Why Inspect Triangle Diagonals?
A loss triangle arranges observations by accident period and development age. Reading across a row follows one accident period as it develops; reading down a column compares accident periods at the same development age. A diagonal follows cells associated with the same calendar period. That third view can help flag a possible calendar-year distortion, such as a change in claims handling, inflation, or the operating environment that affects several accident periods at once.
A diagonal is a diagnostic clue, not proof of a cause. The cells available in a triangle are limited, and a single unusual cell belongs simultaneously to one row, one column, and one diagonal. The key task is therefore to ask which grouping best describes a repeated pattern, and to say when the observations do not distinguish the explanations.
For example, accident year 2022 at development age 3 is associated with calendar year \(2022+3-1=2024\). So are accident year 2023 at age 2 and accident year 2024 at age 1. These three cells form one diagonal. The calendar-year label follows from the stated age convention; if a data source defines development periods differently, use its convention consistently.
What to Compare on a Diagonal
Raw amounts in different cells are not usually directly comparable: expected loss levels can differ by accident period, and amounts typically develop with age. A useful diagnostic is a ratio of observed to expected amounts, provided the observed and expected quantities refer to the same measure and cell. For this tutorial, call that ratio \(R_{a,j}\). A value near 1 means the observation is near its benchmark; a value above 1 is higher than the benchmark, and a value below 1 is lower.
The benchmark should be defined consistently across the cells being compared. For instance, if the triangle contains cumulative paid amounts, compare a cumulative observed amount with a cumulative expected amount at that same age—not with an incremental expectation. The ratios help make relative deviations visible, but do not by themselves establish that the benchmark is correct or that a particular calendar-year cause is responsible.
A Short Justification for the Diagonal Mapping
The diagonal grouping follows directly from the way the cell is indexed. Moving one accident year later increases \(a\) by 1. To stay in the same calendar year, development age must decrease by 1, so \(j\) decreases by 1. The sum \(a+j-1\) is unchanged. More generally, every pair of cells with the same value of that sum belongs to the same calendar-year diagonal.
This identity explains why inspecting diagonals can reveal a shared calendar-period signal: the cells combine different accident periods and different development ages, while retaining the same calendar-year label. It also explains the main limitation. Each cell is grouped three ways at once, so a pattern cannot be assigned to a diagonal merely because one cell on it is unusual.
Worked Diagonal Diagnostics
In the examples, all values are hypothetical observed-to-expected ratios from a consistent benchmark. The purpose is to compare the shape of the deviations, not to estimate a reserve or prove a causal explanation.
Worked Example: A Repeated High Pattern on One Diagonal
A hypothetical insurer reviews these ratios for accident years 2021–2024. Determine whether the pattern aligns more clearly with a calendar-year diagonal or with a development-age column.
| Accident year | Age 1 | Age 2 | Age 3 | Age 4 |
|---|---|---|---|---|
| 2021 | 1.00 | 1.02 | 1.35 | 1.03 |
| 2022 | 1.01 | 1.30 | 1.00 | — |
| 2023 | 1.28 | 1.02 | — | — |
| 2024 | 1.03 | — | — | — |
Method. Map the unusual cells to calendar years, then compare the pattern with the other observed cells in their development-age columns and accident-year rows.
Step 1: Identify the candidate diagonal. The ratio 1.35 is for accident year 2021, age 3, giving calendar year \(2021+3-1=2023\). The ratio 1.30 is for accident year 2022, age 2, giving \(2022+2-1=2023\). The ratio 1.28 is for accident year 2023, age 1, giving \(2023+1-1=2023\). These three elevated ratios lie on the same calendar-year diagonal.
Step 2: Check other cells in those age columns. At age 3, the other observed ratio is 1.00; at age 2, the other observed ratios are 1.02 and 1.02; at age 1, the other observed ratios are 1.00, 1.01, and 1.03. Those comparison values are near 1, unlike the three ratios on the 2023 diagonal.
Step 3: Compare the conclusion. The three elevated values occur across three ages and three accident years, but all share calendar year 2023. The diagonal pattern is more evident than a single age-column or accident-year-row pattern. This supports flagging a possible calendar-year distortion for investigation; it does not establish its cause.
Check. The mean of the three diagonal ratios is \((1.35+1.30+1.28)/3=3.93/3=1.31\). The calculation confirms the cluster is materially above 1, while the nearby comparison ratios are close to 1. A common mistake avoided here is to focus on the largest value alone; the diagnosis rests on the three-cell alignment.
Worked Example: A Development-Age Pattern Rather Than a Diagonal
A hypothetical triangle has the following observed-to-expected ratios. Determine whether the strongest repeated pattern is associated with age 2 or with one calendar year.
| Accident year | Age 1 | Age 2 | Age 3 | Age 4 |
|---|---|---|---|---|
| 2021 | 1.00 | 1.32 | 1.01 | 0.99 |
| 2022 | 1.02 | 1.29 | 1.00 | — |
| 2023 | 0.98 | 1.34 | — | — |
| 2024 | 1.01 | — | — | — |
Method. Compare like development ages and map the elevated cells to their calendar years before interpreting the cluster.
Step 1: Summarize age 2. The three age-2 ratios are 1.32, 1.29, and 1.34. Their sum is \(1.32+1.29+1.34=3.95\), so their arithmetic mean is \(3.95/3=1.3167\) to four decimal places.
Step 2: Compare with other ages. The age-1 mean is \((1.00+1.02+0.98+1.01)/4=4.01/4=1.0025\). The two observed age-3 ratios have mean \((1.01+1.00)/2=1.005\). Both are close to 1, unlike the age-2 values.
Step 3: Check the calendar-year mapping. The three age-2 cells correspond to calendar years 2022, 2023, and 2024, respectively: \(2021+2-1=2022\), \(2022+2-1=2023\), and \(2023+2-1=2024\). The elevated ratios are spread over three diagonals, rather than concentrated on one diagonal. The pattern is therefore more consistent with a development-age association than with a single calendar-year diagonal.
Check. The observed comparison supports investigating age 2 as the grouping with the clearest repeated elevation. It does not, by itself, explain why age 2 differs. A common mistake avoided here is to label any cluster of high ratios a calendar-year effect without first calculating the calendar years of the cells.
Worked Example: When the Pattern Does Not Distinguish the Groupings
In a hypothetical triangle, age-2 ratios for accident years 2022, 2023, and 2024 are 0.95, 1.40, and 1.05. The ratio at accident year 2022, age 3, is 1.02, and the ratio at accident year 2024, age 1, is 0.98. What can be concluded about the elevated age-2 ratio at accident year 2023?
Method. Compare the focal cell with both its development-age peers and the other observed cells on its calendar-year diagonal.
Step 1: Find its age and calendar year. The focal ratio 1.40 is at accident year 2023, age 2. Its calendar year is \(2023+2-1=2024\).
Step 2: Compare within age 2. The other age-2 ratios are 0.95 and 1.05. The value 1.40 is an isolated high value in that age column.
Step 3: Compare along calendar year 2024. The other stated cells on that diagonal are 1.02 at accident year 2022, age 3, and 0.98 at accident year 2024, age 1. Both are near 1. The value 1.40 is also an isolated high value on that diagonal.
Conclusion and check. These observations do not clearly favor a calendar-year explanation over an age-2 explanation. The high value is isolated in both groupings, so more observations or other relevant information would be needed to distinguish the patterns. A common mistake avoided is to claim a calendar-year effect because the other two diagonal values are near 1—or to claim an age effect solely because the age-2 peers are near 1. Both comparisons point to the same isolated cell and neither resolves its explanation.
Using the Diagnostic Carefully
A practical inspection can proceed in a few deliberate steps. The goal is to describe the observed pattern accurately before proposing explanations.
Identify the accident period, development age, and age convention. Compute the calendar-year label consistently.
Compare like measures and, where useful, use observed-to-expected ratios based on a consistent benchmark.
Check whether several unusual cells share a diagonal, a row, or a column. Do not treat one cell as repeated evidence.
Describe the clearest alignment, note sparse or conflicting comparisons, and avoid claiming a cause that the triangle alone does not establish.
A diagonal pattern can be consistent with a calendar-year influence, but coincidence is possible, especially when only a few cells are available. Conversely, calendar-year conditions may affect cells in different ways, so ratios need not be identical. The diagnostic is most useful as a prompt to review the affected periods and the assumptions behind the benchmark, not as an automatic adjustment to a projection.
Common Exam Trap
A frequent error is to see one high ratio on a diagonal and assign it to calendar year because the other diagonal cells are near 1. That same high cell may also be unusual within its development-age column or accident-period row. The correct question is whether a repeated pattern aligns more clearly with one grouping than with the others. If the same isolated cell creates each apparent pattern, the evidence is ambiguous.
Another error is to confuse a diagonal with a development-age column. A column keeps \(j\) fixed; a diagonal keeps \(a+j-1\) fixed. When accident year increases by one along a diagonal, development age decreases by one. Checking the indices explicitly prevents a visually plausible but incorrectly grouped comparison.
Check Your Understanding
For each problem, identify the relevant grouping before drawing a conclusion.
- A cell is for accident year 2022 at development age 4. What calendar year does it represent? (Answer: 2025.)
- Ratios of 1.27, 1.31, and 1.29 occur at accident year 2020 age 4, accident year 2021 age 3, and accident year 2022 age 2. What feature do the three cells share? (Answer: They are all on the calendar-year 2023 diagonal.)
- Age-3 ratios for three accident years are 0.99, 1.26, and 1.02. The elevated value is the only unusual value among the cells examined on its calendar-year diagonal. Is a calendar-year pattern established? (Answer: No. The high value is isolated on that diagonal; the evidence does not establish a calendar-year pattern.)
- Ratios at age 2 are 1.25, 1.30, and 1.28 for three consecutive accident years; the cells lie on three different calendar-year diagonals, and nearby ages are near 1. Which grouping is more strongly indicated? (Answer: Development age 2, because the repeated elevation occurs across different calendar years.)
- A ratio of 1.40 is the only high value among its age-2 peers and also the only high value among the other observed cells on its calendar-year diagonal. What is the most defensible conclusion? (Answer: The cell is an isolated departure; the data do not clearly distinguish an age-2 explanation from a calendar-year explanation.)