What Equity Diagnostics Ask
An experience-rating plan can produce different premiums for different risks. That fact alone does not establish either unfairness or sound risk classification. The relevant question is whether risks with similar expected costs receive comparable treatment, and whether material premium differences correspond to credible differences in predicted costs.
Equity is not the same as identical premiums, identical modifications, or identical realized losses. A risk expected to generate more covered loss may appropriately receive a higher premium. Conversely, an aggregate comparison can conceal an issue: two groups may have the same average modification even though risks with similar predictions are treated differently within those groups. Useful diagnostics therefore compare risks on a consistent basis, first within predictive-risk segments and then across the distribution of modifications.
A meaningful comparison requires comparable units. Align the evaluation period, coverage, limits, deductibles, exposure basis, and premium basis. Where exposure differs, compare per exposure unit or use a consistent expected-cost measure. Use predictions developed without relying on the loss outcomes being evaluated when possible; otherwise, apparent agreement can reflect reuse of the same information rather than genuine predictive performance.
Compare Premiums With Predicted Cost
Suppose a portfolio is divided into segments \(g\), such as bands of predicted loss cost. Let \(\hat L_g\) be the aggregate predicted loss for segment \(g\), \(P_g\) its charged premium, and \(q\) the target loss ratio used to convert expected loss into premium. The following calculation assumes the same target loss ratio applies to every segment, predictions are calibrated for the evaluation period, and premium and loss are measured on consistent coverage and exposure bases.
Here \(P_g^*\) is the indicated premium and \(A_g\) is charged premium divided by indicated premium. A value of 1 means the segment’s charged premium equals the indication under the stated assumptions. A value above 1 means charged premium is higher than that indication; a value below 1 means it is lower. If target loss ratios differ because of documented differences in expenses or other rating provisions, compare premiums against the appropriate segment-specific target rather than imposing one common \(q\).
The observed-to-expected loss ratio provides a related, but different, diagnostic. Let \(L_g\) be observed loss for the evaluation period. Then \(L_g/\hat L_g\) compares actual with predicted loss. It helps assess prediction calibration, but it does not by itself show whether premiums were equitable: realized losses can vary around expected losses, especially in small segments.
The distinction matters. A segment can have a high observed-to-expected ratio because its experience happened to be adverse, while its charged premium was consistent with its expected cost. Conversely, a segment can have a low observed-to-expected ratio and still have a premium that was not supported by the prediction used to rate it. Review pricing indications and prediction calibration as separate questions.
Worked Example: Comparing Indicated and Charged Premium by Risk Segment
A hypothetical insurer evaluates two segments, each with 100 exposure units. The predicted aggregate losses are $800,000 for Segment A and $1,200,000 for Segment B. Charged premiums are $1,000,000 and $1,500,000, respectively. The plan uses a common target loss ratio of 0.80. Determine whether the premium difference is supported by predicted costs, and calculate observed-to-expected ratios if observed losses are $840,000 and $1,140,000.
Method and assumptions. Use the common target loss ratio to calculate indicated premium. Assume the predictions and observed losses use the same coverage, exposure basis, and evaluation period. The comparison tests aggregate indications; it does not establish that every individual risk within a segment is treated consistently.
For Segment A, indicated premium is:
Its charged-to-indicated ratio is \(\$1{,}000{,}000/\$1{,}000{,}000=1.00\). For Segment B:
The predicted loss ratio between B and A is \(\$1{,}200{,}000/\$800{,}000=1.50\). The charged premium ratio is \(\$1{,}500{,}000/\$1{,}000{,}000=1.50\). Thus, under the common target loss ratio, the premium difference matches the predicted cost difference.
Now calculate observed-to-expected ratios. Segment A has \(\$840{,}000/\$800{,}000=1.05\). Segment B has \(\$1{,}140{,}000/\$1{,}200{,}000=0.95\). Actual losses were 5% above prediction in A and 5% below prediction in B. These results do not change the indicated-premium calculation; they are evidence about calibration over this period and should be interpreted with the segments’ uncertainty and volume in mind.
Check and mistake avoided. Both charged-to-indicated ratios equal 1.00, and the charged premium ratio equals the predicted loss ratio. Comparing the premium amounts alone and concluding that B was treated inequitably because it paid more would ignore the predicted cost difference.
Distributional Diagnostics Within Comparable Risks
Segment averages can hide substantial differences among risks that have similar predicted costs. To examine that possibility, compare modification distributions within reasonably narrow predictive-risk bands. Useful summaries include the mean, median, selected quantiles, range, and the share of risks receiving modifications above or below a specified threshold. A plot or empirical cumulative distribution can reveal a broad shift, unusual tails, or a subgroup that an average conceals.
The comparison should use the same definition of modification and the same starting premium basis. If one group’s modification is measured relative to a different base rate, the numerical distributions are not directly comparable. Also check whether caps, floors, eligibility rules, or different plan terms truncate one group’s modifications; a distributional difference can reflect those rules rather than a difference in predicted cost.
Worked Example: Comparing Modification Distributions at the Same Predicted Risk
Two hypothetical underwriting territories contain risks in the same predicted-cost band. Each territory has four risks with the same base premium basis and the same predicted cost per exposure. The modifications in Territory North are 0.80, 0.90, 1.00, and 1.10. Those in Territory South are 1.00, 1.10, 1.20, and 1.30. Compare their means and medians, then state what the result does and does not establish.
Method and assumptions. Compare like with like because the risks are stipulated to have the same predicted-cost band, base premium basis, and plan terms. Treat each listed risk equally for these summaries; if exposures differ, an exposure-weighted comparison would also be needed.
North’s mean modification is:
Its two middle values are 0.90 and 1.00, so its median is \((0.90+1.00)/2=0.95\). South’s mean is:
Its median is \((1.10+1.20)/2=1.15\). Both the mean and median are 0.20 higher in South. In fact, each South modification is 0.20 greater than the corresponding North value when the values are ordered.
The consistent shift is a reason to investigate why comparable risks receive different modifications. It does not, by itself, prove that territory is the cause or that the difference is unjustified. Confirm that the predicted-cost band captures relevant risk differences and check for omitted rating factors, exposure or coverage mismatches, plan eligibility, and data-quality issues.
Check and mistake avoided. The distributions are not equivalent merely because they have the same number of risks. Reporting only the counts would miss the shift. Equally, claiming that the territory difference is definitively unfair would go beyond what these summaries show.
Standardize Before Comparing Groups
Two broad groups can have different average modifications simply because they contain different proportions of low- and high-predicted-cost risks. To distinguish composition from within-segment differences, calculate each group’s result within risk strata, then apply the same set of weights to each group. This direct standardization answers a specific question: what would the average modification be in each group if both had the chosen common risk mix?
where \(\bar m_{gs}\) is group \(g\)’s average modification in stratum \(s\), and \(w_s\) is the common weight assigned to that stratum. Choose weights that reflect a clear reference population, such as the combined portfolio, and use the same weights for every group. Standardization does not remove differences within a stratum; it makes the composition used in the group comparison consistent.
Worked Example: Separating Risk Mix From Within-Band Differences
Two hypothetical groups have low- and high-predicted-cost risks. In Group A, 90 exposure units are low risk with average modification 0.80, and 10 are high risk with average modification 1.40. In Group B, 20 units are low risk with average modification 0.80, and 80 are high risk with average modification 1.40. Calculate each raw average and then standardize both groups to a 50% low-risk, 50% high-risk mix.
Method and assumptions. Calculate exposure-weighted averages because the supplied counts are exposure units. Assume the two risk bands are sufficiently comparable across groups and the reported modifications use the same base. The chosen 50/50 mix is a diagnostic reference, not a claim about the actual portfolio.
Group A has 100 total exposure units. Its raw average is:
Group B also has 100 units. Its raw average is:
The raw difference is \(1.28-0.86=0.42\). But within each risk band, the groups have identical average modifications. Apply the common 50/50 weights:
After standardization, the difference is \(1.10-1.10=0\). The raw difference arose from Group B having more high-risk exposure, not from a modification difference within either stated band. This supports a composition explanation for this comparison, subject to whether the bands adequately capture predicted risk.
Check and mistake avoided. The calculation verifies both raw averages and both standardized averages. Comparing 0.86 directly with 1.28 without accounting for the different risk mixes would incorrectly attribute the entire difference to group treatment.
Why Standardization Works
The raw average for group \(g\) can be written as a weighted sum of its within-stratum averages, using that group’s own stratum proportions \(w_{gs}\). If those proportions differ between groups, the raw comparison combines two effects: differences within strata and differences in composition. Standardization replaces group-specific weights with the same \(w_s\), so a remaining difference must come from the within-stratum averages under the selected partition and weights.
This identity follows directly from averaging: each stratum contributes its average multiplied by the share of exposure assigned to that stratum. Substituting identical weights into both standardized averages and subtracting leaves only the weighted within-stratum differences. The method is informative only if strata are defined consistently and contain comparable risks; overly broad bands can leave important risk differences uncontrolled, while very narrow bands can leave too little data for stable estimates.
Interpret Results With Care
Equity diagnostics are strongest when several views tell a consistent story: premium-to-indication ratios by segment, observed-to-expected results, modification distributions within predicted-risk bands, and standardized group comparisons. None should be interpreted without checking data credibility and rating mechanics. Review exposure completeness, loss development and trend, changes in coverage terms, claim handling, and the effects of plan caps or floors. A small segment’s extreme ratio may be a signal to investigate, but it may also reflect ordinary volatility.
A useful operational sequence is:
Use consistent periods, coverage, exposure units, premium bases, and modification definitions.
Group risks by predicted cost using a measure appropriate to the coverage being evaluated.
Assess charged premium against indicated premium and compare modification summaries within strata.
Apply common risk-mix weights, identify differences that remain, and test plausible data or plan explanations before drawing conclusions.
Common exam trap: comparing raw averages as if the groups had the same risk mix. In the preceding example, the incorrect conclusion would be: “Group B’s average modification is 0.42 higher, so comparable Group B risks receive higher modifications.” That calculation uses \(1.28-0.86=0.42\), but it does not compare comparable risks. The correction is to compare within the low- and high-risk bands or standardize both groups to common weights. Both within-band differences are zero, and both standardized averages are 1.10. The raw difference is explained by the different mix in this example, not by a within-band modification difference.
When a difference remains after careful standardization, it is a finding to explain—not an automatic verdict. The predictive model may omit relevant cost variation, the strata may be too coarse, or plan provisions may produce a legitimate but poorly understood distinction. A defensible conclusion connects the observed premium difference to a documented predictive-cost difference and makes clear what the available diagnostics cannot establish. This approach is aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.
Check Your Understanding
Work each problem using the stated assumptions and bases.
- A hypothetical segment has predicted losses of $640,000, charged premium of $800,000, and a target loss ratio of 0.80. Calculate indicated premium and charged-to-indicated ratio. (Answer: indicated premium $800,000; ratio 1.00.)
- A segment has observed losses of $990,000 and predicted losses of $1,100,000. Calculate its observed-to-expected ratio and interpret it. (Answer: 0.90; observed losses were 10% below prediction.)
- Two groups in the same risk band have modifications 0.90, 1.00, 1.10 and 1.00, 1.10, 1.20, respectively. Calculate each mean and their difference. What diagnostic question does the difference raise? (Answer: means 1.00 and 1.10; difference 0.10, prompting investigation of why comparable risks receive different modifications.)
- Group C has 75 low-risk and 25 high-risk exposure units, with modifications 0.70 and 1.30. Group D has 25 low-risk and 75 high-risk units, with the same respective modifications. Calculate both raw averages and both standardized averages using 50% weights for each risk band. (Answer: raw averages 0.85 for C and 1.15 for D; standardized averages 1.00 for both.)
- A comparison finds that one territory has a higher raw average modification, but also has a larger share of high-predicted-cost risks. State the next diagnostic calculation and one condition needed for its interpretation. (Answer: standardize both territories to the same risk-band weights and compare within-band results; the bands must represent comparable risks across territories.)