Why claims need a clock as well as a column
In the previous guide you met the problem in the abstract: a non-life insurer's largest liability is the money it still owes on claims that have not yet been fully paid, including claims that have happened but have not even been reported yet — the famous IBNR. The whole craft of reserving is putting an honest number on that pile. But before we can estimate it, we have to *see* it, and to see it we need to confront the awkward truth that makes the whole problem hard: a claim is not an instant. A car crashes today, but the injured passenger's medical bills, the lawsuit, the final settlement — those can take five, ten, fifteen years to fully resolve and pay.
So every claim carries two different dates, and you must never confuse them. One is *when the loss happened* — the accident year (the year of the crash, the fire, the injury). The other is *when the money about it moved or was recorded* — how far along the claim is in its slow journey to settlement. A single dollar of loss therefore has a fixed accident year but is paid out across many later calendar years. To make sense of the whole book, we cannot just total losses by the year they were paid, nor just by the year they happened — we need a table whose two axes are exactly these two clocks.
Building the triangle: accident year down, age across
Here is the construction, and it is wonderfully concrete. Make a grid. Each *row* is one accident year — 2021, 2022, 2023, and so on, stacked top to bottom. Each *column* is a development age: 12 months of maturity, 24 months, 36 months — how much time has elapsed *since the start of that row's accident year*. The number you write in each cell is the *cumulative* losses for that accident year, as seen at that age — that is, everything paid (or reported) on those claims from the start up to that point in their life. Read across a row and you watch a single year's claims grow up.
Now ask: why is it a *triangle* and not a full rectangle? Because of where we stand in time. Today is the end of 2024. The oldest accident year, 2021, has had four years to mature — we can fill its 12-, 24-, 36-, and 48-month cells. But accident year 2024 was just born; we have only seen its first 12 months, so only its leftmost cell is filled. Each newer row has had less time to develop, so each row is one cell shorter than the one above. The filled cells form a staircase — a right triangle — and the empty lower-right corner is precisely the future we have not yet lived through. That empty corner is the reserve we must estimate.
Reading a real (tiny) triangle
Let us make it real with the smallest honest example. Below is a cumulative paid-loss triangle for four accident years, in thousands of dollars, evaluated at the end of 2024. Each row starts small at 12 months and swells as more of the year's claims get paid; each row is shorter than the one above because younger years have had fewer evaluations. Take a moment to find the three features we just named: a row growing left to right, the columns lining up the same age, and the latest diagonal (the last filled number in each row) being our newest view of each year.
Cumulative PAID losses ($000), evaluated at end of 2024
Accident | Development age (months)
year | 12 24 36 48
---------+--------------------------------
2021 | 500 850 1020 1100 <- fully grown? (4 evals)
2022 | 520 900 1080 .
2023 | 560 960 . .
2024 | 600 . . .
^^^ \__ empty corner = the future
newest year, only 12mo seen = what we must estimate
Age-to-age (link) factor 12->24, using the column we can see:
(850 + 900 + 960) / (500 + 520 + 560) = 2710 / 1580 = 1.715
Reading it: a typical accident year's paid losses at 24 months are
about 1.715x what they were at 12 months. Apply such factors across
the empty cells, and each row marches toward its ULTIMATE value.Notice how the numbers in a row grow fast at first and then slow down — 500 to 850 is a big jump, but 1020 to 1100 is a gentle one. That deceleration is the signature of claims *maturing*: early on, lots of claims are reported and paid quickly; later, only the slow, complicated, litigated tail is still moving. The whole reserving exercise is really one question dressed up in a grid: for each row, how much further will it grow before it stops? The final, settled figure each row is heading toward has a name we will use constantly — the ultimate loss.
Age-to-age factors: the engine that moves a row forward
How do we predict the unwritten cells? We borrow from history. An age-to-age factor (also called a *link ratio*) measures, for a given step like 12-to-24 months, how much losses typically multiply as a year matures from one age to the next. The simple recipe: for the 12-to-24 step, add up all the 24-month figures you can see and divide by the sum of the matching 12-month figures. In our triangle that was 2710 / 1580 = 1.715, meaning losses at 24 months tend to be about 72% larger than at 12 months. Each step has its own factor, and they shrink toward 1.0 as the year matures and slows.
There are several honest ways to pick the factor for a step — a straight all-years average, a volume-weighted average (the column-sum recipe above, which lets bigger years count more), or an average of just the most recent few years if the pattern has been shifting. None is uniquely "correct"; choosing among them is judgement, not arithmetic, and it is one of the places where a reserving actuary earns their keep. Multiply a row's latest figure by its remaining age-to-age factors, one after another, and you march it across the empty cells all the way to its ultimate. Stringing these factors together is exactly the chain ladder method of the next guide — but the triangle and its factors are the bedrock under it.
One more honest caveat about the very edge of the table. Even at the oldest age we can observe — 48 months in our example — a long-tailed line like liability may *still* not be fully settled. The losses keep creeping up beyond the right edge of the triangle. To reach the true ultimate we therefore need a tail factor: an extra multiplier, often estimated from industry data or curve-fitting rather than read off our own short triangle, that carries the row from its last observed age out to true completion. Forgetting the tail is one of the most common ways young analysts under-reserve.
What kind of dollars? Paid versus reported, and the honest limits
We slipped past a choice that quietly shapes everything: *what number goes in each cell?* The two classic answers are paid losses (cash actually handed over so far) and reported (incurred) losses (cash paid plus the claims adjusters' case estimates of what is still owed on known claims). A paid triangle and a reported triangle of the same book look different and develop differently — the paid one starts lower and climbs longer; the reported one jumps quickly once claims are known, then can even *drift down* if adjusters were initially over-cautious. Mature actuaries build both and compare them; when the two tell different stories, that disagreement is itself information.
Now the most important honesty of the whole guide. The triangle is a *picture of the past projected onto the future*, and that projection rests on one heroic assumption: that older accident years are a fair guide to how younger ones will develop — that the pattern is stable. Reality often refuses. A change in how fast claims are settled, a court ruling that reshapes liability, a burst of inflation on repair costs, a shift in your mix of business — any of these can bend the development pattern, and a triangle that blindly applies yesterday's loss development factors will quietly mislead you. The grid is a tool for *seeing*, not an oracle.
What you carry forward
Hold the whole picture in one breath. Every loss has two clocks — *when it happened* (accident year, going down the rows) and *how far it has matured* (development age, going across the columns). Stack cumulative losses in that grid and the data you have forms a triangle; the empty lower-right corner is the future you must estimate, and that estimate *is* the reserve. Age-to-age factors read from the historical columns tell you how much a row typically grows from one age to the next, and chaining them carries each row toward its ultimate.
And carry the warnings as carefully as the method. The triangle assumes the past predicts the future; it does not, whenever settlement speed, the law, inflation, or your book of business changes — so read the raw development, do not just trust an average. Decide deliberately whether your cells hold paid or reported dollars. Remember the tail beyond the last observed age. With the triangle now in hand, the next guide turns it into a working estimate by formalising the chaining of factors — the chain ladder — and the one after that confronts the cases where the chain ladder alone is dangerously naive.