Checking The Constant Process-Variance Assumption

Why Check the Exposure-Variance Relationship?

In the Bühlmann-Straub model, a risk-period summary based on more exposure is assumed to be more precise. Specifically, its process variance is inversely proportional to its exposure. That relationship matters because it determines how observations from periods of different sizes contribute to credibility estimates. A serious departure can make the usual exposure weights misleading, even if the calculations that follow are mechanically correct. Checking this assumption is aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.

The diagnostic question is not whether every observation follows a perfectly smooth pattern. Random variation can obscure a relationship in a small data set. Instead, ask whether the observed within-risk variation is broadly consistent with a variance that decreases in proportion to exposure, and whether the conclusion is supported by enough periods to be useful.

Definition: Let $X_{ij}$ be the observed summary for risk $i$ in period $j$, and let $w_{ij}>0$ be its exposure in a consistent unit. The Bühlmann-Straub constant process-variance assumption is $$ E[X_{ij}\mid \Theta_i]=\mu_i,\qquad \operatorname{Var}(X_{ij}\mid \Theta_i)=\frac{\sigma^2}{w_{ij}}, $$ where $\Theta_i$ represents the risk’s underlying characteristics, $\mu_i$ is its conditional mean, and $\sigma^2$ is a common process-variance parameter. Under the usual model, process fluctuations from different periods are conditionally independent. The assumption concerns process variation around a risk’s mean; it does not say that different risks have the same mean.

For example, quadrupling exposure predicts one-quarter the process variance, not one-quarter the standard deviation. If $w$ changes from 1 to 4, the standard deviation changes by a factor of $1/2$, since standard deviation is the square root of variance. The exposure unit must be defined and used consistently; changing the unit changes the numerical value of $\sigma^2$, though not the underlying inverse relationship.

What to Compare

If the risk mean were known, the simplest diagnostic would be to form the process residual $e_{ij}=X_{ij}-\mu_i$. The model predicts that $w_{ij}e_{ij}^2$ has expectation $\sigma^2$. Thus raw squared residuals should, on average, be smaller at larger exposures, while exposure-weighted squared residuals should have broadly similar averages across exposure groups.

In practice, $\mu_i$ is usually unknown. A natural estimate is the exposure-weighted mean for risk $i$:

\bar X_i=\frac{\sum_j w_{ij}X_{ij}}{W_i}, \qquad W_i=\sum_j w_{ij}.

The centered residual $r_{ij}=X_{ij}-\bar X_i$ is not distributed like a raw process residual: estimating the mean uses the same observations and reduces the residual variance. Comparing $w_{ij}r_{ij}^2$ directly across exposures can therefore be misleading, especially when a risk has few periods.

Result: Under the model above, with conditionally independent period errors and a weighted mean estimated from the same $n_i$ periods, $$ \operatorname{Var}(r_{ij}\mid \Theta_i) =\sigma^2\left(\frac{1}{w_{ij}}-\frac{1}{W_i}\right). $$ Consequently, an exposure-adjusted squared residual for checking scale is $$ q_{ij}=\frac{r_{ij}^2}{1/w_{ij}-1/W_i}. $$ Its expected value is $\sigma^2$. This comparison requires at least two positive-exposure periods for a risk; a single period cannot provide a within-risk residual.

Why the Centered-Residual Adjustment Works

Write $X_{ij}=\mu_i+\varepsilon_{ij}$, where $E[\varepsilon_{ij}\mid\Theta_i]=0$ and $\operatorname{Var}(\varepsilon_{ij}\mid\Theta_i)=\sigma^2/w_{ij}$. The weighted mean error is $\bar\varepsilon_i=\sum_jw_{ij}\varepsilon_{ij}/W_i$. Independence gives $\operatorname{Var}(\bar\varepsilon_i\mid\Theta_i)=\sigma^2/W_i$. It also gives $\operatorname{Cov}(\varepsilon_{ij},\bar\varepsilon_i\mid\Theta_i)=\sigma^2/W_i$, because the $j$th term contributes $w_{ij}(\sigma^2/w_{ij})/W_i$. Therefore:

\begin{aligned} \operatorname{Var}(r_{ij}\mid\Theta_i) &=\operatorname{Var}(\varepsilon_{ij}-\bar\varepsilon_i\mid\Theta_i)\\ &=\frac{\sigma^2}{w_{ij}}+\frac{\sigma^2}{W_i} -2\frac{\sigma^2}{W_i}\\ &=\sigma^2\left(\frac{1}{w_{ij}}-\frac{1}{W_i}\right). \end{aligned}

A related overall diagnostic is the within-risk weighted sum of squares:

Q=\sum_i\sum_j w_{ij}(X_{ij}-\bar X_i)^2.

For a risk with $n_i$ periods, its expected contribution is $(n_i-1)\sigma^2$. Thus, if the model is appropriate and the required periods are available, a pooled method-of-moments estimate is $Q/\sum_i(n_i-1)$. This summarizes overall process variation; it does not by itself check whether the inverse-exposure pattern holds. A scale diagnostic should also compare exposure-adjusted residuals across exposure levels.

Practical Diagnostic Sequence

Check the data definition.
Confirm that each $X_{ij}$ is a comparable summary and that exposures use the same unit. A change in coverage, limits, or measurement method can create apparent variance changes unrelated to exposure.

Estimate risk means.
For each risk with multiple observed periods, calculate its exposure-weighted mean and residuals from that mean.

Adjust for fitting the mean.
Calculate $q_{ij}=r_{ij}^2/(1/w_{ij}-1/W_i)$. Compare averages or plots of these values across exposure bands. Under the assumption, there should be no systematic exposure trend in their scale.

Interpret cautiously.
Look for a persistent pattern across risks, not just one unusually large residual. Sparse periods and a few influential observations make visual and numerical comparisons unstable.

A useful plot places exposure on the horizontal axis and either $q_{ij}$ or the absolute adjusted residual $|r_{ij}|/\sqrt{1/w_{ij}-1/W_i}$ on the vertical axis. Under the assumed model, the adjusted residuals should not become systematically more or less dispersed with exposure. Raw residuals, by contrast, are expected to shrink as exposure increases.

Worked Example: Comparing Known-Mean Residuals

A hypothetical portfolio has an independently established expected loss cost of $200 per exposure unit for the risks in a small diagnostic sample. Four periods have exposure 1 and residuals of $5, −$5, $10, and −$10. Four other periods have exposure 4 and residuals of $2.50, −$2.50, $5, and −$5. Compare raw and exposure-adjusted dispersion.

Method. Because the means are supplied independently, use $e=X-\mu$ directly rather than estimating a separate mean from these eight observations. For exposure 1, the mean squared residual is $(25+25+100+100)/4=62.50$. For exposure 4, it is $(6.25+6.25+25+25)/4=15.625$. The high-exposure raw mean square is one-quarter of the low-exposure mean square, as the model predicts.

Now multiply each squared residual by its exposure. For exposure 1, the four values $we^2$ are 25, 25, 100, and 100, with mean $250/4=62.50$. For exposure 4, they are 25, 25, 100, and 100, also with mean 62.50. The two groups have matching exposure-adjusted dispersion in this sample.

Check and common mistake. The raw mean squares have ratio $15.625/62.50=0.25$, matching the inverse exposure ratio $1/4$. Comparing the raw mean squares and concluding that high-exposure periods are unusually stable would ignore the expected effect of exposure. The adjusted comparison is consistent with the assumption, though a small sample cannot establish it.

Worked Example: Adjusting Residuals After Fitting a Risk Mean

For one hypothetical risk, three periods have exposures 1, 2, and 4 and observed loss costs of $92, $102, and $101. Calculate the weighted mean and the centered-residual scale diagnostics. Round diagnostic values to two decimals.

Method. The total exposure is $W=1+2+4=7$. The weighted total is $1(92)+2(102)+4(101)=92+204+404=700$, so $\bar X=700/7=\$100$. The residuals are −$8, $2, and $1.

For exposure 1, the variance factor is $1/1-1/7=6/7$. Thus $q=64/(6/7)=74.67$. For exposure 2, the factor is $1/2-1/7=5/14$, giving $q=4/(5/14)=11.20$. For exposure 4, the factor is $1/4-1/7=3/28$, giving $q=1/(3/28)=9.33$.

The adjusted values differ substantially, but they are only three observations from one risk. The pattern could reflect ordinary process fluctuation; it is not enough evidence by itself to reject the assumption. In particular, with just two periods, the weighted-mean fit forces the two adjusted squared residuals to be equal, so that case cannot reveal an exposure trend within that risk.

Check and common mistake. The weighted residual sum is $1(64)+2(4)+4(1)=76$. The three diagnostic values sum to $74.67+11.20+9.33=95.20$; they need not sum to the weighted residual sum because each uses a different variance factor. Using $r^2/(1/w)$ instead would omit the fitted-mean adjustment and would not have the stated expected value.

Worked Example: Estimating Overall Process Variation

Two hypothetical risks each have four periods with exposures 1, 1, 4, and 4. Risk A’s observed summaries are $90, $110, $95, and $105; risk B’s are $195, $205, $205, and $195. Calculate the pooled within-risk estimate of $\sigma^2$, then assess whether the calculation alone checks inverse scaling.

Method. For risk A, total exposure is 10 and the weighted total is $90+110+4(95)+4(105)=1{,}000$, so its weighted mean is $100. Its residuals are −$10, $10, −$5, and $5. The weighted sum of squared residuals is $100+100+4(25)+4(25)=400$.

For risk B, the weighted total is $195+205+4(205)+4(195)=2{,}000$, so its weighted mean is $200. Its residuals are −$5, $5, $5, and −$5. Its weighted sum of squares is $25+25+4(25)+4(25)=250$. Therefore $Q=400+250=650$. There are eight periods and two fitted risk means, so the residual degrees of freedom are $8-2=6$. The pooled estimate is $\hat\sigma^2=650/6=108.33$, in squared summary units times exposure units.

This pooled value estimates an overall process scale under the model; by itself it does not show that variance is inversely proportional to exposure. For example, one must still compare residual dispersion across exposure levels, using the centered-residual adjustment or an equivalent model-based diagnostic.

Check and common mistake. The divisor is 6, not 8, because two risk means were estimated. Dividing by the number of observations would give $650/8=81.25$ and would not be the pooled within-risk estimator described here.

Interpreting Departures and Avoiding Traps

A clear exposure trend in adjusted residual dispersion is a reason to investigate the model or the data, not an automatic instruction to substitute a different credibility formula. Possible explanations include inconsistent exposure measurement, changes in the definition of the observed summary, influential losses, or genuine variation in process behavior. Those possibilities should be evaluated before a model is changed.

Do not compare unadjusted residual sizes as if they should match. Under the assumption, raw variance falls as exposure rises.
Do not confuse variance with standard deviation. Four times the exposure predicts half the standard deviation, not one-quarter.
Do not overlook fitted-mean effects. Centered residuals have variance $\sigma^2(1/w_{ij}-1/W_i)$, not simply $\sigma^2/w_{ij}$.
Do not treat a few residuals as a formal verdict. Residual comparisons are diagnostics; a limited number of periods can give a misleading impression.

Key takeaway: The model predicts constant variance for exposure-weighted raw errors. When risk means are estimated from the same periods, compare centered residuals after adjusting for the variance reduction caused by fitting those means, and interpret any pattern in light of the amount and quality of data.

Check Your Understanding

For each problem, use the stated assumptions and distinguish raw dispersion from exposure-adjusted dispersion.

A known-mean residual is $12 at exposure 1 and −$6 at exposure 4. Calculate $we^2$ for each period. Are the values consistent with equal adjusted squared residuals? (Answer: 144 and 144; yes.)
If $\sigma^2=36$, calculate the process variances at exposures 2 and 8. (Answer: 18 and 4.5, respectively.)
A risk has two periods with exposures 1 and 3 and observed summaries $40 and $50. Find its weighted mean and each centered-residual diagnostic $q_{ij}$. (Answer: weighted mean $47.50; both $q_{ij}=75$.)
Three risks each have three observed periods, and their combined within-risk weighted sum of squares is $Q=540$. Calculate the pooled method-of-moments estimate of $\sigma^2$, assuming the model conditions hold. (Answer: $540/[3(3-1)]=90$.)
Across two exposure bands, the observed root-mean-square raw residual is twice as large in the band with exposure 1 as in the band with exposure 2. Is that pattern consistent with the inverse-variance assumption? (Answer: No. The model predicts a raw standard-deviation ratio of $\sqrt{2}$, not 2; the observed ratio suggests a steeper decline, though further data and checks are needed.)

What You'll Learn