Bayes’ Rule for Posterior Risk Estimates

Why Update a Risk Estimate?

An insurer may begin with a collective view of the possible claim risk for a policyholder or class, then observe experience or another signal about that risk. The central Bayesian question is how to combine those two sources of information consistently. The prior distribution represents uncertainty about the risk before observing the new data; the likelihood describes how probable the observed data would be at each possible risk level.

Bayes’ rule converts these ingredients into a posterior distribution. That distribution is the updated uncertainty about the risk parameter after the data are observed. A posterior mean or a future-period prediction is then calculated from the posterior, rather than from the prior alone or from the observed data alone. The examples below use hypothetical insurance risks.

Definition: Let $\theta$ denote an unknown risk parameter, let $\pi(\theta)$ be its prior density, and let $y$ be observed data with likelihood $f(y\mid\theta)$. When the marginal density of the observation is positive and finite, the posterior density is $$ \pi(\theta\mid y)=\frac{f(y\mid\theta)\pi(\theta)}{m(y)}, \qquad m(y)=\int f(y\mid t)\pi(t)\,dt. $$ The denominator $m(y)$ is the marginal density of the observed data and normalizes the product so the posterior density integrates to one.

For a risk parameter with a finite set of possible values $\theta_1,\ldots,\theta_k$, the equivalent formula is

P(\theta=\theta_j\mid y) = \frac{P(y\mid\theta_j)P(\theta=\theta_j)} {\sum_{r=1}^{k}P(y\mid\theta_r)P(\theta=\theta_r)}.

The numerator is the prior probability of a risk level multiplied by the probability of the observed data at that level. These products are called unnormalized posterior weights. Dividing each one by their sum produces posterior probabilities that add to one. The likelihood alone is not the posterior: the prior probabilities also matter.

Why the Normalized Product Is Bayes’ Rule

The continuous formula follows by expressing the joint density of the parameter and data in two ways. By the definition of conditional density, the joint density at $(\theta,y)$ is $\pi(\theta)f(y\mid\theta)$. It is also the marginal density of the data times the conditional density of the parameter given the data:

\pi(\theta)f(y\mid\theta)=m(y)\pi(\theta\mid y).

Solving for the posterior gives $\pi(\theta\mid y)=f(y\mid\theta)\pi(\theta)/m(y)$. Integrating the joint density over all parameter values gives $m(y)=\int f(y\mid t)\pi(t)\,dt$, which establishes the normalizing denominator. In the discrete case, integration is replaced by summation. This derivation also explains why posterior weights must be normalized: they are joint probabilities or densities expressed relative to the observed data, and the posterior must have total probability one.

List the prior risk levels and probabilities.
Check that the prior probabilities sum to one.

Evaluate the likelihood at each risk level.
Use the probability of the entire observation under the stated data model.

Form and normalize the weights.
Multiply each likelihood by its prior probability, sum the products, and divide each product by that total.

Calculate the requested estimate or prediction.
For example, take a posterior mean or average a future claim probability over the posterior distribution.

Posterior Distributions and Posterior Estimates

A posterior distribution is the full collection of updated probabilities or densities. A posterior estimate is a summary of that distribution, and the appropriate summary depends on the purpose and loss function. Under squared-error loss for estimating a numerical parameter, the posterior mean minimizes posterior expected squared error. For any proposed estimate $a$,

E[(\theta-a)^2\mid y] = \operatorname{Var}(\theta\mid y) + \left(a-E[\theta\mid y]\right)^2.

The first term does not depend on $a$, and the second is minimized by $a=E[\theta\mid y]$. The posterior mode, by contrast, is the risk value with the greatest posterior probability in a discrete model. These summaries answer different questions and need not agree.

If $\theta=p$ is a claim probability and a future claim outcome is conditionally independent of the observed data given $p$, the future claim probability is the posterior mean of $p$:

P(\text{future claim}\mid y) = \sum_j P(\text{future claim}\mid p_j)P(p=p_j\mid y) = \sum_j p_jP(p=p_j\mid y).

The conditional-independence assumption matters: it says that once the underlying risk parameter is known, the past observation provides no additional information about the future outcome. It is not generally valid to treat the future claim as independent of past data without conditioning on the risk parameter.

Worked Example: One Claim Updates a Two-Level Risk

A hypothetical portfolio model assigns a policyholder a claim probability $p=0.02$ with prior probability $0.75$, or $p=0.10$ with prior probability $0.25$. A single period is observed, and a claim occurs. Assume the period’s outcome is Bernoulli conditional on $p$. Find the posterior probabilities, posterior mean claim probability, and predictive probability of a claim in one future period, assuming conditional independence given $p$.

Method. Apply the discrete form of Bayes’ rule. For one Bernoulli period, the likelihood of observing a claim is $P(y\mid p)=p$.

Unnormalized weights. At $p=0.02$, the weight is $0.75(0.02)=0.015$. At $p=0.10$, it is $0.25(0.10)=0.025$. Their sum is $0.015+0.025=0.040$.

Posterior probabilities. The posterior probability of $p=0.02$ is $0.015/0.040=0.375$. The posterior probability of $p=0.10$ is $0.025/0.040=0.625$. They sum to $1.000$, as required.

Posterior mean. The updated mean claim probability is $0.02(0.375)+0.10(0.625)=0.0075+0.0625=0.0700$.

Future-period prediction. By conditional independence given $p$, the predictive probability of a claim is the posterior mean $0.0700$, or $7.00\%$. Thus the posterior distribution, rather than either possible risk level alone, determines the prediction.

Check and common mistake avoided. The posterior probabilities add to one, and the posterior mean lies between $0.02$ and $0.10$. Using the likelihoods $0.02$ and $0.10$ as posterior probabilities would be wrong because it omits the prior weights and does not normalize.

Likelihoods Must Match the Observation

The likelihood is the probability of the data actually observed, given a candidate risk parameter. For $n$ conditionally independent Bernoulli periods with $x$ claims, the likelihood is

P(X=x\mid p)=\binom{n}{x}p^x(1-p)^{n-x}.

The factor $\binom{n}{x}$ is the same for every candidate $p$ when $n$ and $x$ are fixed. It therefore cancels when the posterior weights are normalized. The $p$-dependent terms do not cancel and must be evaluated at each risk level. If the observed data are a particular sequence rather than just a claim count, use the probability of that sequence; for a fixed count, the binomial likelihood correctly accounts for all sequences with that count.

Worked Example: Two Claim-Free Periods

A hypothetical risk has $p=0.02$ with prior probability $0.70$ and $p=0.08$ with prior probability $0.30$. Two independent Bernoulli periods are observed and neither has a claim. Find the posterior probabilities and posterior mean claim probability.

Method and likelihood. Given $p$, the probability of two claim-free periods is $(1-p)^2$. Thus the likelihoods are $0.98^2=0.9604$ and $0.92^2=0.8464$.

Unnormalized weights. For $p=0.02$, the weight is $0.70(0.9604)=0.67228$. For $p=0.08$, it is $0.30(0.8464)=0.25392$. The normalizing total is $0.67228+0.25392=0.92620$.

Posterior probabilities. The posterior probability of $p=0.02$ is $0.67228/0.92620\approx0.7258$. The posterior probability of $p=0.08$ is $0.25392/0.92620\approx0.2742$. The unrounded probabilities sum to one.

Posterior mean. Using the unrounded ratios, the mean is $0.02(0.67228/0.92620)+0.08(0.25392/0.92620)\approx0.036449$. Rounded to five decimal places, the posterior mean is $0.03645$, or approximately $3.645\%$.

Check and common mistake avoided. The result is between $0.02$ and $0.08$, and claim-free experience shifts posterior probability toward the lower risk. Raising $1-p$ to the second power is essential: using only $1-p$ would treat two periods as one.

Signals and Other Evidence

Bayesian updating is not limited to claim counts. A risk classification inspection, a loss-prevention indicator, or another signal can be used if its conditional probabilities at each risk level are specified. The observation model must describe how the signal depends on the underlying risk. If a signal and future claim outcome are conditionally independent given the risk parameter, the same posterior can be used to predict the future claim probability.

Worked Example: Updating From a Positive Inspection Signal

A hypothetical risk has claim probability $p=0.02$ with prior probability $0.70$, or $p=0.08$ with prior probability $0.30$. An inspection gives a positive signal with conditional probability $0.20$ at $p=0.02$ and $0.60$ at $p=0.08$. After observing a positive signal, calculate the posterior probabilities, posterior mean, and future-period predictive claim probability. Assume the future outcome and signal are conditionally independent given $p$.

Method. Treat the positive signal as the observed data and multiply each signal likelihood by its prior probability.

Unnormalized weights. The weight at $p=0.02$ is $0.70(0.20)=0.14$. The weight at $p=0.08$ is $0.30(0.60)=0.18$. The total is $0.14+0.18=0.32$.

Posterior probabilities. They are $0.14/0.32=7/16=0.4375$ for $p=0.02$ and $0.18/0.32=9/16=0.5625$ for $p=0.08$. The probabilities sum to $1$.

Posterior mean and prediction. The mean is $0.02(7/16)+0.08(9/16)=0.00875+0.04500=0.05375$. Conditional independence makes the future-period predictive claim probability equal to this mean, $0.05375$, or $5.375\%$.

Check and common mistake avoided. The positive signal is more likely at the higher risk, so the higher-risk posterior probability increases from $0.30$ to $0.5625$. The unnormalized weights must be $0.14$ and $0.18$; dividing them by their sum gives $7/16$ and $9/16$, not probabilities obtained by mixing or reversing the prior and likelihood terms.

Common Traps in Posterior Calculations

Using likelihoods as posterior probabilities. A likelihood ranks how compatible the observation is with each risk value, but it does not include prior probability or normalization.
Forgetting the prior. Even if the data are more likely at one risk level, that level need not have the largest posterior probability if it had a sufficiently small prior probability.
Using the wrong observation probability. For multiple independent periods, use the likelihood of all observed outcomes. For example, two no-claim periods contribute $(1-p)^2$, not $1-p$.
Confusing the posterior mean with a posterior risk level. A mean such as $0.05375$ can be an appropriate estimate even when it is not one of the possible values of $p$.
Predicting without stating the dependence assumption. Averaging $p$ over its posterior gives a future claim probability when the future outcome is conditionally independent of the data given $p$.

For exam-style calculations aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus, a reliable presentation shows the likelihood, prior-weighted products, normalizing total, posterior probabilities, and requested estimate separately. Keeping these steps visible makes arithmetic and modeling errors easier to detect.

Check Your Understanding

For each problem, identify the likelihood, normalize the posterior weights, and then calculate the requested estimate.

A hypothetical risk has $p=0.01$ with prior probability $0.60$ and $p=0.05$ with prior probability $0.40$. One period produces a claim. Find both posterior probabilities and the posterior mean of $p$. (Answer: posterior probabilities $3/13$ and $10/13$; posterior mean $0.04077$.)
A hypothetical risk has $p=0.02$ with prior probability $0.80$ and $p=0.10$ with prior probability $0.20$. Two independent periods have no claims. Find the posterior mean of $p$. (Answer: approximately $0.03393$.)
A risk has $p=0.02$ or $p=0.08$, with prior probabilities $0.70$ and $0.30$. A positive signal has likelihoods $0.20$ and $0.60$, respectively. Find the posterior probabilities and, assuming conditional independence, the future-period claim probability. (Answer: posterior probabilities $7/16$ and $9/16$; predictive probability $0.05375$.)
A hypothetical severity parameter $\theta$ equals $100$, $300$, or $700$, with prior probabilities $0.20$, $0.50$, and $0.30$. An observed inspection signal has likelihoods $0.10$, $0.40$, and $0.80$ at those values. Find the posterior probabilities and posterior mean of $\theta$. (Answer: posterior probabilities $1/23$, $10/23$, and $12/23$; posterior mean $500$.)
A hypothetical risk has $p=0.10$ or $p=0.30$, with equal prior probabilities. Among three independent Bernoulli periods, exactly two claims occur. Find the posterior probability of $p=0.30$ and the posterior mean of $p$. (Answer: posterior probability $7/8$ for $p=0.30$; posterior mean $0.275$.)

What You'll Learn