RESEARCH DOSSIER
Bornhuetter-Ferguson Claims Reserving
Abstract. The Bornhuetter–Ferguson (BF) method estimates ultimate insurance claims by combining observed claim emergence with an independently selected expectation of the portion yet to emerge. It is a foundational technique for property and casualty reserving, especially where experience is immature, volatile, or sparse. This monograph reconstructs its historical argument from Bornhuetter and Ferguson (1972), establishes precise paid and incurred formulations, derives its relationship with chain-ladder and expected-loss methods, and tests the interpretability of emergence factors, prior expected loss, valuation data, and exposure definitions. Reproducible numerical examples address accident-year ultimates, unpaid amounts, calendar-year changes, a multi-cohort illustration, prior bias, development-pattern stress, and the distinction between estimated ultimate claims and financial-reporting liabilities. A deliberately specified Poisson model isolates the conditions under which the method's variance advantage disappears when its prior is biased; the calculation is an analytical illustration rather than a new reserving theory. The synthesis incorporates standard actuarial guidance, subsequent credibility and stochastic literature, and the distinct requirements of US actuarial practice, IFRS 17, and Solvency II. The central conclusion is that BF's apparent stability is earned through a substantive prediction about future claims: the quality of its reserve depends simultaneously on the independently supportable prior, the expected development pattern, the current emergence measure, and the purpose for which the estimate is used.
Keywords: Bornhuetter–Ferguson; loss reserving; claims reserves; incurred but not reported; IBNR; paid loss; incurred loss; expected ultimate claims; development factors; chain ladder; expected loss ratio; insurance triangles; reserving bias; model uncertainty; actuarial standards; IFRS 17; Solvency II.
Introduction
Insurance losses are incurred before they are fully known. A claim can be reported long after its occurrence, reported promptly but paid over many years, or initially valued at an amount that changes as information accumulates. Financial statements and managerial decisions therefore need estimates of obligations that are still emerging through a claims process. In a portfolio with numerous reasonably homogeneous claims, historical development patterns can be informative. In a new, rapidly growing, long-tailed, or highly volatile portfolio, recent reported or paid amounts can be an unstable foundation for extrapolating the total.
Ronald L. Bornhuetter and Ronald E. Ferguson's 1972 analysis of the actuary and IBNR identified a central conceptual problem: a currently observed claim amount and the expected remaining claims are different pieces of information. Their procedure values the part not yet emerged using an independently assessed expected ultimate amount and a factor representing the undeveloped proportion. The original paper explicitly considers sparse experience and long-tail business and treats adverse or favorable development on reported claims as part of the wider provision for incurred but unresolved losses (Bornhuetter and Ferguson, 1972). The later CAS basic reserving textbook provides a sustained educational treatment of the development, expected claims, BF, Cape Cod, and diagnostic techniques (Friedland, 2010).
A methodological account must identify its estimand. Let ultimate claims for cohort i mean the final nominal amount incurred by that cohort, including the specified claim expenses and recoveries only if the defined loss measure includes them. Let observed cumulative development mean the amount recognized at an evaluation age on a specified paid or incurred basis. The actuarial unpaid amount is ultimately incurred claims less the cumulative paid amount, regardless of whether an incurred development triangle was used to estimate ultimate. The monetary unit, origin period, coverage perimeter, gross/net convention, accounting date, and claims expense definition must remain explicit.
The US unpaid-claim estimating standard makes the distinction between an actuarial unpaid-claim estimate and an amount booked as a financial statement reserve, while requiring consideration of the intended measure and appropriate methods (ASB, ASOP 43). The BF method supplies an actuarial estimate within those choices; it does not itself assign accounting recognition, risk margin, discount rate, or regulatory capital. This monograph considers global practice while stating the precise jurisdiction of each cited professional or reporting rule.
The analysis is deductive and documentary. It checks historical primary literature and official standards, reconstructs the classical calculations, and presents synthetic examples reproducible from stated values. It makes no claim to have analyzed confidential insurer data, generated independent empirical calibration, performed external peer review, or introduced a new original reserving doctrine.
Expected Ultimate Loss Assumptions
Historical development and the estimand
The traditional development method projects the currently observed aggregate to ultimate using an empirically selected age-to-ultimate factor. That procedure implicitly lets the current observation determine the size of all future claims. A single large early claim can then determine an implausibly large ultimate for an immature origin period. The traditional expected-claims method, by contrast, estimates a cohort's total liability before considering its emergence. BF retains current observation as observed and predicts only the portion that remains, thereby separating the observed and unobserved components.
The original 1972 reserving study emphasizes that prospective claims and development on already reported claims create complex IBNR obligations, especially in casualty excess and reinsurance. The paper's original exhibit uses a development factor to calculate the proportion still unreported, then multiplies that fraction by expected ultimate claims. The phrase IBNR in historical usage can include changes in the estimated values of reported claims; contemporary analytics should identify when true unreported claims and incurred-but-not-enough-reported (IBNER) are combined.
Write E_i for the selected prior expectation of ultimate claims for origin period i; C_i for cumulative claims observed on a chosen development basis; and p_i for the expected fraction of ultimate that has emerged by the valuation date, with 0 < p_i ≤ 1. The BF ultimate estimate is
The additional amount beyond the observed development measure is (1-p_i)E_i. It is tempting to describe E_i as a forecast with fixed authority. In fact, E_i is an actuarial assumption subject to loss-cost trend, exposure, coverage and rate changes, calendar effects, claims inflation, changes in limits, and selection of historical reference years.
Constructing the prior expected ultimate
For a premium-based approach, an initial ultimate expectation can be expressed as E_i=P_i r_i, where P_i is an appropriate premium measure and r_i an anticipated ultimate claims ratio. A premium base must be on-level when rates have changed; an expected ratio based on earned premiums of a different rate level, expense convention, or coverage mix is not portable without adjustment. Gross and net earned premiums may imply quite different ultimate claims proportions where reinsurance structures change. Premium cannot be confused with the volume of exposure when rate level or premium adequacy differs sharply across periods.
An exposure-based approach is E_i=N_i q_i s_i, where N_i is exposure units, q_i an expected claim frequency per unit, and s_i an expected severity conditional on claim. Frequency and severity need consistent units and gross/net definitions; an appropriate extension may stratify by perils, claim sizes, attachment points or geography. This decomposition can reveal which portion of an expected loss ratio change originates in exposure, propensity to claim, or claim cost. It remains an estimate, not an independent fact of future claims.
Expected ultimate claims may be calibrated from older mature origin periods, policy pricing, external benchmarks, underwriting studies, or detailed individual-claim modeling. The CAS reserving framework treats selection of expected claims and selection of development as distinct analytical tasks (Friedland, 2010). The expected-loss and BF comparison underscores that a credible prior is especially significant for immature experience. An independently selected prior can be informed by earlier mature experience, but repeated use of the same immature observations in both E_i and C_i may undermine any claim of independent evidence.
Selection should document the valuation basis, source of the experience, adjustments for inflation and rate changes, historical credibility, loss cost trend, changes in underwriting or claim reporting, the ultimate claim cost definition, and reasons that the selected assumption remains plausible. A prior is operationally credible when its provenance and expected performance can be explained and challenged. Statistical precision alone cannot settle its relevance in a changed portfolio.
Independence, exposure, and operational judgment
BF is frequently presented as resistant to early randomness. That description is conditional on the prior and the development pattern being useful. If the original expected ratio was selected using the same observed losses, the purported stabilization may conceal double use of information. If recent business includes a materially different deductible, coverage limit, legal jurisdiction, or concentration, the past prior may fail even while the observed triangle appears numerically smooth.
An especially important distinction concerns the observational base. In a paid BF calculation, C_i is cumulative paid; E_i is expected ultimate paid loss on the same ultimate basis, and p_i is the expected paid proportion. In an incurred BF calculation, C_i is reported incurred (paid plus case outstanding), and p_i is the expected reported-incurred emergence proportion. The case outstanding amount is already incorporated in reported incurred and must not be added again. When an insurer's case reserving policy changes, p_i for the incurred triangle can change without an equivalent change in the economic ultimate.
The loss reserve adequacy investigation by Berquist and Sherman (1977) remains central to the diagnosis of changing claims settlement and reserving operations. Claim closure speed and the level of case estimates can affect development factors. Their paper argues for knowledge of the underlying insurance operations in interpreting reserves, an observation that applies forcefully to BF because a structurally distorted p_i becomes a direct input to future expected losses.
Unreported and Emerged Loss
Paid, reported, case, and ultimate amounts
Suppose losses of a cohort have ultimate U, cumulative payments P, and a reported-incurred valuation I=P+K, where K denotes the case outstanding reserve recognized for reported claims. The ultimate unpaid amount is U-P. Under an incurred analysis, the additional required amount beyond reported incurred is U-I. Under a paid analysis, all unpaid claims, including already known case amounts, remain to be estimated from the cumulative paid measure.
For clarity, consider a synthetic portfolio with paid claims of $4.0 million, reported incurred of $7.0 million, an independently determined expected ultimate of $10.0 million, 40% estimated paid emergence, and 70% estimated reported-incurred emergence. The paid BF ultimate is 4.0+(1-0.40)×10.0=$10.0 million; the paid BF unpaid amount is $6.0 million. The reported-incurred BF ultimate is 7.0+(1-0.70)×10.0=$10.0 million; the development beyond recorded incurred is $3.0 million. Since case outstanding equals 7.0−4.0=$3.0 million, the latter approach implies unpaid 3.0+3.0=$6.0 million. Both paths reconcile to the same ultimate in this deliberately aligned example.
Now suppose case incurred is $7.3 million while all other inputs remain unchanged. Incurred BF gives U=7.3+3.0=$10.3 million, and an unpaid estimate of $6.3 million after subtracting paid $4.0 million. Paid BF remains at $10.0 million. The disagreement is a prompt to investigate the case reserve level, development patterns, and expected loss base; it does not establish that one accounting number is automatically superior.
A reserving table must identify whether “IBNR” means true claims not yet reported or the entire amount beyond reported incurred, including development on known claims. The original authors' discussion of the IBNR definition explicitly includes adverse and favorable development of reported claims. If definitions differ across internal reports, the same term can refer to materially different amounts. Consistent decomposition is essential to prevent case reserve double counting.
Unpaid loss and IBNR reconciliations
The reconciliation from an ultimate estimate to an unpaid estimate is mechanical once the claims measure is fixed:
On the incurred basis, if I_i=P_i+K_i, the corresponding amount beyond case incurred is (\widehat U_i-I_i), and unpaid is (K_i+(\widehat U_i-I_i)). For paid BF, the unpaid estimate is simply (1-p_i^{paid})E_i if C_i=P_i. These identities make the selection of paid or incurred emergence a major substantive decision: the paid-development percentage gives a forecast of all remaining payments; the incurred-development percentage gives a forecast beyond case-incurred amounts.
For allocated loss adjustment expenses and other costs, the same method may be applied to an appropriately defined expense development base, possibly separately from indemnity payments. The ASOP 43 scope guidance distinguishes claims from associated unpaid claim adjustment expenses and requires clarity about components included. A claims-only BF ultimate cannot automatically be called a complete claims-and-expenses liability.
Gross, ceded, and net measurements need matching data. Ceded claims often have different reporting lags, attachment-point effects, reinstatement provisions, and collectible reinsurance timing from gross claims. A numerical difference between two gross and net BF ultimates is not necessarily a coherent ceded estimate unless their experience, emergence assumptions, and recoveries reconcile. A separate ceded method, or a sufficiently detailed gross-to-ceded allocation, may be required.
Emergence under a changing claims environment
Observed emergence is shaped by reporting practices, litigation, claim settlement, inflation, reopenings, large-loss recognition, and data-processing lags. A catastrophic event may produce prompt notifications but slow final payments; a liability portfolio may exhibit late reports and subsequent settlement escalation. For this reason the incurred and paid triangles should be reviewed side by side, even when only one ultimately determines the selected reserve.
Calendar-year conditions may also alter claim amounts at every origin age. A sudden change in medical-cost inflation or legal settlement practice cannot always be represented by a stationary development pattern. The actuarial choice is then to adjust expected ultimate amounts, adjust development proportions, segment the portfolio, or employ an alternative method with documented reasoning. Keeping an obsolete p_i merely because the arithmetic is familiar transfers an unrecognized model error to the unpaid estimate.
The US data quality standard requires appropriate examination and use of data, including material limitations and reliance on others (ASB, ASOP 23). The same principles apply beyond the United States as research-quality controls even though the US standard is binding only within its own professional applicability. Analysts should identify duplicate transactions, negative recoveries, claim reclassifications, incomplete accident dates, accounting cutoffs and claim changes posted after the valuation date.
Development Proportion Estimates
Triangles and age-to-ultimate factors
A run-off triangle organizes cumulative paid or incurred amounts by origin period and development age. In a deterministic chain-ladder workflow, selected age-to-age factors are multiplied to produce an age-to-ultimate factor F_i at the observed age of each origin period. When F_i≥1 and corresponds to the chosen development basis, the emerged proportion is p_i=1/F_i and the future proportion is 1−1/F_i. The BF ultimate becomes
The relationship is exact algebra, independent of how F_i was selected. A factor of 4.00 means 25% emerged and 75% remaining, not a 400% remaining fraction. A factor of 1.25 means 80% emerged and 20% remaining. An age-to-ultimate factor of 1.00 implies no future development through the selected terminal horizon; that claim requires explicit confidence in any omitted tail.
In a simplified hypothetical age schedule, selected cumulative ages at 12, 24, and 36 months have factors F=(4.0000, 1.81818, 1.28205), hence p≈(0.25,0.55,0.78). The corresponding remaining proportions are 0.75, 0.45, and 0.22. An individual age-to-age factor from 12 to 24 months would be about 4/1.81818≈2.20; from 24 to 36 months it is about 1.81818/1.28205≈1.41818. These are illustrative values, not a sampled insurer development pattern.
If a tail factor T>1 is added beyond the last observed age, the cumulative factor F_i rises and p_i falls, increasing BF's expected future amount. This effect may be substantial for very immature or long-tailed business. It follows directly that reserve conclusions based on a terminal factor set to 1.00 cannot be represented as fully developed without defending the assumed tail.
Choice, calibration, and diagnostics
The standard volume on basic unpaid-claims techniques explains development triangles as both calculation tools and diagnostics (Friedland, 2010). Factor selection may use volume-weighted average age-to-age ratios, simple averages, medians, regression, or judgment where individual origin periods contain anomalous development. When portfolio composition changes, a selected factor from historical averages should be checked against claim count, average severity, closure rates, and the age structure of the claims inventory.
Ratio-based factors can be unstable when prior observed cumulative amounts are close to zero or negative. Negative increments may arise from salvage, subrogation, refunds or case estimate reductions. There is no universally legitimate mechanical fix that transforms every such triangle into a positive, monotone maturity curve. Analysts should examine the meaning of the loss basis and choose an appropriate model or segmentation. A BF p_i representing a share emerged is conceptually bounded in [0,1]; using a factor below 1 without explanation can imply a negative future amount and may signal a mismatch between a net runoff pattern and a conventional positive BF emergence model.
For validation, compare successive evaluations of the same origin period. If one period had prior E=10.0 million, p=0.40, and observed incurred 4.4 million, the BF ultimate is $10.4 million. At the next valuation, if the proportion rises to 0.60 and the observed cumulative reaches $6.4 million, BF still estimates $10.4 million: the $2.0 million of newly observed development exactly offsets a $2.0 million reduction in expected outstanding emergence. If instead the observed cumulative reaches $7.2 million, BF estimates $11.2 million. The change is attributable to excess emergence beyond the unchanged expected runoff, not to arbitrary repricing of a fixed reserve.
Back-testing can compare previously predicted one-year emergence with realized incremental paid or incurred claims, both by segment and across years. A good fit of one portfolio aggregate may hide offsetting miscalibration. Specific residual checks can compare actual emergence (C_{i,t+1}-C_{i,t}) with (E_i(p_{i,t+1}-p_{i,t})). This comparison keeps the stated prior fixed during the diagnostic interval; if the prior itself changed, the test should show that revision separately. The formula offers an auditable attribution of reserve movement and supports a review of development factors and expected ultimate losses.
Data quality and tail uncertainty
The relevant data are claims and their emergence, not merely the geometry of a spreadsheet triangle. The ASOP 56 modeling standard provides general professional guidance about model purpose, design, input, validation and limitations. For BF, these controls include versioning of selected factors, exposure bases, boundary conditions, trend adjustments, and reconciliation of source claims to financial-control totals.
Tail development deserves its own research decision. In volatile long-tail litigation or occupational liability, the remaining fraction may remain material many years after an accident period. A truncated triangle can yield seemingly stable early factors while missing late emergence. Conversely, changing laws, claim limits, and settlement behavior can make very old tails inapplicable to current claims. Independent tail studies, scenario analysis, and paid-incurred comparisons are substantive safeguards against false maturity.
The actuarial record should preserve the exact cutoffs for accounting, observed transactions and later review. Under ASOP 43's disclosures, these dates can differ and have distinct roles. A post-valuation development observed before a later review date should be explicitly identified as subsequent information rather than silently inserted into a historical triangle. Otherwise successive model versions cannot be replicated.
BF Formula and Derivation
Algebraic identities and interpretation
Let the deterministic chain-ladder ultimate be (\widehat U_i^{CL}=C_i/p_i) for p_i>0. Then BF may be rewritten as
This exact identity explains a frequently used credibility interpretation: the current chain-ladder extrapolation receives weight p_i, while the independently selected ultimate receives weight 1−p_i. These weights sum to one and are nonnegative when 0≤p_i≤1. The identity is algebraic; it does not establish that p_i is a statistically optimal credibility weight or that E_i is a Bayesian posterior prior under arbitrary probability models.
The distinction is important. The alternative representation (C_i+(1-p_i)E_i) tells what is actually projected: future claims are linked to the expected ultimate, not multiplied out of current observed claims. As p_i approaches 0, BF tends to the prior E_i for small observed losses; as p_i approaches 1, BF approaches observed cumulative claims. The classical expected-claims method uses E_i throughout, while chain ladder uses C_i/p_i throughout.
Another exact identity is
The right-hand side is the difference between observed emergence and the emergence expected from the selected prior and pattern. Therefore BF changes the prior ultimate only in proportion to how much actual emergence exceeds or falls short of the expected emerged portion. This equation provides a concise reserve movement explanation that is more informative than saying the estimate is an average of two methods.
Paid and incurred numerical applications
Consider a hypothetical cohort with expected ultimate E=$12.0 million, expected proportion emerged p=0.35, and observed cumulative incurred C=$3.9 million. The BF ultimate is 3.9+0.65×12.0=$11.7 million. Chain ladder produces 3.9/0.35≈$11.143 million, while the expected-claims approach produces $12.0 million. The BF result is 0.35×11.143+0.65×12.0≈$11.7 million. Current emergence was $0.3 million below the expected 0.35×12=$4.2 million, so BF is exactly $0.3 million below the prior ultimate.
If the same observed amount were cumulative paid rather than reported incurred, the BF $11.7 million ultimate would imply unpaid $7.8 million. If the $3.9 million is incurred, $7.8 million is development beyond reported incurred; a separate cumulative paid amount is required to establish the complete unpaid liability. This difference is economically decisive and often obscured by generic references to “reserves.”
An alternative calculation with premium base P=$15 million and expected ultimate loss ratio r=80% leads to E=$12 million. The BF result above is unchanged. However, if premium increased because rates rose while exposure remained stable, a simple prior based on unadjusted earned premium can overstate expected ultimate claims. The $12 million should therefore be tied to an explicitly defensible on-level premium and loss ratio, or to independent exposure and expected loss costs.
Consecutive valuations and a simple stochastic model
A deliberately minimal stochastic example clarifies stability and its cost. Suppose an origin cohort has an actual but random ultimate claim count U=C+F, where observed emergence C and future emergence F are independent Poisson counts with means pm and (1−p)m. The true expected ultimate is m and the selected BF prior is a fixed E_0. BF estimates the total count as C+(1−p)E_0. Its error relative to the realized ultimate is ((1-p)E_0-F). The mean error is ((1-p)(E_0-m)), and its mean squared error equals
The corresponding chain-ladder count estimate is C/p. Its error relative to the realized U is (((1-p)/p)C-F), so independence yields
When E_0=m, BF has lower mean squared error than chain ladder whenever 0<p<1 under these specific Poisson assumptions. When the prior is biased, the second term of BF's error grows quadratically. Equating the two MSE formulas gives the simple boundary (|E_0-m|=\sqrt{m/p}) for 0<p<1. The result applies to a Poisson count setting with fixed p, independent emergence, and the stated estimators; it does not establish a universal superiority rule for monetary claims where claim severities, dependence, and parameter estimation matter.
For example, with m=100 expected ultimate claims, p=0.20, and a prior E_0=130, BF has MSE 80+0.64×900=656 claim-count squared units; chain ladder has MSE 100×0.8/0.2=400. If the prior were correctly centered at 100, BF MSE would fall to 80. The boundary for a prior error is √(100/0.2)≈22.36 claims. The calculation demonstrates that strong resistance to early observations can be advantageous when the prior is credible and costly when an inherited prior is materially wrong.
This is a transparent pedagogical derivation, not an estimation of empirical industry MSE. It also separates model conditionality from data reality: real monetary triangles may exhibit process and parameter uncertainty, inflation, claim count/severity dependence, and imperfect emergence patterns. Stochastic reserving research investigates prediction uncertainty and full outcome distributions beyond such a simple illustration (England and Verrall, 2002).
Numerical Comparison with Chain Ladder
Three accident-year example
The following internally consistent synthetic portfolio uses the same loss basis for observed cumulative claims C, prior E, and selected emergence p. Amounts are in thousands of currency units. These are exact illustrations, not observed insurer results.
| Origin period | Emerged fraction p | Observed C | Prior E | Chain ladder C/p | BF ultimate C+(1-p)E | BF future beyond C |
|---|---|---|---|---|---|---|
| AY1 | 0.80 | 820 | 1,000 | 1,025 | 1,020 | 200 |
| AY2 | 0.50 | 580 | 1,200 | 1,160 | 1,180 | 600 |
| AY3 | 0.20 | 260 | 1,500 | 1,300 | 1,460 | 1,200 |
| Total | — | 1,660 | 3,700 | 3,485 | 3,660 | 2,000 |
The portfolio BF ultimate of 3,660 consists of 1,660 observed and 2,000 expected future claims. Chain ladder produces 3,485, so the methods differ by 175, or about 5.02% of the chain-ladder ultimate. This difference is concentrated in the youngest period: its low emerged amount of 260, multiplied by 1/0.20, implies chain-ladder ultimate 1,300, whereas BF retains 80% weight on the independent prior of 1,500 and estimates 1,460. An aggregate-only comparison could conceal the large dependence of the youngest year's result on prior quality.
For AY1 the observed 820 exceeds 80% of the 1,000 prior (800) by 20, and BF accordingly estimates 1,020. For AY2 observed 580 is below 50% of 1,200 (600) by 20, yielding BF 1,180. For AY3 observed 260 is below 20% of 1,500 (300) by 40, yielding BF 1,460. The sum of deviations is −40: the total BF ultimate of 3,660 is precisely 40 below the 3,700 prior. These identity checks independently reconcile the table.
If C is cumulative paid, the column “BF future beyond C” is the implied unpaid amount by origin period. If C is reported incurred, it is additional development beyond reported incurred, and the financial unpaid amount must still add the booked case outstanding. This distinction applies to every row and should appear in any operational reserve report.
When the methods converge or diverge
BF and chain ladder coincide exactly when E_i=C_i/p_i, because the prior ultimate and the chain-ladder ultimate match. They also coincide as p_i reaches 1, assuming a complete and adequately defined run-off. BF and the pure expected-claims method coincide when C_i=p_iE_i, as observed emergence exactly meets the prior's prediction. These equalities give useful diagnostic thresholds, and can be evaluated independently of a particular company's accounting choices.
The difference between BF and chain ladder follows
Thus the divergence is greatest, all else equal, where little of the ultimate has emerged and the prior disagrees with the chain-ladder implied ultimate. The expression also shows why a late-maturing cohort can exhibit little BF-versus-chain-ladder difference even if the prior remains poor: the factor (1−p_i) suppresses the prior's role near maturity.
If the current cumulative amount includes a single early catastrophe claim, chain ladder may extrapolate a disturbance across future emergence. BF allows the experienced loss to enter fully while assigning remaining losses using E_i and the residual fraction. This treatment may be more reasonable for an isolated loss if the prior already anticipates the catastrophe exposure and future claims are not mechanically scaled to a single event. However, if catastrophe occurrence changes the expected ultimate frequency or severity of future claims, the unchanged BF prior may be seriously deficient. The method does not diagnose the causal structure of an event merely by maintaining stable results.
Stochastic, credibility, and likelihood extensions
England and Verrall's stochastic reserving survey places BF in a broader landscape including Bayesian treatments, parametric curves, and prediction errors (England and Verrall, 2002). Mack's distribution-free chain-ladder uncertainty analysis derives standard-error estimates for chain-ladder reserve estimates under its specified assumptions (Mack, 1993). These findings explain why a deterministic BF point estimate should not be mistaken for a complete predictive distribution or a calibrated risk margin.
A related Benktander credibility approach combines BF and chain ladder and is discussed using a simple stochastic model (Mack, 2000). The paper's performance findings are model-dependent. The original BF identity can be viewed as a weighted combination, but the “weights” in alternative credibility methods may be recomputed from data and assumptions in different ways. A controlled comparison should name the exact formula, rather than treating all blended reserving techniques as equivalent.
The 2019 likelihood analysis of Elpidorou, Margraf, Martínez-Miranda, and Nielsen develops an approach that incorporates external relative-ultimate information into a Poisson generalized linear model through likelihood constraints. It is a subsequent statistical extension with its own assumptions and estimation object; it does not retroactively transform classical BF into a generally likelihood-optimal estimate. The authors analyze how externally supplied expected ultimate information affects forecasts, illustrating the continuing relevance of the core BF problem.
The stochastic reserving reference work by Wüthrich and Merz (2008) emphasizes claims reserve uncertainty and stochastic models. In applications, process variation, parameter estimation, prior-selection uncertainty, development-factor uncertainty, and correlations among origin periods need consideration. Many standard bootstrap methods assume a statistical structure that may be compromised by major operational or portfolio changes; producing a distribution is only useful when its data-generating assumptions can be defended.
Parameter Sensitivity
Prior and proportion stress analysis
BF has transparent partial derivatives when its three inputs are treated as locally independent:
The identities mean one extra unit of observed cumulative claims increases ultimate by one unit, one extra unit of prior ultimate increases it by the un-emerged fraction, and an increase in the assumed emergence proportion reduces the future expected amount by E times the change. They are derivatives of the fixed-input BF function; empirical calibration of E and p may create dependence requiring total rather than partial derivatives.
Return to E=$12.0 million, C=$3.9 million, and p=0.35, for an $11.7 million BF ultimate. Raising E by 10% to $13.2 million increases the result by 0.65×1.2=$0.78 million, to $12.48 million. Increasing p from 0.35 to 0.40 with C and E unchanged reduces BF ultimate by 0.05×12=$0.60 million, to $11.10 million. Simultaneously raising E to 13.2 and p to 0.40 produces 3.9+0.60×13.2=$11.82 million. The joint change is not merely the sum of the two separate one-factor changes because the formula contains the interaction term −pE.
For a development factor F, sensitivity follows from p=1/F:
A higher factor, representing slower recognized emergence and a larger un-emerged proportion, increases the BF ultimate. For an immature cohort with F=5 and E=10 million, a rise from F=5 to F=5.5 changes remaining emergence from 0.80 to about 0.81818 and adds about $0.1818 million to estimated ultimate, assuming C and E fixed. Treating a factor change as inconsequential without checking the implied p_i can be a serious reserve error.
A complete sensitivity exercise stresses prior loss ratio, exposure, reporting/payment speed, calendar-year severity trend, loss adjustment expense, tail factor, discounting where applicable, and reinsurance recoveries. Report the effect on each cohort and on portfolio total; stress both directions where economically plausible. A symmetric +/- percentage on every variable need not represent a plausible scenario because some quantities are constrained, and conditions such as p≤1 impose real mathematical boundaries.
Portfolio segmentation and external change
A full portfolio may comprise short-tail property losses and long-tail professional liability claims, each with different maturation patterns. Computing a single weighted-average emergence proportion without matching the corresponding expected ultimate amounts can distort BF. For segments j, the defensible portfolio aggregate is (\sum_j[C_j+(1-p_j)E_j]). Its effective aggregate un-emerged fraction, if desired for presentation, is (\sum_j(1-p_j)E_j/\sum_j E_j), provided the expected ultimate bases are comparable and positive. This weighting is by expected ultimate claims, not automatically by policy count or earned premium.
For instance, consider a short-tail segment with E=8 million and p=0.90 and a long-tail segment with E=2 million and p=0.20. The aggregate residual fraction equals (0.10×8+0.80×2)/10=0.24. A naive arithmetic average of the two p values is 0.55, implying 45% remaining, almost twice the properly weighted 24%. The loss-weighted aggregate emerged fraction is 0.76. This simple example explains why heterogeneous portfolios need segmented calculation even when top-line triangles are available.
Long-tail settlement can change with judicial decisions, economic inflation and legal reforms. Exposure can change with retentions, coverage forms, and product distribution. Ceded reinsurance can have a different claims notification and settlement cadence from gross business. Each change can affect E, p or both. An appropriate analysis distinguishes the mechanisms and documents why historical data remain usable. Assumptions drawn from a different country or insurance class cannot be treated as transferable merely because both are described as claims reserving.
Under the US ASOP 43 method-selection provisions, actuarial selection, sensitivity, material external conditions, and communication of significant limitations are substantive considerations. ASOP 23 data guidance and ASOP 56 modeling guidance complement these requirements. ASOP 41 on communications addresses form, content, identification of responsibilities and disclosures. These standards describe professional practice within their scope and do not constitute a universal reserve formula for all jurisdictions.
International reporting, governance, and auditability
The BF output is ordinarily a point estimate of ultimate claims or related future claims. Under IFRS 17 Insurance Contracts, reporting obligations are measured by the applicable insurance-contract accounting framework, including current estimates of fulfillment cash flows and a risk adjustment for nonfinancial risk; the IFRS definition of incurred-claims liability identifies a distinct obligation relating to services already provided. A nominal BF estimate may be one building block in estimating those cash flows, but does not by itself produce the total IFRS 17 liability for incurred claims or determine eligibility for measurement simplifications.
Under Solvency II Article 77, technical provisions are based on a best estimate and risk margin, and the best estimate corresponds to probability-weighted future cash flows discounted using the relevant risk-free term structure. Article 82 on data quality addresses suitability, completeness and accuracy of technical-provision data and allows appropriate approximations when suitable data are insufficient under the stated conditions. A BF ultimate can inform a cash-flow projection but cannot replace discounting, risk margins or requirements governing eligible methods.
For US P/C practice, the current ASOP 20 cash-flow standard was revised in December 2025 and is effective for covered work products from June 1, 2026. It addresses cash-flow analysis including discounting and does not require every claim estimate to be discounted. Application depends on the assignment, legal reporting basis, and relevant actuarial practice scope. A nominal BF estimation should therefore be separated in the working papers from the subsequent cash-flow and discounting assessment.
Operational governance starts with a versioned record of the portfolio definition and valuation date, paid/incurred triangle, exposure and rate histories, selected prior E and its provenance, factors F and emergence shares p, selection rationale, stress tests, method comparisons, and management decisions. At each successive valuation, a bridge should separate new claims emergence, changes in prior expected ultimate, changes in the development pattern, changes in accounting scope, and any individually assessed large losses. This bridge makes reserve movements traceable without claiming that each source of uncertainty is statistically independent.
Source documentation has the same importance as numerical replication. The analyst should be able to show, for every selected external benchmark or standard, what claim it supports, what jurisdiction or portfolio it covers, and why it is relevant. The manuscript's short contextual links identify the original works at the moment their claims are used, and the full bibliographic records appear below. This research approach supports an auditable international comparison while preserving distinctions among actuarial method, professional standard, financial-reporting obligation, and published scientific literature.
Integrated research findings
The combined historical, mathematical and regulatory inquiry supports six analytical conclusions. First, BF predicts remaining development through E(1−p), so the apparent “blend” is derived from a forecast about unobserved claims. Second, the prior E and the emergence factor p must be defended together, because either can distort the remaining amount even when current observations are accurate. Third, the actuarial distinction between paid and incurred observations determines whether the residual is complete unpaid loss or additional development beyond case-incurred amounts. Fourth, comparison with chain ladder identifies disagreements but cannot establish correctness without inspecting exposures, claim operations and prior information. Fifth, stochastic uncertainty measures require explicit probabilistic assumptions, and the method's stability advantage may disappear under prior bias. Sixth, the route from BF ultimate to booked insurance obligations depends on the applicable reporting purpose, cash-flow timing, risk adjustments and regulatory context.
A decision-ready BF analysis accordingly has four reproducible numerical layers: an observed emergence ledger, an independently justified expected ultimate basis, a selected age-to-ultimate pattern with diagnostics, and a resulting ultimate-to-unpaid reconciliation. It has two interpretative layers: a comparison against alternative methods and a reporting-framework bridge. These are organizational elements of a rigorous analysis, not newly named models or theories. The scientific contribution of this monograph is the integrated derivation, source-based differentiation, and reproducible demonstration of their relationships.
Conclusion
Bornhuetter–Ferguson claims reserving remains central because it operationalizes a simple but consequential distinction: what claims have already emerged can be recognized directly, while the remaining obligation is estimated from a prior view of ultimate claims and a selected development pattern. Its equation, (\widehat U=C+(1-p)E), is straightforward. The scientific work lies in establishing the meaning, provenance, and adequacy of C, E, and p; interpreting them consistently for paid and reported-incurred claims; and testing their behavior as a portfolio evolves.
The reconstructed 1972 primary source, subsequent methods literature, numerical comparisons, and professional standards establish the method's appropriate use and its limits. BF gains stability in immature origin periods when a prior expected ultimate is supportable and the expected pattern of future emergence remains relevant. An obsolete prior or distorted emergence estimate can convert that stability into persistent error. Internal reconciliation, repeated diagnostics, alternative estimates, and explicit uncertainty disclosure are therefore part of the method's proper scientific application.
The final numerical reserve required for financial reporting is a separate valuation conclusion. BF may contribute evidence to the estimate of future claims; jurisdiction-specific rules determine what further recognition, timing, risk adjustment, discounting, and disclosure are necessary. The resulting actuarial analysis should remain reproducible, documented, and proportionate to the risks being measured. No new authorial reserving theory is proposed here; the work is a foundational analytical synthesis grounded in the historical method and subsequent actuarial science.
References
Actuarial Standards Board (ASB). (2010, revised). Actuarial Standard of Practice No. 41: Actuarial Communications. Effective May 1, 2011. https://www.actuarialstandardsboard.org/asops/actuarial-communications/
Actuarial Standards Board (ASB). (2011 revision). Actuarial Standard of Practice No. 43: Property/Casualty Unpaid Claim Estimates. Original effective date September 1, 2007. https://www.actuarialstandardsboard.org/asops/propertycasualty-unpaid-claim-estimates/
Actuarial Standards Board (ASB). (2016). Actuarial Standard of Practice No. 23: Data Quality. https://www.actuarialstandardsboard.org/asops/data-quality/
Actuarial Standards Board (ASB). (2019). Actuarial Standard of Practice No. 56: Modeling. Effective October 1, 2020. https://www.actuarialstandardsboard.org/asops/modeling-3/
Actuarial Standards Board (ASB). (2025). Actuarial Standard of Practice No. 20: Analysis of Property/Casualty Cash Flows, Including Discounting. Revised December 2025; effective June 1, 2026. https://www.actuarialstandardsboard.org/asops/asop-no-20-analysis-of-property-casualty-cash-flows-including-discounting/
Berquist, J. R., & Sherman, R. E. (1977). Loss reserve adequacy testing: A comprehensive systematic approach. Proceedings of the Casualty Actuarial Society, 64, 123–184. https://www.casact.org/abstract/loss-reserve-adequacy-testing-comprehensive-systematic-approach
Bornhuetter, R. L., & Ferguson, R. E. (1972). The actuary and IBNR. Proceedings of the Casualty Actuarial Society, 59, 181–195. Primary text: https://www.casact.org/sites/default/files/database/proceed_proceed72_72181.pdf . CAS bibliographic record: https://www.casact.org/abstract/actuary-and-ibnr
Elpidorou, V., Margraf, C., Martínez-Miranda, M. D., & Nielsen, B. (2019). A likelihood approach to Bornhuetter–Ferguson analysis. Risks, 7(4), Article 119. https://doi.org/10.3390/risks7040119
England, P. D., & Verrall, R. J. (2002). Stochastic claims reserving in general insurance. British Actuarial Journal, 8(3), 443–518. https://doi.org/10.1017/S1357321700003809
European Union. (2009, as consolidated). Directive 2009/138/EC (Solvency II), Article 77: Calculation of technical provisions. EIOPA Single Rulebook. https://www.eiopa.europa.eu/rulebook/solvency-ii-single-rulebook/article-2160_en
European Union. (2009, as consolidated). Directive 2009/138/EC (Solvency II), Article 82: Data quality and application of approximations. EIOPA Single Rulebook. https://www.eiopa.europa.eu/rulebook/solvency-ii-single-rulebook/article-2170_en
Friedland, J. F. (2010). Estimating Unpaid Claims Using Basic Techniques (Version 3, July 30, 2010). Casualty Actuarial Society. https://www.casact.org/sites/default/files/2021-03/5_Friedland.pdf . CAS bibliographic record: https://www.casact.org/abstract/estimating-unpaid-claims-using-basic-techniques
IFRS Foundation. (2017, subsequently amended). IFRS 17 Insurance Contracts. Effective for annual reporting periods beginning on or after January 1, 2023. https://www.ifrs.org/issued-standards/list-of-standards/ifrs-17-insurance-contracts/
IFRS Foundation. (n.d.). IFRS 17 Insurance Contracts—Key Terms. https://www.ifrs.org/supporting-implementation/supporting-materials-by-ifrs-standards/ifrs-17/key-terms/
Mack, T. (1993). Distribution-free calculation of the standard error of chain ladder reserve estimates. ASTIN Bulletin, 23(2), 213–225. https://doi.org/10.2143/AST.23.2.2005092
Mack, T. (2000). Credible claims reserves: The Benktander method. ASTIN Bulletin, 30(2), 333–347. https://doi.org/10.2143/AST.30.2.504639
Wüthrich, M. V., & Merz, M. (2008). Stochastic Claims Reserving Methods in Insurance. John Wiley & Sons. ISBN 978-0-470-72346-3. https://www.wiley.com/en-us/stochastic-claims-reserving-methods-in-insurance-p-9780470723463