Why Triangle Cells Are Not All Expected to Contain Data
A run-off triangle is often displayed as a rectangular grid, but the data behind it are not a complete rectangle. At a given valuation date, early accident periods have had more time to develop than recent ones. As a result, some cells refer to development that has not yet occurred by the valuation date. Those cells are structurally unobserved, not missing observations. Other cells correspond to development that should already be observable but have no recorded value. Those blanks require investigation.
This distinction matters before analyzing a triangle. Treating an unobservable cell as missing data can prompt an unnecessary search or an inappropriate fill-in. Treating a genuinely missing observable value as structural can hide an incomplete data extract. The classification depends on the accident period, the development age, and the valuation cutoff—not on whether the cell happens to be blank.
The Observable Shape of a Triangle
For accident period \(i\), the latest development age available by period \(T\) is found by solving the observability condition for \(j\):
Thus, when \(i \leq T\), the observable ages in row \(i\) are \(1,2,\ldots,T-i+1\). Each later accident period has one fewer observable age than the preceding period. For a triangle with accident periods \(1\) through \(T\), the number of observable cells is
This count is a useful check when a complete triangular extract is expected. It is not a rule that every observed cell must contain a nonzero amount: a value of zero can be a valid recorded observation. Nor does the count prove that the data are complete if the extract intentionally omits accident periods or uses a different observation window.
Identify the last calendar period included, and confirm that all rows use that same cutoff.
For accident period \(i\), calculate \(T-i+1\). Ages above that maximum are structural blanks.
Check the source records, valuation snapshots, period labels, and data-extract rules before deciding why a value is absent.
The condition assumes that accident periods and calendar periods are indexed consecutively and that development age 1 falls in the accident period itself. A different labeling convention changes the index formula, so the period mapping must be established first. If a valuation is made partway through a period, or if different rows have different valuation dates, use the actual observation schedule rather than forcing a single common \(T\).
Worked Classifications
Worked Example: Classifying Blanks in a Recent Accident Period
A hypothetical triangle is valued through calendar period \(T=9\). In accident period \(i=6\), the cells for ages 1 through 4 contain recorded values, while ages 5 and 6 are blank. Classify each blank.
Method. Calculate the largest observable development age for row 6:
Therefore, ages 1, 2, 3, and 4 are observable in this row. Age 5 corresponds to calendar period \(6+5-1=10\), which is after the cutoff \(T=9\). Age 6 corresponds to calendar period \(6+6-1=11\), also after the cutoff.
Classification. Both blanks, ages 5 and 6, are structurally unobserved. There are zero missing observable cells among the two blanks.
Check. The row has four observable ages, exactly matching the four recorded cells. The first unobservable age is 5, as \(6+5-1=10>9\). The common mistake is to assume age 5 is observable just because it immediately follows the last recorded age; the cutoff calculation shows that it is not.
Worked Example: Finding an Incomplete Observable Cell
A hypothetical triangle is valued through period \(T=10\), with displayed age columns 1 through 9. In row \(i=4\), ages 1, 2, 4, 5, 6, and 7 have recorded values; ages 3, 8, and 9 are blank. Classify the three blanks.
Method. The maximum observable age in row 4 is
So ages 1 through 7 are observable. The blank at age 3 is within that range. Its calendar period is \(4+3-1=6\), which is no later than period 10. It is therefore an observable cell with a missing value that needs investigation.
For age 8, the calendar period is \(4+8-1=11\), and for age 9 it is \(4+9-1=12\). Both exceed the cutoff. Those two blanks are structurally unobserved.
Conclusion. Of the three blanks, one is a potentially missing observable value and two are structural. Check: the last observable age is 7; the blank at age 3 lies within ages 1–7, while ages 8 and 9 lie beyond them. The mistake to avoid is labeling every blank in a row the same way.
Worked Example: Reconciling Blanks Across Several Rows
A hypothetical extract is valued through period \(T=9\) and includes accident periods \(i=3,4,5,6,7\). All displayed blank cells are listed below. For row 3, ages 1–7 are recorded. For row 4, ages 1, 2, 4, 5, and 6 are recorded, and age 3 is blank. For row 5, ages 1–5 are recorded. For row 6, ages 1, 2, and 4 are recorded, while ages 3 and 5 are blank. For row 7, ages 1–3 are recorded. How many observable cells should the extract contain, how many are missing, and how many blanks are structural?
Step 1: Count the observable cells expected in each included row. The maximum ages are \(9-3+1=7\), \(9-4+1=6\), \(9-5+1=5\), \(9-6+1=4\), and \(9-7+1=3\). The expected total is therefore
Step 2: Check the listed blanks against those row limits. Row 4, age 3 is observable because it is no greater than the row’s maximum age of 6. Row 6, age 3 is also observable because it is no greater than the row’s maximum age of 4. These are two potentially missing observations. In row 6, age 5 exceeds the maximum age of 4, so it is structurally unobserved.
Step 3: Reconcile the counts. Row 3 has 7 recorded cells; row 4 has 5; row 5 has 5; row 6 has 3; and row 7 has 3. That gives \(7+5+5+3+3=23\) recorded cells. Of the 25 expected observable cells, \(25-23=2\) are missing, matching the two identified observable blanks.
Conclusion. The extract has 25 expected observable cells, 23 recorded cells, two missing observable cells, and one listed structural blank. Check: each included row’s recorded cells plus its observable blank equals its expected row count. The structural blank is outside the expected total and must not be counted as a missing observation.
Why the Classification Is a Data-Quality Check
The triangular pattern follows from the calendar timing of the observations. The calendar period for cell \((i,j)\) is \(i+j-1\): age 1 falls in accident period \(i\), age 2 falls one period later, and each further age advances one more period. A valuation through \(T\) cannot contain a later calendar period. That is why cells beyond the diagonal are structural rather than absent data.
The formula also gives a direct audit. For any blank cell, calculate \(i+j-1\) and compare it with \(T\). If the result exceeds \(T\), the cell was not yet observable under the stated setup. If it is at most \(T\), the cell belongs to the expected triangle and should be explained. Possible explanations include an incomplete source extract, a missing valuation snapshot, a period-labeling error, or a record that has not been processed. The classification identifies what to investigate; it does not by itself establish the cause.
A complete data review should also verify that the stated cutoff is appropriate for every row, that accident periods and development ages use the intended units, and that the extract has not omitted a row or an observable column by design. For example, if a file contains only selected accident periods, its expected observable-cell count is the sum of the row limits for those included periods—not the count for a full triangle beginning at period 1.
| Cell status | Timing test | How to treat a blank |
|---|---|---|
| Observable | \(i+j-1 \leq T\) | Investigate as a potentially missing or incomplete value |
| Structural | \(i+j-1 > T\) | Expected under the cutoff; do not count as a missing observation |
| Recorded zero | The cell is observable and has a recorded value of zero | Retain as an observed value; zero is not the same as blank |
Common Mistakes
- Classifying by appearance. A blank is not automatically missing. Calculate the row’s observable-age limit first.
- Using the wrong cutoff. If the valuation includes period 9, a cell in period 10 is structural even if it is the next age after the last recorded value.
- Making an off-by-one error. With age 1 in the accident period, the calendar index is \(i+j-1\), not \(i+j\). Check that age 1 maps to period \(i\).
- Treating zero as absent. A recorded zero in an observable cell is a value. A blank or missing-data code is not automatically a zero.
- Assuming every row has the same maturity. The maximum age changes with the accident period when the rows share a common valuation cutoff.
Check Your Understanding
For each problem, classify only the cells specified and use the stated common valuation cutoff.
- A triangle is valued through period \(T=11\). In row \(i=8\), ages 1–3 have values and ages 4 and 5 are blank. How many blanks are observable, and how many are structural? (Answer: One observable blank, age 4; one structural blank, age 5.)
- A triangle is valued through period \(T=7\). In row \(i=3\), ages 1–3 are recorded and ages 4 and 5 are blank. How many missing observable cells do the blanks represent? (Answer: Two; both ages 4 and 5 are observable.)
- A triangle is valued through period \(T=9\) and includes accident periods 2 through 9. How many observable cells should it contain if all those rows and ages are included? (Answer: 36 observable cells.)
- A triangle is valued through period \(T=12\). The listed blanks are row 4 age 9, row 5 age 7, and row 6 age 8. Classify them. (Answer: Two observable blanks, row 4 age 9 and row 5 age 7; one structural blank, row 6 age 8.)
- A triangle is valued through period \(T=6\). In row \(i=4\), age 3 contains a recorded value of $0 and age 4 is blank. What is the status of each cell? (Answer: Age 3 is an observed zero; age 4 is structurally unobserved.)