Diagnosing Experience-Rating Incentives And Volatility

Why Diagnose Experience-Rating Incentives and Volatility?

A credibility premium combines an individual risk’s observed experience with an experience complement. The weight on the observed experience determines how strongly a change in that experience affects the prospective premium. It also transmits some of the experience measure’s statistical variability into the premium. These two effects are linked: increasing the weight generally makes the premium more responsive to the risk’s own results, but also more variable.

A useful diagnosis therefore asks more than “What is the premium?” It asks whether the premium is volatile because individual results are noisy, because risks genuinely differ, or because the complement is uncertain. It also asks how much a feasible reduction in losses would change the next premium. The calculations below treat the rating rule as given; they do not determine whether a particular weight is appropriate for a rating plan.

Definition: For an experience measure $\bar X$, credibility weight $Z$, and complement $m$, a linear credibility premium is $P=Z\bar X+(1-Z)m$, where $0\leq Z\leq1$. The complement is the non-individual component of the indicated premium. Unless stated otherwise, assume the exposure basis and coverage are consistent, $Z$ is fixed, and the same loss-cost units are used for $\bar X$, $m$, and $P$.

The formula is a prospective indication: it uses past experience to help set a future premium. It does not say that the entire premium equals expected claims, nor does it account by itself for expenses, profit provisions, limits, or other rating adjustments. Comparisons below isolate the credibility component so that its effects are visible.

Separate Premium Level from Premium Volatility

Suppose a risk has an underlying expected loss cost $\theta$, and its experience measure satisfies $\bar X=\theta+\varepsilon$, where $E[\varepsilon\mid\theta]=0$ and $\operatorname{Var}(\varepsilon\mid\theta)=\sigma^2/n$. Here $n$ is the amount of comparable exposure, and $\sigma^2$ is the variance per unit of exposure. Assume $m$ and $Z$ are fixed for this risk. Then the prospective premium’s conditional expectation and variance are:

E[P\mid\theta]=Z\theta+(1-Z)m,\qquad \operatorname{Var}(P\mid\theta)=Z^2\frac{\sigma^2}{n}.

The first equation shows that a credibility premium is not generally conditionally unbiased for $\theta$: its conditional bias is $(1-Z)(m-\theta)$. The second shows how experience noise is transmitted. Since the premium multiplies $\bar X$ by $Z$, its variance contribution is multiplied by $Z^2$, and its standard deviation by $Z$.

Conditional volatility describes repeated experience for a given risk. It is different from the variation in premiums across a portfolio of risks, which also reflects differences in underlying risks. For example, if $\theta$ varies across risks with variance $\tau^2$, $\varepsilon$ is independent of $\theta$, and every risk has the same fixed complement $m$, then:

\operatorname{Var}(P)=Z^2\left(\tau^2+\frac{\sigma^2}{n}\right).

This portfolio variance includes both genuine risk differences and noise in observed experience. It should not be described as uncertainty about one particular risk without specifying which source of variation is being measured.

Deriving the Volatility and Incentive Effects

For fixed $m$ and $Z$, consider two possible experience values, $\bar X_1$ and $\bar X_2$. Subtracting their indicated premiums eliminates the complement:

P_2-P_1 =\left[Z\bar X_2+(1-Z)m\right] -\left[Z\bar X_1+(1-Z)m\right] =Z(\bar X_2-\bar X_1).

Thus, if a risk’s experience measure falls by $\Delta>0$, its indicated premium falls by $Z\Delta$, provided the complement and weight remain unchanged. This is the direct rating incentive: $Z$ is the marginal premium response to a one-unit change in the experience measure. It is not necessarily the response of actual losses to an intervention, and it does not establish that the insured can control every component of the measured loss.

The volatility result follows in the same way. With $P=Z\bar X+(1-Z)m$ and fixed $m$, adding a constant does not affect variance, while multiplying a random variable by $Z$ multiplies its variance by $Z^2$. Hence $\operatorname{Var}(P\mid\theta)=Z^2\operatorname{Var}(\bar X\mid\theta)$. A higher $Z$ therefore increases both sensitivity to experience changes and the premium’s response to noisy experience.

Worked Example: Comparing Weights Under Noisy Experience

A hypothetical commercial equipment account has an experience measure $\bar X$. Conditional on its underlying expected loss cost $\theta$, assume $\bar X\sim N(\theta,100^2)$. Across accounts, assume $\theta\sim N(1{,}000,150^2)$, independently of the conditional experience error. The complement is fixed at $m=1{,}000$. For this account, $\bar X=1{,}300$. Compare weights $Z=0.40$ and $Z=0.70$: calculate the premium under each, the conditional standard deviation of a future premium indication for this same risk, and the across-account standard deviation under the stated model. All amounts are dollars per exposure unit.

Method. Apply the linear credibility formula for each premium. Use the conditional variance formula for repeated experience on the same risk, then the portfolio variance formula for variation across risks.

Weight $Z=0.40$. The premium is $0.40(1{,}300)+0.60(1{,}000)=520+600=\$1{,}120$. The conditional standard deviation is $0.40(100)=\$40$. Across accounts, $\operatorname{Var}(\bar X)=150^2+100^2=22{,}500+10{,}000=32{,}500$. Therefore $\operatorname{Var}(P)=0.40^2(32{,}500)=0.16(32{,}500)=5{,}200$, and the across-account standard deviation is $\sqrt{5{,}200}=\$72.11$.

Weight $Z=0.70$. The premium is $0.70(1{,}300)+0.30(1{,}000)=910+300=\$1{,}210$. The conditional standard deviation is $0.70(100)=\$70$. Across accounts, $\operatorname{Var}(P)=0.70^2(32{,}500)=0.49(32{,}500)=15{,}925$, giving a standard deviation of $\sqrt{15{,}925}=\$126.19$.

The higher weight raises the indicated premium for this above-complement account by $\$90$. It also raises both measures of volatility. Common mistake avoided: using $Z$, rather than $Z^2$, to scale variance. The standard deviation scales by $Z$; the variance scales by $Z^2$. The square roots above check the reported standard deviations.

The Complement Can Change Level and Uncertainty

For a fixed weight, changing a fixed complement from $m_1$ to $m_2$ changes the premium by $(1-Z)(m_2-m_1)$. It does not change the marginal response to the risk’s own experience, which remains $Z$. A complement can, however, be estimated rather than fixed. Its uncertainty then contributes to premium volatility.

More generally, if the complement is a random quantity $M$, then:

\operatorname{Var}(P) =Z^2\operatorname{Var}(\bar X) +(1-Z)^2\operatorname{Var}(M) +2Z(1-Z)\operatorname{Cov}(\bar X,M).

This follows by applying the variance formula to the sum $Z\bar X+(1-Z)M$. The covariance term matters: experience and complement estimates from overlapping data need not be independent. If they are independent, that term is zero. A segment-specific complement may also change the premium level systematically across segments; its effect should not be mistaken for individual experience response.

Worked Example: Comparing a Fixed and an Uncertain Complement

For a hypothetical specialty liability account, assume $\bar X\sim N(1{,}200,100^2)$, $Z=0.50$, and a complement with mean $\$900$. Compare a fixed complement of $\$900$ with an estimated complement $M\sim N(900,120^2)$, independent of $\bar X$. Calculate the expected premium and standard deviation in each case.

Method. The mean premium is $Z E[\bar X]+(1-Z)E[M]$. For volatility, use the variance of a weighted sum; in the estimated-complement case, independence makes covariance zero.

With a fixed complement, the expected premium is $0.50(1{,}200)+0.50(900)=600+450=\$1{,}050$. Its variance is $0.50^2(100^2)=0.25(10{,}000)=2{,}500$, so its standard deviation is $\sqrt{2{,}500}=\$50.00$.

With the estimated complement, the expected premium remains $0.50(1{,}200)+0.50(900)=\$1{,}050$. Its variance is $0.50^2(10{,}000)+0.50^2(120^2)=2{,}500+0.25(14{,}400)=2{,}500+3{,}600=6{,}100$. Its standard deviation is $\sqrt{6{,}100}=\$78.10$.

The complement uncertainty raises volatility without changing the expected premium. Common mistake avoided: treating a complement estimated from data as though it were a known constant. The two standard deviations are checked by squaring them: $50.00^2=2{,}500$ and $78.10^2$ is approximately $6{,}100$, with the displayed standard deviation rounded to cents.

Measure the Incentive Without Overstating It

A credibility weight describes how much the rating formula responds to a change in the experience measure, not how much losses will change when an insured takes an action. To connect an intervention to a premium, state how the intervention changes the measure. For example, if a loss-control action is expected to reduce the prospective experience measure by $\Delta$, and the complement and weight are unchanged, the expected indicated-premium reduction is $Z\Delta$.

Worked Example: Premium Reward for a Loss Reduction

A hypothetical fleet’s historical loss-cost measure is $\bar X=1{,}500$, with complement $m=1{,}100$. Two proposed rating designs use $Z=0.25$ and $Z=0.65$. Assume a safety intervention is expected to reduce the next comparable experience measure by $\$200$ per exposure unit, with unchanged exposure, complement, and weight. The experience measure before the intervention is modeled as $N(1{,}500,120^2)$; after the intervention it is $N(1{,}300,120^2)$. Calculate both current indicated premiums and the expected premium reduction under each design.

Method. Calculate $Z\bar X+(1-Z)m$ before and after the expected experience change. Equivalently, use the derived marginal response $Z\Delta$. The stated normal distributions have the same standard deviation, so the intervention changes the mean, not the assumed experience volatility.

At $Z=0.25$, the current indicated premium is $0.25(1{,}500)+0.75(1{,}100)=375+825=\$1{,}200$. After the expected reduction, it is $0.25(1{,}300)+0.75(1{,}100)=325+825=\$1{,}150$. The premium reduction is $\$50$, also $0.25(\$200)=\$50$.

At $Z=0.65$, the current indicated premium is $0.65(1{,}500)+0.35(1{,}100)=975+385=\$1{,}360$. After the expected reduction, it is $0.65(1{,}300)+0.35(1{,}100)=845+385=\$1{,}230$. The premium reduction is $\$130$, also $0.65(\$200)=\$130$.

The larger weight gives a larger premium reward for the same assumed reduction in experience. Common mistake avoided: multiplying the reduction by $1-Z$, which is the complement weight rather than the experience weight. The calculation concerns premium only; it does not measure the intervention’s claim savings, implementation costs, or effect on total expected cost.

Exam Traps and Practical Interpretation

When comparing credibility designs, keep the following distinctions explicit:

Variance versus standard deviation. A weight $Z$ scales a random premium’s standard deviation by $Z$, but its variance by $Z^2$.
Conditional versus portfolio volatility. Conditional volatility holds the risk’s underlying expected cost fixed. Portfolio volatility includes differences among risks as well as observation noise.
Experience response versus complement response. The slope with respect to experience is $Z$; the slope with respect to a complement is $1-Z$. A change in the complement can shift premiums even if individual experience is unchanged.
Premium incentive versus loss-control effect. The formula quantifies the premium change for a specified change in the experience measure. It does not establish that an action will produce that change.
Fixed versus estimated complements. If a complement is estimated, its variance and possible covariance with experience can affect premium volatility.

A useful additional check is whether the comparison holds the exposure basis, coverage, experience period, and complement definition constant. If these change at the same time as $Z$, the observed premium difference cannot be attributed to credibility alone. These distinctions support a disciplined diagnosis of experience-rating incentives and volatility, aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.

Check Your Understanding

For each problem, treat the credibility formula and stated assumptions as applying directly, and keep dollars per exposure unit throughout.

A hypothetical account has $\bar X=980$, $m=1{,}100$, and $Z=0.35$. What is its prospective premium? (Answer: $\$1{,}058$)
Conditional on a risk’s underlying cost, its experience measure has standard deviation $\$80$. If $Z=0.60$ and the complement is fixed, what is the conditional standard deviation of the premium? (Answer: $\$48$)
Assume $m=E[\theta]$, $\operatorname{Var}(\theta)=1{,}600$, and $\operatorname{Var}(\bar X\mid\theta)=900$. For $Z=0.70$, what is the mean squared error of $P$ as an estimator of $\theta$? (Answer: $585$ dollars squared)
Let $Z=0.40$, $\operatorname{Var}(\bar X)=2{,}500$, and let an independent estimated complement $M$ have variance $10{,}000$. What is the variance and standard deviation of $P$? (Answer: variance $4{,}000$ dollars squared; standard deviation $\$63.25$)
A fixed-complement rating has $Z=0.28$. If a hypothetical intervention reduces the experience measure by an expected $\$75$, with the weight and complement unchanged, what is the expected premium reduction? (Answer: $\$21$)

What You'll Learn