The loss development triangle is the single most important artifact in property and casualty reserving. Almost every method you will see in a reserve report starts from one. If you can read a triangle with confidence, you can ask the right questions about the report that comes out of it.
What the shape means
Two ideas are running at right angles to each other in a triangle. Once you see them, the rest of the report follows.
Accident year is the year the loss happened. When the car crashed, when the worker got hurt, when the property burned. Each row in a triangle is one accident year. A claim stays in its accident-year row no matter when the bill arrives or the case reserve is set.
Development age is how long you have been watching that accident year. “12 months” means evaluated at the end of the accident year itself. “24 months” means a year later. “72 months” means six years of payment and case-reserve history have accumulated.
Each cell, then, answers a single question: for accident year X, how much loss is on the books at development age Y? In the figure above, the cell at AY 2022 / age 36 reads 8,720. Plain English: for claims that happened in 2022, three years after the start of that accident year, $8.72 million had been paid out.
The reason the shape is a triangle, not a rectangle, is simple. Accident year 2025 cannot have 24 months of development until the end of 2026, and accident year 2024 cannot have 48 months until the end of 2028. The most recent evaluation date, usually the prior year-end, sits on the diagonal. The diagonal is what you have. Everything below it is the future, and a reserving method’s job is to fill those cells in.
| Accident year | 12 | 24 | 36 | 48 | 60 | 72 |
|---|---|---|---|---|---|---|
| Development age (months) | ||||||
| 2020 | 4,200 | 6,090 | 7,180 | 7,750 | 7,990 | 8,150 |
| 2021 | 4,650 | 6,740 | 7,950 | 8,590 | 8,850 | |
| 2022 | 5,100 | 7,400 | 8,720 | 9,420 | ||
| 2023 | 5,580 | 8,090 | 9,550 | |||
| 2024 | 6,120 | 8,870 | ||||
| 2025 | 6,710 | |||||
| Age-to-age | 1.450 | 1.179 | 1.080 | 1.031 | 1.020 | |
Paid losses, $ thousands, evaluated at 12/31/2025. The shaded cells on the diagonal are the most recent evaluation: what you have actually paid out. The hatched cells below the diagonal are the future, to be filled in by a reserving method. The bottom row shows volume-weighted age-to-age factors.
Two triangles, not one
A report will usually show at least two triangles side by side: one for paid losses and one for reported losses (paid plus case reserves). The figure above is a paid triangle. The two views tell you different things.
- Paid triangle. The most objective data. Paid losses do not depend on the adjuster’s opinion. But the paid pattern for long-tailed lines can be slow and volatile because settlements come in lumps.
- Reported triangle. Includes the adjuster’s case estimate on each open claim. Reacts faster to changes in severity and frequency. But it is sensitive to case-reserving philosophy and adequacy, which can drift as claims leadership changes.
Whenever the paid triangle and the reported triangle tell different stories, you have a puzzle worth asking about. A few common patterns and what they typically mean:
- Reported runs higher than paid in early periods and the two converge by maturity. Normal claim emergence. Case reserves are set on notification, claims pay over time, and the two patterns close.
- Paid is catching up faster than expected in recent diagonals. Settlements are coming in sooner than case-reserve adequacy assumed. Could be a settlement initiative or a softening tort environment. Either way, a reported-pattern projection is probably running too high.
- Reported is flat while paid keeps growing. Case-reserve runoff, offset by payments. The ultimate is barely moving; the composition of the reserve is just shifting toward paid. Usually fine.
- Reported jumps in a single calendar-year column without a matching paid jump. Reserve strengthening, often after a new claims leader takes over or after an internal reserve study. The question is whether the correction is a one-time event or the start of a trend.
Paid and reported are two views on the same book of claims. When they agree, you can rely on either projection. When they disagree, the gap is where the interesting questions live.
Rows, columns, diagonals
Reading a triangle is pattern recognition along three axes. The same data tells a different story depending on which direction you look.
- Rows (one accident year over time). Each row follows one accident year as it develops. You want a smooth, stable pattern: losses grow quickly at first, then slow as claims close and case reserves run off. A sharp jump or drop in a row relative to its neighbors is worth understanding. In the figure, the AY 2020 row climbs from 4,200 at 12 months to 8,150 at 72 months and is clearly flattening out, which is typical of a mid-tail line.
- Columns (one development age across accident years). Each column shows how the same point in development has behaved year after year. You want the age-to-age factors to be stable. If the 12-to-24 factor was 1.45 for AY 2020 but is 1.55 for AY 2024, something about recent experience has changed: severity, frequency, mix, or case reserving practice.
- Diagonals (one calendar year across accident years). Each diagonal is one calendar year of actual claim activity. Diagonals pick up calendar-year phenomena: a claims-system change, a settlement initiative, a reserving strengthening, an inflation shock. Calendar-year effects are the hardest to see in a routine chain ladder and the easiest to miss.
Age-to-age factors, in English
Below the triangle you will usually see a row of development factors: 1.45, 1.18, 1.08, 1.03, and so on. These are age-to-age factors. Each one is the ratio of one column to the previous column for a given accident year, then averaged across accident years in some way.
A worked example using the figure. AY 2020 went from 4,200 at 12 months to 6,090 at 24 months. That row’s 12-to-24 factor is 6,090 / 4,200 = 1.450. Do the same for AY 2021 (6,740 / 4,650 = 1.449), AY 2022 (7,400 / 5,100 = 1.451), AY 2023 (8,090 / 5,580 = 1.450), and AY 2024 (8,870 / 6,120 = 1.449). The 1.450 in the bottom row of the loss triangle is the volume-weighted average of those five row ratios.
Do that for every column and every accident year, and you get a triangle of factors of its own:
| Accident year | 12-24 | 24-36 | 36-48 | 48-60 | 60-72 |
|---|---|---|---|---|---|
| Age-to-age period (months) | |||||
| 2020 | 1.450 | 1.179 | 1.079 | 1.031 | 1.020 |
| 2021 | 1.449 | 1.180 | 1.081 | 1.030 | |
| 2022 | 1.451 | 1.178 | 1.080 | ||
| 2023 | 1.450 | 1.180 | |||
| 2024 | 1.449 | ||||
| Vol-weighted | 1.450 | 1.179 | 1.080 | 1.031 | 1.020 |
Age-to-age factors by accident year. Each cell is one column of the loss triangle divided by the previous column for that accident year. The bottom row is the volume-weighted average across accident years, the same set of factors shown at the bottom of the loss triangle.
The row factors cluster tightly around the column average: the five 12-to-24 factors all sit between 1.449 and 1.451; the four 24-to-36 factors run 1.178 to 1.180; the remaining periods are similarly tight. That clustering is what a stable, predictable line of business looks like, and it is the property that makes a chain-ladder projection trustworthy.
The usual averaging choices are:
- All-year simple average. Equal weight to every year.
- Volume-weighted average (also called dollar-weighted). Larger accident years count more. This is the most common default in software.
- Three-year, five-year, or ex-high-ex-low. Attempts to down-weight old or anomalous years.
None of these is inherently right. A good reserving report tells you which average was selected for each period and why. A report that does not tell you is asking you to trust the default in the software.
Selecting the factors
The averaging choices above are the mechanical menu. The harder question is which one to pick, and where to override the average with judgment. A few principles experienced reviewers apply:
- Use the longest stable history you have. All-year volume-weighted is a reasonable default. But if the book was bought from a different carrier seven years ago, or if a major program redesign happened five years ago, the data before the break does not belong in the average.
- Drop a year, do not silently down-weight it, when something genuinely broke. A claim-system conversion year, a one-time settlement push, a court reform that retroactively changed reserve adequacy: those years are noise, not signal. The ex-high-ex-low average is a blunt tool that often hides the real problem rather than fixing it.
- For the earliest periods (12-to-24, 24-to-36), trust the data less. Emergence is volatile when only a few months of activity have come in. Many reserve reviews replace the very first factor with an industry benchmark, or override the chain-ladder result with a Bornhuetter-Ferguson estimate that does not lean as heavily on immature data.
- For the latest periods (60-to-72 and beyond), trust the data less for a different reason. You have only one or two observations. A single large claim that closed unusually can swing a factor materially. Smoothing across recent periods, fitting a curve, or borrowing from industry data is usually more defensible than picking a raw average.
- Selected factors should bias toward stability, not toward the most recent year. Reserve estimates that move sharply year to year are usually doing the wrong thing. Let evidence accumulate across two or three quarters before re-selecting a factor that has drifted.
A report that publishes only the volume-weighted average without commentary on the selection is hiding the analyst’s judgment. The selected factor and the reason it differs from the mechanical average should be visible for every transition.
Diagonal to ultimate
Once you have a set of selected age-to-age factors, the rest of a basic chain-ladder reserve estimate is mechanical. The recipe has three steps.
- Multiply the selected age-to-age factors together, working backward from the last observed age, to get cumulative development factors (CDFs). The CDF at age N is the ratio of ultimate losses to losses observed at age N. For the figures here, CDF at 72 = 1.000 (no tail), CDF at 60 = 1.020 × 1.000 = 1.020, CDF at 48 = 1.031 × 1.020 = 1.052, and so on back through CDF at 12 = 1.450 × 1.339 = 1.942.
- Multiply each accident year’s diagonal value by its CDF to get projected ultimate losses for that year.
- Subtract the diagonal from ultimate to get the reserve. This is the total dollars still to be paid out after the evaluation date. It covers the payments on claims already in the case-reserve inventory, claims that will develop above their current case estimates, and IBNR on claims that have not yet been reported.
Applied to the loss triangle, using the volume-weighted age-to-age factors and assuming no tail beyond 72 months, the calculation looks like this:
| Accident year | Age (mo.) | Diagonal | CDF | Ultimate | Reserve |
|---|---|---|---|---|---|
| 2020 | 72 | 8,150 | 1.000 | 8,150 | 0 |
| 2021 | 60 | 8,850 | 1.020 | 9,027 | 177 |
| 2022 | 48 | 9,420 | 1.052 | 9,910 | 490 |
| 2023 | 36 | 9,550 | 1.136 | 10,849 | 1,299 |
| 2024 | 24 | 8,870 | 1.339 | 11,877 | 3,007 |
| 2025 | 12 | 6,710 | 1.942 | 13,028 | 6,318 |
| Total | 51,550 | 62,841 | 11,291 |
Diagonal multiplied by CDF gives Ultimate. Subtracting the diagonal from Ultimate gives the Reserve. CDFs are running products of the volume-weighted age-to-age factors above, working backward from age 72. For simplicity, this exhibit assumes a tail factor of 1.000 at 72 months; the next section explains why that is almost never right. All dollar columns in $ thousands.
The indicated reserve is $11.29 million on top of $51.55 million already paid, for a projected ultimate of $62.84 million. Most of the reserve sits in the youngest accident years. AY 2024 and AY 2025 alone account for $9.33 million of the $11.29 million total, because they have the longest way to go before they are mature.
This is a deliberately simple example. Real reserve reviews layer judgment on top of the mechanics: selecting averages other than volume-weighted, smoothing or replacing volatile factors, blending with a Bornhuetter-Ferguson estimate for green accident years, and, most importantly, adding a tail.
The tail
The exhibit above assumed a tail factor of 1.000 at 72 months. That is almost never right. For long-tailed lines there are payments still to come long after the last observed development period. The estimate for those payments is the tail factor, and it is where most of the subjective judgment in a reserve review lives.
In the loss triangle, the last observed factor (60-to-72) is 1.020. Development is slowing but not zero. A tail factor extends that pattern out to ultimate. Whether the right tail multiplier is 1.05 or 1.15 depends on the line of business and on judgment about how thick the remaining tail really is.
A tail factor multiplies straight through every projected ultimate. Plug a 5% tail (1.05) into the exhibit above and the projected ultimate goes up by 5% across the board, from $62.84 million to $65.98 million. But the reserve goes from $11.29 million to $14.43 million, an increase of about 28%. The diagonal does not change, so a small change in the tail leverages directly into a much larger change in the indicated reserve. That is why the tail gets so much attention in any serious review.
When reviewing a report, look for how the tail was selected: from industry curves, from a fitted curve extrapolating the observed pattern, from a judgmental pick, or from external benchmarking. Ask what the answer changes by if the tail factor changes by a reasonable amount. If a small change in the tail moves the indicated reserve by ten percent, that is important to know.
When the chain ladder is the wrong tool
The mechanical projection in the exhibit above assumes the future will look like the past. That assumption holds for stable books of business with consistent claim handling, no structural changes, and adequate maturity in every accident year you care about. When it does not hold, the chain ladder produces the wrong answer with full mathematical confidence, which is a dangerous combination.
The chain ladder breaks in a few specific situations, and a serious reserve review will name which one is in play:
- Green accident years. AY 2025 in the exhibit has only 12 months of data. Multiplying 6,710 by a 1.942 CDF amplifies whatever noise is in that 12-month value straight into the indicated reserve. A Bornhuetter-Ferguson method blends the chain-ladder projection with an a-priori expected-loss-ratio estimate, weighting the chain ladder more heavily only as the year matures. Most reports use BF for the most recent one or two accident years and the chain ladder for the older, more credible rows.
- Calendar-year shocks. When inflation, a court ruling, or a settlement push hits, the diagonal covering that calendar year carries elevated losses across every accident year at once. A chain ladder applied naively treats that elevation as part of the development pattern and projects it forward forever. The fix is a calendar-year trend adjustment, usually visible in the report as a separate inflation pick or trend factor layered on top of the chain-ladder result.
- Mix shifts. If the book grew rapidly in commercial auto, or contracted out of workers comp, the historical development pattern reflects the old mix, not the current one. The chain ladder cannot see the mix shift on its own. A segmented review by class, state, or program is more defensible than a single all-book triangle in this case.
- Case-reserve adequacy changes. A new claims leader who reserves more conservatively than the prior team will distort a reported triangle, even if the underlying losses have not changed. Paid triangles are less sensitive to this distortion, which is one of the practical reasons every serious review looks at both views.
- Soft data at the bottom-right. The factors for 60-to-72 and beyond are computed from only one or two observations. A single unusual claim in those cells can swing the indicated reserve materially. Sensitivity testing is the answer, not pretending the precision is real.
In each of these cases, the chain-ladder indication is still useful as a diagnostic, even when it is the wrong primary method. The pattern of disagreement between methods (chain ladder vs. BF, paid vs. reported, this year’s selected vs. last year’s) is often what tells you what is actually going on inside the book of business.
Diagnostic questions
Before you accept the projection that comes out of a triangle, ask:
- Is the development pattern stable from year to year in the columns you care about, and where it is not, what is the explanation?
- Do the paid and reported triangles tell the same story, and if not, which one is closer to the truth and why?
- Are there visible diagonals? If yes, what calendar-year event is driving them, and has the analyst adjusted for it explicitly?
- How was the tail factor selected, and how sensitive is the indicated reserve to a reasonable change in that selection?
- Are there accident years with too few data points to trust the mechanical projection? If so, which method took over, and what a-priori loss ratio was assumed?
- Which selected factors deviate from the mechanical average, and what is the analyst’s stated reason for each override?
- What does the indicated reserve from this method look like next to the prior year’s estimate, and does the analyst explain any movement beyond what new data alone would justify?
A reserving actuary will have an answer to each of these. If the answers are clear, specific, and supported by the exhibits in the report, the triangle is doing its job. If they are not, the number at the bottom of the report is not the number you should be relying on.