Link ratios from a paid triangle, simple, weighted and latest-two averages, a declared tail, and the age-to-ultimate factors that come out of them.
A loss development factor measures how much a cohort of losses grows from one age to the next. From a four-year paid triangle, the 12 to 24 month link ratios read 2.204, 2.195, 2.226; their simple average is 2.208, the weighted average 2.209 and the weighted average of the latest two years 2.211. With a declared tail of 1.050, the weighted factors multiply into an age-to-ultimate factor of 3.542 at 12 months.
Worked in full in The Loss Reserving Handbook by Julian R. Sterling, with every figure reproduced in a free workbook.Get the DRM-free PDF, $19.99 →
The words loss development factor are used for two related numbers, and mixing them up is a common source of error. The age-to-age factor, or link ratio, compares two neighboring ages of the same accident year: losses at 24 months divided by losses at 12 months, for instance. The age-to-ultimate factor, or cumulative development factor, takes a year from its current age all the way to its final cost. It is the product of every age-to-age factor from that age onward, times a tail factor for development beyond the triangle.
The first kind is measured on history. The second is built from choices: which average, which exclusions, which tail. That is why a reserve file shows both, side by side, with the reason for each choice.
A sample book, not the book's case. Cumulative paid losses by accident year, evaluated at December 31, 2024. All figures are teaching assumptions.
| Accident year | 12 months | 24 months | 36 months | 48 months |
|---|---|---|---|---|
| 2021 | $540,000 | $1,190,000 | $1,620,000 | $1,810,000 |
| 2022 | $575,000 | $1,262,000 | $1,731,000 | |
| 2023 | $602,000 | $1,340,000 | ||
| 2024 | $634,000 |
Each link ratio divides a cell by the cell to its left. Rounded to three decimals:
| Accident year | 12 to 24 | 24 to 36 | 36 to 48 |
|---|---|---|---|
| 2021 | 2.204 | 1.361 | 1.117 |
| 2022 | 2.195 | 1.372 | |
| 2023 | 2.226 |
Before averaging, read each column. The 12 to 24 month ratios move within a narrow band; the 24 to 36 ratios too. The single 36 to 48 ratio is all the triangle knows about that age. A ratio far from its neighbors is a question to ask the claims department, not a number to average away: a large claim, a change in payment practice or a coding error each leave a different trace.
The simple average gives each year the same weight. The weighted average divides the sum of the later cells by the sum of the earlier ones, so larger years count for more. The latest-years average keeps only the most recent links, which reacts faster to a change in the pattern and is more exposed to noise.
| Average | 12 to 24 | 24 to 36 | 36 to 48 |
|---|---|---|---|
| Simple average, all years | 2.208 | 1.366 | 1.117 |
| Weighted average, all years | 2.209 | 1.367 | 1.117 |
| Weighted average, latest two years | 2.211 | 1.367 | 1.117 |
Here the three averages nearly agree, which is a comfortable result. At 12 to 24 months the weighted average, 2.209, sits just above the simple one, 2.208, because the 2023 year, which developed fastest, is also the largest. The latest two years read 2.211, a little higher again. When averages disagree more than this, the disagreement is the finding: it says the pattern is changing, or that one year is unusual, and the analyst has to say which.
The example selects the weighted averages. That is a choice for the sample, written down as such. Real selections often exclude a declared link, a year with a known large claim for instance, and then say whether the excluded link is replaced by an older one or simply left out.
The triangle stops at 48 months, but paid losses on a liability line keep coming for years after that. The tail factor stands for all development beyond the last age observed. This example declares a paid tail of 1.050, an assumption of the sample.
A tail is never observed directly, so it needs support. Common supports include the ratio of reported to paid losses on the oldest year, which shows how much of the known cost is still unpaid; a reported tail, if the reported triangle has flattened, multiplied by that ratio; a curve fitted to the later factors, which would be judged on how well it fits and how far it is extrapolated; and outside reference patterns, declared as such. Whatever the support, the tail should be stated in the file, not left at 1.000 by default.
| Age | Age-to-ultimate factor | Built from |
|---|---|---|
| 12 months | 3.542 | 2.209 x 1.367 x 1.117 x tail |
| 24 months | 1.603 | 1.367 x 1.117 x tail |
| 36 months | 1.173 | 1.117 x tail |
| 48 months | 1.050 | tail |
The factor at 12 months, 3.542, says that a year at its first evaluation has paid, on this pattern, well under a third of its final cost. That is the factor a chain ladder would apply to the youngest year, and the one whose complement a Bornhuetter-Ferguson calculation would use as the unpaid share. Each product is computed from the rounded selected factors and rounded once to three decimals, so the table can be rebuilt by hand.
Paid factors are larger and longer than reported factors, because payments lag reports. They are free of case reserving judgment, which is their strength, and they react to any change in how fast claims are settled, which is their weakness. Reported factors are smaller, settle sooner and carry whatever changes in case reserve adequacy the claims department makes. A review usually builds both and reads one against the other.
Public disclosures use development data too. As read in November 2026, an actuarial white paper describes the short-duration contract disclosures of ASU 2015-09 as including incurred and paid claims development tables by accident year, net of reinsurance, for up to 10 years. Those tables are a starting point for an outside reader, not a substitute for the company's own triangles.
The book's case is fictional: Ashmerrow Mutual Insurance Company, a regional mutual writing commercial auto liability in its home state, reviewed at December 31, 2026 over accident years 2017 to 2026, gross of reinsurance, with ALAE kept out of the loss triangles. Every amount, premium, factor and ratio is a teaching assumption, not a market or industry figure.
Chapter 4 computes six averages for every age of the paid and reported triangles and selects the weighted average of the five latest links. Two links are excluded and declared: the reported 12 to 24 month link of accident year 2022, which reads 1.697 because it carries the report of a large claim, and the paid 48 to 60 month link of the same year, 1.366, which carries its payment. After that exclusion, the weighted average of the five latest reported 12 to 24 month links, which is the selected factor, reads 1.401. Chapter 5 declares an incurred tail of 1.005 and a paid tail of 1.043, and checks the second against the first times the reported to paid ratio of the oldest year at 120 months. Using the weighted average of all years instead of the five latest would lower the total unpaid claim estimate by $179,303; dropping both tails would lower it by $625,826.
Sources and limits: the standards named on this page are summarized as read in November 2026: Actuarial Standard of Practice No. 43, Property/Casualty Unpaid Claim Estimates (Actuarial Standards Board), and a Casualty Actuarial Society study text on basic techniques for estimating unpaid claims (Version 3, 2010) and an actuarial white paper on ASU 2015-09. They are paraphrased in principle and they can change. This is not actuarial advice and not a statement of actuarial opinion: every figure is a sample, and only a qualified actuary working on the company's own data can say which method and which factors fit it.
Divide cumulative losses at a later age by cumulative losses at the previous age for the same accident year. That gives an age-to-age factor; averaging those factors by column gives the factor selected for that age.
A simple average gives every year the same weight. A weighted average divides the sum of later cells by the sum of earlier cells, so larger years count for more.
Multiply the selected age-to-age factors from that age onward, then multiply by the tail factor. Rounding the product once, from the printed factors, keeps the exhibits reproducible.
From support such as the reported to paid ratio of the oldest year, a flattened reported triangle, a fitted curve or declared reference patterns. It should always be stated in the file.
This page uses none: every factor here is a sample. As read in November 2026, ASOP No. 43 leaves the choice of method to the actuary, and only a qualified actuary working on the data can say which factors fit.
This article is one step of The Loss Reserving Handbook. The book follows one regional insurer through one year-end review: a claims registry turned into paid, reported and count triangles, diagnostics read before any method, factors chosen with their reasons, two declared tails, the chain ladder, the expected loss ratio method, Bornhuetter-Ferguson and Cape Cod side by side, an estimate selected year by year with its range, ALAE and ULAE added once, the next year compared with what was expected, and a captive with a thin triangle.
Paperback PDF edition, $19.99 Book page and free files
Free: the loss reserving calculator in Excel.
Case figures computed by the book's model; the companies are fictional and every term is a teaching assumption. Figures checked on 21 November 2026. Standards described as read on the dates given; not actuarial advice.