Checking Triangle Arithmetic and Data Consistency

Why Triangle Checks Come Before Projection

A loss triangle organizes observed claims amounts by accident period and development age. Before using its entries to select link ratios or project ultimate losses, an actuary needs to know that the entries represent the intended amounts, dates, and units. A triangle can look orderly and still contain a cumulative-total error, a misplaced cell, or a conversion mistake.

This tutorial concerns validation, not the selection of development factors. The checks are aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus. Each check asks whether the triangle agrees with its construction rules and its underlying source data.

Definition: Let \(I_{i,j}\) denote the incremental amount for accident period \(i\) in development interval \(j\), and let \(C_{i,j}\) denote the corresponding cumulative amount through development age \(j\). On a consistent basis, \(C_{i,j}\) is the sum of the increments from the first development interval through interval \(j\). A blank cell outside the observed portion of a triangle means the amount is not yet observed; it is not a zero.

The labels and basis matter as much as the arithmetic. For example, a paid triangle and an incurred triangle do not represent the same measure. Likewise, a triangle stated in thousands of dollars cannot be compared directly with source records stated in dollars. Before reconciling entries, establish whether both sources use the same valuation date, currency, unit, claim basis, and treatment of recoveries or adjustments.

Core Reconciliation Checks

For each accident period, compare the incremental sequence with the cumulative sequence. The first cumulative amount equals the first increment. Each later cumulative amount equals the previous cumulative amount plus the next increment. Conversely, subtracting consecutive cumulative amounts recovers the increments.

C_{i,1}=I_{i,1}, \qquad C_{i,j}=C_{i,j-1}+I_{i,j}\quad (j\geq 2), \qquad I_{i,j}=C_{i,j}-C_{i,j-1}.

These identities give two independent directions for checking a row: build cumulative amounts from increments, then take differences of the cumulative amounts and compare them with the original increments. A mismatch localizes an error to a cell or to a difference in data definitions. It does not, by itself, show which value should be changed; that requires checking the source records.

Valuation-date check: A triangle cell must be observable by the stated valuation date. Under an annual convention in which development age 12 months covers the accident year ending December 31, the cell for accident year \(i\) and age \(j\) years corresponds to calendar year \(i+j-1\). Thus, at December 31, 2025, accident year 2024 can be observed through age 24 months, but not age 36 months.

The date rule depends on the convention used to define accident periods and development ages. Do not apply the annual rule blindly to quarterly, monthly, or differently aligned data. Use the source documentation to establish the interval represented by each age. Then check that the latest observed diagonal agrees with the valuation date and that no future development period has been populated as if it were already known.

Why the Reconciliation Identity Holds

The cumulative amount is defined as the total of all increments recorded up to that development age. Consequently, the amount through age \(j-1\) plus the new amount in interval \(j\) is exactly the total through age \(j\). Subtracting the earlier total isolates the new interval.

C_{i,j}=\sum_{k=1}^{j} I_{i,k} =\sum_{k=1}^{j-1} I_{i,k}+I_{i,j} =C_{i,j-1}+I_{i,j}.

The same reasoning gives \(C_{i,j}-C_{i,j-1}=I_{i,j}\). This is an accounting identity, not a statistical assumption. If it fails, the entries are not mutually consistent on the stated basis. The identity also explains a useful limitation: adding every cumulative cell in a triangle does not give the total observed loss, because earlier payments are counted again in later cumulative cells.

Fix the data definition.
Confirm whether amounts are paid or incurred, gross or net of recoveries, and which currency and unit are used.

Confirm the time basis.
Record the valuation date and the accident-period and development-interval conventions.

Reconcile the arithmetic.
Build cumulative amounts from increments, or difference cumulative amounts to recover increments.

Trace differences to the source.
Check the relevant records and aggregation rules before correcting any triangle cell.

Worked Reconciliations

Worked Example: Finding a Cumulative-Total Error

A hypothetical paid triangle is stated in thousands of dollars. For accident year 2023, the source file gives incremental payments of 15, 21, and 9 in development intervals 1, 2, and 3. The triangle reports cumulative amounts of 15, 36, and 46. Identify the inconsistency and the value supported by the source increments.

Method. Rebuild the cumulative row from the increments, then compare each result with the reported row. This avoids accepting a plausible-looking cumulative total without checking its components.

Through interval 1: \(C_{2023,1}=15\).

Through interval 2: \(C_{2023,2}=15+21=36\). This agrees with the reported value 36.

Through interval 3: \(C_{2023,3}=36+9=45\). The reported value is 46, so it is 1 thousand dollars higher than the amount implied by the increments.

Check in the reverse direction: the reported cumulative amounts imply interval 3 increment \(46-36=10\), not the source increment 9. The source increments sum to \(15+21+9=45\), so the source supports a cumulative amount of 45 in interval 3. The triangle should be investigated and corrected to 45 if the source data and stated definitions are confirmed. Check: \(15+21+9=45\), and the differences of \(15,36,45\) are \(15,21,9\).

Common mistake avoided: changing the source increment to 10 merely to make it match the triangle. Reconciliation detects a conflict; it does not authorize changing a source value without evidence.

Worked Example: Checking a Valuation Date and Unit Conversion

A hypothetical annual paid triangle is valued at December 31, 2023. Development age 12 months covers the accident year ending on that date; age 24 months covers the following calendar year. The triangle is in thousands of dollars. Source records for accident year 2022 show 92,000 dollars paid in the first development interval and 92,500 dollars paid in the second. The triangle reports 184,500 in its accident-year 2022, age-24 cell. Check the date, unit, and cumulative amount.

Method. First map the cell to its calendar period, then convert the source amounts into the triangle's unit and reconcile the cumulative total.

Date check: Accident year 2022 at age 12 months is the period ending December 31, 2022. At age 24 months it is the period ending December 31, 2023. The second-interval payments therefore belong in this age-24 cell and are observable at the stated valuation date.

Cumulative amount in dollars: \(92{,}000+92{,}500=184{,}500\) dollars.

Convert to thousands: \(184{,}500\div1{,}000=184.5\) thousand dollars. Equivalently, the incremental amounts are 92.0 and 92.5 thousand, and their cumulative total is \(92.0+92.5=184.5\).

The reported triangle value 184,500 is inconsistent with a column stated in thousands: it has not been converted. The supported cell value is 184.5 in thousands of dollars. Check: converting back gives \(184.5\times1{,}000=184{,}500\) dollars, matching the source.

Common mistake avoided: treating 184,500 as a plausible triangle value because it matches the dollar source total. The unit label applies to every cell, so the source total must be converted before comparison.

Worked Example: Reconciling Triangle Rows to Source Totals

A hypothetical paid triangle is valued at December 31, 2025 and shown in thousands of dollars. Source records give accident year 2024 increments of 42 in the first interval and 18 in the second, and accident year 2025 increments of 50 in the first interval. The triangle reports cumulative rows \((42,61)\) for 2024 and \((50)\) for 2025. Check the rows and the total represented by the latest observed cumulative amount in each row.

Method. Reconcile each row to its increments, confirm which cells are observable, and compare the latest cumulative amount for each accident year with the source total. This avoids summing repeated cumulative amounts as if they were separate payments.

Accident year 2024: age 12 months is \(42\). Age 24 months should be \(42+18=60\), but the triangle reports 61. Taking the difference of the reported cells gives \(61-42=19\), while the source second-interval increment is 18. The row is inconsistent by 1 thousand dollars.

Accident year 2025: at December 31, 2025, age 12 months is observed and equals 50. Age 24 months is not yet observable, so its cell should remain blank, not be entered as zero.

Source total across observed increments: \(42+18+50=110\) thousand dollars. The latest observed cumulative amounts supported by the source are 60 for accident year 2024 and 50 for accident year 2025, also totaling \(60+50=110\). The reported latest cumulative amounts total \(61+50=111\), a discrepancy of 1 thousand dollars. The source supports changing the 2024 age-24 value to 60, subject to confirming the underlying records.

Check: the corrected 2024 row \((42,60)\) differences to \((42,18)\), and its latest amount plus the 2025 latest amount is \(60+50=110\), equal to the source total. Common mistake avoided: adding every reported cumulative cell, which would count the 2024 first-interval payments twice.

Practical Controls and Common Traps

A useful validation record states the triangle's valuation date, development-age convention, unit, currency, claim basis, and source extract used. The record should also make clear whether amounts are transactions in a development interval or cumulative totals through an age. Without those labels, a numerical reconciliation can appear to pass while comparing quantities with different meanings.

Do not treat blank future cells as zero. A blank cell may mean that the development period has not occurred by the valuation date. Replacing it with zero changes the meaning of the data.
Do not add cumulative cells to get a source total. Cumulative rows repeat earlier amounts. Compare the source total with increments, or with the latest observed cumulative amount for each accident period.
Do not ignore units or currency. Convert dollars to thousands, or one currency to another when appropriate, before comparing values. Confirm the conversion convention used.
Do not infer the correct value from the triangle alone. A failed check identifies disagreement. Source records and definitions are needed to decide whether the triangle, an increment, or the stated basis is wrong.
Do not assume every development-age convention is identical. Map dates from the documented interval boundaries, especially when periods are not annual or do not end on December 31.

These checks matter because a small data error can alter a link ratio or distort the apparent development pattern. Arithmetic and source validation do not establish that a development pattern is appropriate; they establish that the data supplied for analysis are internally consistent and correctly positioned. That distinction keeps a data problem from being mistaken for a reserving result.

Key takeaway: For each triangle, verify the data basis, valuation date, units, cumulative-versus-incremental arithmetic, and agreement with source totals. Then investigate each discrepancy at its source rather than forcing the numbers to agree.

Check Your Understanding

For each problem, check the arithmetic and state any date or unit issue before accepting the triangle entry.

A hypothetical accident-year row has increments of 8, 13, and 4, all in thousands of dollars. The cumulative row is 8, 21, 25. Is it consistent? (Answer: Yes; the cumulative amounts are 8, 8+13=21, and 21+4=25.)
A source file gives 37,000 dollars in the first interval and 12,000 dollars in the second. A cumulative triangle in thousands reports 49. What is the correct value in thousands? (Answer: 49 thousand dollars; 37,000+12,000=49,000 dollars, or 49 thousand.)
At December 31, 2024, an annual triangle uses age 12 months for the accident year ending that date and age 24 months for the following year. Which age is available for accident year 2024? (Answer: Age 12 months only; age 24 months ends December 31, 2025 and is not yet observable.)
In thousands, a row has cumulative amounts 26 and 43, while source increments are 26 and 16. Identify the discrepancy and the cumulative value implied by the source. (Answer: The source implies 26+16=42; the reported 43 is 1 thousand too high, and its difference from 26 is 17 rather than 16.)
A triangle's latest observed cumulative amounts are 72 for one accident year and 45 for another. The source increments total 117 across those two years. A reviewer instead adds all four cumulative cells, including earlier ages, and obtains 189. Which total is the appropriate reconciliation target, and why? (Answer: 117; it equals 72+45, the latest cumulative amount for each accident year. Adding earlier cumulative cells counts some increments more than once.)

What You'll Learn