Why Valuation Dates and Cohorts Must Be Defined Together
A claim triangle is useful only when each observation has a clear accident period, valuation date, and development age. A change in a cell can reflect the passage of time, new claims being reported, or a genuine change in experience. It can also be an artifact: the coverage included, the unit counted, or the data cutoff may have changed. Before interpreting a movement in incurred or paid claims, an actuary must know which claims are being compared and how mature they are.
The central distinction is between a cohort and a snapshot. A cohort is defined by fixed eligibility rules, such as loss dates in a specified accident year and a specified coverage. A snapshot records the information known for that cohort at one valuation date. As valuation advances, additional claims may be reported and existing claim amounts may change; the cohort’s accident-period definition does not change.
Specify the Cohort Before Comparing Its Snapshots
A practical cohort definition should specify at least four things: the accident-period boundaries, the included coverage, the counting unit, and the valuation cutoff. For example, “unique collision claims with loss dates in calendar year 2023, reported by December 31, 2024” is more precise than “2023 claims.” Depending on the analysis, the counting unit might be a claim identifier, an occurrence, or a claimant. Those units are not interchangeable.
For an annual accident year \(y\), use the same year-end maturity convention at every valuation. At a December 31 valuation in calendar year \(k\), define age in months as:
Thus, accident year 2023 is at 12 months at December 31, 2023, and at 24 months at December 31, 2024. Accident year 2022 is at 24 months at December 31, 2023. Equal-maturity observations lie on the same development age, even though their calendar valuation dates differ.
For a fixed accident year, the set of claims reported by valuation date \(v\) can be represented as:
The reporting condition means that \(C_y(v)\) can grow as \(v\) advances: a claim from the accident year may not yet have been reported at an earlier valuation. This does not move the claim into a later accident year. It means that the earlier snapshot did not yet include it.
How Equal-Maturity Comparisons Work
Suppose the question is whether claims in two accident years look different at 12 months of development. Compare the 12-month observation for each accident year, not the latest available observation for each. The valuation dates will differ, but the maturity is controlled. If exposures differ, interpret claim counts alongside the corresponding exposure measure rather than treating raw counts as directly comparable.
The reason for aligning maturity follows from what a snapshot measures. At an earlier age, some losses have not yet been reported and known claim amounts have had less time to develop. As valuation advances, both the information set and the recorded amounts may change. Comparing different ages therefore mixes calendar-period differences with development. Equal-maturity comparisons remove that specific mismatch, though they do not by themselves control for differences in coverage, exposure, or claim definition.
Worked Example: Aligning Annual Cohorts at 12 Months
Hypothetical data. At December 31, 2022, accident year 2022 has 1,000 earned vehicle-years, 28 reported unique claims, $280,000 incurred, and $90,000 paid. At December 31, 2023, accident year 2023 has 1,100 earned vehicle-years, 31 reported unique claims, $310,000 incurred, and $100,000 paid. The accident-year and coverage definitions, claim-count rules, and valuation cutoffs are consistent. All amounts are nominal dollars; no interest discounting, mortality assumption, or probability distribution is involved.
Step 1: Calculate maturity. Accident year 2022 at December 31, 2022 is \(12(2022-2022+1)=12\) months old. Accident year 2023 at December 31, 2023 is \(12(2023-2023+1)=12\) months old. The observations are at equal maturity.
Step 2: Compare claim counts with exposure. Accident year 2022 has \(28/1{,}000=0.0280\) reported claims per earned vehicle-year. Accident year 2023 has \(31/1{,}100=0.02818\), or approximately 0.0282 claims per earned vehicle-year.
Step 3: Compare recorded amounts at the same age. Incurred claims per earned vehicle-year are \(280{,}000/1{,}000=\$280\) for accident year 2022 and \(310{,}000/1{,}100=\$281.82\) for accident year 2023. Paid claims per earned vehicle-year are \(90{,}000/1{,}000=\$90\) and \(100{,}000/1{,}100=\$90.91\), respectively.
Check. Both observations are at 12 months, and each amount and claim count is divided by the exposure for its own accident year. The differences in the resulting measures are small, but this comparison alone does not establish why they differ.
Method and trap avoided. This is an equal-development-age comparison. A common mistake would be to compare accident year 2022 at December 31, 2023 (24 months old) with accident year 2023 at December 31, 2023 (12 months old). That comparison mixes maturity with accident year. The correction is to use the 12-month accident year 2022 snapshot from December 31, 2022, or to make a deliberate comparison at another common age.
Keep Coverage and Counting Rules Fixed
Equal maturity is necessary for many comparisons, but it is not sufficient. A cohort defined as collision claims cannot be compared directly with a later data extract that also includes glass claims. Similarly, counting claim-level records in one period and claim-feature records in another can inflate the apparent number of claims. A claim can have several records or features while still representing one claim under a unique-claim counting rule.
The same principle applies to the amounts being compared. If a data extract adds a coverage, the total amount may rise because additional claims are now included. That increase is not evidence of a change in the original coverage’s experience. First isolate the common scope, then compare amounts and counts under the same definitions.
Worked Example: Correcting a Coverage and Record-Count Mismatch
Hypothetical data. At 12 months, a prior accident year has 1,600 earned exposure units, 80 unique collision claims, and $400,000 incurred on those claims. A later accident year also has 1,600 exposure units. Its broad data extract contains 100 records: 84 unique collision claims with $420,000 incurred, 12 glass claims with $36,000 incurred, and four additional feature records belonging to the collision claims. The four feature records are not additional claims and do not add incurred dollars. All figures are at 12 months of development and in nominal dollars; no interest discounting, mortality assumption, or probability distribution is involved.
Step 1: Identify the comparable scope. The prior observation is collision-only and counts unique claims. Therefore, the later observation must also be restricted to unique collision claims. Its comparable count is 84, not 100, and its comparable incurred amount is $420,000, not $456,000.
Step 2: Calculate counts per exposure. The prior rate is \(80/1{,}600=0.0500\) claims per exposure unit. On the matching collision basis, the later rate is \(84/1{,}600=0.0525\). The relative increase is \((0.0525/0.0500)-1=0.05\), or 5%.
Step 3: Check incurred amount per unique collision claim. The prior average is \(400{,}000/80=\$5{,}000\). The later average is \(420{,}000/84=\$5{,}000\). On the stated basis, the average incurred amount per counted collision claim is unchanged.
Check. The matched scope produces 84 claims and $420,000 for the later period. The 12 glass claims and four duplicate feature records are excluded from the collision unique-claim comparison.
Method and trap avoided. This is a scope-and-unit reconciliation before comparison. The wrong calculation is \(100/1{,}600=0.0625\), which suggests a 25% increase from 0.0500. It counts glass claims and duplicate records that were absent from the prior definition. The correction is to use 84 unique collision claims, yielding a 5% increase. A comparison of the broad $456,000 amount with the prior $400,000 would likewise mix coverage scopes.
Keep Accident Period Separate From Report Period
The date of loss assigns a claim to an accident period; the report date determines when the claim first appears in a reported snapshot. Confusing these dates shifts late-reported claims into the wrong cohort. This can distort accident-year counts and amounts, as well as indicators built from those quantities.
A consistent historical snapshot also needs a clear “known by” date. A current data file may contain backdated information about a claim, but that does not mean the information was available at an earlier valuation. When reconstructing an earlier snapshot, apply the earlier valuation cutoff to the transaction or report history rather than silently using later knowledge.
Worked Example: Assigning Late-Reported Claims to the Correct Accident Year
Hypothetical data. Accident year 2023 has 1,200 earned exposure units. By December 31, 2023, 40 unique claims with 2023 loss dates have been reported, with $240,000 incurred. By December 31, 2024, 55 unique claims with 2023 loss dates have been reported, with $330,000 incurred; the additional 15 claims were first reported during 2024. Separately, accident year 2024 has 1,000 earned exposure units and 20 unique claims with 2024 loss dates reported by December 31, 2024, with $110,000 incurred. The count and coverage definitions are consistent. All amounts are nominal dollars; no interest discounting, mortality assumption, or probability distribution is involved.
Step 1: Assign claims by loss date. The 40 claims reported by December 31, 2023, and the 15 later-reported claims all belong to accident year 2023 because their loss dates fall in 2023. The 20 claims with 2024 loss dates belong to accident year 2024.
Step 2: Describe the 2023 snapshots accurately. At 12 months, accident year 2023 has 40 reported claims and $240,000 incurred. At 24 months, that same accident-year cohort has 55 reported claims and $330,000 incurred. The increase between snapshots includes the 15 claims reported later; it is not a transfer of those claims to accident year 2024.
Step 3: Compare equal maturity where appropriate. The 12-month claim count per exposure is \(40/1{,}200=0.03333\) for accident year 2023 and \(20/1{,}000=0.02000\) for accident year 2024. The corresponding incurred amounts per exposure are \(240{,}000/1{,}200=\$200\) and \(110{,}000/1{,}000=\$110\). These are equal-maturity comparisons, but differences may still reflect factors other than accident-year experience.
Check. The 24-month 2023 count reconciles as \(40+15=55\). The 2024 accident-year count remains 20; it does not include the 15 late reports for 2023 losses.
Method and trap avoided. This method assigns claims by date of loss and applies each valuation cutoff to the relevant snapshot. The wrong calculation would assign all 2024 reports to accident year 2024, giving it \(20+15=35\) claims. That classification uses report year rather than accident year. It overstates the 2024 cohort by 15 claims and understates the later reported view of accident year 2023 by the same number.
A Practical Construction Sequence
Use the following sequence when extracting or reviewing data for a valuation-date comparison. It makes the key definitions explicit before any result is interpreted.
State the loss-date boundaries and assign each claim using its date of loss, not its report or payment date.
Specify included coverages and whether one counted unit is a unique claim, occurrence, claimant, or another defined unit.
Include only information available by the specified valuation date when reconstructing a historical snapshot.
Use one maturity convention across all accident periods and valuation dates.
Compare equal ages, reconcile exposures and definitions, and label any remaining difference as a result requiring further investigation—not an automatic conclusion.
Why the Definitions Matter
Reserve and settlement analyses rely on comparisons that are sensitive to the population being measured. If one valuation includes a broader coverage, a different claim-count unit, or a different maturity, an apparent change in reserve adequacy or settlement behavior may reflect the data construction rather than an underlying change in claim handling or loss experience. The cohort definitions do not explain the cause of a change; they make sure the comparison is about the intended population.
A useful audit is to reconcile a sample of records from each valuation back to loss date, coverage, claim identifier, report date, and transaction cutoff. Then check that the same rules were applied to every period. A label such as “accident year 2023” is not enough on its own: the analyst must also know which claims qualify, what is counted, and what was known at the valuation date.
Check Your Understanding
For each problem, state the relevant cohort or maturity issue before calculating.
- At December 31, 2024, accident year 2022 is being compared with accident year 2023 at December 31, 2024. What is the development age of each under the stated year-end convention, and are they equal-maturity observations? (Answer: Accident year 2022 is 36 months old; accident year 2023 is 24 months old; they are not equal-maturity observations.)
- A hypothetical 12-month cohort has 45 unique claims and 900 earned exposure units. Calculate reported claims per exposure unit. (Answer: \(45/900=0.0500\) claims per exposure unit.)
- At the same development age and coverage, one year has 72 unique claims per 1,200 exposure units. The next has 78 unique claims per 1,300 exposure units. Calculate both rates and the relative change. (Answer: \(72/1{,}200=0.0600\); \(78/1{,}300=0.0600\); relative change \(=0\%\).)
- A 2025 valuation shows 63 claims with 2024 loss dates and 14 newly reported claims with 2023 loss dates. How many belong to accident year 2023, and how many to accident year 2024, among the claims described? (Answer: 14 belong to accident year 2023 and 63 belong to accident year 2024.)
- A baseline cohort has 50 unique claims and $250,000 incurred. A later, equal-maturity extract has 66 records: 54 unique claims in the same coverage with $270,000 incurred, eight claims from a newly included coverage with $24,000 incurred, and four duplicate feature records. Find the later cohort count and incurred amount comparable to the baseline, and the change in average incurred per claim. (Answer: Comparable later values are 54 unique claims and $270,000; both periods have average incurred of $5,000 per claim, so the change is $0.)