Risk Classification Under Actuarial Standards

Risk Classification as an Actuarial Judgment

A rating variable can look useful in a model and still be a poor classification choice. It may show little relationship to expected costs, be difficult to measure at the time a policy is rated, or be applied differently to otherwise similar risks. A sound review therefore asks more than whether a variable improves a statistical fit.

ASOP 12, Risk Classification (for All Practice Areas), provides actuarial guidance for evaluating risk classification systems. The questions in this tutorial focus on three practical tests: whether a proposed characteristic is related to expected costs, whether it can be used in practice, and whether its definitions and application are consistent. These principles are aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus. An actuarial standard does not replace applicable law, regulation, or the actuary’s professional judgment.

Definition: A risk classification system groups risks using one or more characteristics and uses those groups in setting rates or other terms. A rating characteristic is an attribute of a risk used to distinguish expected outcomes among groups. In the examples below, expected claim cost per exposure is used as an actuarial measure of those outcomes; expenses and other rate components are excluded unless stated.

Test the Relationship to Expected Costs

For a proposed characteristic, begin by defining its categories and the outcome being compared. For a loss-cost analysis, a basic measure is total losses divided by the corresponding exposure. If claim frequency and average claim severity are available, the same quantity can be calculated as frequency per exposure multiplied by average severity.

\text{Claim cost per exposure} = \frac{\text{total losses}}{\text{exposures}} = \frac{\text{claim count}}{\text{exposures}} \times \frac{\text{total losses}}{\text{claim count}}

This measure is meaningful only when the numerator and denominator correspond to comparable coverage, exposure periods, loss definitions, and cost levels. A difference in raw claim counts, for example, could reflect different exposure volumes rather than different expected costs. A fuller analysis may also need to account for trends, changes in coverage, and other relevant rating characteristics.

The underlying justification is the law of total expectation. If $C$ is a classification and $L$ is loss for an exposure, then the overall expected loss is the exposure-weighted average of the conditional expected losses in the classes. A difference between class means can therefore support distinct estimates of expected cost, provided the data and assumptions are appropriate.

E[L] = \sum_c P(C=c)E[L\mid C=c]

This identity explains why class-specific expected costs matter to a rate indication. It does not prove that a particular observed difference will persist, that the data are sufficiently credible, or that the classification is appropriate in every setting. Statistical association is evidence to evaluate, not an automatic approval rule.

Worked Example: Comparing Two Proposed Equipment Classes

A hypothetical insurer is considering whether equipment installation type is related to annual claim cost. In a historical study, Class A has 1,000 earned exposures, 80 claims, and total losses of $120,000. Class B has 500 earned exposures, 60 claims, and total losses of $120,000. Assume the loss amounts have been adjusted to a common cost level and the coverage and exposure definitions are comparable. Calculate the observed costs and relativity, then state what the results do and do not establish.

Method. Calculate claim frequency per exposure and average severity for each class, then multiply them. This checks both the exposure adjustment and the cost calculation.

Class A. Frequency is $80/1{,}000=0.08$ claims per exposure. Average severity is $$120{,}000/80=$1{,}500$. Thus the observed claim cost per exposure is $0.08\times $1{,}500=$120$. Directly, $$120{,}000/1{,}000=$120$.

Class B. Frequency is $60/500=0.12$ claims per exposure. Average severity is $$120{,}000/60=$2{,}000$. Thus the observed claim cost per exposure is $0.12\times $2{,}000=$240$. Directly, $$120{,}000/500=$240$.

The observed Class B to Class A relativity is $$240/$120=2.00$. In this hypothetical data, Class B has twice the observed claim cost per exposure. The calculation supports investigating a cost relationship; it does not by itself establish that a 2.00 prospective relativity is appropriate. The actuary would assess credibility, consistency of the underlying data, other relevant variables, and whether the pattern is expected to continue.

Check. Multiplying each class’s frequency by its severity reproduces the direct loss-per-exposure calculation in both classes. The common mistake is to compare the equal total losses and conclude that the classes have equal costs. Correcting for exposures gives $120 and $240 per exposure, not equal costs.

Assess Whether the Characteristic Is Practical

A characteristic must be usable as well as informative. Practicality concerns the complete process: whether the value can be obtained when needed, whether it can be verified, how missing or disputed values are handled, and whether the administrative effort is proportionate to the expected benefit. A theoretically useful variable may not be suitable if it is unavailable for a substantial share of risks or cannot be measured consistently at rating time.

The information source matters. A value reported by an applicant, measured by an inspection, or obtained from an external data source may have different costs, timing, and error risks. The actuary should understand what is recorded, when it is recorded, how often it is updated, and what happens when it is absent. A fallback rule should be specified rather than left to informal judgment.

Worked Example: Missing Values in a Proposed Inspection Variable

A hypothetical insurer is considering a roof-condition category for a property product. Of 600 earned exposures, a valid inspection category is available for 480: 420 are Class Good, with 42 claims and $63,000 in losses; 60 are Class Needs Repair, with 12 claims and $24,000 in losses. The remaining 120 exposures have no valid category and have 18 claims. For this example, assume the missing exposures have $27,000 in losses. Calculate the observed costs and evaluate the practical issue.

Class Good. The cost per exposure is $$63{,}000/420=$150$. Its frequency is $42/420=0.10$, and severity is $$63{,}000/42=$1{,}500$; the product is $0.10\times $1{,}500=$150$.

Class Needs Repair. The cost per exposure is $$24{,}000/60=$400$. Its frequency is $12/60=0.20$, and severity is $$24{,}000/12=$2{,}000$; the product is $0.20\times $2{,}000=$400$.

The observed ratio for Needs Repair to Good is $$400/$150=2.6667$, or approximately 2.67. The missing group’s observed cost is $$27{,}000/120=$225$ per exposure. The entire book’s cost is $($63{,}000+$24{,}000+$27{,}000)/600=$114{,}000/600=$190$ per exposure.

A valid inspection category is available for $480/600=80\%$ of exposures, leaving $120/600=20\%$ unclassified by this variable. The observed class difference may be actuarially informative, but implementation requires a workable process for the missing 20%: for example, obtaining an inspection, using a defined fallback category, or not using this characteristic for those risks. The best choice depends on the full evidence and constraints.

Check. The two classified groups have 480 exposures and $87,000 of losses; adding the missing group gives 600 exposures and $114,000. A common mistake is to omit the missing group and treat the classified-subset average of $$87{,}000/480=$181.25$ as the overall book average. The correct overall value, given the stated data, is $190 per exposure.

Require Consistent Definitions and Application

A classification is only as stable as its rules. The actuary should be able to explain what each class means, how borderline cases are assigned, which data source controls, and how changes over time are handled. The same documented rule should be applied to comparable risks in comparable circumstances. Otherwise, measured class differences may partly reflect inconsistent assignment rather than genuine differences in expected costs.

Consistency applies to both the data and the operational process. Definitions used in an experience study should correspond to the definitions used in rating. If a value is measured at policy inception in one case and at renewal in another, the actuary should examine whether that difference changes the classification. Recordkeeping, training, system implementation, and audit procedures can all affect whether a rule is applied as intended.

Worked Example: Applying a Mileage Boundary Consistently

A hypothetical commercial auto program proposes two annual-mileage classes for a vehicle: Class 1 for no more than 12,000 miles and Class 2 for more than 12,000 miles. The indicated annual premiums, excluding taxes and fees, are $420 for Class 1 and $510 for Class 2. One vehicle has an applicant-reported estimate of 11,800 miles, while a later verified record for the same rating period shows 12,400 miles. Explain the consistent classification and premium consequence.

Method. Apply the written boundary to the mileage value defined as controlling for the rating period. If the program’s rule specifies verified mileage as controlling, 12,400 exceeds 12,000, so the vehicle is Class 2 and the indicated premium is $510.

If the same policy were instead assigned using the applicant estimate, 11,800 is no more than 12,000, so it would be Class 1 at $420. The difference between the two indicated premiums is $$510-$420=$90$.

The arithmetic does not decide which data source should control. That must be established in a documented, workable rule and applied consistently to comparable policies. If one policy is classified using an estimate and another using a verified value without a defined reason, the resulting assignments may not be comparable.

Check. Under the stated condition that verified mileage controls, 12,400 assigns to Class 2 and $510. A common mistake is to choose whichever value produces the lower premium or to treat exactly 12,000 as Class 2. The stated boundary places exactly 12,000 in Class 1; 12,400 is above it.

Putting the Review Together

A disciplined review makes the evidence and implementation questions explicit. The following sequence keeps a promising statistical result from being mistaken for a complete actuarial case.

Define the characteristic and outcome.
Specify the proposed categories, the exposure basis, and the cost measure being evaluated.

Check the evidence.
Compare exposure-adjusted expected-cost measures using data appropriate to the risks and coverage being rated. Review credibility, data quality, and whether the differences are relevant prospectively.

Test practical use.
Determine whether values can be obtained, verified, and updated at the required time. Document procedures for unavailable, disputed, or out-of-range values.

Test consistent application.
Align study definitions with rating rules, resolve boundary cases, and consider controls that help ensure comparable risks receive comparable treatment under the rule.

Document the judgment.
Summarize the evidence, limitations, operational choices, and any applicable legal or regulatory requirements. A predictive relationship alone does not settle every question about use.

These checks are related but not interchangeable. Strong cost evidence cannot compensate for a classification that cannot be administered reliably. Conversely, a convenient and consistently collected characteristic does not warrant use if it lacks relevant support for expected costs. A sound actuarial recommendation explains both the evidence and the limitations.

Common Exam Traps

Comparing claim totals instead of exposure-adjusted costs. Divide losses by the appropriate exposure, or calculate frequency and severity separately.
Treating observed differences as automatically prospective. Review data quality, credibility, and whether the experience is relevant to future costs.
Ignoring missing or unverifiable values. Quantify the problem and state a workable procedure for those risks.
Using a vague category boundary. Define inclusions, exclusions, the controlling data source, and treatment of boundary values.
Confusing actuarial support with legal permission. A cost relationship is not, by itself, a determination that a proposed use complies with all applicable requirements.

Key takeaway: Evaluate a proposed rating characteristic on the evidence for expected-cost differences, its practical measurability and administration, and the consistency of its definitions and application. State the assumptions and limitations; do not treat any single favorable test as a complete justification.

Check Your Understanding

For each problem, calculate as requested and consider what the result does or does not establish.

A hypothetical product has 300 exposures in Class X, with 24 claims and $30,000 of losses, and 250 exposures in Class Y, with 25 claims and $37,500 of losses. Find the cost per exposure in each class and the Y-to-X relativity. (Answer: X = $100; Y = $150; relativity = 1.50.)
A proposed characteristic has 400 exposures, 32 claims, and $28,800 of losses in one class. Find its frequency, average severity, and cost per exposure. (Answer: frequency = 0.08; severity = $900; cost per exposure = $72.)
Of 900 hypothetical policies, 720 have a valid value for a proposed rating characteristic. What proportion is missing, and what implementation issue should be addressed before relying on that characteristic? (Answer: 20% is missing; define and assess a practical, consistently applied procedure for policies without a valid value.)
Two classes have 800 and 400 exposures, with losses of $96,000 and $72,000, respectively. Calculate each class’s loss cost per exposure and the second-to-first relativity. What additional question should be asked before treating that relativity as prospective? (Answer: $120 and $180 per exposure; relativity = 1.50; assess credibility and the relevance and consistency of the data, among other assumptions.)
A written rule assigns values below 8,000 to Class A and values of 8,000 or more to Class B. The indicated premiums are $420 and $510. One risk has a verified value of exactly 8,000, while its reported estimate is 7,900. If verified values control, identify its class and premium, and calculate the difference from using the reported estimate. (Answer: Class B at $510; the reported estimate would give Class A at $420, a $90 difference.)

What You'll Learn