The chain ladder method is the most widely used technique in non-life (property and casualty) insurance for estimating how much a book of business will finally cost once every claim has been reported and settled. It arranges historical paid or incurred claims in a claims triangle, measures how each cohort of claims has grown from one development age to the next, and applies those growth rates to the less mature cohorts to estimate their ultimate loss.
The method sits at the centre of claims reserving. Its output, ultimate losses minus amounts already paid, gives the outstanding claims reserve, and when it is applied to incurred data, the gap between ultimate and reported losses is the estimate of incurred but not reported (IBNR) claims. Actuaries rarely use it alone: it is usually compared with the Bornhuetter-Ferguson method and the expected loss ratio method, especially for recent years.
How the triangle is built
Claims are grouped by origin period, most often accident year but sometimes underwriting year or reporting year. Each row shows the cumulative amount paid or incurred for that origin year at successive development ages, such as 12, 24 and 36 months. Older years have more columns filled in; the latest year has only one, which gives the data its triangular shape. The diagonal of the triangle represents the latest valuation date, and each new valuation adds a fresh diagonal.
Paid triangles use only cash payments and are free of claims handlers' judgement, but they develop slowly. Incurred triangles add case reserves, so they mature faster but are affected by any change in how adjusters set reserves, a problem examined under case reserve adequacy.
Formula and worked example
The volume-weighted age-to-age factor from development age to age is:
- is the cumulative claims amount for origin year at development age .
- The sums run only over origin years that have values at both ages and .
- is the selected loss development factor for that step.
The ultimate for origin year , latest observed at age , is the latest diagonal value multiplied by all remaining factors, including any tail factor :
Take cumulative incurred claims (in millions) for three accident years: year 1 shows 100, 150 and 165 at 12, 24 and 36 months; year 2 shows 110 and 170; year 3 shows 120. The 12-to-24 factor is . The 24-to-36 factor is . Assuming no development after 36 months, year 1's ultimate stays at 165, year 2's ultimate is , and year 3's ultimate is . The IBNR estimate is therefore 17.0 for year 2 and 81.1 for year 3, a total of 98.1.
Selecting factors in practice
Real triangles hold ten or more years and many columns, and the factors are not simply calculated but selected. Actuaries compare volume-weighted averages, simple averages, averages of the latest three or five years and averages that exclude the highest and lowest values. Large single losses or catastrophe events may be removed and reserved separately so they do not distort the pattern. Where the triangle stops before claims are fully settled, a tail factor is chosen from curve fitting or industry benchmarks.
The method can be run on claim counts as well as amounts, and on gross or net-of-reinsurance data. Results feed the reserve booked in the financial statements, regulatory returns and capital models, and the roll-forward of those estimates between full reviews.
Assumptions and weaknesses
The chain ladder rests on one central assumption: that future development will follow the pattern of the past. Several things break that assumption:
- Changes in claims handling. Faster settlement or a change in case reserving practice alters the shape of the triangle and leads to over- or under-projection.
- Claims inflation. A sudden rise in repair, medical or legal costs is not captured by factors calculated in a calmer period.
- Leverage on immature years. The latest year's estimate multiplies a small, volatile number by a large cumulative factor, so a single early claim can swing the result.
- Changes in mix. Growth in a new line or territory changes the development pattern of the portfolio as a whole.
These weaknesses explain why the Bornhuetter-Ferguson approach is often preferred for recent origin years. The chain ladder is also deterministic: it produces a point estimate without a measure of uncertainty. Thomas Mack's 1993 distribution-free model gave a way to estimate the standard error of chain ladder reserves, and bootstrapping and other stochastic methods are used to produce a distribution of outcomes for capital modelling and for the risk adjustment under IFRS 17.