Why the Observation Unit Matters
A claims summary is meaningful only when its numerator and denominator describe compatible experience. A count of claims is not itself a frequency rate: it must be related to the amount of time during which insurance coverage was actually observed. Similarly, total losses per policy-period and average loss per claim answer different questions and use different denominators.
The central task is to define the observation window and the unit being counted before calculating a summary. A policy may contribute one full year of exposure, only part of a year, or no exposure at all to a particular study window. A claim may also be included according to its occurrence date, report date, or another specified rule. These choices need to be explicit.
A policy-period record is a data unit, not automatically one year of exposure. For instance, a record covering three months contributes one quarter of a policy-year if the chosen exposure unit is a policy-year and coverage applies throughout those months. A policy can have multiple records—for example, when its coverage is split into segments—so records must be checked for overlap or duplication before their exposure is added.
Define the Window Before Counting
An observation window specifies which dates are eligible for inclusion. In an occurrence-based study, a claim is selected because its loss occurred during the window. In a report-based study, it is selected because it was reported during the window. Those rules can produce different claim sets, especially when reporting is delayed.
Exposure must refer to coverage time inside the same observation window used to select claims. If a study includes occurrences from April 1 through September 30, a policy covered for the full six months can contribute at most six months of exposure to that study. Coverage before April or after September does not add exposure to this window.
When coverage is continuous within each record and exposure is measured in policy-years, a period of \(m\) covered months contributes \(m/12\) policy-years. This conversion is appropriate only when months are the intended time measure and the record’s coverage dates truly fall within the study window. For exact dates or irregular exposure patterns, use the defined day-count or exposure convention rather than assuming every interval is a whole number of months.
Match the Denominator to the Quantity
Three summaries are often confused. Claim frequency is claims per exposure unit. Claim severity is loss per claim. Aggregate loss per exposure unit is total loss divided by exposure. The numerator and denominator in each ratio must refer to a compatible set of policies, claims, and dates.
| Quantity | Numerator | Denominator | Interpretation |
|---|---|---|---|
| Claim frequency | Number of eligible claims | Covered exposure | Claims per policy-year, or other exposure unit |
| Claim severity | Total loss for eligible claims | Number of those claims | Average loss per claim |
| Aggregate loss per exposure | Total loss for eligible claims | Covered exposure | Loss per policy-year, or other exposure unit |
The word “loss” also needs a consistent meaning. Depending on the task, it may mean amounts paid, amounts incurred, or another defined claim amount. A claim count based on reported claims should not be combined casually with a loss total that covers a different set of claims or a different development period. State the data basis before comparing or calculating.
Worked Example: Frequency From Partial Policy Periods
A hypothetical insurer studies claim occurrences from January 1 through June 30. Four policy records have coverage within that window: Record A is covered throughout January through June and has two eligible claims; Record B is covered March through June and has one; Record C is covered May through June and has none; Record D is covered January through March and has one. Find the observed frequency per policy-year.
Method. Count claims whose occurrences fall in the six-month window, then divide by covered months within that same window converted to policy-years.
Record A contributes \(6/12=0.5\) policy-year. Record B contributes \(4/12=1/3\) policy-year. Record C contributes \(2/12=1/6\) policy-year. Record D contributes \(3/12=0.25\) policy-year. Every listed coverage interval lies within the January–June window, so none contributes time outside it.
The total exposure is \(0.5+1/3+1/6+0.25=1.25\) policy-years. The claim count is \(2+1+0+1=4\). Therefore the observed frequency is \(4/1.25=3.2\) claims per policy-year.
Check. The four records contribute 15 covered months, or \(15/12=1.25\) policy-years, and the four claims are counted once each. The result uses only exposure and claims from the stated window.
Severity and Aggregate Loss Need Different Units
Severity is calculated over claims, not over policies or policy-period records. A policy with three eligible claims contributes three observations to a per-claim calculation, while a policy with no claims contributes none. In contrast, that policy still contributes its covered exposure to a frequency or aggregate-loss-per-exposure calculation.
Likewise, aggregate loss per policy-year is not the average loss per claim. The first spreads all eligible loss across covered time, including time associated with records having no claims. The second averages loss across eligible claims only.
Worked Example: Separating Frequency, Severity, and Loss per Exposure
A hypothetical portfolio has four fully observed one-year policy-periods, each covered for the full year. Across those periods, three eligible claims have losses of $600, $900, and $1,500. The fourth period has no claims. Find the claim frequency per policy-year, average loss per claim, and aggregate loss per policy-year.
Method. Use the three claims for frequency and severity, but use all four policy-years of covered exposure for frequency and aggregate loss.
Total exposure is \(4 \times 1=4\) policy-years. Total claim count is 3, so frequency is \(3/4=0.75\) claims per policy-year.
Total loss is \(\$600+\$900+\$1{,}500=\$3{,}000\). Average severity is \(\$3{,}000/3=\$1{,}000.00\) per claim. Aggregate loss per policy-year is \(\$3{,}000/4=\$750.00\) per policy-year.
Check. Frequency times severity is \(0.75 \times \$1{,}000.00=\$750.00\) per policy-year, which agrees with aggregate loss divided by exposure. The zero-claim period belongs in the exposure denominator even though it contributes no claim to severity.
A Useful Identity and Its Conditions
When all quantities use the same eligible claims and exposure base, aggregate loss per exposure can be expressed as frequency times average severity. This is an accounting identity, not an additional assumption about the claim process.
To see why it holds, let \(N\) be the number of eligible claims, let \(L_j\) be the loss attached to claim \(j\), and let \(E\) be the matching exposure. If \(N>0\) and \(E>0\), total loss is \(\sum_{j=1}^{N}L_j\). Then:
The first factor is observed claim frequency, and the second is observed average severity. The equality follows because \(N\) in the numerator of the first factor cancels the \(N\) in the denominator of the second. If there are no claims, average severity is undefined, so this product representation cannot be used, even though total loss per exposure may be zero.
The identity also fails as a comparison if the factors use inconsistent definitions. For example, multiplying frequency from a calendar-year claim count by severity from claims selected over a different period need not reproduce either period’s aggregate loss per exposure.
Observation and Reporting Windows in Practice
A claim can occur within the study window but be reported afterward. If the study’s rule is occurrence-based and the data have sufficient reporting follow-up, such a claim belongs in the selected set even though its report date falls later. If the study is report-based, it may not. Neither convention is universally correct; the selection rule must match the question and the available data.
A reporting cutoff can make an occurrence-based dataset incomplete: claims from near the end of the occurrence window may not yet have been reported. In that situation, the observed count may understate the eventual number of claims. That is a data-completeness issue, not a reason to add exposure outside the defined occurrence window. The study should identify its cutoff and the claim-selection basis.
Worked Example: A Late-Reported Claim and a Matched Window
A hypothetical insurer examines occurrences from April 1 through September 30 and has complete reporting follow-up through December 31. Record A is covered for all six months and has one eligible claim. Record B is covered June through September and has one. Record C is covered April and May and has none. Record D is covered for all six months; one claim occurred September 28 and was reported October 10. Under an occurrence-based rule, find the claim frequency and aggregate loss per policy-year if the three eligible claims have losses of $900, $1,200, and $900.
Method. Include claims by occurrence date, use covered time within April–September for exposure, and use the stated claim losses for the same three eligible claims.
Record A contributes \(6/12=0.5\) policy-year; B contributes \(4/12=1/3\); C contributes \(2/12=1/6\); and D contributes \(6/12=0.5\). Total exposure is \(0.5+1/3+1/6+0.5=1.5\) policy-years.
All three claims occurred in the observation window, including the claim reported on October 10. Thus frequency is \(3/1.5=2\) claims per policy-year. Total loss is \(\$900+\$1{,}200+\$900=\$3{,}000\), so aggregate loss per policy-year is \(\$3{,}000/1.5=\$2{,}000.00\).
Check. Average severity is \(\$3{,}000/3=\$1{,}000.00\) per claim, and frequency times severity is \(2 \times \$1{,}000.00=\$2{,}000.00\) per policy-year. This agrees with total loss divided by exposure.
Common Mistakes and a Reliable Workflow
A frequent error is to use the full policy term as exposure even when the study observes only part of it. Another is to count a policy-period record as one policy-year regardless of its duration. Both errors distort the denominator and therefore the resulting frequency or loss rate.
For example, if a six-month occurrence study includes three policies each covered throughout those six months, the exposure is \(3 \times 6/12=1.5\) policy-years—not three policy-years. If there are three eligible claims, the correct frequency is \(3/1.5=2\) claims per policy-year. Dividing by three would incorrectly treat six months of coverage as a full year for each policy.
A second common mistake is to omit zero-claim periods from exposure. They add no claims and no claim loss, but their covered time remains part of the observed risk base. A third is to divide total loss by the number of policy-period records and call the result average severity; severity uses the number of eligible claims.
Specify the dates and whether claims are selected by occurrence, report, or another defined date.
For every record, retain only covered time inside the study window; check for gaps and duplicate or overlapping records.
Apply one consistent claim definition and loss basis to the selected set.
Use exposure for frequency and aggregate loss per exposure; use eligible claim count for average severity.
Confirm that exposure is in the stated unit and, when claims exist, verify that frequency times severity equals aggregate loss per exposure.
These definitions are the foundation for later empirical summaries: a calculation can be arithmetically correct yet answer the wrong question if its observation unit, dates, or denominator are mismatched. Careful bookkeeping makes the resulting estimate interpretable and reproducible. The distinctions are aligned with the SOA Exam FAM (Fundamentals of Actuarial Mathematics) syllabus.
Check Your Understanding
For each problem, state the relevant observation unit and keep the claim set and exposure window consistent.
- A hypothetical study covers a three-month occurrence window. Two policies are covered throughout all three months, and a third is covered during the final two months. There are two eligible claims. Find total exposure in policy-years and observed frequency. (Answer: \(8/12=2/3\) policy-year; 3 claims per policy-year.)
- During a four-month window, three records contribute four, three, and two months of covered time within the window. There are two eligible claims. Find the exposure and frequency per policy-year. (Answer: \(9/12=0.75\) policy-year; \(2/0.75=2.6667\) claims per policy-year.)
- Five fully observed one-year policy-periods include two eligible claims with losses of $800 and $1,200; the other three periods have no claims. Find average severity and aggregate loss per policy-year. (Answer: $1,000.00 per claim and $400.00 per policy-year.)
- A six-month occurrence study has 12 covered policy-months and three eligible claims with total loss $4,500. Find frequency, average severity, and aggregate loss per policy-year. (Answer: 3 claims per policy-year; $1,500.00 per claim; $4,500.00 per policy-year.)
- A claim occurred on the final day of a defined occurrence window but was reported two weeks later. The data have complete follow-up through the reporting date. Should it be counted in an occurrence-based study, and should its later report date add exposure? (Answer: Count it under the occurrence-based rule; no, exposure remains limited to covered time inside the defined occurrence window.)