Comparing Bayesian and Bühlmann Forecasts With Sparse Data

Why Compare the Two Forecasts?

When a risk has little claim experience, its observed rate can be unstable: one claim or no claims at all may have a large effect on the raw estimate. Both Bayesian prediction and Bühlmann credibility temper that experience with information about a broader population. Their forecasts can be identical under particular assumptions, but the methods do not always agree. The distinction matters when selecting a model, interpreting a forecast, or checking a calculation aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.

The comparison below concerns a risk with an unknown claim rate and a future expected claim count. All examples are hypothetical. Claim counts are modeled as Poisson conditional on the risk’s underlying rate; exposure is measured in consistent units, such as policy-years.

Definition: Let \(\Theta\) denote a risk’s unknown claim rate per unit exposure. Given \(\Theta=\theta\), assume the claim count \(N\) over exposure \(W\) has distribution \(\operatorname{Poisson}(W\theta)\). The observed experience rate is \(X=N/W\) when \(W>0\). A Bayesian rate forecast is the posterior mean \(E[\Theta\mid N]\). A Bühlmann rate forecast is a linear credibility estimate \(Z X+(1-Z)\mu\), where \(\mu=E[\Theta]\) is the collective mean and \(Z\) is the credibility factor.

Poisson-Gamma Forecasts

Suppose the prior distribution of \(\Theta\) is Gamma with shape \(\alpha\) and rate \(\beta\), written \(\operatorname{Gamma}(\alpha,\text{ rate }\beta)\). Its mean and variance are \(\alpha/\beta\) and \(\alpha/\beta^2\), respectively. Given total claims \(N\) over exposure \(W\), the posterior distribution is \(\operatorname{Gamma}(\alpha+N,\text{ rate }\beta+W)\), so the Bayesian rate forecast is:

E[\Theta\mid N]=\frac{\alpha+N}{\beta+W}.

For the Bühlmann calculation, treat the risk’s rate as the hypothetical mean. Conditional on \(\Theta\), \(X=N/W\) has mean \(\Theta\) and variance \(\Theta/W\). Therefore the expected process variance is \(EPV=E[\operatorname{Var}(X\mid\Theta)]=\mu/W\), while the variance of hypothetical means is \(VHM=\operatorname{Var}(\Theta)\). The Bühlmann credibility factor is \(Z=VHM/(VHM+EPV)\).

Result: For a Poisson-Gamma model with total exposure \(W>0\), \(\mu=\alpha/\beta\), \(VHM=\alpha/\beta^2\), and \(EPV=\mu/W\). The credibility scale is \(K=EPV\cdot W/VHM=\beta\), and \(Z=W/(W+\beta)\). Here \(K\) is the credibility scale parameter, expressed in exposure units; the complement of credibility is \(1-Z=K/(W+K)\).

Why the Forecasts Coincide Under These Assumptions

The equality follows from the model structure, not from a general identity between Bayesian and Bühlmann methods. First, conditional on \(\Theta=\theta\), the Poisson count likelihood is proportional to \(\theta^N e^{-W\theta}\). Multiplying by a Gamma prior density adds the exponents of \(\theta\) and the rates in the exponential. Thus the posterior shape is \(\alpha+N\), and the posterior rate is \(\beta+W\).

Second, Bühlmann’s forecast uses the same prior mean and the variance components implied by that prior and the Poisson process. Substituting \(K=\beta\) into its credibility formula gives:

\begin{aligned} ZX+(1-Z)\mu &=\frac{W}{W+\beta}\frac{N}{W} +\frac{\beta}{W+\beta}\frac{\alpha}{\beta}\\ &=\frac{N+\alpha}{W+\beta}. \end{aligned}

This is exactly the Bayesian posterior mean. The match requires the specified Gamma prior, a Poisson count process conditional on a constant risk rate, and the corresponding Bühlmann moments. If the prior has a different form, the Bayesian posterior mean may depend on more than the prior mean and variance; Bühlmann’s linear estimate still uses those two moments.

Worked Example: No Claims From a Sparse Risk

A hypothetical insurer models a risk’s annual claim rate \(\Theta\) with prior \(\operatorname{Gamma}(\alpha=2,\text{ rate }\beta=4)\). The risk has no claims in \(W=1\) policy-year. Calculate both rate forecasts and the expected number of claims over the next \(3\) policy-years.

Method and assumptions. Use a Poisson count conditional on \(\Theta\), the stated Gamma prior, and exposure in policy-years. Since \(N=0\), the observed rate is \(X=N/W=0/1=0\).

Bayesian forecast. The posterior parameters are shape \(2+0=2\) and rate \(4+1=5\). Hence the posterior mean rate is \(2/5=0.4000\) claims per policy-year.

Bühlmann forecast. The prior mean is \(\mu=2/4=0.5\). The process variance for the unit-exposure rate is \(EPV=\mu/W=0.5/1=0.5\). The variance of hypothetical means is \(VHM=2/4^2=2/16=0.125\). Thus \(K=EPV\cdot W/VHM=0.5\cdot1/0.125=4\), and \(Z=W/(W+K)=1/(1+4)=0.2\). The Bühlmann forecast is \(0.2(0)+0.8(0.5)=0.4\) claims per policy-year.

Future count and check. For future exposure \(3\), the expected count is \(3(0.4)=1.20\) claims under either forecast. The Bayesian calculation gives the same result by multiplying the posterior mean rate \(0.4\) by \(3\). The two rate estimates agree, as required under this Poisson-Gamma model. The common mistake avoided here is to treat zero observed claims as proof that the underlying rate is zero.

Worked Example: Combining Experience Across Unequal Periods

A hypothetical risk has prior rate \(\Theta\sim\operatorname{Gamma}(3,\text{ rate }6)\). It has one claim in a period with exposure \(0.5\) policy-years and three claims in a separate period with exposure \(1.5\) policy-years. Find the rate forecast and the expected count over a future \(1.5\) policy-years.

Method and assumptions. Assume independent Poisson counts across periods conditional on the same constant \(\Theta\). Add the counts and exposures: \(N=1+3=4\) and \(W=0.5+1.5=2\). The aggregate observed rate is \(X=4/2=2\) claims per policy-year.

Bayesian forecast. The posterior is \(\operatorname{Gamma}(3+4,\text{ rate }6+2)\). Its mean rate is \(7/8=0.875\) claims per policy-year.

Bühlmann forecast. The prior mean is \(\mu=3/6=0.5\), and its variance is \(VHM=3/6^2=3/36=1/12\). The expected process variance of the aggregate rate is \(EPV=\mu/W=0.5/2=0.25\). Therefore \(K=EPV\cdot W/VHM=0.25\cdot2/(1/12)=6\), and \(Z=2/(2+6)=0.25\). The credibility estimate is \(0.25(2)+0.75(0.5)=0.5+0.375=0.875\).

Future count and check. The expected count over \(1.5\) future policy-years is \(1.5(0.875)=1.3125\), or \(1.31\) claims to two decimal places. The Bayesian and Bühlmann rate forecasts agree. A key trap is averaging the two period rates without exposure weights: the correct aggregate rate is total claims divided by total exposure, \(4/2\), not the unweighted average of \(1/0.5\) and \(3/1.5\).

When the Forecasts Can Differ

Bühlmann credibility is a linear estimate based on the collective mean and the two variance components. A Bayesian forecast uses the full prior distribution and likelihood. If the prior is not Gamma, the posterior update for a Poisson rate generally does not reduce to the same linear blend. Sparse data can make this distinction especially visible: with little information, the prior’s shape has a strong influence on the Bayesian forecast.

Worked Example: A Two-Point Prior

Consider a hypothetical risk whose annual rate \(\Theta\) is \(0.2\) or \(1.2\) claims per policy-year, each with prior probability \(0.5\). It has zero claims over \(W=1\) policy-year. Calculate the Bayesian and Bühlmann forecasts, and identify why they differ.

Bayesian method. For zero claims, the Poisson likelihood at rate \(\theta\) is \(e^{-\theta}\). The posterior probability of the high rate is therefore \(\displaystyle \frac{0.5e^{-1.2}}{0.5e^{-0.2}+0.5e^{-1.2}}=\frac{1}{1+e^1}=0.2689\). The posterior mean rate is \(0.2(1-0.2689)+1.2(0.2689)=0.2+0.2689=0.4689\) claims per policy-year.

Bühlmann method. The prior mean is \(\mu=0.5(0.2)+0.5(1.2)=0.7\). Its variance is \(0.5(0.2-0.7)^2+0.5(1.2-0.7)^2=0.25\). For \(W=1\), \(EPV=\mu/W=0.7\), so \(K=EPV\cdot W/VHM=0.7/0.25=2.8\) and \(Z=1/(1+2.8)=0.2632\). The Bühlmann forecast is \(0.2632(0)+0.7368(0.7)=0.5158\) claims per policy-year.

Check and interpretation. The forecasts differ: Bayesian \(0.4689\), Bühlmann \(0.5158\). The Bayesian estimate uses the likelihood to update the two prior probabilities separately. Bühlmann uses the prior mean and variance to form a linear blend. The difference is not an arithmetic error; it reflects the non-Gamma prior and the methods’ different use of distributional information. The common mistake avoided is assuming that any Bayesian prior produces the exact Poisson-Gamma credibility formula.

Calculation Traps and Practical Interpretation

For sparse experience, the raw rate can be extreme, but the credibility weight is usually limited. Keeping exposure units and model assumptions visible helps distinguish a real model difference from a calculation error.

Using the wrong scale parameter. In the Gamma-Poisson model, \(K=\beta\), not \(\beta/\alpha\). For \(\alpha=2\), \(\beta=4\), \(W=1\), and \(N=0\), the incorrect choice \(K=2\) gives \(Z=1/3\) and a Bühlmann forecast of \((1-1/3)(0.5)=0.3333\), which does not equal the Bayesian forecast \(2/5=0.4000\). The correct \(K=4\) gives \(Z=1/5\) and forecast \(0.4000\).
Confusing \(K\) with the complement of credibility. \(K\) is measured in exposure units. The complement is \(1-Z=K/(W+K)\), which is dimensionless.
Mixing up Gamma parameterizations. Here \(\beta\) is a rate, not a scale. The prior mean is \(\alpha/\beta\), and adding exposure changes the posterior rate to \(\beta+W\).
Ignoring the conditioning assumptions. The exact equality assumes a constant underlying rate for the risk and Poisson counts conditional on that rate. If the process or prior changes, verify the variance components and posterior update rather than applying the equality automatically.

These forecasts concern expected rates, not the realized number of future claims. To convert a forecast rate to an expected count, multiply by future exposure. This distinction is important when comparing risks with different amounts of future coverage.

Key takeaway: Under a Poisson likelihood and a Gamma prior with shape \(\alpha\) and rate \(\beta\), the Bayesian posterior mean \((\alpha+N)/(\beta+W)\) equals the Bühlmann forecast with \(K=\beta\). Outside that model, compare assumptions and calculations: equality is not guaranteed.

Check Your Understanding

For each hypothetical scenario, calculate or compare the forecasts using the stated exposure units.

A risk has prior \(\operatorname{Gamma}(4,\text{ rate }8)\), one claim over \(W=2\) policy-years. Find the Bayesian and Bühlmann rate forecasts. (Answer: Both are \(5/10=0.5000\) claims per policy-year.)
A risk has prior \(\operatorname{Gamma}(2,\text{ rate }5)\) and three claims over \(W=1\) policy-year. Calculate the common forecast. (Answer: Both are \(5/6=0.8333\) claims per policy-year.)
Under a prior \(\operatorname{Gamma}(3,\text{ rate }4)\), a risk has one claim over \(0.5\) policy-years and two claims over \(1.5\) policy-years. Find the common forecast and expected claims over the next \(2\) policy-years. (Answer: Rate \(6/6=1.0000\); expected future count \(2.00\).)
A risk’s rate is \(0.1\) or \(0.9\), each with prior probability \(0.5\). It has zero claims over \(W=2\) policy-years. Calculate the Bayesian forecast, then the Bühlmann forecast from the prior moments. (Answer: Bayesian \(0.2344\); Bühlmann \(0.3049\) claims per policy-year.)

What You'll Learn