Tutorials › CAS Exam 8 - Advanced Ratemaking › Estimating Severity Variation for Credibility Standards

Home  /  Tutorials  /  CAS Exam 8 / Credibility and Limited Fluctuation

CAS Exam 8 · Credibility and Limited Fluctuation · Tutorial 312 of 989

Estimating Severity Variation for Credibility Standards

Use observed severity data to estimate aggregate-loss variation, calculate a credibility threshold, and assess the effect of uncertainty in the estimated severity coefficient of variation.

Beginner 10 min read

What You'll Learn

  • Calculate the sample coefficient of variation from observed claim severities.
  • Apply the estimated coefficient of variation in an aggregate-loss credibility standard.
  • Explain the assumptions linking claim counts, severities, and aggregate-loss variance.
  • Assess how plausible severity-variation estimates affect required claim volume.
  • Use a sampling-error approximation cautiously and recognize its limits.

Why Severity Variation Must Be Estimated

An aggregate-loss credibility standard depends on more than the average claim size. Even when two portfolios have the same expected severity, the one with more variable severities will generally need more claim volume to make its aggregate losses sufficiently stable. In practice, the severity coefficient of variation is often estimated from historical claim data rather than known in advance.

This tutorial develops the sample estimate, shows how it enters a limited-fluctuation standard, and examines how uncertainty in that estimate affects the result. The calculations use the aggregate-loss criterion, not a claim-count-only criterion.

Definition: For claim severity \(X\) with finite mean \(\mu=E[X]>0\) and finite variance \(\sigma^2=\operatorname{Var}(X)\), the severity coefficient of variation is \(CV=\sigma/\mu\). Given a sample of \(n\) severities \(X_1,\ldots,X_n\), define the sample mean \(\bar X=\sum_{j=1}^{n}X_j/n\), the sample variance \(s^2=\sum_{j=1}^{n}(X_j-\bar X)^2/(n-1)\), and the plug-in estimate \(\widehat{CV}=s/\bar X\).

The denominator \(n-1\) in \(s^2\) is the usual sample-variance convention when estimating the population variance from a sample. The resulting ratio \(s/\bar X\) is a plug-in estimate; it is not generally an exactly unbiased estimator of \(CV\). A reliable estimate also requires severities that are representative of the risks and coverage being priced. Changes in limits, deductibles, claim handling, inflation, or the mix of risks can make historical paid amounts a poor estimate of the target severity distribution.

How the Estimated CV Enters the Standard

Let \(N\) be a Poisson claim count with mean \(\lambda\), and let \(X_1,X_2,\ldots\) be independent, identically distributed severities with mean \(\mu\) and variance \(\sigma^2\), independent of \(N\). The aggregate loss is \(S=\sum_{j=1}^{N}X_j\), with \(S=0\) when \(N=0\). Assume the severity moments are finite and the claim process and severity model are appropriate for the portfolio under consideration.

For a limited-fluctuation aggregate-loss standard, let \(k\) be the allowed relative deviation from expected aggregate loss and let \(z\) be the standard-normal quantile corresponding to the specified two-sided confidence level. The usual normal-approximation threshold is the expected number of claims needed for the standard to meet the selected confidence and tolerance.

Result: Under the Poisson random-sum assumptions, the aggregate-loss threshold is \(\lambda_{\mathrm{req}}=z^2(1+CV^2)/k^2\). When severity parameters are estimated, the plug-in threshold is \(\widehat{\lambda}_{\mathrm{req}}=z^2(1+\widehat{CV}^{\,2})/k^2\). This is a threshold in expected claim count; converting it to exposure requires a claim frequency per exposure unit.

The formula is a limited-fluctuation approximation. It is not an exact probability statement for every claim-count and severity distribution. In particular, it relies on the model assumptions and a normal approximation for aggregate losses at the volume being assessed. Do not interpret an estimated threshold as protection against model misspecification or changing conditions.

Why the Formula Uses \(1+CV^2\)

The severity contribution is not just the variance of an individual claim. Random variation in the number of claims also matters, which is why the formula contains \(1+CV^2\), rather than \(CV^2\) alone.

$$ E[S]=E[N]\mu=\lambda\mu $$

Conditioning on \(N\), the conditional mean and variance of aggregate loss are \(E[S\mid N]=N\mu\) and \(\operatorname{Var}(S\mid N)=N\sigma^2\). Applying the law of total variance gives:

$$ \begin{aligned} \operatorname{Var}(S) &=E[\operatorname{Var}(S\mid N)]+\operatorname{Var}(E[S\mid N])\\ &=E[N\sigma^2]+\operatorname{Var}(N\mu)\\ &=\lambda\sigma^2+\lambda\mu^2\\ &=\lambda(\sigma^2+\mu^2). \end{aligned} $$

The last equality uses the mean and variance of a Poisson variable, both \(\lambda\). Dividing the aggregate variance by the square of its mean yields the squared coefficient of variation of aggregate loss:

$$ \frac{\operatorname{Var}(S)}{(E[S])^2} =\frac{\lambda(\sigma^2+\mu^2)}{\lambda^2\mu^2} =\frac{1+CV^2}{\lambda}. $$

For the normal-approximation criterion, the allowed deviation is \(kE[S]\). Requiring this deviation to be at least \(z\) aggregate-loss standard deviations gives \(kE[S]\ge z\sqrt{\operatorname{Var}(S)}\). Squaring both nonnegative sides and substituting the expressions above gives \(k^2\lambda^2\mu^2\ge z^2\lambda(\sigma^2+\mu^2)\). For \(\lambda>0\), rearrangement yields \(\lambda\ge z^2(1+CV^2)/k^2\). Replacing the unknown \(CV\) with its sample estimate produces the plug-in threshold.

Worked Example: Estimating CV From Severity Observations

A hypothetical insurer reviews five ground-up claim severities, recorded in thousands of dollars: 80, 100, 120, 140, and 160. Treat the observations as an independent sample from a common severity distribution. Estimate the severity CV and the expected claim-count threshold for a two-sided 95% aggregate-loss standard with relative tolerance \(k=0.10\). Use \(z=1.96\).

Step 1: Calculate the sample mean. The total is \(80+100+120+140+160=600\) thousand dollars, so \(\bar X=600/5=120\) thousand dollars.

Step 2: Calculate the sample variance. The deviations from the mean are \(-40,-20,0,20,40\). Their squared deviations sum to \(1{,}600+400+0+400+1{,}600=4{,}000\), in squared thousands of dollars. Thus \(s^2=4{,}000/(5-1)=1{,}000\), and \(s=\sqrt{1{,}000}=31.6228\) thousand dollars.

Step 3: Estimate the CV. The units cancel in the ratio: \(\widehat{CV}=31.6228/120=0.263523\), approximately \(0.2635\).

Step 4: Calculate the threshold. Here \(z^2=1.96^2=3.8416\), \(\widehat{CV}^{\,2}=1{,}000/120^2=1{,}000/14{,}400=0.0694444\), and \(k^2=0.01\). Therefore,

$$ \widehat{\lambda}_{\mathrm{req}} =\frac{3.8416(1+0.0694444)}{0.01} \approx 410.84. $$

The estimated standard requires about 410.84 expected claims; if a whole-number operational threshold is required, round up to 411 claims. This is a plug-in result based on only five observations, so the arithmetic does not imply that the estimated CV is precise. The method avoids the common mistake of using \(s/\bar X\) without first calculating the sample variance with the \(n-1\) denominator.

Sampling Uncertainty and Threshold Sensitivity

A plug-in threshold treats \(\widehat{CV}\) as if it were known. In reality, a severity sample—especially a small sample—can give a CV that differs materially from the underlying portfolio value. Since \(CV\) is squared in the threshold, an upward revision to severity variation increases the required claim volume. The relationship is monotone: for fixed \(z\) and \(k\), a larger nonnegative \(CV\) gives a larger threshold.

One practical first assessment is sensitivity analysis: calculate the threshold for the point estimate and for one or more plausible alternative CV values. This does not produce a confidence interval; it shows how dependent the decision is on the severity assumption. If the estimate is used for a high-stakes standard, more data or a suitable statistical uncertainty analysis may be warranted.

Worked Example: Sensitivity to a Higher CV

A hypothetical book has a severity CV estimate of \(0.80\), but the analyst wants to assess a plausible value of \(1.00\) because the severity sample is limited. For both cases, use a two-sided 95% standard with \(z=1.96\) and \(k=0.10\). Calculate both thresholds and the increase.

At \(\widehat{CV}=0.80\). The squared CV is \(0.80^2=0.64\), so:

$$ \widehat{\lambda}_{\mathrm{req}} =\frac{1.96^2(1+0.80^2)}{0.10^2} =\frac{3.8416(1.64)}{0.01} =\frac{6.300224}{0.01} =630.0224. $$

At \(CV=1.00\). The squared CV is \(1.00\), giving:

$$ \lambda_{\mathrm{req}} =\frac{3.8416(1+1.00^2)}{0.01} =\frac{3.8416(2)}{0.01} =768.32. $$

The higher-CV scenario increases the threshold by \(768.32-630.0224=138.2976\) expected claims. The difference is consequential even though the confidence level and tolerance have not changed. This sensitivity calculation avoids treating the point estimate as certain; it does not establish that \(1.00\) is a statistically justified upper bound.

If a probability-based allowance for sampling error is needed, the method must match the severity distribution and sampling design. For example, under an independent normal sample with positive mean, a large-sample delta-method approximation for the standard error of \(s/\bar X\) is:

$$ \operatorname{SE}(\widehat{CV}) \approx \widehat{CV} \sqrt{\frac{1+2\widehat{CV}^{\,2}}{2n}}. $$

This approximation is not distribution-free. It follows by approximating the variance of \(s/\bar X\) using the normal-sample variance of \(s\) and \(\bar X\), then applying the first-order delta method to their ratio. Skewed or heavy-tailed claim severities can have substantially different sampling behavior. In those settings, an appropriate resampling or parametric method may be preferable, provided its assumptions are stated and checked.

Worked Example: An Approximate Upper-CV Sensitivity Check

Suppose a hypothetical sample has \(n=25\) independent severities, with \(\bar X=1{,}000\) dollars and \(s=300\) dollars. For an illustrative sampling check, assume the severity sample is reasonably modeled as normal and use a one-sided 95% normal multiplier of \(1.645\). Estimate an approximate upper CV value and compare the associated threshold with the point-estimate threshold for \(z=1.96\) and \(k=0.10\).

Step 1: Calculate the point estimate. \(\widehat{CV}=300/1{,}000=0.30\).

Step 2: Approximate its standard error. Since \(1+2(0.30)^2=1.18\) and \(2n=50\),

$$ \operatorname{SE}(\widehat{CV}) \approx0.30\sqrt{\frac{1.18}{50}} =0.30\sqrt{0.0236} \approx0.04609. $$

Step 3: Form the illustrative upper value. The approximate one-sided upper value is \(0.30+1.645(0.04609)\approx0.37581\).

Step 4: Compare thresholds. At the point estimate, the threshold is \(3.8416(1+0.30^2)/0.01=3.8416(1.09)/0.01=418.7344\). At the approximate upper value, it is \(3.8416(1+0.37581^2)/0.01\approx438.4\). The calculation illustrates how sampling uncertainty can raise a threshold. The upper value is only an approximation under the stated normal-sample assumption; it should not be reported as an exact confidence limit for a skewed severity distribution.

Practical Checks and Common Traps

  • Using variance where CV is required. The threshold uses \(1+CV^2\), not \(1+s^2\). A variance has squared monetary units; a CV is dimensionless.
  • Using the wrong denominator for sample variance. Under the usual sample convention, divide the squared-deviation sum by \(n-1\), not \(n\).
  • Ignoring count variation. For the stated Poisson aggregate-loss model, the factor \(1+CV^2\) includes both severity variation and random claim-count variation. Substituting \(CV^2\) alone omits the count component.
  • Overstating precision. A threshold such as 410.84 is a model-based expected-claim threshold, not a claim that the estimated severity CV is accurate to four decimals.
  • Applying an inappropriate severity sample. Paid claims subject to deductibles or limits do not automatically represent ground-up severity. Use data consistent with the severity definition and coverage basis required by the calculation.
  • Confusing expected claims with exposure units. The formula returns expected claim count. A frequency assumption is needed to convert that count to exposure, and the frequency must use the same exposure basis as the portfolio.
Key takeaway: Estimate severity variation as \(s/\bar X\), substitute that dimensionless estimate into \(z^2(1+\widehat{CV}^{\,2})/k^2\), and then examine how plausible sampling uncertainty changes the threshold. The result is only as credible as the severity data and model assumptions supporting it.

Check Your Understanding

For each problem, use the stated assumptions and show the intermediate calculations before checking the final answer.

  1. A hypothetical sample of severities, in hundreds of dollars, is \(50,100,150\). Calculate the sample mean, sample variance using denominator \(n-1\), sample CV, and the expected-claim threshold for \(z=1.96\) and \(k=0.10\). (Answer: \(\bar X=100\), \(s^2=2{,}500\), \(\widehat{CV}=0.50\), threshold \(=480.20\) expected claims.)
  2. A severity sample has \(\bar X=\$200\) and \(s=\$100\). Using \(z=1.645\) and \(k=0.08\), calculate the plug-in expected-claim threshold. (Answer: \(\widehat{CV}=0.50\), threshold \(=528.52\) expected claims.)
  3. For \(z=1.96\) and \(k=0.12\), compare the thresholds at \(CV=0.60\) and \(CV=0.90\). How many additional expected claims does the higher CV require? (Answer: \(362.82\) and \(482.87\) expected claims, respectively; the increase is \(120.05\) expected claims.)
  4. A sample has \(n=16\), \(\bar X=\$500\), and \(s=\$200\). Calculate the plug-in threshold for \(z=1.96\) and \(k=0.10\). (Answer: \(\widehat{CV}=0.40\), threshold \(=445.63\) expected claims.)
  5. A hypothetical portfolio has claim frequency \(0.25\) claims per exposure unit and severity \(CV=0.80\). For \(z=1.645\) and \(k=0.10\), calculate the expected-claim threshold and the corresponding exposure threshold. (Answer: \(443.79\) expected claims and \(1{,}775.15\) exposure units; at whole-unit implementation, round up to \(1{,}776\) exposure units.)

Previous

Next

Back to Tutorials
Clinical Biostats