caused the difference. The common mistake is treating the crude rates as a clean class effect despite the very different hazard mixes.

Check: Each class is evaluated with weights summing to 1.00. The standardized values are $200 and $225, and applying the same weights makes the comparison use the same 50/50 mix.

Why the Denominator and Standardization Matter

The basic loss-cost calculation follows directly from aggregation. Suppose exposures are divided among groups $j$, with exposure $E_j$ and expected loss cost $r_j$ for each group. Expected aggregate loss is the sum of the group expected losses, $\sum_j E_jr_j$. Dividing that total by the total exposure gives an exposure-weighted average rate:

\frac{\sum_j E_jr_j}{\sum_j E_j} = \sum_j \left(\frac{E_j}{\sum_k E_k}\right)r_j

The weights $E_j/\sum_k E_k$ add to 1. Therefore, a crude aggregate loss cost changes when the exposure mix changes, even if every within-group rate $r_j$ stays fixed. A standardized rate instead replaces the observed weights with the same chosen weights for each class. This is why the crude and standardized comparisons in the example differ, and why the evaluation question must specify whether it concerns the actual portfolio mix or a common mix.

Common Evaluation Traps

Reusing rating-period losses as the outcome. If losses help determine the modification, they cannot serve as an independent later outcome without a design that addresses this overlap.
Mixing loss bases. Comparing capped predicted losses with uncapped actual losses, or immature paid losses with developed ultimate losses, changes the target mid-comparison.
Comparing totals when exposure differs. Large aggregate losses may simply reflect more exposure. Use a compatible exposure denominator when the target is loss cost.
Calling a premium relativity an adequacy result. Relative premium variation does not determine whether the overall premium level covers all intended costs.
Confusing a class mix with a class effect. Different underlying exposure composition can produce different crude rates even when the within-stratum rates are similar.
Ignoring the rating date. Information that was unavailable when the rating was set cannot be used as if it were part of the original prediction.

Key takeaway: Define the target before calculating. A prospective prediction needs a later outcome; premium differentiation compares relative premium variation with relative expected loss costs; risk classification compares groups on a stated, comparable basis. In every case, specify the period, exposure denominator, loss measure, and any adjustments.

Check Your Understanding

For each problem, identify the target or calculate the requested quantity using the stated basis.

A risk is rated using experience through December 31. Its projected loss cost is $900 per vehicle-year. The next calendar year has 40 vehicle-years and $34,200 of developed ultimate losses. Calculate the observed loss cost and state whether the losses are temporally eligible as a prospective outcome. (Answer: $855 per vehicle-year; yes, because the outcome year follows the rating date.)
Class R has an expected loss cost of $500 per unit and premium of $700 per unit. Class S has an expected loss cost of $750 and premium of $945. Calculate S-to-R loss-cost and premium relativities. (Answer: 1.50 and 1.35, respectively.)
A classification study compares classes using 30 and 70 exposure units, with losses of $6,000 and $21,000. Calculate each crude loss cost and explain why the totals alone are not a fair comparison. (Answer: $200 and $300 per unit; totals alone ignore the different exposure volumes.)
Two classes have low- and high-hazard loss costs of $80 and $240 for Class C, and $120 and $280 for Class D. Standardize both to 25% low-hazard and 75% high-hazard exposure. (Answer: Class C, $200; Class D, $240 per unit.)
A modification is calculated from losses in 2022–2024, and an analyst proposes using those same losses to assess future prediction. What is the principal design issue, and what outcome period would be more suitable? (Answer: the outcome overlaps the information used to set the modification, so the assessment is not an independent prospective comparison; use a later period, such as 2025, with a consistent developed loss basis.)

What You'll Learn