Sparse Histories and Zero Exposures

Why Sparse Histories Need Care

A risk may have just one short period of experience, several periods with very different exposures, or no observed exposure at all. Bühlmann-Straub credibility accommodates unequal exposures by giving more weight to periods that contain more experience. The key is to distinguish little experience from no experience, and both from an observed outcome of zero.

The model prices a risk using its exposure-weighted experience together with the collective mean. A period with zero exposure cannot supply an observed rate to average: it contributes neither a rate nor weight. By contrast, a zero outcome recorded over positive exposure is valid experience and can affect the premium.

Definition: For risk $i$ in period $j$, let $w_{ij}$ be its nonnegative exposure and $X_{ij}$ its observed experience rate per unit of exposure. In the Bühlmann-Straub model, conditional on the risk parameter $\Theta_i$, the observations are independent and satisfy $$ E[X_{ij}\mid\Theta_i]=\mu(\Theta_i), \qquad \operatorname{Var}(X_{ij}\mid\Theta_i)=\frac{\sigma^2(\Theta_i)}{w_{ij}} $$ for periods with $w_{ij}>0$. Define the collective mean $m=E[\mu(\Theta_i)]$, the variance of hypothetical means $a=\operatorname{Var}(\mu(\Theta_i))$, the expected process variance per unit exposure $s^2=E[\sigma^2(\Theta_i)]$, and $K=s^2/a$, assuming $a>0$. A risk's total observed exposure is $w_i=\sum_j w_{ij}$. If $w_i>0$, its experience rate is $\bar X_i=\sum_j w_{ij}X_{ij}/w_i$, its credibility factor is $Z_i=w_i/(w_i+K)$, and its Bühlmann-Straub premium per unit exposure is $P_i=Z_i\bar X_i+(1-Z_i)m$. A period with $w_{ij}=0$ has no defined $X_{ij}$ under this model and is omitted from the sums.

How Exposure Enters the Calculation

Exposure might represent policy-years, insured units, or another consistent measure of the amount of risk observed. The experience $X_{ij}$ must be expressed as a rate per unit of that exposure. For example, if a period's aggregate loss is $L_{ij}$, the corresponding rate is $X_{ij}=L_{ij}/w_{ij}$ when $w_{ij}>0$. The premium $P_i$ is then a rate; for a future period with exposure $w^*$, the model's expected aggregate cost is $w^*P_i$, provided the same rate basis and assumptions apply.

The factor $w_{ij}$ is not just a label. It determines how much information a period contributes. If two periods have equal rate observations but one has four times the exposure, that period receives four times the weight in $\bar X_i$. The total exposure $w_i$, in turn, determines how much credibility the risk receives relative to the collective mean.

\bar X_i=\frac{\sum_{j:w_{ij}>0}w_{ij}X_{ij}}{\sum_{j:w_{ij}>0}w_{ij}}, \qquad Z_i=\frac{w_i}{w_i+K}, \qquad P_i=m+Z_i(\bar X_i-m)

The final form makes the adjustment easy to interpret: the premium starts at the collective mean $m$, then moves toward the risk's observed rate by the fraction $Z_i$ of the difference. When experience is sparse, the factor is small and the premium remains closer to the collective mean. It is still important to calculate $\bar X_i$ correctly; the credibility factor does not repair an unweighted or otherwise incorrect average.

Worked Example: A Short History With a Zero-Exposure Period

A hypothetical portfolio has collective mean $m=\$120$ per policy-year, $a=900$, and $s^2=3{,}600$. A risk has the following history: exposure $0.5$ with rate $\$100$, exposure $0$ with no rate recorded, and exposure $1.5$ with rate $\$140$. Calculate its Bühlmann-Straub premium per policy-year and the expected cost for a future exposure of 2 policy-years.

Method: exposure-weight the observed rates, then apply the Bühlmann-Straub factor. First calculate the structural constant: $$ K=\frac{s^2}{a}=\frac{3{,}600}{900}=4 $$ The total exposure is $w_i=0.5+1.5=2$. The middle period has zero exposure, so it contributes nothing and has no rate to include. Thus $$ \bar X_i=\frac{(0.5)(100)+(1.5)(140)}{2} =\frac{50+210}{2} =\frac{260}{2} =\$130 $$ The credibility factor and premium are $$ Z_i=\frac{2}{2+4}=\frac{1}{3}, \qquad P_i=\frac{1}{3}(130)+\frac{2}{3}(120) =\frac{130+240}{3} =\$123.33 $$ rounded to the nearest cent. For 2 future policy-years, the expected aggregate cost is $2(123.33)=\$246.66$ if using the displayed rounded rate. Using the unrounded premium $370/3$, the expected cost is $740/3=\$246.67$ rounded to the nearest cent.

Check: The observed rate is $\$10$ above the collective mean, and the premium is one-third of that difference above the mean: $120+(1/3)(10)=\$123.33$. The separate aggregate-cost calculation also confirms why it is best to carry unrounded values until the final step. The common mistake avoided is treating the zero-exposure period as another observation.

Why the Formula Uses Total Exposure

The credibility factor follows from the relative variation in a risk's weighted average. Conditional independence gives the average a conditional variance that decreases as total exposure increases. Averaging over risks adds the between-risk variation $a$. These two sources of variation explain why the credibility factor depends on $w_i$ through $w_i/(w_i+K)$.

Find the conditional variance of the weighted average.
For positive exposures, conditional independence and the model variance give $\operatorname{Var}(\bar X_i\mid\Theta_i)=\sum_j w_{ij}^2\sigma^2(\Theta_i)/w_{ij}\,/\,w_i^2=\sigma^2(\Theta_i)/w_i$.

Average over risks.
The expected conditional variance is $s^2/w_i$. Since $E[\bar X_i\mid\Theta_i]=\mu(\Theta_i)$, the total variance is $\operatorname{Var}(\bar X_i)=a+s^2/w_i$, while $\operatorname{Cov}(\bar X_i,\mu(\Theta_i))=a$.

Obtain the linear credibility weight.
The least-squares linear predictor of $\mu(\Theta_i)$ from $\bar X_i$, with an intercept, assigns weight $\operatorname{Cov}(\bar X_i,\mu(\Theta_i))/\operatorname{Var}(\bar X_i)=a/(a+s^2/w_i)=w_i/(w_i+K)$ to the experience.

This justification relies on the stated model: a common conditional mean across periods for a risk, conditional independence, and conditional variance inversely proportional to exposure. If experience is dependent across periods or exposures do not represent the relevant volume of risk, this variance argument and the resulting factor need not apply.

Worked Example: No Exposure Versus Zero Observed Loss

For a hypothetical class, $m=\$100$, $a=400$, and $s^2=1{,}600$, so $K=4$. Risk A has zero exposure in every period and therefore no observed rates. Risk B has exposure 1 in one period and 2 in another; its observed rates are both zero. Calculate each risk's premium per unit exposure.

Risk A: no exposure. Since $w_A=0$, there is no $\bar X_A$ to calculate and the formula for the factor gives $Z_A=0/(0+4)=0$. The experience term is absent, so its premium is the collective mean, $P_A=\$100$. This is not a claim that the risk had zero losses; it has no observed experience in this model.

Risk B: positive exposure with zero rates. Its total exposure is $w_B=1+2=3$, and its weighted experience rate is $\bar X_B=((1)(0)+(2)(0))/3=0$. Thus $$ Z_B=\frac{3}{3+4}=\frac{3}{7}, \qquad P_B=\frac{3}{7}(0)+\frac{4}{7}(100) =\frac{400}{7} =\$57.14 $$ rounded to the nearest cent.

Check: Risk A remains at the collective mean because it supplied no information. Risk B's positive exposure and zero rates pull its premium below the mean. The common mistake avoided is assigning a zero rate to Risk A and pricing it as though it had observed zero losses.

Handling Very Small Exposures

A small but positive exposure is valid experience. Calculate its rate and total exposure using the same units as the portfolio parameters, then apply the formula. Do not round a small exposure to zero before calculating: doing so can discard real experience. Conversely, do not inflate a fractional exposure to a whole period just because the observation is reported in an annual format.

The scale of exposure matters. The constant $K=s^2/a$ is expressed in the same exposure units as $w_i$. If exposure is converted from policy-years to months, the variance and exposure convention must be converted consistently. Combining a monthly exposure with a $K$ calibrated in policy-years gives a meaningless credibility factor.

Worked Example: Unequal Sparse Periods and an Unweighted-Average Trap

A hypothetical risk has exposure $0.2$ with an observed loss-cost rate of $\$500$ per unit and exposure $0.3$ with a rate of $\$100$ per unit. A third period has zero exposure and no valid rate. The collective parameters are $m=\$200$, $a=1{,}600$, and $s^2=6{,}400$. Calculate the risk premium per unit.

Method: exclude the zero-exposure period and weight each usable rate by exposure. The structural constant is $K=6{,}400/1{,}600=4$, and total exposure is $w_i=0.2+0.3=0.5$. The weighted experience rate is $$ \bar X_i=\frac{(0.2)(500)+(0.3)(100)}{0.5} =\frac{100+30}{0.5} =\$260 $$ The credibility factor is $Z_i=0.5/(0.5+4)=0.5/4.5=1/9$. Therefore $$ P_i=\frac{1}{9}(260)+\frac{8}{9}(200) =\frac{260+1{,}600}{9} =\$206.67 $$ rounded to the nearest cent.

Check and trap: The weighted rate is $\$60$ above the collective mean, so a one-ninth credibility weight puts the premium $\$60/9=\$6.67$ above $\$200$. An incorrect unweighted average would be $(500+100)/2=\$300$, producing $(300+8(200))/9=\$211.11$. That calculation overweights the high rate from the smaller exposure and is not the Bühlmann-Straub experience average.

Data Checks and Common Exam Traps

Before calculating a premium for a sparse history, verify that the exposure and experience refer to the same risk, period, and basis. A zero-exposure record with an apparent rate is internally inconsistent with the model: a rate per unit exposure cannot be calculated when the denominator is zero. If a claim is reported during a period with no policy exposure, determine whether it belongs to an earlier exposure period or whether the exposure record is incomplete. Do not silently insert the claim as a zero-exposure observation.

Do not confuse missing experience with zero experience. Zero exposure means no rate enters the calculation; a zero rate with positive exposure is an observation.
Do not use a simple average of period rates when exposures differ. The experience rate is weighted by exposure.
Do not confuse a premium rate with an aggregate premium. Multiply the rate by future exposure only when an expected aggregate amount is requested.
Keep units consistent. Exposure and $K$ must use the same exposure scale.
Do not interpret a zero-exposure premium as evidence of low risk. It reflects reliance on the collective mean, not a favorable individual record.

Key takeaway: With positive total exposure, compute the exposure-weighted rate and use $Z_i=w_i/(w_i+K)$. With no exposure, there is no individual experience rate to use, so the premium is the collective mean. These conclusions are aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.

Check Your Understanding

For each problem, use the Bühlmann-Straub model and round monetary premiums to the nearest cent where needed.

A risk has zero observed exposure. The collective mean is $\$75$, $a=225$, and $s^2=900$. What is its premium per unit exposure? (Answer: $\$75.00$.)
A risk has exposure $0.5$ and rate $\$110$. If $m=\$150$, $a=400$, and $s^2=1{,}600$, calculate its premium. (Answer: $\$145.56$.)
A risk has exposure 1 with rate $\$0$ and exposure 2 with rate $\$120$. If $m=\$90$, $a=100$, and $s^2=400$, calculate its weighted experience rate and premium. (Answer: Weighted rate $\$80$; premium $\$85.71$.)
A risk has exposure $0.25$ with rate $\$320$, exposure $0.75$ with rate $\$80$, and a third period with zero exposure. If $m=\$100$, $a=400$, and $s^2=1{,}600$, calculate its premium. (Answer: $\$108.00$.)
A risk's record lists exposure zero with a purported rate of $\$50$, and exposure 1 with a valid rate of $\$160$. With $m=\$200$, $a=400$, and $s^2=1{,}600$, explain how to handle the zero-exposure entry and calculate the premium from the valid experience. (Answer: The zero-exposure rate is invalid and must be reconciled or omitted; using the valid period only, the premium is $\$192.00$.)

What You'll Learn