Loss Development Factors vs. Bornhuetter-Ferguson: When to Trust Your Own Experience

The choice between the loss development factor method and the Bornhuetter-Ferguson method is not a matter of preference — it is a question of how much credibility your own experience deserves relative to an external prior. Actuaries who default to LDF on every line of business without examining the credibility of their development patterns are making a systematic error, usually in the direction of overconfidence in thin data.

The consequence is reserve volatility. LDF projections on immature accident years amplify whatever random variation exists in the early development data. BF projects attenuate that noise by anchoring part of the projection to the a priori expected loss ratio. The tradeoff is explicit: LDF is fully responsive to your own experience; BF blends your experience with an external benchmark. Which is better depends entirely on how credible your experience is.

The Mathematics of the Tradeoff

The LDF method projects ultimate losses by multiplying the current reported (or paid) losses at each accident year by the cumulative development factor from the current maturity age to ultimate. The development factors are estimated from your historical development patterns — typically a volume-weighted average of two to five prior accident years, depending on your actuarial selection criteria.

The Bornhuetter-Ferguson method starts from an a priori expected ultimate loss ratio — typically the current year's prospective rate level loss ratio, benchmarked against industry experience for your lines of business. It then estimates unreported losses as the product of the expected ultimate and the percent unreported (one minus the reciprocal of the cumulative development factor). The ultimate is the sum of reported losses plus this independently estimated unreported component.

The key insight is that BF behaves like LDF when the book is large and development patterns are stable — in that regime, the a priori expected loss ratio converges to what the LDF projection implies, so the methods agree. BF diverges meaningfully from LDF when development is volatile, when the book is small, or when there is a structural break in the development pattern. In all three cases, the BF estimate is more stable, at the cost of being slower to respond to genuine experience changes.

Credibility and Volume: The Governing Factor

The actuarial concept of credibility provides a principled framework for choosing between the two methods. A book with 500 workers compensation claims per accident year, all with clean development patterns, has high credibility — the LDF from that data is reliable. A book with 50 commercial umbrella claims per accident year, with development patterns that vary substantially from year to year, has low credibility — the LDF from that data may be statistically unstable.

The FCAS examination syllabus covers credibility-weighted development factor selection in the context of the Cape Cod method, which provides an alternative a priori loss ratio estimate based on actual book experience. CPCU curriculum takes a more practical approach, but the underlying principle is the same: when your data is thin, blend toward an external benchmark. When your data is credible, let your own experience dominate.

A useful heuristic: if your cumulative development factors at 12-to-24 months are varying by more than 15% across accident years, you do not have enough data to use LDF alone. The BF method — or a credibility-weighted blend of LDF and BF — will produce more stable projections without materially compromising responsiveness to genuine trend changes.

Structural Breaks: When BF Can Also Mislead

BF is not immune to error. The a priori expected loss ratio is the method's core assumption, and if that assumption is wrong, the BF projection inherits the error. In a rapidly hardening market, where premium increases are outpacing loss cost trends, the prior year's expected loss ratio will overestimate losses on the current book — and the BF estimate will be higher than the LDF estimate, which will be tracking the improving experience directly.

This is the situation where actuaries need to update the a priori explicitly, rather than letting it carry forward from the prior year's filing. The a priori loss ratio should reflect current rate adequacy, current trend assumptions, and any structural changes in the mix of business that affect expected frequency or severity. Using a stale a priori in a market where rates have moved significantly is worse than using LDF, because it introduces a systematic bias rather than random noise.

The practical implication: review the a priori assumption with the same rigor you apply to the development factor selection. It is not a constant — it is a judgment that needs to be updated when conditions change. Documenting the basis for the a priori, and the sensitivity of the BF projection to changes in that assumption, is actuarial practice standard for good reason.

Workers Comp vs. Commercial Auto: Two Different Regimes

The appropriate method varies substantially by line of business, and the workers compensation / commercial auto comparison is instructive. Workers compensation has a very long development tail — medical-only claims develop over three to five years, while permanent total disability claims can develop over decades. LDF on workers comp requires age-to-age development factors through high maturity ages, and the uncertainty in those tail factors is large for any single carrier.

The standard practice for workers comp reserving is to blend carrier-specific development patterns with industry benchmarks (available from the NCCI or state rating bureaus for monopolistic-state funds) using a credibility weighting based on developed premium volume. A carrier with $20M in workers comp premium typically does not have enough history to support independent tail factor selection beyond the 48-to-60 month maturity range — the blend with NCCI benchmarks at later ages is both standard practice and actuarially justified.

Commercial auto bodily injury develops faster — most claims close within 24 to 36 months, with the exception of catastrophic injury claims that involve litigation. The development tail is shorter, and the LDF method works adequately on books with moderate volume. The FNOL severity triage discussed in our article on FNOL data can further stabilize commercial auto development by improving initial reserve accuracy on high-severity claims.

Practical Development Triangle Construction

The mechanics of constructing a development triangle for LDF selection are well-established, but the judgment calls within that process are where actuaries introduce meaningful variation. The key decisions are: (1) how many accident years to include in the weighted average, (2) whether to use paid or incurred losses as the basis, and (3) how to handle outlier accident years with anomalous development patterns.

Volume-weighted averages give more weight to higher-premium accident years, which is generally appropriate because they produce more statistically stable development experience. Simple averages treat all years equally, which can amplify the influence of a small recent accident year with random development volatility. Most actuaries use volume-weighted averages as the starting point and apply judgment adjustments when the result is implausible given known book changes.

The paid vs. incurred decision matters when reserve practices have changed. If your claims department implemented a more conservative initial reserve standard three years ago, the incurred development at early maturities will look different from the historical pattern, and using older accident years in your factor selection will produce development factors that overstate the required development on current business. This is a case where segmenting the development triangle by reserving practice era — rather than averaging across it — gives more accurate factor estimates.

Integrating with Pricing Models

Reserve projections and pricing models feed the same combined ratio forecasting process. The reserve actuary's selection of ultimate losses by accident year, and the pricing actuary's estimate of expected losses on in-force business, need to be consistent with each other to produce a coherent combined ratio projection. This sounds obvious, but in practice the two functions often work from different assumptions about trend, development, and rate adequacy.

The connection to underwriting risk scoring is direct: a submission scoring model that predicts loss cost at the policy level needs to be calibrated against the same set of actuarial projections that the reserving team uses. If the reserving team applies a 1.08x development factor to commercial auto accident year 2023, the scoring model's estimated loss cost for commercial auto submissions should implicitly reflect that same development expectation. A disconnect between the model's pricing assumptions and the reserving team's projections is a consistency failure that will show up as unexplained variance in your loss ratio analysis.

Conclusion

LDF and Bornhuetter-Ferguson are tools with defined use cases, not competing philosophies. LDF is appropriate when your development data is credible and patterns are stable. BF is appropriate when data is thin, patterns are volatile, or there is reason to doubt that recent experience is representative of the underlying risk. In most real books, you will use both: LDF on mature lines with stable development, BF on new lines or recent accident years where credibility is limited. The mistake is applying either method uniformly without examining the credibility conditions that determine which is appropriate.