Common Errors in Conjugate Credibility Calculations

Why Conjugate Credibility Calculations Go Wrong

Conjugate models make Bayesian credibility calculations manageable: the posterior distribution has the same family as the prior. But a tidy formula does not protect against using the wrong exposure, reversing a Gamma parameter convention, or updating on the wrong event. These errors can produce a plausible-looking rate or probability that is nevertheless based on a different model from the one intended.

The central discipline is to define the unknown quantity and the observation process before updating. For claim frequency, ask what one unit of exposure represents and how much exposure was observed. For claim incidence, identify the binary event recorded for each trial. Then match the likelihood and prior to those definitions. The checks in this tutorial are aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.

Definition: In a Gamma-Poisson model, the unknown claim rate is \(\lambda\), measured per stated exposure unit. Using the shape-rate convention, write \(\lambda\sim\operatorname{Gamma}(\alpha,\beta)\), where \(\alpha\) is the shape and \(\beta\) is the rate parameter. Conditional on exposure \(e\), the claim count satisfies \(N\mid\lambda\sim\operatorname{Poisson}(\lambda e)\). In a Beta-Binomial model, the unknown probability is \(p\), with \(p\sim\operatorname{Beta}(a,b)\); conditional on \(p\), the number of specified binary events among \(m\) trials is \(Y\mid p\sim\operatorname{Binomial}(m,p)\).

Gamma-Poisson Updates: Exposure and Parameterization

Under the shape-rate convention, the Gamma prior density is proportional to \(\lambda^{\alpha-1}e^{-\beta\lambda}\), and its mean is \(\alpha/\beta\). The posterior after observing \(n\) claims over exposure \(e\) is \(\operatorname{Gamma}(\alpha+n,\beta+e)\), with posterior mean \((\alpha+n)/(\beta+e)\). Here \(\beta\) is measured in exposure units, while \(\lambda\) is measured in claims per exposure unit.

Exposure must use the same time and risk-unit basis as the rate. For \(r\) identical risks each observed for \(t\) years, exposure is \(e=rt\) risk-years. A policy-year is not interchangeable with a policy if coverage periods differ. If exposures vary by risk, add the individual exposures; do not simply count the risks.

Worked Example: Converting Coverage into Exposure

A hypothetical insurer models a class’s annual claim frequency per vehicle-year with a Gamma prior \(\operatorname{Gamma}(3,100)\) in shape-rate form. It observes 8 claims among 80 vehicles, each covered for 18 months. Find the posterior mean claim frequency per vehicle-year.

Method and exposure check. Convert 18 months to \(18/12=1.5\) years per vehicle. Total exposure is \(e=80(1.5)=120\) vehicle-years. The observed count is \(n=8\); the number of vehicles alone is not the exposure.

Update. The posterior shape is \(3+8=11\), and its rate parameter is \(100+120=220\) vehicle-years. Thus the posterior is \(\operatorname{Gamma}(11,220)\), and the posterior mean is \(11/220=0.05\) claims per vehicle-year.

Check. The prior mean is \(3/100=0.03\) claims per vehicle-year; the observed rate is \(8/120=0.0667\) claims per vehicle-year. Since the posterior mean lies between those two rates, the result is directionally sensible. The common mistake avoided is treating 80 vehicles as 80 vehicle-years despite the 18-month coverage period.

A second frequent error is to read the second Gamma parameter as a scale when it is specified as a rate. In shape-scale form, the scale is the reciprocal of the rate: if the shape-rate parameters are \((\alpha,\beta)\), the equivalent shape-scale parameters are \((\alpha,1/\beta)\). The scale has the units of \(\lambda\), not the units of exposure.

Worked Example: Converting Shape-Rate to Shape-Scale

A hypothetical portfolio has an unknown claim rate per unit-year with prior \(\operatorname{Gamma}(2,40)\) in shape-rate form. It records 3 claims over 20 unit-years. Calculate the posterior mean and express the posterior using both parameter conventions.

Method. The prior rate parameter is \(\beta=40\) unit-years, and the prior mean is \(2/40=0.05\) claims per unit-year. The observed exposure is \(e=20\) unit-years, so the posterior shape is \(2+3=5\), and the posterior rate parameter is \(40+20=60\) unit-years.

Shape-rate result. The posterior is \(\operatorname{Gamma}(5,60)\), so its mean is \(5/60=1/12=0.08333\) claims per unit-year.

Shape-scale result. The equivalent scale is \(1/60\) claims per unit-year. Therefore, in shape-scale form the posterior is \(\operatorname{Gamma}(5,1/60)\). The prior scale, correspondingly, was \(1/40\) claims per unit-year.

Check. The observed rate is \(3/20=0.15\) claims per unit-year, above the prior mean of \(0.05\); the posterior mean of \(0.08333\) moves toward it without exceeding it. The common mistake avoided is treating 60, the posterior rate parameter, as a scale. Doing so would imply a mean of \(5(60)=300\), rather than \(5/60\).

Why the Gamma-Poisson Formula Holds

The update follows by multiplying the prior density by the likelihood. For a Poisson count \(n\) over exposure \(e\), the likelihood as a function of \(\lambda\) is proportional to \(e^{-e\lambda}\lambda^n\). Multiplying by the Gamma prior kernel gives

\lambda^{\alpha-1}e^{-\beta\lambda}\times \lambda^n e^{-e\lambda} = \lambda^{\alpha+n-1}e^{-(\beta+e)\lambda}.

This is the kernel of a Gamma distribution with shape \(\alpha+n\) and rate \(\beta+e\). The derivation also explains the units: \(\beta\) and \(e\) are added, so they must be in the same exposure units. If the rate is defined per vehicle-year, the prior rate parameter and observed exposure must both be expressed in vehicle-years.

When risks have different exposures \(e_j\), with conditionally independent counts \(N_j\mid\lambda\sim\operatorname{Poisson}(\lambda e_j)\), the likelihood kernel depends on the data through the total count \(\sum_jN_j\) and total exposure \(\sum_je_j\). The update therefore uses \(n=\sum_jN_j\) and \(e=\sum_je_j\), not the number of records alone.

Beta-Binomial Updates: Define the Event Before Counting

A Beta-Binomial model applies when each trial contributes a binary outcome for one specified event. With prior \(p\sim\operatorname{Beta}(a,b)\), \(y\) observed events among \(m\) trials gives posterior \(\operatorname{Beta}(a+y,b+m-y)\). Its mean is \((a+y)/(a+b+m)\). The event must be counted once per trial: for claim incidence, a policy with several claims is still one policy with at least one claim.

Justification: The Binomial likelihood is proportional to \(p^y(1-p)^{m-y}\), and the Beta prior kernel is \(p^{a-1}(1-p)^{b-1}\). Their product is proportional to \(p^{a+y-1}(1-p)^{b+m-y-1}\), the kernel of \(\operatorname{Beta}(a+y,b+m-y)\). Its posterior mean follows from the Beta mean formula.

Worked Example: Incidence Is Not the Same as Claim Count

A hypothetical insurer assigns a prior claim-incidence probability \(p\sim\operatorname{Beta}(2,18)\) to a class of policies. Among 40 one-year policies, 7 have at least one claim, and those policies generate 11 claims in total. Find the posterior mean incidence probability. Also calculate the separate frequency estimate under a specified Gamma-Poisson model.

Incidence update. The binary event is “this policy has at least one claim.” Thus \(m=40\) and \(y=7\), even though the event policies produced 11 claims altogether. The posterior is \(\operatorname{Beta}(2+7,18+40-7)=\operatorname{Beta}(9,51)\). Its mean is \(9/(9+51)=9/60=0.15\).

Frequency update. For comparison, suppose the frequency prior is \(\operatorname{Gamma}(2,20)\) in shape-rate form, measured per policy-year. The total exposure is \(40\) policy-years, and the total claim count is 11. The posterior is \(\operatorname{Gamma}(2+11,20+40)=\operatorname{Gamma}(13,60)\), with mean \(13/60=0.2167\) claims per policy-year.

Check. Incidence is a probability of at least one claim per policy; frequency is an expected number of claims per policy-year. Their values need not be equal. The common mistake avoided is substituting 11 total claims for 7 policies with the binary incidence event.

Posterior Weighting and Denominator Checks

The Beta posterior mean can be written as a weighted average of the prior mean and the observed event proportion. Let \(m\) be the number of trials and \(y/m\) the observed proportion. Then

\frac{a+y}{a+b+m} = \frac{a+b}{a+b+m}\frac{a}{a+b} + \frac{m}{a+b+m}\frac{y}{m}.

The two weights sum to one. The prior contributes weight \(a+b\), and the data contribute weight \(m\). A common denominator error is to divide the updated success parameter by the posterior failure parameter alone. For \(\operatorname{Beta}(9,51)\), the mean is \(9/(9+51)\), not \(9/51\): both posterior parameters contribute to the total.

For the Gamma-Poisson model, the posterior mean is \((\alpha+n)/(\beta+e)\). It is not generally a weighted average of the prior mean and observed rate using weights \(\alpha\) and \(e\), because the prior mean is \(\alpha/\beta\) and the observed rate is \(n/e\); the numerator’s prior contribution is \(\alpha\), while the denominator’s prior contribution is \(\beta\). Check the actual posterior formula rather than transferring the Beta weighting expression mechanically.

Define the target.
State whether the unknown is a rate per exposure unit or a probability of a specified binary event.

Match the data to the likelihood.
Use total claims and total exposure for a Poisson frequency model; use event flags and number of trials for a Binomial incidence model.

Check conventions and units.
For Gamma shape-rate form, add exposure to the rate parameter. For shape-scale form, use the reciprocal scale. Keep time units consistent.

Verify the posterior.
Check that parameters are positive, the posterior mean has the target’s units, and the result responds sensibly to the observed experience.

Common Exam Traps

Counting risks instead of exposure. A portfolio of 80 vehicles covered for 18 months contributes 120 vehicle-years, not 80 vehicle-years.
Reversing Gamma rate and scale. In shape-rate form the mean is \(\alpha/\beta\); in shape-scale form it is \(\alpha\theta\), where \(\theta=1/\beta\). The scale’s units are the units of the unknown rate.
Updating on the wrong event. For claim incidence, one policy is one trial and “at least one claim” is one success, regardless of how many claims occurred.
Omitting part of the posterior total. A Beta mean divides the posterior success parameter by the sum of both posterior parameters.
Mixing frequency and incidence interpretations. A Poisson claim rate can exceed one; a probability cannot. State which quantity is being estimated before interpreting the answer.

Key takeaway: Conjugacy is useful only after the likelihood, exposure basis, and parameter convention are correct. Before accepting a posterior estimate, verify what one observation means, how much exposure it represents, and which parameterization determines the mean.

Check Your Understanding

For each problem, identify the model and check its units before calculating.

A Gamma prior is \(\operatorname{Gamma}(2,30)\) in shape-rate form, measured per unit-year. There are 4 claims over 30 unit-years. Find the posterior distribution and mean. (Answer: \(\operatorname{Gamma}(6,60)\); mean \(0.1000\) claims per unit-year.)
A frequency prior is \(\operatorname{Gamma}(3,12)\) in shape-rate form per policy-year. Eighteen policies are each observed for six months, with 2 claims in total. Find the posterior mean. (Answer: \(\operatorname{Gamma}(5,21)\); mean \(5/21\approx0.2381\) claims per policy-year.)
A claim-incidence prior is \(\operatorname{Beta}(4,16)\). Among 50 policies, 9 have at least one claim. Find the posterior distribution and mean. (Answer: \(\operatorname{Beta}(13,57)\); mean \(13/70\approx0.1857\).)
Thirty policies each contribute one policy-year. Three policies have at least one claim, while the group has 5 claims total. Use prior \(\operatorname{Beta}(1,9)\) for incidence and prior \(\operatorname{Gamma}(2,18)\), shape-rate form, for frequency. Find both posterior means. (Answer: incidence \(4/40=0.1000\); frequency \(7/48\approx0.1458\) claims per policy-year.)
A prior is \(\operatorname{Gamma}(5,0.02)\) in shape-scale form for a rate per exposure unit. There are 6 claims over 50 exposure units. Find the posterior mean and state the posterior in shape-rate form. (Answer: prior rate parameter \(50\); posterior \(\operatorname{Gamma}(11,100)\) in shape-rate form; mean \(0.1100\) claims per exposure unit.)

What You'll Learn