Predictive Mean for a Future Exposure Period

From Past Experience to a Future Mean

A credibility estimate often describes a loss rate, such as dollars per vehicle-year. A decision-maker may instead need an expected total for a future period, perhaps because the risk is expected to have a specified number of vehicle-years of exposure. These are related quantities, but they are not interchangeable: the estimated rate is first calculated from historical experience and collective information; the expected future total is that rate multiplied by the future exposure.

This distinction is central to the Bühlmann-Straub model. Historical exposure determines how much credibility to give the risk’s own experience. Once that credibility-weighted rate has been found, a separately specified future exposure scales it to a predicted total. The future exposure does not increase the credibility assigned to historical data.

Definition: In the Bühlmann-Straub model, let $X_j$ be the observed loss rate for period $j$, based on exposure $w_j>0$, and let $\Theta$ denote the risk’s underlying parameter. Assume that, conditional on $\Theta$, the observations are independent, with $E[X_j\mid\Theta]=\mu(\Theta)$ and $\operatorname{Var}(X_j\mid\Theta)=\sigma^2(\Theta)/w_j$. Define the collective mean $m=E[\mu(\Theta)]$, the expected process variance per unit exposure $v=E[\sigma^2(\Theta)]$, and the variance of hypothetical means $a=\operatorname{Var}(\mu(\Theta))$. For historical total exposure $W=\sum_j w_j$, the experience rate is $\bar X_w=\sum_j w_jX_j/W$. When $a>0$, the credibility constant is $K=v/a$ and the credibility factor is $Z=W/(W+K)$.

The model assumes that the risk’s underlying mean remains relevant for the future period and that exposure is measured on a consistent basis. An exposure might be policy-years, insured units, or another defined measure of risk volume. The rate and exposure units must match: multiplying dollars per policy-year by policy-years gives dollars. If a period has zero exposure, it contributes no observed rate or weight to these formulas.

The Predictive Rate and the Future Total

The Bühlmann-Straub credibility premium for the rate is a weighted average of the risk’s historical rate and the collective mean. If $w_*$ is the specified future exposure and $T_*$ is the future aggregate loss, the corresponding predicted mean total is obtained by multiplying the rate estimate by $w_*$.

\widehat{\mu}=Z\bar X_w+(1-Z)m,\qquad \widehat{E[T_*]}=w_*\widehat{\mu}.

Here $\widehat{\mu}$ is a rate, in loss units per unit exposure. The second expression is a total loss in dollars if the rate is expressed in dollars per exposure unit and $w_*$ is measured in those exposure units. This is a prediction of the conditional mean under the credibility model, not a claim that the actual future loss will equal the predicted amount.

Key distinction: The credibility factor uses historical exposure $W$, not $W+w_*$. The future exposure $w_*$ determines the scale of the predicted total after the rate has been estimated; it is not additional historical information about the risk.

Why the Credibility Factor Has This Form

The formula can be justified as the best linear predictor of the risk’s underlying mean using the historical weighted rate. Conditional independence and the variance assumption imply that the weighted rate is unbiased for $\mu(\Theta)$ conditional on the risk parameter. Its unconditional variance has one between-risk component and one within-risk component.

Find the mean and covariance.
Because each historical rate has conditional mean $\mu(\Theta)$, $E[\bar X_w\mid\Theta]=\mu(\Theta)$. Thus $E[\bar X_w]=m$ and $\operatorname{Cov}(\mu(\Theta),\bar X_w)=a$.

Calculate the variance of the weighted rate.
Conditional independence gives $\operatorname{Var}(\bar X_w\mid\Theta)=\sigma^2(\Theta)/W$. The law of total variance therefore gives $\operatorname{Var}(\bar X_w)=a+v/W$.

Choose the linear coefficient.
For a predictor $m+c(\bar X_w-m)$, the coefficient minimizing mean squared error for predicting $\mu(\Theta)$ is covariance divided by variance: $c=a/(a+v/W)$.

Rewrite the coefficient.
Multiplying numerator and denominator by $W/a$ gives $c=W/(W+v/a)=W/(W+K)=Z$. Therefore the best linear predictor is $Z\bar X_w+(1-Z)m$.

The derivation shows why additional historical exposure increases the weight on the risk’s own rate: it reduces the process-variance contribution $v/W$. It also clarifies the meaning of the formula. In the Bühlmann model, this is the best linear predictor. It is not necessarily the exact conditional expectation $E[\mu(\Theta)\mid\text{history}]$ for every possible underlying distribution; equality with that posterior mean requires additional conditions. The future total mean follows by linearity of expectation when the future aggregate is exposure times its future rate.

Worked Calculations

Worked Example: Predicting Both a Rate and a Total

A hypothetical risk has collective mean $m=\$120$ per exposure unit, expected process variance $v=360$, and variance of hypothetical means $a=90$, using consistent rate units. Its historical rate was $\$100$ at exposure 2 and $\$160$ at exposure 4. Find the credibility predictive rate and expected total for a future exposure of 3 units.

Method. First find the historical weighted rate and credibility factor; then calculate the predicted rate and scale it to the future exposure.

Historical rate. Historical exposure is $W=2+4=6$. The exposure-weighted loss is $2(100)+4(160)=200+640=840$, so \[ \bar X_w=\frac{840}{6}=\$140 \] per exposure unit.

Credibility. The credibility constant is $K=v/a=360/90=4$. Thus \[ Z=\frac{6}{6+4}=0.6. \]

Predicted rate. \[ \widehat{\mu}=0.6(140)+0.4(120)=84+48=\$132 \] per exposure unit.

Predicted total. For future exposure $w_*=3$, \[ \widehat{E[T_*]}=3(132)=\$396.00. \] The result is internally consistent: $\$132$ per unit multiplied by 3 units gives $\$396.00$. The common mistake avoided here is reporting $\$132$ as the future total; it is the predicted rate.

Worked Example: Starting With Historical Aggregate Losses

A hypothetical portfolio risk has $m=\$200$ per unit, $v=800$, and $a=200$. Its aggregate historical losses were $\$180$ at exposure 1 and $\$720$ at exposure 3. Find its predicted future rate and expected total for a future exposure of 2 units.

Convert totals to rates. The first historical rate is $180/1=\$180$ per unit. The second is $720/3=\$240$ per unit. Equivalently, the weighted historical rate is \[ \bar X_w=\frac{180+720}{1+3}=\frac{900}{4}=\$225 \] per unit.

Calculate credibility. The historical exposure is $W=1+3=4$, and $K=800/200=4$. Therefore $Z=4/(4+4)=0.5$.

Predict the rate. \[ \widehat{\mu}=0.5(225)+0.5(200)=112.50+100=\$212.50 \] per unit.

Predict the future total. \[ \widehat{E[T_*]}=2(212.50)=\$425.00. \] Thus the rate is $\$212.50$ per unit and the expected total for exposure 2 is $\$425.00$. A common error is to treat the historical aggregate losses as though they were rates; the rates must be weighted by their exposures, or the aggregate losses can be summed and divided by total exposure.

Worked Example: Unequal Historical Exposures and a Larger Future Period

A hypothetical risk has $m=\$80$ per unit, $v=240$, and $a=60$. Historical rates were $\$60$, $\$100$, and $\$120$, at exposures 1, 2, and 3 respectively. Find the future predicted rate and expected total when future exposure is 5 units.

Historical weighted rate. Total historical exposure is $W=1+2+3=6$. The weighted amount is $1(60)+2(100)+3(120)=60+200+360=620$. Hence $\bar X_w=620/6=\$103.333\ldots$ per unit.

Credibility factor. The credibility constant is $K=240/60=4$, so $Z=6/(6+4)=0.6$.

Predicted rate. Using the unrounded historical rate, \[ \widehat{\mu}=0.6\left(\frac{620}{6}\right)+0.4(80)=62+32=\$94 \] per unit.

Predicted total. The expected total is $5(94)=\$470.00$. The rate is exact to the dollar, and multiplying it by the future exposure gives the stated total. The common mistake avoided is taking an unweighted average of the three rates: the exposures differ, so the rate at exposure 3 must receive three times the weight of the rate at exposure 1.

Common Exam Traps and Interpretation

Keep the calculation in two stages. First, use historical exposure and the collective parameters to estimate a future rate. Second, multiply that rate by the specified future exposure to obtain an expected total. Maintain units at each step; this simple check often reveals an incorrect answer.

Using future exposure in $Z$. The credibility factor is $W/(W+K)$, where $W$ is historical exposure. Replacing $W$ with historical plus future exposure falsely treats planned future volume as observed experience.
Confusing total and rate. A result in dollars per exposure unit is not a predicted dollar total. Multiply it by $w_*$ only when the question asks for a future aggregate mean.
Ignoring unequal exposures. The historical rate is $\sum_j w_jX_j/W$, not the ordinary arithmetic mean of period rates unless all exposures are equal.
Rounding too early. Retain the unrounded historical rate and credibility factor through the calculation, then round the requested monetary result to the cent.

For example, in the first worked example, incorrectly including the future exposure in the credibility factor would give $Z=9/(9+4)=9/13$, rather than $6/(6+4)=0.6$. That incorrect factor would produce a rate of $(9/13)(140)+(4/13)(120)=\$133.85$ per unit (rounded), and a total of about $\$401.54$. The correction is to use historical exposure 6 for credibility, giving the correct rate of $\$132$ and expected total of $\$396.00$.

Key takeaway: The Bühlmann-Straub predictive mean for a future exposure is the credibility-weighted rate $Z\bar X_w+(1-Z)m$ multiplied by future exposure $w_*$. Historical exposure sets credibility; future exposure scales the predicted rate into a total.

Check Your Understanding

For each problem, identify whether the requested result is a rate, a future total, or both. Keep units consistent and round monetary answers to the nearest cent.

A risk has $m=\$50$ per unit, $v=180$, and $a=60$. Its rates are $\$40$ at exposure 2 and $\$70$ at exposure 3. Find the predicted rate and future total for exposure 4. (Answer: $\$55.00$ per unit; $\$220.00$ total.)
A risk has $m=\$250$ per unit, $v=600$, and $a=150$. Its historical aggregate losses are $\$220$ at exposure 1 and $\$720$ at exposure 3. Find the predicted rate and expected total for future exposure 2. (Answer: $\$242.50$ per unit; $\$485.00$ total.)
A risk has $m=\$100$ per unit, $v=500$, and $a=100$. Its historical rates are $\$90$ at exposure 2 and $\$120$ at exposure 4. Find the predicted rate and total for future exposure 5. (Answer: $\$105.45$ per unit; $\$527.27$ total.)
A risk has $m=\$75$ per unit, $v=400$, and $a=100$. Historical rates are $\$60$ at exposure 1 and $\$90$ at exposure 2. Find the predicted rate and total for future exposure 7. (Answer: $\$77.14$ per unit; $\$540.00$ total.)
A risk has $m=\$120$ per unit, $v=320$, and $a=80$. Historical aggregate losses are $\$260$ at exposure 2 and $\$700$ at exposure 5. Find the predicted future total for exposure 1.5. (Answer: $\$196.36$ total.)

What You'll Learn