GI Reserving Methods Explained: Chain Ladder vs. Bornhuetter-Ferguson
Two of the most widely used general insurance reserving methods approach the same problem from opposite directions. Here's how each one actually works, and why most books end up needing both.
Every general insurer carries a reserve for claims that have happened but haven't fully worked their way through the system yet — some not yet reported, some reported but not yet settled. Getting that reserve right shapes the balance sheet, the loss ratio, and increasingly, under IFRS 17, the timing of profit recognition itself. Two methods dominate how actuaries get there: the chain ladder method and the Bornhuetter-Ferguson method. They're often taught as a pair because they solve the same problem with opposite instincts.
The chain ladder method: let the data speak
Chain ladder is the older, more mechanical of the two. It starts from a loss triangle — claims for each accident or underwriting period, laid out across development periods as they mature. The method calculates development factors from that history: on average, how much do claims in month 12 grow by month 24, by month 36, and so on. Those factors get applied to the most recent, still-immature periods to project them out to an assumed "ultimate" value.
The appeal is that chain ladder makes almost no external assumptions. It doesn't ask an actuary to estimate an expected loss ratio or lean on pricing assumptions — it simply extrapolates from how claims in this book have historically developed. For long-tail, high-volume, reasonably stable lines with years of consistent claims history, that's a real strength.
The weakness shows up exactly where the data is thin. In the most recent accident periods — the ones with the least development observed — a small, early swing in reported claims gets multiplied by a large development factor. A single unusually large early claim, or an unusually quiet one, can swing the projected ultimate for that period dramatically. Chain ladder trusts the data completely, even when there isn't much of it yet.
Bornhuetter-Ferguson: blend the data with a prior view
Bornhuetter-Ferguson (BF) was built specifically to fix that immature-period problem. Instead of relying purely on development factors, BF starts with an independent, a priori estimate of the expected loss ratio for the period — usually derived from pricing, exposure, or industry benchmarks. It then splits the ultimate loss into two pieces: the portion already reported, taken at face value, and the portion still expected to emerge, calculated by applying the "percentage unreported" (the inverse of the chain ladder development factor) to that a priori expected loss.
The effect is that BF anchors immature periods to a stable, judgment-informed expectation rather than letting a handful of early claims drive the whole projection. As the period matures and more claims are reported, the method naturally leans more on actual experience and less on the prior — the reported portion grows, the "still to emerge" portion shrinks.
The trade-off is that BF is only as good as its a priori loss ratio. If that assumption is stale or badly calibrated — built on outdated pricing, or ignoring a rate change that just went through — BF will happily project a wrong number with a great deal of apparent stability. It trades chain ladder's volatility for a new, quieter kind of risk: confident-looking numbers built on a shaky assumption.
The short version: chain ladder trusts the triangle; Bornhuetter-Ferguson trusts a prior view and lets the triangle earn its way in as data matures. Neither is "more correct" — they're built for different amounts of data maturity.
Comparing the two side by side
| Chain Ladder | Bornhuetter-Ferguson | |
|---|---|---|
| Best suited for | Mature accident periods, stable long-tail lines | Immature periods, new products, thin volumes |
| Main input | Historical development factors only | Development factors + a priori expected loss ratio |
| Main risk | Volatile in early development periods | Sensitive to a poorly chosen prior |
| Behaves as data matures | No change in mechanics | Converges toward chain ladder as reported losses grow |
Why most reserving reviews use both
In practice, few actuarial teams pick one method for an entire book. It's far more common to run chain ladder on mature accident periods where there's enough development history to trust it, and BF on the newest one or two periods where chain ladder would otherwise be driven by noise. Some teams go a step further and blend the two methods across a transition band, so the reserve doesn't jump discontinuously as a period crosses from "BF-led" to "chain-ladder-led."
None of that is exotic — it's standard actuarial practice. What's harder in most insurers isn't picking the method, it's keeping the triangle itself current. Loss triangles are frequently rebuilt only at quarter-end, in a spreadsheet, from a claims extract someone had to remember to pull. Between refreshes, the reserve view management is looking at is already out of date.
Where automation actually helps
The value of automating reserving isn't replacing actuarial judgment — it's removing the lag between claims activity and the triangle reflecting it. When claims data feeds the triangle continuously rather than quarterly, development factors and BF priors can be recalculated as soon as new information lands, and any period where the two methods start to diverge sharply can be flagged for actuarial review immediately, rather than discovered at the next close.