CAS Exam 8 · Bayesian Credibility Foundations · Tutorial 434 of 989
Encoding Collective Rates as Bayesian Priors
Represent a collective claim rate as a Bayesian prior mean and use prior dispersion to determine how individual experience changes the estimate.
Intermediate
10 min read
What You'll Learn
Translate a collective expected claim rate into Gamma prior parameters.
Distinguish the prior mean from the prior variance across risks.
Update a rate prior using claim counts and exposure.
Explain how greater prior dispersion changes the posterior estimate.
Keep rate and exposure units consistent in Poisson-Gamma calculations.
Identify parameterization and unit-conversion errors.
From a Collective Rate to an Individual Prior
A collective claim rate summarizes experience across a group, while an individual risk’s rate may differ from that group average. A Bayesian model can represent this distinction by making the individual rate a random variable whose prior mean equals the collective rate. The prior variance then describes how much rates are expected to vary across risks before observing that individual’s own claims.
This tutorial develops that encoding for Poisson claim counts and a Gamma prior. The collective rate is the prior center, not a claim that every risk has that exact rate. Choosing the prior dispersion is consequential: holding the collective mean fixed, a more dispersed prior generally allows a given individual’s experience to move the estimate further.
Definition: Let \(\lambda\) be an individual risk’s claim rate per unit of exposure. Suppose \(N\mid\lambda\sim\operatorname{Poisson}(e\lambda)\), where \(e\) is known exposure measured in the same units used to define \(\lambda\). A Gamma prior with shape \(\alpha>0\) and rate \(\beta>0\) has density
$$
\pi(\lambda)=\frac{\beta^\alpha}{\Gamma(\alpha)}\lambda^{\alpha-1}e^{-\beta\lambda},\qquad \lambda>0.
$$
Its mean is \(\alpha/\beta\), and its variance is \(\alpha/\beta^2\). Here \(\beta\) is the rate parameter, not the scale parameter.
Let \(m>0\) be the collective expected claim rate, and let \(\tau^2>0\) be the chosen prior variance of individual rates around that mean. To make the prior mean equal to \(m\) and its variance equal to \(\tau^2\), solve
$$
\frac{\alpha}{\beta}=m,\qquad \frac{\alpha}{\beta^2}=\tau^2.
$$
The resulting parameters are
$$
\alpha=\frac{m^2}{\tau^2},\qquad \beta=\frac{m}{\tau^2}.
$$
The variance \(\tau^2\) describes the assumed distribution of underlying rates among risks, not the sampling variance of a collective rate estimate. This specification treats \(m\) and \(\tau^2\) as fixed prior inputs. It does not, by itself, model uncertainty in the collective mean or account for trends, rating variables, or changes in the claim process.
Updating the Prior With Individual Experience
For a known exposure \(e\), the Poisson likelihood for observing \(N=n\) claims is proportional, as a function of \(\lambda\), to \(\lambda^n e^{-e\lambda}\). Multiplying this likelihood by the Gamma prior gives a posterior Gamma distribution with shape \(\alpha+n\) and rate \(\beta+e\). The posterior mean is therefore:
This calculation requires a conditionally Poisson count with mean \(e\lambda\), known exposure, and a prior defined on the same rate scale as \(\lambda\). The model assumes the risk’s rate is constant over the observation period, conditional on \(\lambda\). If exposure varies by time, \(e\) can still be total exposure when the rate is constant and the count process has the stated Poisson model. If the rate definition or exposure unit changes, the prior and likelihood must be put on a consistent scale before updating.
Why the update works: The prior density contributes the factor \(\lambda^{\alpha-1}e^{-\beta\lambda}\), while the Poisson likelihood contributes \(\lambda^n e^{-e\lambda}\), omitting terms that do not depend on \(\lambda\). Their product is proportional to
$$
\lambda^{\alpha+n-1}e^{-(\beta+e)\lambda}.
$$
This is the kernel of a Gamma density with shape \(\alpha+n\) and rate \(\beta+e\). Its mean follows from the Gamma mean formula, giving \((\alpha+n)/(\beta+e)\).
Substituting the prior parameters also gives a useful expression in terms of the prior mean \(m\) and observed rate \(n/e\), for \(e>0\):
The two coefficients are nonnegative and sum to one. Thus the posterior mean lies between the collective rate and the individual observed rate. The dispersion affects the balance: increasing \(\tau^2\) decreases \(m/\tau^2\), so the observed rate receives more influence, all else equal. If \(e=0\), there is no individual claim experience to contribute and the posterior mean remains \(m\).
Worked Applications
Worked Example: Encoding a Collective Claim Frequency
A hypothetical portfolio has a collective claim frequency of \(0.08\) claims per policy-year. For individual policyholder rates, an actuary chooses a prior variance of \(0.0016\) in squared claim-frequency units. One policyholder has 3 claims over 20 policy-years. Find the prior parameters and posterior mean.
Method. Match the Gamma prior’s mean and variance to the selected collective mean and prior variance, then apply the Poisson-Gamma update. The exposure is measured in policy-years, so the rate is claims per policy-year.
The prior inputs are \(m=0.08\) and \(\tau^2=0.0016\). Therefore,
$$
\alpha=\frac{m^2}{\tau^2}=\frac{0.08^2}{0.0016}
=\frac{0.0064}{0.0016}=4,
$$
and
$$
\beta=\frac{m}{\tau^2}=\frac{0.08}{0.0016}=50.
$$
The prior mean check is \(4/50=0.08\), and the variance check is \(4/50^2=4/2500=0.0016\).
With \(n=3\) and \(e=20\), the posterior parameters are shape \(4+3=7\) and rate \(50+20=70\). Hence the posterior mean is
$$
\frac{7}{70}=0.10
$$
claims per policy-year. The individual observed rate is \(3/20=0.15\), so the posterior mean of \(0.10\) lies between the collective rate \(0.08\) and the observed rate \(0.15\), as expected. Rates are reported to four decimal places: \(0.1000\).
Check. The posterior mean is \((50/70)(0.08)+(20/70)(0.15)=0.0571429+0.0428571=0.1000\). A common mistake is to use the observed rate \(0.15\) alone and ignore the collective prior.
Worked Example: Changing Dispersion While Holding the Collective Rate Fixed
Two hypothetical classifications share a collective rate of \(0.10\) claims per exposure unit. For the first, the selected prior variance is \(0.0004\); for the second, it is \(0.0016\). For an individual in either classification, the experience is 4 claims over 25 exposure units. Calculate both posterior means and identify which prior permits the experience to move the estimate further.
Method. Encode each variance separately while keeping \(m=0.10\), then use the same individual count and exposure in both updates. The observed rate is \(4/25=0.16\).
For the first prior,
$$
\alpha_1=\frac{0.10^2}{0.0004}=25,\qquad
\beta_1=\frac{0.10}{0.0004}=250.
$$
The posterior mean is \((25+4)/(250+25)=29/275=0.1054545\), or \(0.1055\) claims per exposure unit to four decimal places.
For the second prior,
$$
\alpha_2=\frac{0.10^2}{0.0016}=6.25,\qquad
\beta_2=\frac{0.10}{0.0016}=62.5.
$$
The posterior mean is \((6.25+4)/(62.5+25)=10.25/87.5=0.1171429\), or \(0.1171\) claims per exposure unit to four decimal places.
The second prior has greater dispersion and its posterior mean is farther from \(0.10\), moving toward the observed rate \(0.16\). The results check against the weighted-rate expression: the larger variance reduces \(m/\tau^2\), increasing the influence of the individual experience. A common mistake is to change the prior mean when changing the variance; here both classifications retain the same collective rate.
Worked Example: Keeping Exposure Units Consistent
A hypothetical vehicle portfolio has a collective claim rate of 18 claims per 1,000 vehicle-years. The selected prior standard deviation of individual rates is 6 claims per 1,000 vehicle-years. An insured fleet produces 60 claims over 2,500 vehicle-years. Find the prior parameters and the posterior mean, retaining rates per 1,000 vehicle-years.
Method. Express both the rate and exposure in matching units. Let one exposure unit equal 1,000 vehicle-years. Then \(m=18\), \(\tau^2=6^2=36\), and \(e=2500/1000=2.5\). The observed rate is \(60/2.5=24\) claims per 1,000 vehicle-years.
The prior parameters are
$$
\alpha=\frac{18^2}{36}=\frac{324}{36}=9,\qquad
\beta=\frac{18}{36}=0.5.
$$
The prior checks are \(9/0.5=18\) and \(9/0.5^2=36\). Updating with \(n=60\) and \(e=2.5\) gives shape \(9+60=69\) and rate \(0.5+2.5=3\). Thus the posterior mean is \(69/3=23\) claims per 1,000 vehicle-years.
The result is reported as \(23.00\) claims per 1,000 vehicle-years. As a check, \((0.5/3)(18)+(2.5/3)(24)=3+20=23\). A common mistake is to insert \(e=2500\) while leaving the rate in claims per 1,000 vehicle-years. The corrected calculation uses \(e=2.5\) on the 1,000-vehicle-year scale, or alternatively converts all rates and prior parameters to a per-vehicle-year scale before updating.
Choosing and Interpreting the Prior Dispersion
The collective mean and dispersion answer different actuarial questions. The mean specifies the prior center; the variance specifies how broadly individual rates are distributed around that center. Two classifications can have the same collective mean but different expected heterogeneity, and therefore respond differently to the same observed count and exposure.
A defensible dispersion should be based on relevant actuarial information, such as the degree of variation among comparable risks, while respecting the model’s assumptions and units. It should not be selected merely to force a desired posterior result. In particular, an estimated collective mean does not automatically provide a suitable prior variance. The variance is an additional modeling choice or estimate.
The Gamma-Poisson specification is not appropriate without qualification if counts are not reasonably modeled as Poisson conditional on the rate, if exposure is measured inconsistently, or if important risk differences are omitted in a way that makes a single prior unsuitable. A prior with positive Gamma shape also requires \(m>0\) and \(\tau^2>0\); a zero collective mean or a zero variance requires a different limiting or model treatment.
Common Errors to Avoid
Confusing rate and scale. In this tutorial, \(\beta\) is the Gamma rate parameter, so the prior mean is \(\alpha/\beta\), not \(\alpha\beta\).
Using a variance formula for the wrong parameterization. With shape \(\alpha\) and rate \(\beta\), the variance is \(\alpha/\beta^2\). With a scale parameter, the formulas would be written differently.
Mixing exposure units. A rate per 1,000 exposure units must be paired with exposure measured in thousands, unless all quantities are converted to another consistent scale.
Treating the collective rate as the posterior answer. The collective rate is the prior mean; individual count experience can move the posterior mean.
Assuming greater prior variance means a larger prior mean. Changing \(\tau^2\) while holding \(m\) fixed changes both Gamma parameters but leaves the prior mean unchanged.
Key takeaway: To encode a positive collective rate \(m\) and prior variance \(\tau^2\) in a Gamma-Poisson model, use \(\alpha=m^2/\tau^2\) and \(\beta=m/\tau^2\). Keep exposure units aligned, then update with posterior parameters \(\alpha+n\) and \(\beta+e\). This formulation is aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.
Check Your Understanding
For each problem, use the Gamma shape-rate convention and keep claim-rate and exposure units consistent.
A collective rate is \(0.05\) claims per exposure unit, and the selected prior standard deviation is \(0.01\). Find \(\alpha\) and \(\beta\). (Answer: \(\alpha=25,\ \beta=500\))
A collective rate is \(0.20\) per exposure unit with prior variance \(0.01\). An individual has 2 claims over 5 exposure units. Find the posterior mean. (Answer: \(0.2400\) claims per exposure unit)
A collective rate is \(0.03\) claims per exposure unit with prior variance \(0.0009\). An individual has no claims over 10 exposure units. Find the posterior mean. (Answer: \(0.0231\) claims per exposure unit)
A collective rate is 12 claims per 1,000 exposure units, and the prior standard deviation is 4 claims per 1,000 exposure units. An account has 30 claims over 3,000 exposure units. Find the posterior mean per 1,000 exposure units. (Answer: \(10.4000\) claims per 1,000 exposure units)
Two priors have collective mean \(0.08\) claims per exposure unit, with variances \(0.0004\) and \(0.0016\). For an individual with 1 claim over 8 exposure units, compute both posterior means and state which moves farther toward the observed rate. (Answer: \(0.0817\) and \(0.0862\), respectively; the prior with variance \(0.0016\) moves farther toward the observed rate \(0.1250\))