Why Gamma Parameterization Matters
In a Gamma-Poisson credibility model, the prior distribution describes uncertainty about a risk’s claim frequency. Gamma distributions are commonly written using either a shape-rate pair or a shape-scale pair. The shape parameter is the same in both conventions, but the second parameter means something different. Confusing the two can change a posterior mean and its credibility weight, even when the underlying prior is intended to be identical.
Let \(\lambda>0\) be a risk’s claim frequency per exposure unit. In the shape-rate convention, the second parameter controls the exponential decay in the density. In the shape-scale convention, the second parameter is its reciprocal. The distinction is especially important when adding observed exposure to a prior parameter during a Poisson update.
Converting Between Rate and Scale
The safest conversion is to match the exponential terms in the two densities. The shape-rate density contains \(e^{-\beta\lambda}\), while the shape-scale density contains \(e^{-\lambda/\theta}\). They match when \(\beta=1/\theta\). The normalizing factors then match as well: \(\beta^\alpha=(1/\theta)^\alpha\). Thus the shape remains unchanged, and only the second parameter is reciprocated.
For a claim frequency, \(\lambda\) is measured in claims per exposure unit. The rate parameter \(\beta\) is measured in exposure units, while the scale \(\theta\) is measured in claim-frequency units. Their reciprocal relationship is also a useful unit check: adding exposure to \(\beta\) is dimensionally appropriate; adding exposure to \(\theta\) is not.
| Quantity | Shape-rate notation | Shape-scale notation |
|---|---|---|
| Parameters | \(\alpha,\beta\) | \(\alpha,\theta\) |
| Conversion | \(\beta=1/\theta\) | \(\theta=1/\beta\) |
| Mean | \(\alpha/\beta\) | \(\alpha\theta\) |
| Variance | \(\alpha/\beta^2\) | \(\alpha\theta^2\) |
| Poisson update parameter | Add exposure to \(\beta\) | Take the reciprocal of \(\beta+\text{exposure}\) |
Worked Example: Convert a Prior and Check Its Moments
A hypothetical fleet model assigns a risk’s annual claim frequency the prior \(\operatorname{Gamma}(6,3)\) in shape-rate notation, where \(\lambda\) is measured per vehicle-year. Convert it to shape-scale notation and calculate its mean and variance. Assume no mortality model is involved and no financial discounting is applied (\(i=0\)). The distributional assumption is \(\lambda\sim\operatorname{Gamma}(\alpha=6,\beta=3)\) in shape-rate form.
Method. Keep the shape and take the reciprocal of the rate: \(\theta=1/\beta=1/3\). The equivalent prior is therefore \(\operatorname{Gamma}(6,1/3)\) in shape-scale notation.
Mean. In rate notation, \(E[\lambda]=\alpha/\beta=6/3=2\) claims per vehicle-year. In scale notation, \(E[\lambda]=\alpha\theta=6(1/3)=2\), confirming the conversion.
Variance. In rate notation, \(\operatorname{Var}(\lambda)=\alpha/\beta^2=6/3^2=6/9=2/3\). In scale notation, the same result is \(\alpha\theta^2=6(1/3)^2=6/9=2/3\).
Check and trap. Both conventions give the same mean and variance. Treating 3 as a scale instead would give mean \(6(3)=18\), not 2. The error is using a rate as though it were a scale.
Gamma-Poisson Updating in Both Conventions
Suppose observed claims \(N\), conditional on the risk’s frequency \(\lambda\), follow a Poisson distribution with known exposure \(E\): \(N\mid\lambda\sim\operatorname{Poisson}(E\lambda)\). Assume the exposure is fixed and measured in the same units used for \(\lambda\), and that claim counts are conditionally independent over disjoint exposure. If the prior is Gamma shape-rate, \(\operatorname{Gamma}(\alpha,\beta)\), the posterior is shape-rate \(\operatorname{Gamma}(\alpha+N,\beta+E)\). This update is specific to the Poisson likelihood with exposure; it is not a general rule for every likelihood or for an unknown exposure.
The reason for the update follows directly from the likelihood. Apart from factors that do not depend on \(\lambda\), the Poisson likelihood is proportional to \(\lambda^N e^{-E\lambda}\). Multiplying it by the Gamma prior adds \(N\) to the power of \(\lambda\) and \(E\) to the coefficient of \(-\lambda\) in the exponential.
The shape-rate prior is proportional to \(\lambda^{\alpha-1}e^{-\beta\lambda}\).
The likelihood is proportional to \(\lambda^N e^{-E\lambda}\), so the product is proportional to \(\lambda^{\alpha+N-1}e^{-(\beta+E)\lambda}\).
This kernel is Gamma shape-rate with shape \(\alpha+N\) and rate \(\beta+E\). The updated scale is \(1/(\beta+E)\), not the prior scale plus exposure.
If the prior is instead given in shape-scale form as \(\operatorname{Gamma}(\alpha,\theta)\), first convert its scale to the rate \(\beta=1/\theta\). The posterior shape-rate parameters are \(\alpha+N\) and \(1/\theta+E\), so the posterior scale is \(1/(1/\theta+E)\). Exposure is added to the prior rate, not to the prior scale.
Posterior Mean as a Credibility-Weighted Average
For \(\beta>0\) and \(E>0\), the posterior mean can be written as a weighted average of the prior mean and the observed claim rate \(N/E\). This form makes the role of exposure explicit.
To verify the identity, put the two terms on the right over the common denominator \(\beta+E\). Their numerator is \(N+\alpha\), which gives the posterior mean. The observed rate receives weight \(Z=E/(\beta+E)\); the prior mean receives weight \(1-Z=\beta/(\beta+E)\). These weights are nonnegative and sum to one. This interpretation presumes positive exposure; with \(E=0\), the data provide no observed rate and the posterior mean remains the prior mean.
The rate parameter \(\beta\) acts as prior exposure in this expression: for fixed prior mean, larger \(\beta\) gives the prior more weight against a given observed exposure. In shape-scale notation this prior strength is \(1/\theta\), not \(\theta\). The credibility-weight formula is therefore another place where confusing rate and scale changes the result.
Worked Example: Update a Rate Prior and Calculate Credibility
A hypothetical group of delivery vans has prior claim frequency \(\lambda\sim\operatorname{Gamma}(4,2)\) in shape-rate notation, per van-year. During three van-years of exposure, the group has nine claims. Calculate the posterior mean and its credibility-weighted form. Assume \(N\mid\lambda\sim\operatorname{Poisson}(E\lambda)\), with \(N=9\) and \(E=3\); there is no mortality assumption, and no financial discounting is applied (\(i=0\)).
Method. Update shape and rate separately: \(\alpha'=4+9=13\), \(\beta'=2+3=5\). Thus the posterior is \(\operatorname{Gamma}(13,5)\), and its mean is \(13/5=2.6\) claims per van-year.
The prior mean is \(4/2=2\), and the observed rate is \(9/3=3\). The credibility weight on experience is \(Z=E/(\beta+E)=3/(2+3)=3/5=0.6\). The posterior mean in credibility form is \(0.6(3)+0.4(2)=1.8+0.8=2.6\), agreeing with the direct update.
Check and trap. The weights sum to \(0.6+0.4=1\), and the result lies between the prior mean and observed rate. A common error is to use the prior scale \(1/2\) as though it were the rate and add exposure to it, obtaining \((4+9)/(0.5+3)=13/3.5\approx3.714\). That calculation updates the wrong parameter; the correct rate is \(2\), giving \(13/(2+3)=2.6\).
Worked Example: Update a Prior Supplied in Scale Notation
A hypothetical equipment portfolio has prior claim frequency \(\lambda\sim\operatorname{Gamma}(3,0.25)\) in shape-scale notation, measured per policy-year. It records five claims over \(1.5\) policy-years. Find the posterior scale, posterior mean, and experience credibility weight. Assume \(N\mid\lambda\sim\operatorname{Poisson}(E\lambda)\), with \(N=5\), \(E=1.5\); no mortality model applies and no financial discounting is used (\(i=0\)).
Method. Convert the prior scale to a rate: \(\beta=1/0.25=4\). The prior mean is \(\alpha/\beta=3/4=0.75\) claims per policy-year. The posterior shape is \(3+5=8\), and its rate is \(4+1.5=5.5\). Therefore its scale is \(1/5.5=2/11\), and its mean is \(8/5.5=16/11\approx1.4545\) claims per policy-year.
The observed rate is \(5/1.5=10/3\). The experience weight is \(Z=1.5/(4+1.5)=3/11\), and the prior weight is \(8/11\). The weighted mean is \((3/11)(10/3)+(8/11)(3/4)=10/11+6/11=16/11\), matching the direct posterior mean.
Check and trap. The posterior scale is \(2/11\), less than the prior scale \(0.25\), as added information increases the rate and reduces the scale. Adding exposure directly to the prior scale would give \(0.25+1.5=1.75\), which is not the posterior scale; multiplying that value by the posterior shape would incorrectly give a mean of \(8(1.75)=14\).
Common Mistakes and Practical Checks
When a Gamma prior is stated without naming its parameterization, do not assume what the second parameter means. Check the author’s density or an explicitly stated mean. If the density includes \(e^{-\beta\lambda}\), the second parameter is a rate. If it includes \(e^{-\lambda/\theta}\), it is a scale. If only a pair of numbers is supplied and the convention is not specified, the prior is ambiguous.
- Do not add exposure to the scale. Convert the scale to a rate, add exposure to that rate, and then take the reciprocal if the posterior scale is needed.
- Do not mix mean formulas. Use \(\alpha/\beta\) for a rate and \(\alpha\theta\) for a scale.
- Check exposure units. A prior rate expressed in policy-years must be combined with exposure in policy-years, not with a count of months unless that exposure is converted consistently.
- Check the credibility weights. \(E/(\beta+E)\) and \(\beta/(\beta+E)\) should be between zero and one and sum to one.
- Check the posterior mean’s location. With positive exposure, it is a weighted average of the prior mean and observed rate, so it cannot fall outside the interval between them.
These checks are useful in exam calculations and in practical model review: a plausible-looking number can still come from an inconsistent parameter convention. Keeping the density, units, and conversion visible prevents that error. This treatment of Gamma parameterization and conjugate updating is aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.
Check Your Understanding
For each problem, identify the parameterization before updating or interpreting the Gamma distribution.
- A claim-frequency prior is \(\operatorname{Gamma}(5,2)\) in shape-rate notation. Convert it to shape-scale notation and calculate its prior mean. (Answer: shape 5, scale 0.5, mean 2.5)
- A prior is \(\operatorname{Gamma}(4,0.2)\) in shape-scale notation. Find the equivalent rate and prior mean. (Answer: rate 5, mean 0.8)
- A shape-rate prior has \(\alpha=3\), \(\beta=1.5\). The risk has six claims over two exposure units. Find the posterior mean and the credibility weight on the observed rate. (Answer: posterior mean \(18/7\approx2.5714\); experience weight \(4/7\approx0.5714\))
- A prior is \(\operatorname{Gamma}(2,0.5)\) in shape-scale notation. Given four claims over three exposure units, find the posterior rate, posterior scale, and posterior mean. (Answer: rate 5, scale 0.2, mean 1.2)
- A prior is \(\operatorname{Gamma}(5,0.4)\) in shape-scale notation. There are seven claims over two exposure units. Calculate the correct posterior mean and the incorrect mean obtained by treating 0.4 as a rate and adding the exposure to it. (Answer: correct mean \(8/3\approx2.6667\); incorrect mean 5)