Tutorials › CAS Exam 8 - Advanced Ratemaking › Choosing Bayesian Priors As Experience Complements

Home  /  Tutorials  /  CAS Exam 8 / Credibility in Experience Rating

CAS Exam 8 · Credibility in Experience Rating · Tutorial 479 of 989

Choosing Bayesian Priors As Experience Complements

Set a Bayesian experience complement from collective data, then quantify how its mean and strength affect an individual risk’s posterior rate.

Advanced 12 min read

What You'll Learn

  • Select a collective portfolio whose risks are relevant to the risk being rated
  • Calibrate a Gamma prior mean and concentration using portfolio frequency moments
  • Interpret the Gamma rate parameter as equivalent exposure under a Poisson model
  • Calculate posterior rates and express them as weighted averages of experience and complement
  • Calibrate a Beta prior for fixed-period claim probabilities
  • Avoid confusing a prior’s strength with its mean or reversing the Gamma rate and scale

Why Prior Calibration Matters

A Bayesian experience rating combines an individual risk’s observed claims with a prior distribution representing the collective complement. The prior is not just a mathematical starting point: its mean determines the expected risk level before individual experience is considered, and its concentration determines how strongly that complement influences the result. A prior that is too broad can let sparse experience dominate; one that is too concentrated can largely suppress relevant risk-specific information.

The collective portfolio should be chosen to represent risks comparable to the one being rated, using the rating basis and exposure units appropriate to the problem. A broad portfolio may have a convenient amount of data but an unsuitable expected frequency. Conversely, a highly specific group can give a more relevant mean but an unstable estimate of variation between risks. The calibration therefore requires both an actuarial choice of complement and a statistical method for setting prior strength.

Definition: In a Bayesian experience rating, the prior mean is the expected value of a risk parameter before observing that risk’s own claims. The prior concentration describes how tightly the prior distribution is centered around that mean. In conjugate models, concentration can often be interpreted as an amount of equivalent exposure or equivalent prior observations.

Gamma Prior for Poisson Claim Frequency

Suppose a risk has an unknown claim frequency \( \theta \), measured as expected claims per exposure unit. Assume that, conditional on \( \theta \), the claim count \(N\) over exposure \(e\) has a Poisson distribution with mean \(e\theta\). Across risks, assume \( \theta \) follows a Gamma distribution with shape \( \alpha>0 \) and rate \( \beta>0 \). Here, “rate” means the parameter in the exponential term of the density, not a claim frequency:

$$ f(\theta)=\frac{\beta^\alpha}{\Gamma(\alpha)}\theta^{\alpha-1}e^{-\beta\theta},\qquad \theta>0. $$

Under this shape-rate convention, the prior mean is \( \alpha/\beta \), and the prior variance is \( \alpha/\beta^2 \). The prior can therefore be calibrated from a target collective mean \( \mu \) and a target between-risk variance \( \tau^2 \):

$$ \mu=\frac{\alpha}{\beta},\qquad \tau^2=\frac{\alpha}{\beta^2} \quad\Longrightarrow\quad \alpha=\frac{\mu^2}{\tau^2},\qquad \beta=\frac{\mu}{\tau^2}. $$

The parameter \( \beta \) has exposure units in this model and acts as the prior’s equivalent exposure. A larger \( \beta \), with the same prior mean, requires a proportionally larger \( \alpha \) and produces a more concentrated prior. This interpretation follows directly from the posterior formula: after observing \(n\) claims over \(e\) exposure units, the posterior is Gamma with shape \( \alpha+n \) and rate \( \beta+e \). Its mean is

$$ E[\theta\mid N=n]=\frac{\alpha+n}{\beta+e} =\frac{\beta}{\beta+e}\mu+\frac{e}{\beta+e}\frac{n}{e}. $$

Thus the posterior mean is a weighted average of the collective mean and the individual observed frequency. The complement receives weight \( \beta/(\beta+e) \); individual experience receives weight \( e/(\beta+e) \). These weights add to one. The formula assumes \(e>0\); when \(e=0\), there is no individual frequency \(n/e\) to calculate and the posterior remains the prior.

Worked Example: Calibrating a Gamma Prior From Portfolio Moments

In a hypothetical portfolio of comparable risks, each risk has 100 exposure units. The average claim count per risk is 8, and the variance of claim counts across risks is 12. Use a Gamma prior for the risk-specific frequency \( \theta \), and calibrate its shape and rate by the method of moments.

Method and assumptions. Conditional on each risk’s frequency, its count is Poisson. The frequencies vary across risks with mean \( \mu \) and variance \( \tau^2 \). The reported portfolio moments are treated as method-of-moments estimates for this model.

For a risk with 100 exposure units, the expected count is \(100\mu\). Equating this to the portfolio average gives

$$ 100\mu=8 \quad\Longrightarrow\quad \mu=0.08. $$

To use the variance, apply the law of total variance. Conditional Poisson variance equals the conditional mean, so \(E[\operatorname{Var}(N\mid\theta)]=100\mu=8\). The variance of the conditional mean \(100\theta\) is \(100^2\tau^2\). Therefore,

$$ \operatorname{Var}(N)=100\mu+100^2\tau^2 =8+10{,}000\tau^2. $$

Set this equal to the observed count variance of 12:

$$ 12=8+10{,}000\tau^2 \quad\Longrightarrow\quad \tau^2=\frac{4}{10{,}000}=0.0004. $$

Now convert the mean and variance of \( \theta \) to Gamma parameters:

$$ \alpha=\frac{0.08^2}{0.0004} =\frac{0.0064}{0.0004}=16, \qquad \beta=\frac{0.08}{0.0004}=200. $$

The calibrated prior is Gamma with shape 16 and rate 200. Its mean checks: \(16/200=0.08\). Its variance checks: \(16/200^2=0.0004\). The rate parameter represents 200 exposure units of prior strength. This moment calibration is valid here because the observed count variance exceeds the Poisson process contribution of 8, leaving a nonnegative estimate of between-risk variation.

Common mistake avoided: using the variance of observed counts as if it were entirely between-risk variation would give \(12/100^2=0.0012\). That incorrectly includes the Poisson process variance. The correction is to subtract the process component, \(8\), before dividing by \(100^2\).

Using Prior Strength With Sparse Experience

Once calibrated, the prior can be applied to individual experience using the conjugate update. The result also shows precisely how concentration controls the influence of sparse data. If exposure is small relative to \( \beta \), the complement’s weight is large. As exposure grows relative to \( \beta \), the individual’s observed frequency receives more weight. This is a direct consequence of the model, not an additional adjustment.

Worked Example: Updating a Frequency Rating After One Claim

Use the prior Gamma(16, 200) from the hypothetical portfolio above. A risk has 10 exposure units and one claim. Calculate its posterior mean frequency and the expected claim count over a future 12 exposure units.

Method and assumptions. The risk’s count is Poisson conditional on its frequency; exposure is measured in the same units used to calibrate the prior. The 12 future exposure units are assumed to have the same frequency parameter, with no trend or other adjustment.

The individual observed frequency is \(n/e=1/10=0.10\). The posterior shape is \(16+1=17\), and the posterior rate is \(200+10=210\). Hence the posterior mean frequency is

$$ E[\theta\mid n=1,e=10] =\frac{17}{210} =0.08095238\ldots $$

The equivalent weighted-average calculation gives the same result:

$$ \frac{200}{210}(0.08)+\frac{10}{210}(0.10) =0.07619048\ldots+0.00476190\ldots =0.08095238\ldots. $$

For 12 future exposure units, the posterior expected count is \(12\) times the posterior mean:

$$ 12\left(\frac{17}{210}\right) =\frac{204}{210} =0.97142857\ldots. $$

Rounded only at the final step, the posterior mean is 0.0810 claims per exposure unit and the future expected count is 0.97 claims. As a check, the posterior frequency lies between the collective mean of 0.08 and the observed frequency of 0.10, as a positive weighted average must.

Common mistake avoided: assigning the individual the weight \(200/210\) reverses the roles of experience and complement. The posterior rate is the weight on the individual frequency, while the prior equivalent exposure \(200\) determines the complement’s weight.

Calibrating a Prior for a Fixed-Period Claim Probability

Not every rating problem uses a Poisson frequency parameter. Suppose the outcome for each fixed policy period is whether at least one claim occurs. If, conditional on an unknown probability \(p\), the period outcomes are independent Bernoulli trials, a Beta prior is conjugate:

$$ p\sim\operatorname{Beta}(a,b),\qquad E[p]=\frac{a}{a+b}. $$

Let \(m=a+b\) denote the prior concentration. If the collective probability is \(p_0\), set \(a=mp_0\) and \(b=m(1-p_0)\). After observing \(x\) claim periods out of \(r\), the posterior is Beta(\(a+x,b+r-x\)), and its mean is

$$ E[p\mid x,r]=\frac{a+x}{m+r} =\frac{m}{m+r}p_0+\frac{r}{m+r}\frac{x}{r}. $$

Here, \(m\) is equivalent prior policy periods, not an exposure measured in arbitrary frequency units. This model is appropriate for fixed-period yes-or-no outcomes; it is not a substitute for a Poisson model when the number of claims per period matters.

Worked Example: A Beta Prior for Claim Occurrence

A hypothetical selected complement has a 4% probability of at least one claim in a policy period. The actuary represents the prior strength as 96 equivalent policy periods. For a new risk, 1 of 10 observed policy periods had a claim. Calculate the posterior mean probability of a claim in one period.

Method and assumptions. Each period is a Bernoulli trial with the same risk-specific probability \(p\), and periods are conditionally independent given \(p\). Set the Beta concentration to \(m=96\) and the prior mean to \(p_0=0.04\).

The prior parameters are

$$ a=mp_0=96(0.04)=3.84, \qquad b=m(1-p_0)=96(0.96)=92.16. $$

After \(x=1\) claim period and \(r-x=9\) no-claim periods, the posterior parameters are \(a+x=4.84\) and \(b+r-x=101.16\). The posterior mean is

$$ E[p\mid x=1,r=10] =\frac{4.84}{4.84+101.16} =\frac{4.84}{106} =0.04566038\ldots. $$

Equivalently, the collective mean has weight \(96/106\), and the observed proportion \(1/10=0.10\) has weight \(10/106\):

$$ \frac{96}{106}(0.04)+\frac{10}{106}(0.10) =\frac{3.84+1.00}{106} =0.04566038\ldots. $$

The posterior claim probability is 0.0457, or 4.57% to two decimal places as a percentage. It is between 4% and 10%, as expected. This example uses policy periods as the exposure unit; treating the ten observations as ten Poisson exposure units would change the model and would not answer the stated probability question.

Common mistake avoided: adding the observed 1 claim only to the Beta parameter \(a\) and then retaining the old denominator \(m\) gives \(4.84/96\), which is not a posterior mean. The denominator must also include all 10 observed periods, giving \(m+r=106\).

Calibration Checks and Exam Traps

Prior calibration requires more than substituting values into a formula. Before accepting the result, verify the model, units, and interpretation:

  • Match the complement to the risk. The prior mean should be the expected rate or probability for a relevant collective group, not simply an overall average chosen for convenience.
  • Separate process variation from risk variation. Under a Poisson-Gamma mixture, the variance of observed counts includes both Poisson process variance and variation in the underlying frequencies. Subtract the former when using count moments to estimate the latter.
  • State the Gamma convention. With the shape-rate convention used here, \(E[\theta]=\alpha/\beta\). A shape-scale parameterization uses a different second parameter and a different mean formula.
  • Keep exposure units consistent. If frequency is per policy-year, the exposure added to the prior rate must be policy-years. A time-unit conversion changes the numerical rate and exposure together.
  • Check feasibility. If the observed count variance is below the Poisson process contribution, the simple moment estimate of between-risk variance is negative. A negative value cannot define a Gamma prior variance; the model or estimation procedure needs reconsideration rather than mechanically using negative parameters.

A prior’s concentration does not repair a poor choice of complement. It only governs how much the chosen complement contributes relative to individual experience. For a Poisson-Gamma model, the posterior mean’s weights are determined by prior equivalent exposure and observed exposure; for a Beta-Bernoulli model, they are determined by prior equivalent periods and observed periods. The distinction is useful when translating an actuarial judgment about prior strength into model parameters.

Key takeaway: Calibrate the prior mean to the selected collective complement, then choose concentration using a defensible measure of between-risk variation or an explicit equivalent-exposure judgment. Verify that the model and exposure units match the data before applying the Bayesian update. This treatment is aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.

Check Your Understanding

For each problem, identify the model and keep the stated exposure units throughout.

  1. A Poisson-Gamma frequency prior has collective mean \(0.10\) claims per exposure unit and between-risk variance \(0.0004\). Find its shape and rate parameters and equivalent exposure. (Answer: \(\alpha=25\), \(\beta=250\) exposure units, equivalent exposure \(250\))
  2. A risk has \(n=2\) claims in \(e=15\) exposure units. Its prior is Gamma with shape 6 and rate 75. Find the posterior mean frequency and check that it lies between the prior mean and observed frequency. (Answer: posterior mean \(8/90=0.0889\); prior mean \(0.0800\), observed frequency \(0.1333\))
  3. A fixed-period claim probability has a Beta prior with mean \(0.03\) and concentration 50. The risk has no claims in 20 periods. Find the prior parameters and posterior mean probability. (Answer: \(a=1.5\), \(b=48.5\), posterior mean \(1.5/70=0.0214\))
  4. In a Poisson-Gamma portfolio, each risk has 50 exposure units, mean count 6, and variance of counts 10. Use the moment method to find the collective frequency mean, between-risk variance, and Gamma parameters. (Answer: \(\mu=0.12\), \(\tau^2=0.0016\), \(\alpha=9\), \(\beta=75\) exposure units)
  5. A Poisson portfolio has equal exposure \(e_0=20\), average count 4, and variance of counts 3. Can the stated moments produce a positive method-of-moments estimate of Gamma between-risk variance? Explain. (Answer: No; the Poisson process contribution is \(4\), so \(\tau^2=(3-4)/20^2=-0.0025\), which is infeasible as a variance.)

Previous

Next

Back to Tutorials
Clinical Biostats