Why Model Differences Between Risks?
Two insureds with the same exposure may have different expected claim frequencies because their underlying risks differ. A model can represent that variation by assigning each risk a claim-rate parameter, then describing how those parameters vary across a portfolio. This distribution is important even before a particular insured’s own experience is considered: its mean describes the portfolio’s average rate, and its dispersion describes how varied the risks are around that average.
The focus here is the distribution of risk-specific rates, not the updating of a rate using observed claims. For a risk with a fixed underlying rate, claim outcomes still vary from period to period. That outcome variation is distinct from the differences in underlying rates across risks.
The word “prior” does not mean that the risk’s rate changes randomly from one period to the next. In this model, a risk has a particular underlying rate, while the prior represents how rates differ across the population of risks being modeled. It is a probability model for the risk parameter before conditioning on that risk’s own experience.
Suppose that, conditional on a risk’s rate \(\Theta\), its claim count \(N\) over exposure \(e\) follows the Poisson model:
Here \(e\) is measured in the exposure units used to define \(\Theta\). For example, if \(\Theta\) is claims per vehicle-year, \(e\) must be measured in vehicle-years. Conditional on the rate, the expected count and variance are both \(e\Theta\). Across risks, however, the count varies both because of claim-process randomness and because different risks have different values of \(\Theta\).
Reading the Mean and Dispersion Actuarially
Write \(\mu=E[\Theta]\) and \(\tau^2=\operatorname{Var}(\Theta)\). The mean \(\mu\) is the average underlying claim rate under the modeled portfolio mix. It is not necessarily the rate of every risk. The variance \(\tau^2\) measures the spread of risk-specific rates around that average, in squared rate units. The standard deviation \(\tau\) is in the same units as the rate and is often easier to interpret directly.
A prior with a small variance concentrates risk rates near the portfolio mean. A prior with a larger variance places more weight on rates far from that mean, representing greater heterogeneity. Comparing variances is most useful when the rates use the same exposure unit and the means are comparable. The coefficient of variation, \(\tau/\mu\) when \(\mu>0\), gives a dimensionless measure of dispersion relative to the mean.
One useful consequence follows from the conditional Poisson model. The expected count at exposure \(e\), averaging over the rate distribution, is \(e\mu\). Its variance includes both the average conditional Poisson variance and the variation in conditional means:
The first variance term, \(e\mu\), is process variation: counts fluctuate even for a fixed rate. The second, \(e^2\tau^2\), reflects between-risk variation: risks with higher underlying rates tend to have higher expected counts. This distinction matters when interpreting portfolio count variability. The unconditional count distribution is generally more variable than a Poisson count with rate \(e\mu\) whenever \(\tau^2>0\).
Worked Examples
Worked Example: A Two-Group Prior for Claim Rates
A hypothetical insurer models annual claim frequency per vehicle-year using two risk groups. A randomly selected risk has rate \(\Theta=0.20\) with probability \(0.75\) and rate \(\Theta=0.80\) with probability \(0.25\). Find the prior mean, variance, standard deviation, and the mean and variance of the claim count for two vehicle-years of exposure. Use a conditional Poisson count model. No interest rate or mortality assumption is involved.
Method. Calculate the discrete prior moments, then apply the conditional Poisson mixture formulas. The rate values are claims per vehicle-year.
Mean rate. \(E[\Theta]=0.75(0.20)+0.25(0.80)=0.15+0.20=0.35\) claims per vehicle-year.
Variance of rates. First calculate the second moment: \(E[\Theta^2]=0.75(0.20^2)+0.25(0.80^2)=0.75(0.04)+0.25(0.64)=0.03+0.16=0.19\). Thus \(\operatorname{Var}(\Theta)=0.19-(0.35)^2=0.19-0.1225=0.0675\). The standard deviation is \(\sqrt{0.0675}=0.2598\) claims per vehicle-year, to four decimals.
Count moments. With \(e=2\), \(E[N]=2(0.35)=0.70\) claims. The variance is \(2(0.35)+2^2(0.0675)=0.70+0.27=0.97\) claims squared.
Check and trap avoided. The count variance exceeds its mean by \(0.27\), the between-risk component. Treating every risk as if it had the mean rate would incorrectly give a Poisson variance of \(0.70\), omitting heterogeneity. The result is consistent with the formulas: process variance is \(0.70\), and rate variation contributes \(0.27\).
Worked Example: Interpreting a Gamma Prior
For a hypothetical portfolio, the annual claim rate per policy-year has a gamma prior with shape \(\alpha=3\) and rate parameter \(\beta=6\). In this parameterization, the density is proportional to \(\theta^{\alpha-1}e^{-\beta\theta}\) for \(\theta\geq0\). Find its mean, variance, coefficient of variation, and the probability that a risk’s rate exceeds \(1\) claim per policy-year. Also find the claim-count mean and variance for four policy-years of exposure under a conditional Poisson model. No interest rate or mortality assumption is involved.
Method. For this gamma parameterization, \(E[\Theta]=\alpha/\beta\) and \(\operatorname{Var}(\Theta)=\alpha/\beta^2\). These formulas describe the prior distribution across risk rates; they do not update a particular risk’s rate.
Prior moments. The mean is \(3/6=0.50\) claims per policy-year. The variance is \(3/6^2=3/36=0.0833\) in squared rate units, to four decimals. The standard deviation is \(\sqrt{3/36}=0.2887\), and the coefficient of variation is \(0.288675/0.50=0.5774\), to four decimals.
Exceedance probability. Since the shape is the integer 3, the gamma survival probability at rate parameter \(\beta\) and threshold \(t\) is \(e^{-\beta t}[1+\beta t+(\beta t)^2/2]\). At \(t=1\), \(\beta t=6\), so \(P(\Theta>1)=e^{-6}[1+6+36/2]=25e^{-6}=0.0620\), to four decimals.
Count moments. For \(e=4\), \(E[N]=4(0.50)=2.00\). The variance is \(4(0.50)+4^2(3/36)=2+16/12=3.3333\) claims squared, to four decimals.
Check and trap avoided. The exceedance probability is a probability about risk rates, not the probability that a particular policy will have a claim. The count variance is larger than the count mean because the prior variance is positive; the extra component is \(16(3/36)=1.3333\).
Worked Example: Calibrating a Gamma Prior From Actuarial Targets
An actuary is constructing a hypothetical prior for claim frequency per equipment-year. The target portfolio mean is \(0.40\) claims per equipment-year, and the target coefficient of variation across risk-specific rates is \(0.50\). Use a gamma prior with shape \(\alpha\) and rate parameter \(\beta\). Find its parameters, variance, and the probability that a rate exceeds \(0.80\). No interest rate or mortality assumption is involved.
Method. For a gamma prior, \(\mu=\alpha/\beta\) and \(\tau^2=\alpha/\beta^2\), so the squared coefficient of variation is \(\tau^2/\mu^2=1/\alpha\). Thus \(\alpha=1/\mathrm{CV}^2\), and \(\beta=\alpha/\mu\).
Shape and rate. \(\alpha=1/(0.50)^2=4\). Then \(\beta=4/0.40=10\), with units consistent with the rate being per equipment-year. The prior variance is \(\alpha/\beta^2=4/100=0.04\) in squared rate units; the standard deviation is \(0.20\) claims per equipment-year.
Exceedance probability. Here \(\beta t=10(0.80)=8\). For integer shape 4, \(P(\Theta>0.80)=e^{-8}[1+8+8^2/2+8^3/6]\). The bracket is \(1+8+32+85.3333=126.3333\), so the probability is approximately \(126.3333(0.0003354626)=0.0424\), to four decimals.
Check and trap avoided. The fitted mean is \(\alpha/\beta=4/10=0.40\), and the fitted coefficient of variation is \(1/\sqrt{4}=0.50\), so both calibration targets are met. Using \(\alpha=\mathrm{CV}^2\) instead would reverse the relationship and fail the specified dispersion target.
Choosing and Communicating a Prior
A prior should reflect the population of risks to which the model will be applied. The target portfolio mean may come from a collective rate estimate, while the dispersion may be informed by variation among comparable risk groups, prior modeling assumptions, or professional judgment. The basis should be stated clearly. A prior mean and variance alone do not determine the entire distribution: two distributions can share those moments but assign different probabilities to unusually high rates.
The gamma distribution is one common choice for a nonnegative rate because it has support on \(\Theta\geq0\) and can represent a range of right-skewed shapes. A discrete prior can instead make distinct risk groups explicit. Neither distribution is automatically appropriate in every setting. The choice should be plausible for the risks and rate definition at hand; a convenient formula is not a substitute for a suitable model.
Keep the exposure unit fixed throughout. A rate per vehicle-year cannot be combined directly with exposure measured in policy-months unless the units are converted. If the rate is per policy-year, for instance, 6 policy-months correspond to \(e=0.5\) policy-years. A consistent unit convention is essential both for interpreting \(\Theta\) and for evaluating \(e\Theta\).
Also distinguish an individual risk’s parameter from the portfolio’s observed average count. A portfolio average count can reflect both the mean rate and the mix of risks. The prior describes that modeled risk mix through \(G\); it does not imply that all risks have rate \(E[\Theta]\). When risks are dependent or the portfolio has materially different subpopulations, a single common distribution may not adequately describe the intended population.
Check Your Understanding
For each problem, distinguish the distribution of risk rates from claim-count randomness. Rates and exposures use the units specified in the question.
- A prior assigns rate \(0.10\) with probability \(0.60\) and rate \(0.60\) with probability \(0.40\), in claims per policy-year. Find its mean and variance. (Answer: mean \(0.30\); variance \(0.06\).)
- A gamma prior has shape \(2\) and rate parameter \(5\) for claims per unit exposure. Find the mean, variance, and coefficient of variation. (Answer: mean \(0.40\); variance \(0.08\); coefficient of variation \(0.7071\).)
- Suppose \(E[\Theta]=0.20\) and \(\operatorname{Var}(\Theta)=0.01\) claims per exposure-year and squared rate units, respectively. Under a conditional Poisson model, find the mean and variance of the count at exposure \(e=3\). (Answer: mean \(0.60\); variance \(0.69\).)
- Calibrate a gamma prior for a risk rate with mean \(0.25\) claims per unit exposure and coefficient of variation \(1/\sqrt{2}\). Find the shape and rate parameters and the prior variance. (Answer: shape \(2\); rate \(8\); variance \(0.03125\).)
- A discrete prior assigns \(\Theta=0.20\) with probability \(0.75\) and \(\Theta=0.80\) with probability \(0.25\), in claims per vehicle-year. For exposure \(e=2\), calculate the count variance and identify its process and between-risk components. (Answer: total variance \(0.97\); process component \(0.70\); between-risk component \(0.27\).)