Credibility Interpretation of Bornhuetter-Ferguson

Why Interpret Bornhuetter-Ferguson as a Blend?

Bornhuetter-Ferguson (BF) combines two sources of information: an a priori estimate of ultimate losses and the losses observed to date. Its interpretation becomes clearer when the formula is written as a weighted average. The weight on observed experience is the expected proportion of ultimate losses that has emerged; the weight on the a priori estimate is the proportion expected to remain unreported or unpaid.

This interpretation explains why BF is often described as having a credibility feature. As claims develop, the observed-experience weight grows and the a priori weight shrinks. However, those weights come from the selected emergence pattern, not from a statistical credibility calculation based on process variance. The analogy is useful, but it does not make BF a formal credibility model.

Definition: Let $E$ be the a priori expected ultimate loss, $C$ the cumulative loss observed to date on a specified basis, and $p$ the expected proportion of ultimate losses emerged on that same basis. The unreported proportion is $1-p$. For $0<p\leq 1$, the BF ultimate estimate is $U_{\mathrm{BF}}=C+(1-p)E$. The BF reserve on that basis is $U_{\mathrm{BF}}-C=(1-p)E$.

The quantities must be on consistent bases. For example, a paid cumulative amount is paired with an expected paid emergence percentage, while an incurred cumulative amount is paired with an expected incurred emergence percentage. The a priori ultimate and observed losses must also use compatible coverage, valuation, and loss definitions. A percentage selected for one basis cannot be transferred automatically to another.

The Blend Representation

For $p>0$, the development-based estimate obtained by dividing observed losses by the expected emerged proportion is $C/p$. This is the ultimate estimate that would result from scaling the observed amount by the emergence pattern. The BF formula can then be rewritten as a weighted blend.

U_{\mathrm{BF}} =C+(1-p)E =p\left(\frac{C}{p}\right)+(1-p)E

The two weights are $p$ and $1-p$, which are nonnegative and sum to one when $0\leq p\leq 1$. Thus, for $p>0$, BF places weight $p$ on the development-based estimate $C/p$ and weight $1-p$ on the a priori estimate $E$. This is an algebraic interpretation; it does not require $C/p$ to be a statistically unbiased estimate in every situation.

Interpretation: When $p$ is close to zero, little loss is expected to have emerged, so the a priori estimate has most of the weight. When $p$ is close to one, most expected losses have emerged, so the observed cumulative amount has most of the weight. At $p=1$, BF equals $C$; at $p=0$, use the direct formula $C+E$, not the blend expression involving $C/p$.

Why the Formula Has This Form

A short derivation highlights the role of the unreported proportion. Under the selected emergence pattern, the portion of the a priori ultimate expected to remain unreported is $(1-p)E$. BF retains the observed cumulative amount $C$ and adds that expected remaining portion. Therefore, the ultimate estimate is the sum of observed losses and expected unreported losses:

U_{\mathrm{BF}} =C+\underbrace{(1-p)E}_{\text{expected losses not yet emerged}}

Subtracting $E$ from both sides gives another useful identity:

U_{\mathrm{BF}}-E=C-pE

The BF estimate therefore differs from the a priori estimate by the amount observed to date minus the amount expected to have emerged to date under the a priori estimate. If $C=pE$, the observed amount exactly matches expected emergence and BF remains at $E$. If $C>pE$, BF is above $E$; if $C<pE$, it is below $E$. This result follows directly from the formula and does not require dividing by $p$.

The derivation also explains the method’s intended logic: observed losses are not extrapolated to ultimate with full weight at early development. Instead, the expected unreported portion is supplied by the a priori estimate. This is particularly important when the observed amount is small or volatile, although the method still depends on the quality of the a priori estimate and the emergence pattern.

Worked Example: Reading the Blend at One Valuation

A hypothetical specialty insurer has an a priori ultimate loss estimate of $E=\$1{,}200{,}000$ for an accident year. At the valuation date, the selected paid emergence percentage is $p=0.60$, and cumulative paid losses are $C=\$780{,}000$. Calculate the paid BF ultimate and reserve, then express the ultimate as a blend. Use the paid BF method. The common mistake to avoid is treating $C$ as though it were already the developed ultimate.

Step 1: Find the unreported proportion. The paid emergence percentage is 60%, so the unreported proportion is $1-0.60=0.40$.

Step 2: Calculate the BF reserve. The reserve is the unreported proportion multiplied by the a priori ultimate:

(1-p)E=0.40(\$1{,}200{,}000)=\$480{,}000

Step 3: Calculate the BF ultimate.

U_{\mathrm{BF}}=C+(1-p)E =\$780{,}000+\$480{,}000 =\$1{,}260{,}000

Step 4: Verify the blend interpretation. The development-based ultimate is $C/p=\$780{,}000/0.60=\$1{,}300{,}000$. Weighting it by 0.60 and the a priori ultimate by 0.40 gives:

0.60(\$1{,}300{,}000)+0.40(\$1{,}200{,}000) =\$780{,}000+\$480{,}000 =\$1{,}260{,}000

The arithmetic checks: the blend and direct formula both give an ultimate of $\$1{,}260{,}000$, and subtracting paid losses gives a reserve of $\$480{,}000$. The method avoids the mistake of using the paid amount alone as the ultimate.

How the Blend Shifts as Claims Develop

As the expected emergence percentage increases, the weight on the development-based estimate increases and the weight on the a priori estimate decreases. This describes the changing structure of BF, but it does not guarantee that the BF ultimate estimate will rise or fall over time. Both $C$ and $p$ can change between valuations, and the developing observed amount may differ from what the earlier emergence pattern anticipated.

The reserve has a simpler mechanical relationship when the a priori ultimate and emergence pattern are held fixed: it equals $(1-p)E$, so it falls as $p$ increases. The ultimate estimate, however, also includes $C$. New observed losses can offset or exceed the reduction in the expected unreported portion.

Worked Example: Comparing Two Development Stages

For a hypothetical accident year, suppose the a priori ultimate is $E=\$1{,}000{,}000$. At an early valuation, $p=0.40$ and $C=\$360{,}000$. At a later valuation, $p=0.80$ and cumulative losses are $C=\$760{,}000$. Calculate the BF ultimate, the reserve, and the blend at each stage. Use the paid BF method and treat each stated emergence percentage as applicable to the corresponding cumulative paid amount. The common mistake to avoid is assuming that a larger emergence percentage alone determines the direction of the ultimate estimate.

Early valuation. The unreported proportion is $1-0.40=0.60$, so the reserve is $0.60(\$1{,}000{,}000)=\$600{,}000$. The ultimate is $\$360{,}000+\$600{,}000=\$960{,}000$. The development-based estimate is $\$360{,}000/0.40=\$900{,}000$, and the blend verifies the result:

0.40(\$900{,}000)+0.60(\$1{,}000{,}000) =\$360{,}000+\$600{,}000 =\$960{,}000

Later valuation. The unreported proportion is $1-0.80=0.20$, so the reserve is $0.20(\$1{,}000{,}000)=\$200{,}000$. The ultimate is $\$760{,}000+\$200{,}000=\$960{,}000$. The development-based estimate is $\$760{,}000/0.80=\$950{,}000$, and:

0.80(\$950{,}000)+0.20(\$1{,}000{,}000) =\$760{,}000+\$200{,}000 =\$960{,}000

The reserve declines from $\$600{,}000$ to $\$200{,}000$, while the ultimate stays at $\$960{,}000$ in this particular example. That outcome is not guaranteed in general; it occurs here because the later observed amount and remaining expected amount sum to the same total. The calculations avoid the mistake of concluding that a change in blend weights, by itself, determines the direction of the ultimate.

Paid and Incurred Interpretations

The credibility interpretation applies to either paid or incurred BF, but the inputs must stay on their respective bases. For paid BF, $C$ is cumulative paid loss and $p$ is the expected paid emergence percentage. For incurred BF, $C$ is cumulative incurred loss and $p$ is the expected incurred emergence percentage. The development-based estimate $C/p$, where $p>0$, and the reserve measured relative to $C$ are consequently basis-specific.

Worked Example: Keeping Paid and Incurred Weights on Their Own Bases

A hypothetical liability portfolio has an a priori ultimate of $E=\$500{,}000$. At the valuation date, cumulative paid losses are $C_{\mathrm{paid}}=\$180{,}000$ with paid emergence $p_{\mathrm{paid}}=0.30$. Cumulative incurred losses are $C_{\mathrm{incurred}}=\$330{,}000$ with incurred emergence $p_{\mathrm{incurred}}=0.60$. Calculate both BF ultimate estimates and their corresponding reserves. The common mistake to avoid is applying a paid emergence percentage to incurred losses.

Paid estimate. The unreported paid proportion is $1-0.30=0.70$, giving a paid BF reserve of $0.70(\$500{,}000)=\$350{,}000$. Thus:

U_{\mathrm{BF,paid}} =\$180{,}000+\$350{,}000 =\$530{,}000

The blend check uses the paid development-based estimate, $\$180{,}000/0.30=\$600{,}000$:

0.30(\$600{,}000)+0.70(\$500{,}000) =\$180{,}000+\$350{,}000 =\$530{,}000

Incurred estimate. The unreported incurred proportion is $1-0.60=0.40$, giving an incurred BF reserve of $0.40(\$500{,}000)=\$200{,}000$. Thus:

U_{\mathrm{BF,incurred}} =\$330{,}000+\$200{,}000 =\$530{,}000

The incurred development-based estimate is $\$330{,}000/0.60=\$550{,}000$, and its blend is $0.60(\$550{,}000)+0.40(\$500{,}000)=\$330{,}000+\$200{,}000=\$530{,}000$. Both calculations check, but the equal ultimate estimates are a feature of these particular inputs, not a general rule. The paid and incurred reserves differ because they are measured relative to different cumulative amounts. Keeping each emergence percentage with its matching loss basis avoids a basis-mismatch error.

Boundary Cases and Limits of the Credibility Analogy

The blend representation $p(C/p)+(1-p)E$ requires $p>0$. The direct BF formula $C+(1-p)E$, however, is still defined at $p=0$. If $p=0$, it gives $C+E$; it gives $E$ only when $C=0$ as well. The normalized development-based estimate $C/p$ is undefined at zero emergence, so it must not be used to justify the boundary result.

Worked Example: Zero Expected Emergence

A hypothetical program has a prior expected ultimate of $E=\$250{,}000$. At a valuation date, the selected emergence percentage is $p=0$, while a small amount $C=\$40{,}000$ has been recorded. Calculate the direct BF ultimate and reserve. The common mistake to avoid is claiming that substituting $p=0$ into the direct formula always gives $E$.

Step 1: Apply the direct formula. Since $1-p=1$:

U_{\mathrm{BF}}=C+(1-p)E =\$40{,}000+1(\$250{,}000) =\$290{,}000

Step 2: Calculate the reserve relative to the recorded amount.

U_{\mathrm{BF}}-C =\$290{,}000-\$40{,}000 =\$250{,}000 =(1-p)E

The direct formula gives $\$290{,}000$, not $\$250{,}000$, because it includes the observed $C$ as well as the full a priori amount. If $C$ were zero, the same formula would give $U_{\mathrm{BF}}=E=\$250{,}000$. In either case, $C/p$ is undefined, so the blend representation does not apply. This boundary case may also signal that the selected emergence percentage and the observed data deserve review.

Finally, the weights $p$ and $1-p$ should not be mistaken for statistically estimated credibility parameters. The BF method uses a selected emergence pattern to allocate weight algebraically. A larger $p$ means a larger coefficient on $C/p$ in the blend representation, but it does not by itself establish the variance, reliability, or independence properties required by a formal credibility model. Nor does it guarantee that the BF ultimate will move in a particular direction as development proceeds.

Key takeaway: For $p>0$, BF is $p$ times the development-based estimate $C/p$, plus $1-p$ times the a priori estimate $E$. Equivalently, it is $E$ plus the difference between actual cumulative losses and expected emerged losses, $C-pE$. The credibility interpretation comes from the emergence-based weights; it is not a formal statistical credibility calculation.

Check Your Understanding

For each problem, use the stated paid or incurred basis consistently. Treat all amounts as hypothetical.

A hypothetical accident year has $E=\$900{,}000$, paid emergence $p=0.50$, and cumulative paid losses $C=\$510{,}000$. Find the BF ultimate and reserve, and state the weight on the prior in the blend. (Answer: ultimate $\$960{,}000$; reserve $\$450{,}000$; prior weight $0.50$)
A hypothetical portfolio has $E=\$640{,}000$, incurred emergence $p=0.75$, and cumulative incurred losses $C=\$520{,}000$. Find the incurred BF ultimate and compare it with $E$. (Answer: ultimate $\$680{,}000$, which is $\$40{,}000$ above $E$)
For a hypothetical accident year, $E=\$1{,}500{,}000$, paid emergence is $p=0.40$, and cumulative paid losses are $C=\$540{,}000$. Calculate the BF ultimate two ways: directly and as a blend with $C/p$. (Answer: direct $=\$1{,}440{,}000$; blend $=0.40(\$1{,}350{,}000)+0.60(\$1{,}500{,}000)=\$1{,}440{,}000$)
A hypothetical line has $E=\$720{,}000$. At an early valuation, $p=0.25$ and $C=\$160{,}000$; at a later valuation, $p=0.50$ and $C=\$400{,}000$. Find both BF ultimates and reserves. (Answer: early ultimate $\$700{,}000$, reserve $\$540{,}000$; later ultimate $\$760{,}000$, reserve $\$360{,}000$)
A hypothetical claim portfolio has $E=\$300{,}000$, $p=0$, and $C=\$12{,}000$. Apply the direct formula, state whether the normalized blend is available, and give the reserve. (Answer: ultimate $\$312{,}000$; normalized blend unavailable because $C/p$ is undefined; reserve $\$300{,}000$)

What You'll Learn

Why Interpret Bornhuetter-Ferguson as a Blend?

The Blend Representation

Why the Formula Has This Form

Worked Example: Reading the Blend at One Valuation

Worked Example: Reading the Blend at One Valuation

How the Blend Shifts as Claims Develop

Worked Example: Comparing Two Development Stages

Paid and Incurred Interpretations

Worked Example: Keeping Paid and Incurred Weights on Their Own Bases

Boundary Cases and Limits of the Credibility Analogy

Worked Example: Zero Expected Emergence

Check Your Understanding

Previous

Next