From a pattern to a prophecy
In the previous guide you built a loss development triangle: accident years stacked down the left, development ages running across, and inside it the slow blossoming of losses as claims get reported, re-estimated, and paid. You also learned to read the age-to-age factors — the ratios that say 'losses at age 24 months are typically 1.15 times what they were at age 12.' That triangle is a *description* of the past. The chain-ladder method turns it into a *prediction* of the future: it takes each accident year, still only partly grown, and walks it forward along the average growth pattern until it reaches its full, settled size — its ultimate.
The name is a vivid picture. Each age-to-age factor is one *rung*; multiply a year's current losses by the rung from where it stands, then the next rung, then the next, and you have climbed the *ladder* — a chain of multiplications carrying you from today's immature figure to the ultimate at the top. Subtract what has already been paid (or, for a reported triangle, the reserve you still hold), and the remainder is the money you must set aside: the unpaid claim estimate, the heart of which is the IBNR for claims still to emerge.
Walking a year up the ladder
Let us do it once, with real arithmetic, on a tiny triangle. Suppose from your historical data you have *selected* development factors of 1.50 from age 12 to 24 months, 1.20 from 24 to 36, and 1.10 from 36 to 48, after which you believe claims are fully settled. Multiply those together and you get the cumulative factors — the single number that carries a year all the way from a given age to ultimate. From age 12, that is 1.50 × 1.20 × 1.10 = 1.98: a year seen at 12 months is, on this pattern, only about half-grown.
Acc. Dev age (cumulative paid) Cum. Ultimate Reserve
year 12mo 24mo 36mo 48mo factor = latest x f = ult - latest
--------------------------------------------------------------------------
2021 1,000 1,500 1,800 1,980 1.000 1,980 0 (done)
2022 1,100 1,650 1,980 . 1.10 2,178 198
2023 1,200 1,800 . . 1.32 2,376 576
2024 1,300 . . . 1.98 2,574 1,274
--------------------------------------------------------------------------
selected age-to-age: 1.50 1.20 1.10 (tail 1.00)
cumulative to ult: 1.98 1.32 1.10
Total reserve (IBNR + future paid) = 198 + 576 + 1,274 = 2,048Read accident year 2024: only its 12-month figure of 1,300 is known, so multiply by the full cumulative factor 1.98 to get an ultimate of 2,574. It has paid 1,300, so 1,274 is still owed. Year 2023, seen at 24 months, needs only 1.20 × 1.10 = 1.32, and so on. Each younger year leans more heavily on the factors and less on its own thin data — which is exactly why the newest year carries the most guesswork. Add the four reserves and you have developed the whole triangle: this is developing losses to ultimate, and the 2,048 is the loss-development-method estimate of total unpaid claims.
The word *selected* is doing quiet, important work. The triangle hands you several historical age-to-age ratios at each step, and they never agree perfectly. The actuary chooses one — a simple average, a volume-weighted average, a five-year average, the average excluding an odd year, or a number nudged by judgment when a court decision or a process change has made the recent past untypical. Chain-ladder is mechanical once the factors are chosen, but *choosing* them is the craft, and two honest actuaries can land on different reserves from the same triangle.
Beyond the data: the tail factor
Notice the quiet assumption hiding in the example: that claims were fully settled by 48 months. For short-tailed lines — a windscreen claim, a fender bender — that is roughly true. But for long-tailed casualty lines such as liability, workers' compensation, or asbestos, claims keep crawling upward for ten, twenty, even forty years, long past the right edge of any triangle you can build. Your triangle simply *stops*, but the losses do not. The piece of development that happens after the last column your data can see is captured by the tail factor.
A tail factor is just one more rung on the ladder — a single multiplier (say 1.05) that you tack onto the end to cover all development from the last observed age out to true ultimate. Where does it come from, if the data stops short? From industry development patterns, from fitting a smooth curve to the factors you *can* see and extrapolating it, or from a benchmark study. Every one of those is a borrowed assumption, not a measurement, so the tail is a place where reserving leans most on judgment and where it can be most badly wrong.
Two triangles: paid and reported
Chain-ladder does not care what you put inside the cells — only that they grow on a stable pattern. So actuaries almost always run it on *two* different triangles and compare. The paid triangle holds only cash actually disbursed; it is clean and tamper-proof (a cheque is a cheque) but slow, because the biggest claims pay out last. The reported (or *incurred*) triangle holds paid losses *plus* the adjusters' case reserves — their best estimate of what each open claim will eventually cost. It reacts faster, but it inherits all the optimism, pessimism, and reserving habits of the claims department.
When the two answers agree, you sleep well. When they diverge, the gap is a story worth reading, and learning to read it is a real reserving skill — the heart of the paid-versus-reported diagnostic. If the reported ultimate sits well above the paid one and stays there, claims may be reserved conservatively. If reported losses are *falling* down a column — case reserves being released as claims settle for less than feared — your incurred factors may even dip below 1.00. And if adjusters quietly tightened or loosened their case-reserving standard a few years ago, the reported triangle's pattern shifts under your feet while the paid triangle keeps its older rhythm.
The one assumption — and where it breaks
Strip the chain-ladder method to its bones and one assumption holds the whole thing up: that *every accident year develops along the same proportional pattern*, regardless of how large the year is. The future development of a year depends only on its current cumulative amount and the selected factors — not on which year it is, not on how the losses got there. That is a strong claim, and the method's famous weakness is that it believes it utterly. Multiply a wrong number by a factor and you get a confidently wrong answer.
This is why the method goes spectacularly wrong on the *youngest* accident year. That year is multiplied by the largest cumulative factor (1.98 in our example), so it has the most leverage — yet it rests on the least data and the most random first diagonal. One unusually large early claim, or one unusually quiet quarter, gets blown up nearly twofold. A green year with thin, noisy data is precisely where chain-ladder is least trustworthy, and precisely where the Bornhuetter-Ferguson method in the next guide will steady the wheel by leaning on a priori expectations instead.
The assumption also shatters whenever the past stops resembling the future. A change in claims-handling speed, a shift in the mix of business, runaway inflation on medical or repair costs, a court ruling that re-prices old claims, a one-off catastrophe sitting inside an otherwise ordinary year — each one bends the development pattern, and a method that assumes the pattern is fixed will march straight off the cliff. The honest reserving actuary therefore never treats chain-ladder as an oracle. They run it, run a paid and a reported version, read the diagnostics, ask what changed, and treat the answer as one informed estimate among several.