Experience Rating and Credibility

Why Credibility Matters in Treaty Pricing

A treaty’s own adjusted loss history contains information about its expected future cost, but a short history can be dominated by random variation. Experience rating addresses this by weighting the treaty’s observed loss cost against a complement: an expected loss cost for the relevant risk class. Credibility is the weight assigned to the experience, not a judgment that the observations are either wholly reliable or wholly irrelevant.

The history used for this calculation should be on a consistent basis. For example, historical losses may need to be developed, trended, on-leveled, and adjusted to the treaty terms being priced. This tutorial treats those adjustments as complete and focuses on how to use the resulting experience. If exposure units or coverage terms differ materially across years, a simple average may not be appropriate without further normalization.

Definition: Let $C_j$ be the adjusted annual loss cost for observation $j$, expressed per consistent exposure unit, and let $\bar{C}$ be the average of $n$ such observations. A credibility estimate is a weighted estimate of the form $Z\bar{C}+(1-Z)m$, where $Z$ is the credibility assigned to the observed experience, $0\le Z\le1$, and $m$ is the complement or expected loss cost for the relevant class. The corresponding indicated premium is this loss cost multiplied by the applicable exposure base, before any additional pricing provisions.

The complement is needed because the data do not, by themselves, tell us how much random fluctuation is present or how the treaty compares with its broader class. In an application, $m$ and the parameters used to calculate $Z$ must be based on a suitable model or supplied assumptions. Credibility does not replace judgment about whether the historical observations are comparable to the prospective treaty.

Bühlmann Credibility

Bühlmann credibility is a classical model for repeated observations of a risk. Treat the treaty as belonging to a risk type with an unknown expected loss cost $\Theta$. Conditional on that type, annual observations are independent and have the same conditional mean and variance. Different treaty risk types may have different values of $\Theta$.

Model assumptions: For a randomly selected risk type, $E[C_j\mid\Theta]=\Theta$ and $\operatorname{Var}(C_j\mid\Theta)$ has an average value called the expected process variance, denoted by EPV. The variance of $\Theta$ across risk types is the variance of hypothetical means, denoted by VHM. Observations are conditionally independent with the same conditional moments, and the complement is $m=E[\Theta]$. Under these assumptions, Bühlmann credibility is $Z=n/(n+K)$, where $K=\mathrm{EPV}/\mathrm{VHM}$, and the estimate is $Z\bar{C}+(1-Z)m$. If VHM is zero, the model provides no differentiation among risk types and this formula is not used with a zero denominator.

EPV measures variation from year to year for the same risk type; VHM measures variation in expected loss cost across risk types. A large EPV relative to VHM means an observed year is noisy compared with the differences the model is trying to identify, so the experience receives less weight. More observations increase $Z$, because repeated years help distinguish persistent risk characteristics from random variation.

Why the Bühlmann Formula Has This Form

The formula can be justified as the best linear estimate of $\Theta$ based on the observed sample mean. Consider estimates of the form $a+b\bar{C}$, and choose $a$ and $b$ to minimize mean squared error. Under the model, $E[\bar{C}]=m$ and $\operatorname{Var}(\bar{C})=\mathrm{VHM}+\mathrm{EPV}/n$. Also, $\operatorname{Cov}(\Theta,\bar{C})=\mathrm{VHM}$, because the sample mean contains the risk type’s expected value plus mean-zero process variation.

Center the variables.
Write the linear estimate as $m+b(\bar{C}-m)$, which has the correct overall mean $m$. Its error relative to $\Theta$ is $m-\Theta+b(\bar{C}-m)$.

Minimize the mean squared error.
Differentiating the expected squared error with respect to $b$ gives $b=\operatorname{Cov}(\Theta,\bar{C})/\operatorname{Var}(\bar{C})$.

Substitute the model moments.
$b=\mathrm{VHM}/(\mathrm{VHM}+\mathrm{EPV}/n)=n/(n+\mathrm{EPV}/\mathrm{VHM})=n/(n+K)$.

Thus the linear estimate is $m+Z(\bar{C}-m)=Z\bar{C}+(1-Z)m$. The result is a minimum-mean-square-error linear estimate under the stated model. It is not a claim that every possible nonlinear estimate has the same optimality property.

Worked Example: Bühlmann Indication From Five Treaty Years

A hypothetical excess-of-loss treaty has five annual adjusted layer loss costs of $0.72, $1.08, $0.90, $1.30, and $1.00 thousand per 1,000 exposure units. Assume these are comparable annual observations under the Bühlmann model. The complement is $m=\$0.85$ thousand per 1,000 units; EPV is $0.16$ and VHM is $0.04$, in squared units of thousand dollars per 1,000 exposure units. Calculate the credibility indication. No discounting or mortality assumption is used; a parametric distribution is not required because the model is specified through moments.

Method: calculate the sample mean, then apply Bühlmann credibility. The total of the five observations is $0.72+1.08+0.90+1.30+1.00=5.00$. Therefore $\bar{C}=5.00/5=\$1.00$ thousand per 1,000 exposure units.

The Bühlmann parameter is $K=\mathrm{EPV}/\mathrm{VHM}=0.16/0.04=4$. With $n=5$, credibility is $Z=5/(5+4)=5/9=0.5556$, rounded to four decimal places.

The indication is $Z\bar{C}+(1-Z)m=(5/9)(1.00)+(4/9)(0.85)=0.5556+0.3778=0.9333$ thousand dollars per 1,000 exposure units before rounding. This is $\$933.33$ per 1,000 exposure units, rounded to the nearest cent.

Check: The indication lies between the experience mean of $1,000.00 and the complement of $850.00, as a weighted average must. The common mistake avoided is assigning full weight to a five-year mean without considering its process variation.

Limited-Fluctuation Credibility and Exposure Volume

A second approach asks how much exposure is needed for observed losses to fall within a specified tolerance of their expected value with a specified probability. This is limited-fluctuation credibility. It directly connects volume and volatility, and can be used to set a full-credibility standard. A common approximation assumes independent exposure units with mean loss cost $\mu$ and variance $\sigma^2$, and uses a normal approximation for their aggregate.

Limited-fluctuation condition: For $n$ independent, identically distributed exposure units, the aggregate loss has mean $n\mu$ and variance $n\sigma^2$. To achieve probability approximately $p$ that aggregate loss is within $h$ of its mean as a proportion of that mean, require $n\ge n_0$, where $n_0=\left(z_{(1+p)/2}\sigma/(h\mu)\right)^2$. Here $z_{(1+p)/2}$ is the standard normal quantile. This is a normal-approximation criterion; it may be inaccurate for small samples, highly skewed losses, or dependent observations.

The criterion follows by standardizing the aggregate: $(S_n-n\mu)/(\sqrt{n}\sigma)$ is approximately standard normal. The desired event is $|S_n-n\mu|\le hn\mu$; after dividing by $\sqrt{n}\sigma$, its approximate probability is $P(|Z|\le h\mu\sqrt{n}/\sigma)$. Requiring this probability to be at least $p$ gives $h\mu\sqrt{n}/\sigma\ge z_{(1+p)/2}$, which rearranges to the stated threshold.

For a partial-credibility indication under this approach, a conventional volume-based factor is $Z=\min(1,\sqrt{n/n_0})$. It assigns full credibility at or above the chosen full-credibility volume and increases with the square root of volume below it. This factor is a prescribed approximation, not the Bühlmann factor; the two methods have different assumptions and should not be combined without a model that supports doing so.

Worked Example: Full-Credibility Volume and a Partial Indication

For a hypothetical treaty portfolio, assume independent exposure units with annual loss cost mean $\mu=\$100$ and standard deviation $\sigma=\$40$. Use a normal approximation, target $p=95\%$, and require aggregate losses to be within $h=20\%$ of expected aggregate losses. The observed experience consists of $448 in losses over four exposure units. There is no discounting or mortality assumption. Find the full-credibility volume and a partial-credibility indication using the rounded full-credibility volume.

Method: calculate the limited-fluctuation threshold, round it up, and use the square-root volume factor. For a 95% two-sided probability, $z_{0.975}=1.96$. Thus $n_0=(1.96(40)/(0.20(100)))^2=(78.4/20)^2=3.92^2=15.3664$. The full-credibility volume must be a whole number of exposure units, so round up to $16$.

The observed loss cost is $\bar{C}=\$448/4=\$112$ per exposure unit. With $n=4$, $Z=\sqrt{4/16}=0.5$. Taking the supplied expected loss cost of $100 per unit as the complement, the indicated cost is $0.5(112)+0.5(100)=\$106$ per exposure unit.

Check: At 16 units, the expected aggregate is $1,600, its standard deviation is $40\sqrt{16}=\$160$, and the 20% tolerance is $320, or two standard deviations; the approximate coverage is about 95.45%, exceeding 95%. The common mistake avoided is using $n/n_0$ rather than its square root for partial credibility.

Volume, Volatility, and Practical Interpretation

Both approaches formalize a practical principle: more credible experience generally requires more information relative to the randomness in the losses. Under limited fluctuation, required volume is proportional to $\sigma^2/\mu^2$. If the standard deviation doubles while the mean and target remain fixed, the required volume becomes four times as large. Tightening the tolerance or increasing the target probability also increases the required volume.

Under Bühlmann credibility, a higher EPV reduces credibility at a fixed number of observations, while a higher VHM increases it. This comparison matters: experience is useful when it helps distinguish a treaty’s persistent expected cost from random annual variation. A large amount of data does not automatically create high credibility if the observations are not comparable or if the assumed model fails.

Worked Example: How Process Volatility Changes the Indication

A hypothetical treaty has four adjusted annual loss costs of $0.80, $1.10, $1.20, and $1.30 thousand per 1,000 exposure units. The complement is $0.90 thousand per 1,000 units and VHM is $0.03$. Compare indications when EPV is $0.09$ and $0.36$. Assume independent, comparable annual observations under the Bühlmann model. No discounting or mortality assumption applies, and only the specified moments—not a parametric distribution—are used.

Method: hold experience, complement, volume, and VHM fixed; change only EPV. The experience total is $0.80+1.10+1.20+1.30=4.40$, so $\bar{C}=4.40/4=\$1.10$ thousand per 1,000 units.

With EPV $0.09$, $K=0.09/0.03=3$, so $Z=4/(4+3)=4/7=0.5714$. The indication is $(4/7)(1.10)+(3/7)(0.90)=0.6286+0.3857=1.0143$ thousand, or $\$1,014.29$ per 1,000 units.

With EPV $0.36$, $K=0.36/0.03=12$, so $Z=4/(4+12)=0.25$. The indication is $0.25(1.10)+0.75(0.90)=0.275+0.675=0.9500$ thousand, or $\$950.00$ per 1,000 units.

Check: Both indications lie between $900.00 and $1,100.00 per 1,000 units. The higher process variance reduces the experience weight from 0.5714 to 0.25, moving the indication closer to the complement. The common mistake avoided is treating a more volatile history as more informative merely because its observed mean differs from the complement.

Common Traps and Model Checks

Using unadjusted experience. Losses from different treaty terms, exposure bases, or cost levels may not be comparable. Apply appropriate adjustments before calculating the mean, or explain why the observations are already comparable.
Confusing process variance and variance of hypothetical means. EPV describes within-risk variation; VHM describes variation in underlying means across risks. Reversing them reverses the credibility conclusion.
Applying a credibility factor to the wrong quantity. The weighted quantity should be a loss cost on the same basis as the complement. If a premium is required, any additional pricing provisions must be handled separately.
Treating full credibility as certainty. It means a stated fluctuation criterion is met under a specified model or approximation, not that the next year’s result is predictable exactly.
Ignoring dependence or changing exposure. The independent-observation formulas can overstate information when years are correlated, exposure units are not comparable, or coverage changes materially.

Key takeaway: Experience rating starts with a consistent adjusted loss-cost history and a clearly defined complement. Bühlmann credibility weights the experience according to process and between-risk variation; limited-fluctuation credibility relates the required volume to a tolerance, probability, mean, and volatility. The indicated estimate is only as reliable as the data and assumptions supporting those inputs.

Check Your Understanding

Use the stated model and units in each problem. Round monetary answers to cents where needed.

A treaty has three comparable annual adjusted loss costs of $0.80, $1.00, and $1.20 thousand per 1,000 exposure units. Under Bühlmann assumptions, EPV is $0.12$, VHM is $0.06$, and the complement is $0.90 thousand. Find $K$, $Z$, and the indication. (Answer: $K=2$, $Z=0.6000$, indication $=\$0.96$ thousand per 1,000 units.)
For independent exposure units, $\mu=\$80$, $\sigma=\$40$, $h=25\%$, and $p=90\%$. Using $z_{0.95}=1.645$ and the normal approximation, find the minimum whole-number full-credibility volume. (Answer: $n_0=11$ exposure units.)
A treaty has $n=8$, $\bar{C}=\$1.40$ thousand, $m=\$1.00$ thousand, EPV $=0.32$, and VHM $=0.04$. Calculate the Bühlmann indication. (Answer: $K=8$, $Z=0.5000$, indication $=\$1.20$ thousand.)
A limited-fluctuation standard uses $n_0=25$ units. A treaty has four years of experience totaling $600 per unit-year across four equal exposure units, and its complement is $120 per unit. Using $Z=\min(1,\sqrt{n/n_0})$, calculate the partial-credibility indication. (Answer: $Z=0.4000$, indication $=\$132$ per unit.)
For a treaty with $n=6$, $\bar{C}=\$0.80$ thousand, and $m=\$1.10$ thousand, compare the indications when (a) EPV $=0.20$, VHM $=0.05$, and (b) EPV $=0.80$, VHM $=0.05$. (Answer: (a) $K=4$, $Z=0.6000$, indication $=\$0.92$ thousand; (b) $K=16$, $Z=0.2727$, indication $=\$1.0182$ thousand.)

What You'll Learn