Parameterization Traps in Bayesian Credibility Calculations

Why Parameterization Errors Matter

Gamma conjugacy makes Bayesian updating efficient, but a correct-looking calculation can still produce a wrong result if the gamma parameterization is misunderstood. Two choices are common: the second parameter can be a rate or a scale. These parameters are reciprocals, not interchangeable labels. A second, separate trap is treating the number of claims as the amount of exposure. In a Poisson model, the observed count updates the shape parameter, while exposure updates the rate parameter.

Severity calculations introduce another distinction. If a gamma prior is placed on an exponential severity rate, the expected severity is an expectation of the rate's reciprocal. It is generally not the reciprocal of the posterior mean rate. Confusing these quantities can distort a pure premium even when the frequency calculation is correct.

Definition: In the shape-rate convention, $\Theta\sim\operatorname{Gamma}(\alpha,\beta)$ has density $$ f_\Theta(\theta)=\frac{\beta^\alpha}{\Gamma(\alpha)}\theta^{\alpha-1}e^{-\beta\theta},\qquad \theta>0, $$ with $E[\Theta]=\alpha/\beta$ and $\operatorname{Var}(\Theta)=\alpha/\beta^2$. In the shape-scale convention, $\Theta\sim\operatorname{Gamma}(\alpha,s)$ has density proportional to $\theta^{\alpha-1}e^{-\theta/s}$, with mean $\alpha s$ and variance $\alpha s^2$. The rate and scale satisfy $\beta=1/s$.

Always write down which convention is being used before applying an updating formula. In particular, in a gamma distribution for a Poisson intensity, the gamma rate has units of exposure (such as years); the intensity itself has units of claims per exposure unit. For a gamma prior on an exponential severity rate, the gamma rate has units of severity.

Gamma-Poisson Updating: Count and Exposure Have Different Jobs

Suppose a risk has a constant claim frequency intensity $\lambda$ per exposure unit. Conditional on $\lambda$, the claim count $N$ observed over exposure $e$ has a Poisson distribution with mean $e\lambda$. Assume the prior is $\lambda\sim\operatorname{Gamma}(\alpha,\beta)$ in the shape-rate convention, with $\alpha,\beta>0$.

Theorem: Given $N=n$ claims over exposure $e$, the posterior distribution is $$ \lambda\mid N=n\sim\operatorname{Gamma}(\alpha+n,\beta+e). $$ Its posterior mean is $(\alpha+n)/(\beta+e)$. If the prior is instead stated with scale $s$, first use $\beta=1/s$; equivalently, the posterior scale is $s/(1+es)$.

The likelihood as a function of $\lambda$ is proportional to $\lambda^n e^{-e\lambda}$. Multiplying it by the gamma prior density adds $n$ to the power $\alpha-1$, giving the posterior shape $\alpha+n$, and adds $e$ to the exponent coefficient $\beta$, giving the posterior rate $\beta+e$. This shows why claim count and exposure cannot be swapped: they enter different parts of the posterior.

Worked Example: Converting a Scale Prior Before Updating

A hypothetical portfolio has a constant annual claim intensity $\lambda$. Its prior is $\operatorname{Gamma}(3,2)$ in the shape-scale convention. The portfolio produces 4 claims over 2 exposure-years. Find the posterior mean annual intensity. Assume a Poisson count model and no time variation in intensity.

Method. Convert the prior to shape-rate form, apply the gamma-Poisson update, and check the answer in the scale convention.

The prior scale is $s=2$, so the prior rate is $\beta=1/2=0.5$. Thus the prior mean is $3(2)=6$, or equivalently $3/0.5=6$, claims per year.

The posterior shape is $3+4=7$. The posterior rate is $0.5+2=2.5$, since the exposure is 2 years. Therefore,

E[\lambda\mid N=4]=\frac{7}{2.5}=2.8

For a check using scale, the posterior scale is $2/(1+2(2))=2/5=0.4$, and its mean is $7(0.4)=2.8$. The posterior mean is therefore 2.8 claims per year.

Common mistake avoided. Treating the stated scale 2 as if it were a rate gives the incorrect posterior rate $2+2=4$ and mean $7/4=1.75$. The prior mean under that mistaken interpretation would also be $3/2$, contradicting the stated shape-scale prior mean of 6.

Worked Example: Using Exposure Rather Than Claim Count

A hypothetical risk has annual intensity $\lambda\sim\operatorname{Gamma}(2,1)$ in the shape-rate convention. It has 6 claims over 3 exposure-years. Find the expected number of claims during a future half-year, assuming its intensity is unchanged.

Method. Update the annual intensity using the observed count and exposure separately, then multiply the posterior mean intensity by the future exposure.

The posterior shape is $2+6=8$. The posterior rate is $1+3=4$, so the posterior mean annual intensity is $8/4=2$ claims per year. For future exposure $h=0.5$ years, the posterior expected count is

E[N_{\mathrm{future}}\mid\text{data}]=hE[\lambda\mid\text{data}]=0.5(2)=1\text{ claim}.

Check. The posterior mean is 2 claims per year, and half a year at that expected intensity gives 1 expected claim.

Common mistake avoided. Substituting the observed count for exposure would give a posterior rate $1+6=7$, posterior mean $8/7$, and half-year expected count $4/7\approx0.5714$. The observed count belongs in the shape update; the 3 exposure-years belong in the rate update.

Severity Rates and Expected Severity

A related conjugate model applies when claim severities, measured in a specified monetary unit, are conditionally exponential. Let each severity $X_j$, given an unknown rate $\theta$, have density $\theta e^{-\theta x}$ for $x>0$. Then $E[X_j\mid\theta]=1/\theta$. Assume the prior on $\theta$ is $\operatorname{Gamma}(\alpha,\beta)$ in the shape-rate convention, and that the observed severities are conditionally independent given $\theta$.

Theorem: For $m$ observed severities $x_1,\ldots,x_m$, the posterior rate parameter is $$ \theta\mid x_1,\ldots,x_m\sim\operatorname{Gamma}\left(\alpha+m,\beta+\sum_{j=1}^{m}x_j\right). $$ When $\alpha+m>1$, the posterior expected severity of a future claim is $$ E[X_{\mathrm{future}}\mid x_1,\ldots,x_m] =E[1/\theta\mid x_1,\ldots,x_m] =\frac{\beta+\sum_{j=1}^{m}x_j}{\alpha+m-1}. $$

The severity likelihood is proportional to $\theta^m e^{-\theta\sum x_j}$. Multiplying by the prior gives the stated posterior shape and rate. To obtain expected severity, integrate the reciprocal of the posterior rate parameter, rather than taking the reciprocal of its mean. For any $\Theta\sim\operatorname{Gamma}(a,b)$ in the shape-rate convention with $a>1$,

E[1/\Theta] =\frac{b^a}{\Gamma(a)}\int_0^\infty \theta^{a-2}e^{-b\theta}\,d\theta =\frac{b^a}{\Gamma(a)}\frac{\Gamma(a-1)}{b^{a-1}} =\frac{b}{a-1}.

The integral is finite only for $a>1$. In contrast, the reciprocal of the posterior mean rate is $b/a$, which is not the same value. The posterior predictive expected severity requires $E[1/\theta]$ because the conditional expected severity is $1/\theta$.

Putting Frequency and Severity Together

Suppose the frequency intensity $\lambda$ and severity rate $\theta$ have independent priors, and the frequency and severity data update them separately under the models just stated. For future exposure $h$, let $S_{\mathrm{future}}$ be the sum of future claim severities. Conditional on $\lambda,\theta$, the expected number of claims is $h\lambda$, and the expected severity per claim is $1/\theta$. Thus the conditional expected aggregate loss is $h\lambda/\theta$. Posterior independence then gives

E[S_{\mathrm{future}}\mid\text{data}] =hE[\lambda\mid\text{frequency data}]E[1/\theta\mid\text{severity data}].

This is an undiscounted expected aggregate loss, or pure premium, for the stated future exposure. It relies on posterior independence of the two parameters; if their posterior distributions are dependent, the relevant expression is $hE[\lambda/\theta\mid\text{data}]$, which need not factor into two expectations.

Worked Example: Separating Frequency and Severity Posterior Means

A hypothetical book has annual Poisson intensity $\lambda\sim\operatorname{Gamma}(2,1)$, shape-rate, and observed 3 claims over 2 exposure-years. Independently, claim severities in thousands of dollars are exponential with unknown rate $\theta$, where $\theta\sim\operatorname{Gamma}(3,2)$, also shape-rate. The three observed severities are $0.5$, $1.0$, and $2.5$ thousand dollars. Find the posterior expected aggregate loss over one future year. Assume independent frequency and severity priors, conditional independence within each model, and no discounting.

Method. Update the frequency and severity models separately, calculate the predictive mean severity as a reciprocal-rate expectation, and multiply by the expected future count.

For frequency, the posterior is $\operatorname{Gamma}(2+3,1+2)=\operatorname{Gamma}(5,3)$. Hence $E[\lambda\mid\text{data}]=5/3$ claims per year.

For severity, the observed total is $0.5+1.0+2.5=4.0$ thousand dollars. The posterior is $\operatorname{Gamma}(3+3,2+4)=\operatorname{Gamma}(6,6)$, where the posterior gamma rate 6 is measured in thousands of dollars. Since the posterior shape is greater than 1,

E[1/\theta\mid\text{severity data}]=\frac{6}{6-1}=1.2\text{ thousand dollars per claim}.

For one future year, the expected aggregate loss is therefore

1\left(\frac{5}{3}\right)(1.2)=2.0\text{ thousand dollars}=\$2{,}000.00.

Check. The posterior expected count is $5/3$ claims, and each future claim has posterior expected severity $1.2$ thousand dollars. Their product is $2.0$ thousand dollars.

Common mistake avoided. The posterior mean severity rate is $6/6=1$ per thousand dollars. Taking its reciprocal gives 1.0 thousand dollars, not the required posterior expected severity of 1.2 thousand dollars. That error would produce the incorrect pure premium $(5/3)(1.0)=1.6667$ thousand dollars, or approximately $\$1{,}666.67$.

Diagnostic Checks for Exam Calculations

The calculation should be checked for both algebra and units. A posterior parameter can be numerically plausible yet still have the wrong units or represent the wrong quantity. Before reporting a result, verify the following:

Write the convention. Record “shape-rate” or “shape-scale” beside every gamma prior. Convert scale $s$ to rate $1/s$ before using shape-rate formulas.
Separate count from exposure. For a Poisson intensity, add the claim count to the shape and exposure to the rate. The future expected count is future exposure times posterior expected intensity.
Keep severity units consistent. If severities are recorded in thousands of dollars, their sum and the gamma rate update must use thousands of dollars too. Convert the final premium to dollars only after calculating it in the stated unit.
Identify the requested expectation. For exponential severity, the posterior expected severity is $E[1/\theta\mid\text{data}]$, not $1/E[\theta\mid\text{data}]$.
Check existence. The posterior expected reciprocal rate is finite only when the posterior gamma shape exceeds 1.

These checks are useful because conjugacy can make an incorrect update look tidy. A parameterization label, a dimensional check, and a final posterior-predictive interpretation together help locate the error. This parameterization discipline is aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.

Check Your Understanding

For each problem, state the gamma convention and keep exposure and monetary units explicit before calculating.

A hypothetical risk has annual intensity $\lambda\sim\operatorname{Gamma}(2,3)$ in shape-scale form. It produces 1 claim in 4 exposure-years. Find the posterior mean annual intensity. (Answer: $9/13\approx0.6923$ claims per year.)
A hypothetical risk has $\lambda\sim\operatorname{Gamma}(4,2)$ in shape-rate form and produces 5 claims over 2 years. Find the expected claim count in a future 0.4-year period. (Answer: $0.9$ claims.)
Claim severities, in thousands of dollars, are exponential with unknown rate $\theta\sim\operatorname{Gamma}(4,3)$ in shape-rate form. Two observed claims total 5 thousand dollars. Find the posterior expected severity of a future claim. (Answer: $1.6$ thousand dollars, or $\$1{,}600.00$.)
A hypothetical portfolio has $\lambda\sim\operatorname{Gamma}(3,1)$ in shape-rate form and observes 2 claims over 1 exposure-year. Independently, severities in thousands of dollars are exponential with rate $\theta\sim\operatorname{Gamma}(3,1)$, shape-rate, and the two observed severities total 2 thousand dollars. Find the posterior expected aggregate loss for 0.5 future exposure-years. (Answer: $0.9375$ thousand dollars, or $\$937.50$.)
A hypothetical frequency prior is $\lambda\sim\operatorname{Gamma}(3,0.5)$ in shape-scale form. The observed count is 2 over 1 exposure-year. A calculation reports posterior mean $5/(0.5+1)$. Diagnose the parameterization error and give the correct posterior mean. (Answer: $5/3\approx1.6667$ claims per year; the prior scale must first be converted to rate $1/0.5=2$, giving posterior rate 3.)

What You'll Learn