Beta-Binomial Estimates for Classification Rates

From Classification Experience to an Indicated Rate

A classification rate is often estimated from binary outcomes: for each exposure unit, either a claim occurs or it does not. A class with little experience can have a volatile observed rate, especially when its number of trials is small. A Beta-Binomial model combines that experience with a stated portfolio complement to produce an indicated probability for the class.

The calculation has two distinct inputs. The portfolio complement is the prior mean claim probability chosen for the class before observing its experience. The class’s observed number of claims and number of trials then update that prior. The resulting posterior mean is an indicated class probability; dividing it by the stated complement gives a class relativity on that reference scale.

Definition: Suppose the unknown claim probability for a class is $p$, with prior distribution $p\sim\operatorname{Beta}(\alpha,\beta)$, where $\alpha>0$ and $\beta>0$. Conditional on $p$, let $X$ be the number of claims in $n$ independent, identical binary trials, so $X\mid p\sim\operatorname{Binomial}(n,p)$. If $x$ claims are observed, the posterior distribution is $\operatorname{Beta}(\alpha+x,\beta+n-x)$, and the posterior mean is $$ E[p\mid X=x]=\frac{\alpha+x}{\alpha+\beta+n}. $$

The trials must represent comparable opportunities for the binary outcome, with a common probability $p$ conditional on the class risk. For example, a trial might be one policy-year when the outcome is whether that policy-year has at least one claim. This model is not a count-frequency model for multiple claims per exposure; such counts require a different likelihood.

Portfolio Complement and Credibility Weights

Write $s=\alpha+\beta$ for the prior strength and $q=\alpha/(\alpha+\beta)$ for the prior mean. The quantity $q$ is the stated portfolio complement for this update. Since $\alpha=sq$, the posterior mean can be rewritten as

\frac{\alpha+x}{s+n} =\frac{s}{s+n}q+\frac{n}{s+n}\frac{x}{n},\qquad n>0.

Thus the indicated class probability is a weighted average of the complement $q$ and the observed class rate $x/n$. The observed rate receives weight $n/(s+n)$, while the complement receives weight $s/(s+n)$. The weights sum to one. When $n=0$, the posterior mean is simply $q$; the observed-rate form is not used because $x/n$ is undefined.

Definition: For a stated portfolio complement $q>0$, the indicated relativity for a class with posterior mean $\hat p=E[p\mid X=x]$ is $$ R=\frac{\hat p}{q}. $$ A relativity above 1 means the posterior mean exceeds the stated complement; a relativity below 1 means it is lower. The denominator is the complement specified for the comparison, not automatically the posterior mean or the raw class rate.

A portfolio complement is an input to the update, not a claim that every class has that probability. If different classes use the same prior mean and strength, their observed counts can produce different posterior means and relativities. If the complement is changed, the indicated estimate and its relativity may also change.

Why the Posterior Formula Holds

The Binomial likelihood for the observed count $x$, apart from a factor that does not depend on $p$, is proportional to $p^x(1-p)^{n-x}$. The Beta prior density is proportional to $p^{\alpha-1}(1-p)^{\beta-1}$. Multiplying the prior density by the likelihood gives a posterior density proportional to

p^{\alpha+x-1}(1-p)^{\beta+n-x-1}.

This is the kernel of a Beta distribution with parameters $\alpha+x$ and $\beta+n-x$. A Beta distribution with parameters $a,b>0$ has mean $a/(a+b)$; applying that mean formula gives $(\alpha+x)/(\alpha+\beta+n)$. The update adds the observed claims to the first parameter and the observed nonclaims, $n-x$, to the second. It does not add the total trial count to both parameters.

Establish the complement and prior strength.
Set $q=\alpha/(\alpha+\beta)$ and $s=\alpha+\beta$. Check that the stated prior parameters reproduce the complement.

Record the trial counts.
Use $n$ for all eligible binary trials and $x$ for the number with the event. Confirm $0\le x\le n$.

Update and calculate the indicated rate.
Use posterior parameters $(\alpha+x,\beta+n-x)$, then calculate their mean.

Calculate the relativity if requested.
Divide the posterior mean by the specified complement $q$, not by a different class’s posterior estimate.

Worked Classification Updates

Worked Example: A Small Class with a High Observed Rate

A hypothetical class has a stated portfolio complement of $0.10$ and prior strength $45$. It has 15 eligible policy-year trials, of which 5 have at least one claim. Calculate the posterior mean claim probability and its relativity to the complement. Identify the weight-reversal trap.

Method and prior. The prior parameters are $\alpha=45(0.10)=4.5$ and $\beta=45(0.90)=40.5$. Their sum is $45$, and their mean is $4.5/45=0.10$, as required.

Update. The observed rate is $5/15=1/3$. The posterior parameters are $\alpha+x=4.5+5=9.5$ and $\beta+n-x=40.5+(15-5)=50.5$. Their sum is $60$, so

\hat p=\frac{9.5}{60}=0.158333\ldots

The credibility weights give the same result: the observed-rate weight is $15/60=0.25$, and the complement weight is $45/60=0.75$. Therefore $0.25(1/3)+0.75(0.10)=0.083333\ldots+0.075=0.158333\ldots$. The relativity is $0.158333\ldots/0.10=1.58333\ldots$. To four decimal places, the indicated probability is $0.1583$ and the relativity is $1.5833$.

Common mistake avoided. Reversing the weights gives $0.75(1/3)+0.25(0.10)=0.25+0.025=0.275$, which is incorrect: it gives the small class’s observed rate most of the weight. The 15 trials, not the prior strength of 45, determine the observed-rate weight.

Worked Example: Comparing Two Class Relativities

A hypothetical insurer uses a Beta prior with mean complement $0.12$ and strength $30$ for each of two classifications. Class A has 30 trials and 3 claims; Class B has 60 trials and 9 claims. Find each posterior mean and relativity, and state which class has the lower indicated probability.

Prior parameters. The common prior has $\alpha=30(0.12)=3.6$ and $\beta=30(0.88)=26.4$. The sum is 30 and the mean is $3.6/30=0.12$.

Class A. The posterior parameters are $3.6+3=6.6$ and $26.4+(30-3)=53.4$. Their sum is 60, giving $\hat p_A=6.6/60=0.1100$. Its relativity is $0.1100/0.12=0.916666\ldots$, or $0.9167$ to four decimal places.

Class B. The posterior parameters are $3.6+9=12.6$ and $26.4+(60-9)=77.4$. Their sum is 90, giving $\hat p_B=12.6/90=0.1400$. Its relativity is $0.1400/0.12=1.166666\ldots$, or $1.1667$ to four decimal places.

Class A has the lower posterior mean, $0.1100<0.1400$. Class A’s observed rate, $3/30=0.10$, is below the complement, so its indicated probability is pulled below $0.12$. Class B’s observed rate, $9/60=0.15$, is above the complement, so its indicated probability is pulled above $0.12$.

Common mistake avoided. Do not reuse Class A’s posterior mean for Class B or assume equal complements imply equal indicated rates. Class B’s own counts yield $12.6/90=0.1400$, not $0.1100$.

Worked Example: A Higher Complement Weight for Limited Experience

A hypothetical specialty classification uses a complement of $0.08$ and prior strength $50$. In 25 binary trials, 4 have claims. Calculate the posterior mean, the credibility weights, and the relativity.

Prior. The parameters are $\alpha=50(0.08)=4$ and $\beta=50(0.92)=46$. Their sum is 50 and their mean is $4/50=0.08$.

Posterior mean. The posterior parameters are $4+4=8$ and $46+(25-4)=67$. The total is $75$, so $\hat p=8/75=0.106666\ldots$. Equivalently, the observed rate is $4/25=0.16$; the observed weight is $25/75=1/3$ and the complement weight is $50/75=2/3$. Thus $(1/3)(0.16)+(2/3)(0.08)=0.053333\ldots+0.053333\ldots=0.106666\ldots$. The relativity is $0.106666\ldots/0.08=1.333333\ldots$. To four decimal places, the posterior mean is $0.1067$ and the relativity is $1.3333$.

Check. The observed rate $0.16$ exceeds the complement $0.08$, and the posterior mean $0.1067$ lies between them, as a weighted average with positive weights must. The mistake to avoid is using the raw rate $0.16$ as the indicated probability; that ignores the prior complement and its weight.

Interpretation and Exam Traps

The posterior mean is a class-level estimate that balances the stated complement with the class’s observed binary experience. A larger trial count increases the observed-rate weight, while a larger prior strength increases the complement weight. These statements concern the posterior mean; they do not mean that the class’s true probability is known exactly.

Use the correct trial definition. If the outcome is “at least one claim during a policy-year,” a policy-year with several claims still contributes one event to this Binomial count. Do not substitute the number of claims for the number of event-positive trials.
Count nonclaims in the second parameter. The updated Beta parameters are $\alpha+x$ and $\beta+n-x$. Using $\beta+n$ incorrectly treats claims as nonclaims as well.
Keep the weights attached to the right quantities. The observed-rate weight is $n/(s+n)$, and the complement weight is $s/(s+n)$. Reversing them overweights whichever component is less reliable.
Use the specified reference for the relativity. A relativity to $q$ is $\hat p/q$. A relativity to another benchmark would be a different calculation and should be labelled accordingly.
Do not confuse a rate estimate with a full predictive distribution. This calculation gives the posterior mean probability. It does not by itself give the distribution of future claim counts.

For classification, the relativity provides a compact comparison: $R=1$ indicates equality with the stated complement, $R>1$ indicates a higher posterior mean, and $R<1$ indicates a lower one. A valid calculation should also pass a basic check: when $n>0$, the posterior mean lies between $q$ and $x/n$, because it is a weighted average of those two values.

Key takeaway: Specify the Beta prior and binary trial definition, update both parameters using claims and nonclaims, and calculate the posterior mean before forming a relativity. The observed experience receives weight $n/(s+n)$; the stated portfolio complement receives weight $s/(s+n)$.

Check Your Understanding

For each problem, use the stated complement as the relativity denominator. Round probabilities and relativities to four decimal places where needed.

A class has prior $\operatorname{Beta}(2,18)$, and records 0 claims in 10 binary trials. Find the posterior mean and its relativity to the prior mean. (Answer: posterior mean $0.0667$; relativity $0.6667$.)
A class has complement $0.10$, prior strength 50, and 4 claims in 20 trials. Find the posterior mean and relativity. (Answer: posterior mean $0.1286$; relativity $1.2857$.)
A class has complement $0.15$, prior strength 40, and 3 claims in 30 trials. Find the credibility weight on the observed rate, the posterior mean, and the relativity. (Answer: observed-rate weight $30/70=0.4286$; posterior mean $0.1286$; relativity $0.8571$.)
Two classes share complement $0.08$ and prior strength 25. Class A has 0 claims in 15 trials; Class B has 7 claims in 35 trials. Find both posterior means and relativities, and identify the lower indicated rate. (Answer: A $0.0500$, relativity $0.6250$; B $0.1500$, relativity $1.8750$; Class A is lower.)
A class has complement $0.12$, prior strength 50, and 3 claims in 10 trials. Calculate the posterior mean using the correctly assigned credibility weights, and show the two weights. (Answer: observed-rate weight $10/60=0.1667$, complement weight $50/60=0.8333$; posterior mean $0.1500$.)

What You'll Learn