Why Prior Weight Needs a Scale
A prior mean says where a model places its center; it does not, by itself, say how strongly the estimate should resist new experience. That strength must be interpreted on the scale contributed by the likelihood. In a Poisson rate model, data contribute exposure. In a binomial probability model, data contribute trials. Conjugate-prior parameters can therefore be read as equivalent prior exposure or prior trials, making the balance between prior information and observed experience explicit.
The word “equivalent” matters. A prior with an effective weight of 10 trials is not necessarily based on 10 actual policies or observed outcomes. It is a convenient interpretation of how its parameters enter the posterior estimate. Effective prior weight helps explain credibility, compare prior choices, and spot errors involving units or denominators.
Gamma-Poisson Prior Weight Is Exposure
Suppose a risk has an unknown claim frequency \(\lambda\), measured as expected claims per unit of exposure. Conditional on \(\lambda\), a count \(N\) observed over exposure \(E\) follows a Poisson distribution with mean \(\lambda E\). Assume the prior is Gamma with shape \(\alpha>0\) and rate \(\beta>0\), written \(\lambda\sim\operatorname{Gamma}(\alpha,\beta)\). Its density is proportional to \(\lambda^{\alpha-1}e^{-\beta\lambda}\), and its prior mean is \(\alpha/\beta\). The rate parameter \(\beta\) is in exposure units, not claims-per-exposure units.
Under this model, one observation contributes \(N\) to the shape and \(E\) to the rate. Thus the posterior distribution is \(\operatorname{Gamma}(\alpha+N,\beta+E)\), and the posterior mean is \((\alpha+N)/(\beta+E)\). Define the prior mean rate \(m=\alpha/\beta\) and, when \(E>0\), the observed rate \(r=N/E\). Then:
The prior therefore has weight \(\beta\) exposure units, and the data have weight \(E\) exposure units. Its credibility weight in the posterior mean is \(\beta/(\beta+E)\). The shape parameter \(\alpha\) has a different, useful interpretation: it is the prior expected number of claims, because \(\alpha=\beta m\). It is not the prior exposure weight.
Why the Gamma-Poisson Weight Formula Holds
The weighting formula follows directly from the likelihood and prior. For a count \(N=n\), the Poisson likelihood as a function of \(\lambda\) is proportional to \(\lambda^n e^{-E\lambda}\); factors that do not involve \(\lambda\) can be omitted when finding the posterior. Multiplying by the Gamma prior kernel gives:
This is the kernel of a Gamma distribution with shape \(\alpha+n\) and rate \(\beta+E\). Its mean is \((\alpha+n)/(\beta+E)\). Substituting \(\alpha=\beta m\) and \(n=E(n/E)\) gives the exposure-weighted form. The result requires the specified conditional Poisson model and a Gamma prior; it is not a general formula for every frequency model.
Worked Example: Combining a Gamma Prior With Claim Experience
A hypothetical portfolio uses policy-years as its exposure unit. Its claim frequency has prior distribution \(\lambda\sim\operatorname{Gamma}(6,3)\), using the shape-rate convention. A risk produces 11 claims over 4 policy-years. Find the prior mean, effective prior exposure, equivalent expected prior claims, and posterior mean frequency. Assume the conditional count is Poisson with mean \(\lambda E\); no interest or mortality assumption is relevant.
Method and prior interpretation. The prior mean is \(\alpha/\beta=6/3=2\) claims per policy-year. The rate parameter gives effective prior exposure \(\beta=3\) policy-years. The equivalent prior expected claim count is \(\alpha=6\), also equal to \(3\times2=6\).
Observed experience. The observed frequency is \(11/4=2.75\) claims per policy-year. The posterior shape is \(6+11=17\), and the posterior rate is \(3+4=7\) policy-years.
Rounding only at the final step to four decimal places, the posterior mean frequency is 2.4286 claims per policy-year. The weights are \(3/7\) for the prior mean and \(4/7\) for the observed rate; they sum to 1. This calculation avoids the common mistake of treating the shape parameter 6 as the prior exposure. The correct exposure weight is the rate parameter 3.
Beta-Binomial Prior Weight Is Trials
Now suppose each trial has a binary outcome, such as whether a hypothetical policy has at least one claim during a defined period. Let \(p\) be the common probability of success. Conditional on \(p\), assume the outcomes are independent Bernoulli trials, and use the prior \(p\sim\operatorname{Beta}(a,b)\), with \(a>0\) and \(b>0\). The prior mean is \(m=a/(a+b)\).
If \(x\) successes are observed in \(n\) trials, the posterior distribution is \(\operatorname{Beta}(a+x,b+n-x)\). Its mean is:
Thus the prior concentration \(a+b\) is the effective prior number of trials. The parameter \(a\) acts as equivalent prior successes and \(b\) as equivalent prior failures. The data have weight \(n\), the actual number of trials. These equivalent counts can be fractional; a Beta prior does not have to correspond to an integer number of historical observations.
To see the weighting, the likelihood kernel for \(x\) successes in \(n\) trials is proportional to \(p^x(1-p)^{n-x}\). Multiplying by the Beta prior kernel \(p^{a-1}(1-p)^{b-1}\) gives \(p^{a+x-1}(1-p)^{b+n-x-1}\), the kernel of the stated posterior Beta distribution. Taking its mean gives \((a+x)/(a+b+n)\). Writing \(a=(a+b)m\) and \(x=n(x/n)\) produces the weighted average. The formula relies on the Beta-Binomial setup; it does not automatically apply when trials have differing probabilities or dependence.
Worked Example: Translating a Beta Prior Into Equivalent Trials
A hypothetical insurer models whether a policy has at least one claim during a year. The common annual claim probability has prior distribution \(p\sim\operatorname{Beta}(3,57)\). In a new group of 100 independent policies, 7 have claims. Find the prior mean, effective prior trials, equivalent prior successes and failures, and posterior mean probability. No interest or mortality assumption is relevant; the outcome model is Bernoulli with common probability \(p\).
Method and prior interpretation. The prior mean is \(3/(3+57)=3/60=0.05\). The effective prior trial weight is \(3+57=60\) trials, comprising 3 equivalent successes and 57 equivalent failures.
Observed experience. The observed claim proportion is \(7/100=0.07\). The posterior parameters are \(a+x=3+7=10\) and \(b+n-x=57+(100-7)=150\).
The posterior mean is 0.0625, or 6.25%. The prior contributes 60 effective trials and the data contribute 100 actual trials. The weighted numerator is \(3+7=10\) successes, while the denominator is \(60+100=160\) trials. This avoids the common error of using \(a\) alone as the prior trial weight; the correct concentration is \(a+b\).
Choosing and Comparing Effective Weights
Effective weight is useful when a prior is specified through an actuarial judgment such as a collective rate or probability and a desired amount of prior strength. For a Gamma-Poisson prior, choose a prior mean \(m\) and exposure weight \(\beta\), then set \(\alpha=\beta m\). For a Beta prior, choose a prior mean \(m\) and trial weight \(s\), then set \(a=sm\) and \(b=s(1-m)\). The resulting parameters preserve the selected mean while determining how rapidly later experience can move the estimate.
For a Poisson rate, use exposure in the same units as the rate. For a binomial probability, use the number of comparable trials.
For Gamma shape-rate parameters, calculate \(m=\alpha/\beta\) and prior exposure \(\beta\). For Beta parameters, calculate \(m=a/(a+b)\) and prior trials \(a+b\).
Weight each rate by its exposure, or each proportion by its number of trials. Divide the total weighted amount by total weight.
The resulting rate or probability must retain its intended meaning, and equivalent prior counts must not be described as actual observed data unless that is independently true.
The prior mean alone does not determine the posterior response to new data. Two Gamma priors can have the same mean but different \(\beta\), and hence different effective exposure. Likewise, two Beta priors can have the same mean but different \(a+b\). The larger effective weight produces a larger prior share in the posterior mean for a fixed amount of new information.
| Model | Prior mean | Effective prior weight | Equivalent prior outcomes | Data weight |
|---|---|---|---|---|
| Gamma-Poisson | \(\alpha/\beta\) | \(\beta\) exposure units | \(\alpha\) expected claims at the prior mean | Observed exposure \(E\) |
| Beta-Binomial | \(a/(a+b)\) | \(a+b\) trials | \(a\) successes and \(b\) failures | Observed trials \(n\) |
Common Mistakes and Model Limits
- Confusing a shape parameter with the prior weight. In the Gamma-Poisson model, \(\alpha\) represents prior expected claims, while \(\beta\) represents prior exposure. In the Beta-Binomial model, the trial weight is \(a+b\), not just \(a\) or \(b\).
- Mixing exposure units. If \(\lambda\) is claims per policy-year, then \(E\) and \(\beta\) must both be in policy-years. Converting observed exposure to a different unit without converting the prior rate and prior exposure gives an invalid weighted average.
- Treating equivalent outcomes as actual records. Fractional values of \(a\), \(b\), or \(\alpha\) are allowed. They summarize prior strength within a model; they do not establish that fractional claims or trials occurred.
- Ignoring the assumptions behind the likelihood. A single Poisson rate assumes the count’s conditional mean is proportional to exposure. A single binomial probability assumes comparable independent trials with a common probability. If these assumptions are unsuitable, the simple effective-weight calculation may not describe the data-generating process.
Check Your Understanding
For each problem, identify the prior weight on the appropriate scale before calculating the posterior mean.
- A hypothetical frequency model has \(\lambda\sim\operatorname{Gamma}(4,5)\), using shape-rate parameters and policy-years as exposure units. If 5 claims are observed over 5 policy-years, find the prior effective exposure and posterior mean frequency. (Answer: 5 policy-years; 0.9000 claims per policy-year)
- A hypothetical portfolio has collective rate 2.4 claims per exposure unit, and the selected Gamma prior weight is 2.5 exposure units. It observes 4 claims over 1.5 exposure units. Find the equivalent prior expected claims and posterior mean frequency. (Answer: 6 prior expected claims; 2.5000 claims per exposure unit)
- A hypothetical binary claim indicator has prior mean 0.03 and effective prior weight 40 trials. If 4 claims occur in 60 independent trials, find the posterior mean probability. (Answer: 0.0520)
- A prior is \(\operatorname{Beta}(2,38)\). For 20 independent trials, 3 are successes. Find the effective prior trial weight and posterior mean success probability. (Answer: 40 trials; 0.0833)
- Two hypothetical frequency priors have the same mean, 1.2 claims per exposure unit: \(\operatorname{Gamma}(4.8,4)\) and \(\operatorname{Gamma}(12,10)\). Each is updated with 3 claims over 2 exposure units. Calculate both posterior mean frequencies and state which prior has the larger weight. (Answer: 1.3000 and 1.2500 claims per exposure unit, respectively; the second prior has the larger weight, 10 exposure units)