Why Frequency Calculations Go Wrong
A claim frequency is easy to calculate only after the underlying quantities have been defined. A ratio such as “claims divided by policies” can be misleading if the numerator counts claim transactions while the denominator represents policy-years, or if claims cover a different observation period from the exposure. Even when each segment’s rate is calculated correctly, averaging segment rates without regard to their exposure can produce the wrong portfolio rate.
The central discipline is to make four items agree: what counts as a claim, which period is observed, what exposure unit is used, and how segment rates are combined. The methods here are aligned with the CAS Exam 5 (Basic Ratemaking and Estimated Claim Liabilities) syllabus.
For example, a rate may be stated as claims per earned policy-year, claims per vehicle-year, or claims per policy-month. These are different measures. A numerical rate without its denominator unit is incomplete: \(0.20\) claims per vehicle-year is not interchangeable with \(0.20\) claims per policy-year.
Four Checks Before Dividing
Decide whether the count represents distinct claims, reported notices, paid claims, or another defined event. Apply that rule consistently to every record and segment.
Identify the dates covered by the claim count. Claims observed over six months cannot be compared directly with exposure accrued over a full year without an appropriate conversion or alignment.
Use earned exposure for the same risks and time period represented by the claims. If policies have different terms, sum their earned exposure rather than treating every policy as one full year.
When combining segments with different exposure amounts, divide total claims by total exposure. Do not give a small segment the same weight as a large segment unless the intended statistic is explicitly the unweighted mean of segment rates.
Claim Definition and Record Counting
Operational data often contain rows that are not distinct claim events. A single claim can generate multiple notices, transactions, or status changes. Conversely, a reporting rule may intentionally count multiple events associated with one policy. The analyst must therefore identify the unit being counted before using a database total.
A practical check is to ask what one unit in the numerator represents. If it is a distinct reported claim, repeated transactions for that claim should not silently become additional claims. If the study instead counts claim transactions, then that is the measure, but the result must not be described as frequency of distinct claims. Likewise, claim counts and exposure should refer to the same coverage population; claims from one coverage cannot be divided by exposure from a broader, different coverage without an explicit reason.
Worked Example: Removing Non-Claim Records From a Count
A hypothetical insurer reviews one year of records for a defined coverage. Its system contains 32 claim-related rows. Four rows are duplicate notices for claims already listed, and three other rows are reopening transactions for existing claims. These categories do not overlap. The portfolio earned 125 policy-years. Calculate the frequency of distinct reported claims per policy-year.
Method: define and reconcile the numerator first. The 32 rows include four duplicates and three reopening transactions. Neither category represents a new distinct claim event under the stated definition.
The denominator is already expressed in policy-years and covers the same year and coverage. Therefore,
Check. At \(0.20\) claims per policy-year, 125 policy-years correspond to \(0.20(125)=25\) claims, matching the reconciled count. A common mistake is to calculate \(32/125=0.256\), treating all claim-related rows as new claims. That answers a different question—the rate of claim-related rows—not the frequency of distinct claims.
Matching Periods and Exposure Units
Frequency is a rate over time only when its exposure unit includes time. If a calculation uses policy-months, the result is claims per policy-month. To express it per policy-year, convert the exposure consistently: 12 policy-months equal one policy-year. Annualizing a shorter observation period is appropriate only when the exposure and claim counts cover that same period and the rate is intended to represent a per-year rate. It does not, by itself, establish that future frequency will equal the annualized value.
Worked Example: Converting Policy-Month Frequency to an Annual Rate
A hypothetical product has 28 claims during the first six months of a study and 24 during the next six months. The corresponding earned exposures are 420 and 360 policy-months. Calculate the frequency per policy-month and per policy-year, using the full year of experience.
Method: combine counts and exposure over the same period. The total claim count is \(28+24=52\). Total exposure is \(420+360=780\) policy-months. Thus the observed monthly frequency is
There are 12 policy-months in a policy-year, so the annual rate is
An equivalent check converts the denominator: \(780/12=65\) policy-years, and \(52/65=0.80\). The two calculations agree. Rounded to four decimal places, the monthly rate is \(0.0667\); the annual rate is \(0.8000\).
Common mistake. Dividing 52 by 780 and reporting \(0.0667\) claims per policy-year is a unit error: the denominator was policy-months. Another error is to annualize each six-month count while retaining an unannualized denominator. Always write the exposure unit beside the denominator and carry it through the calculation.
Policy counts are not automatically earned policy-years. A policy active for half of a year contributes one-half policy-year if policy-years are the selected exposure unit. If a risk has several covered units, such as multiple vehicles, a vehicle-year denominator may be appropriate for a measure defined as claims per vehicle-year. The selected denominator must reflect the question and the population underlying the claim count.
Worked Example: Converting Policy Counts to Earned Policy-Years
A hypothetical program has 120 policies during a calendar year. Each policy is in force for an average of nine months of that year, and the program records 27 claims during the same policy exposure period. Calculate claims per earned policy-year. Also calculate the rate that would result from using the number of policies as the denominator, and explain the difference.
Method: calculate earned exposure before the rate. Each policy contributes \(9/12=0.75\) policy-year. Total exposure is
The frequency per earned policy-year is
Using the policy count directly gives \(27/120=0.225\) claims per policy in the stated calendar-year listing. That quotient is not the same annual exposure rate because the 120 policies did not each contribute a full year. The check is that \(0.30(90)=27\), whereas applying \(0.225\) to 90 earned policy-years would imply only \(20.25\) claims. The common error is to treat every policy listed during the year as one full policy-year.
Pooling Subgroup Frequencies Correctly
Suppose the portfolio is divided into segments. For segment \(j\), let \(N_j\) be its claim count, \(E_j\) its exposure, and \(f_j=N_j/E_j\) its observed frequency. The pooled frequency is found from the totals. It is not generally the simple arithmetic mean of the segment rates.
This identity is exact for observed counts and exposures. It does not require the segment rates to be equal. If a separate statistical model is used to estimate an underlying rate, its assumptions must be considered in addition to this arithmetic. In particular, an exposure-weighted observed rate is a descriptive pooled rate; it does not eliminate sampling variation in small segments.
Worked Example: Correcting an Unweighted Segment Average
A hypothetical insurer observes 18 claims over 300 policy-years in one territory and 42 claims over 100 policy-years in another. Both counts use the same claim definition and calendar period. Calculate the territory rates and the pooled rate. Compare the pooled rate with the unweighted average.
Method: calculate each rate, then pool counts and exposure. The first territory’s frequency is \(18/300=0.06\) claims per policy-year. The second territory’s frequency is \(42/100=0.42\) claims per policy-year.
The correct portfolio rate is based on 60 claims and 400 policy-years:
The exposure-weighted calculation confirms this result. The first territory has \(300/400=0.75\) of exposure and the second has \(100/400=0.25\):
The unweighted average is \((0.06+0.42)/2=0.24\), which is not the portfolio frequency. It gives the 100-policy-year territory the same weight as the 300-policy-year territory. As a check, applying the pooled rate to total exposure gives \(0.15(400)=60\) claims. The unweighted average would imply \(0.24(400)=96\), inconsistent with the observed total.
Common Errors and a Reliable Final Check
A frequency calculation should be reported with its claim definition, time period, and exposure unit. These details make it possible to determine whether two rates are comparable and whether a change in the result reflects claims, exposure, or a change in measurement.
| Error | Why it fails | Correction |
|---|---|---|
| Counting duplicate records or status transactions as new claims | The numerator no longer represents the stated claim event. | Reconcile records to the chosen event definition before dividing. |
| Dividing a partial-period count by full-year exposure, or the reverse | The numerator and denominator refer to different periods. | Align the dates, or convert the exposure and state the rate’s time unit. |
| Using policies as though each earned a full year | Policy count ignores differences in time in force. | Sum earned policy-years, or another explicitly chosen exposure unit. |
| Averaging segment rates equally | Segments with very different exposure receive the same weight. | Divide total claims by total exposure, or use the equivalent exposure-weighted rate. |
A useful final audit is to multiply the reported rate by its exposure denominator. The result should recover the claim count, subject only to stated rounding. Also verify that the rate’s unit survives every conversion. A calculation that passes both checks is less likely to contain a hidden period or denominator mismatch.
Check Your Understanding
For each problem, identify the relevant unit before calculating.
- A hypothetical portfolio has 16 distinct claims during a period with 80 earned policy-years. What is the frequency per policy-year? (Answer: \(0.20\) claims per policy-year)
- A study records 54 claims over 900 policy-months, with counts and exposure covering the same nine-month period. What are the observed frequencies per policy-month and per policy-year? (Answer: \(0.0600\) claims per policy-month and \(0.7200\) claims per policy-year)
- Two segments have 25 claims over 500 policy-years and 35 claims over 100 policy-years. What is the pooled frequency, and what is the unweighted mean of their segment frequencies? (Answer: pooled frequency \(0.1000\) claims per policy-year; unweighted mean \(0.2000\) claims per policy-year)
- A hypothetical book has 200 policies earning a full year each and 100 policies earning one-half year each. It records 75 claims over those same exposures. What is the frequency per earned policy-year? (Answer: \(0.3000\) claims per policy-year)
- A claims extract has 32 rows, including four duplicate entries and three reopening transactions; the categories do not overlap, and neither type is a new claim under the stated definition. Exposure is 125 policy-years. What is distinct-claim frequency? (Answer: \(0.2000\) claims per policy-year)