Where the chain ladder breaks its leg
In the previous guide you saw the chain ladder take a loss development triangle and march each accident year up to its ultimate by multiplying through a chain of development factors. It is elegant, and for mature years it works beautifully. But notice what it actually *does*: it takes whatever a year has reported so far and scales it up. The reported figure is the whole foundation; the factors merely magnify it.
For a brand-new accident year, that is dangerous. Suppose a year is only three months old and a single large claim has come in. Its loss development factor might be 8.0 — meaning eight-ninths of the eventual losses have not yet appeared. The chain ladder will multiply that one early claim by 8.0 and treat the answer as the ultimate. One unlucky claim, or one quiet quarter, and the estimate swings wildly. The leverage that makes the chain ladder powerful for old years makes it hysterical for green ones.
Starting from what pricing told you: the expected loss ratio
Begin with the simplest of the blending family, the expected loss ratio (ELR) method. The idea is almost rude in its bluntness: ignore the year's reported losses entirely, and estimate its ultimate from premium alone. You take the earned premium for the year and multiply it by a loss ratio you expected up front — the a-priori loss ratio — typically the permissible loss ratio baked into the rates.
So if a year earned 10 million in premium and your pricing assumed a 65% loss ratio, your estimate of ultimate losses is simply 6.5 million — no matter what the triangle says so far. This is wonderfully stable: a single freak claim cannot budge it, because the claim is not even in the formula. But that stability is also its sin. The ELR method is *completely deaf* to experience. If the year is clearly turning into a disaster, with losses already pouring in far above plan, the pure ELR method shrugs and still says 6.5 million. It trusts the prior absolutely and the data not at all.
So we have two extremes facing each other. The chain ladder believes the data and nothing else; the ELR method believes the prior and nothing else. Neither is right for a half-grown year, where you genuinely have *some* signal in the data but not enough to trust it alone. The honest answer must live somewhere in between — and finding that middle is exactly the Bornhuetter-Ferguson idea.
Bornhuetter-Ferguson: trust the emerged, predict the rest
The Bornhuetter-Ferguson (BF) method makes one beautifully simple move. It splits a year's ultimate losses into two parts — what has *already emerged* and what is *yet to come* — and treats each part with the tool it deserves. For the losses already reported, it simply takes them as a fact: they happened, you can see them, no estimation needed. For the losses *not yet* reported, it does not extrapolate from the thin data; instead it predicts them from the a-priori expectation.
How much is 'yet to come'? That is where the development pattern returns, but used backwards. If the chain ladder says a year is 1/8 developed — its development factor to ultimate is 8.0 — then 1/8 of the losses are reported and the remaining 7/8 are still in the future. So BF says: ultimate = (losses reported so far) + (a-priori expected losses) × (the 7/8 still unreported). The reported portion comes straight from the data; the unreported portion is filled in from pricing, scaled by exactly the fraction of development still outstanding.
Green year: earned premium = 10,000,000
a-priori loss ratio = 65% -> expected ult = 6,500,000
reported losses so far = 900,000
chain-ladder dev factor = 8.0 -> % developed = 1/8 = 12.5%
% UNreported = 7/8 = 87.5%
Chain ladder ult = 900,000 * 8.0 = 7,200,000 (volatile!)
ELR ult = 6,500,000 = 6,500,000 (deaf)
BF ult = 900,000 + 6,500,000 * 0.875 = 6,587,500 (blend)
BF reserve (IBNR) = 6,500,000 * 0.875 = 5,687,500Notice what falls out of the formula at the two ends of a year's life. When a year is brand new and essentially 0% developed, the unreported fraction is ~100%, reported losses are tiny, and BF lands almost exactly on the ELR answer. When a year is fully mature and 100% developed, the unreported fraction is 0, the prior term vanishes, and BF lands exactly on the actual reported losses — which is also where the chain ladder lands. BF is a *credibility-weighted bridge*: it leans on the prior when there is little data, and slides smoothly toward the data as the year ripens.
Why this is so much steadier
The stability comes from where the leverage sits. In the chain ladder, your one observed number gets multiplied by a big factor, so any wobble in that number is *amplified*. In BF, the big chunk of the answer — the unreported part — is pinned to the prior, which does not move with this year's noise at all. Only the small, already-emerged piece responds to the data, and it responds one-for-one, not magnified. Less of the answer is exposed to the volatile early signal, so the answer holds steadier from one valuation to the next.
There is a clean way to say it. The chain ladder estimates the *unreported* losses as (a multiple of) the reported ones; BF estimates the unreported losses from the prior. So a useful rule of thumb: use the chain ladder for old, settled years where the data is rich and trustworthy, and use BF for the newest one or two years where the data is too thin to lead. Many reserving exercises do exactly this — chain ladder on the bottom of the triangle, BF along the top, sometimes blending the two year by year.
Cape Cod: let the data choose the prior
BF has one soft spot: where does the a-priori loss ratio come from? In the pure method, you hand it in by judgement — usually the pricing assumption. But what if pricing was simply wrong, the same way across every year? Then BF carries that error faithfully into every reserve. The Cape Cod method (also called the Stanard-Bühlmann method) fixes this by *estimating the a-priori loss ratio from the triangle itself*, rather than importing it.
The trick is gentle. Instead of comparing each year's reported losses to its full premium — which would be unfair, since young years have only reported a sliver — Cape Cod compares total reported losses across all years to the *used-up* premium: each year's earned premium scaled down by how developed it is. A year that is only 1/8 developed contributes only 1/8 of its premium to the denominator, so it is weighted to match how much of its losses you can actually see. The resulting ratio is one blended a-priori loss ratio, derived from the data, which you then feed into the BF machinery.
So you can picture the family as a single dial. Turn it all the way to 'trust the prior' and you get the ELR method. Turn it all the way to 'trust the data' and you get the chain ladder. BF sits in the middle, weighted by maturity. Cape Cod is BF with the prior itself learned from the pooled experience instead of asserted. None of them is the 'true' reserve — they are four honest readings of the same fog, and a good reserving actuary looks at all of them before committing to a number.
Honest limits
BF is steadier, not omniscient. It still borrows the *development pattern* from the chain ladder — the fraction reported at each age — so if that pattern is wrong (the business is settling faster or slower than history), BF inherits the error in how it splits reported from unreported. It only protects you from the leverage on the reported number, not from a mis-estimated development curve, nor from a tail factor that is too short.
There is also a real risk of double comfort. If your a-priori loss ratio came from the same pricing study that, deep down, leaned on the same old triangle, then BF's 'independent prior' is not so independent — you may just be agreeing with yourself. Guard against this: the prior should reflect genuinely external information (rate changes, exposure trends, on-leveling) and not be a quiet echo of the data it is supposed to balance. And, as always in reserving, every one of these methods produces a *point estimate*. The genuine uncertainty around it — the range a real reserve could plausibly land in — is the subject of the next guide on stochastic reserving.