Interpreting Uncertainty in Frequency-Severity Estimates

Why Component Uncertainty Matters

A frequency-severity estimate combines two quantities: an expected claim count and an average cost per claim. A change in either quantity changes the estimated total, but the two components need not be known with equal precision. If the count estimate is stable while severity is uncertain, refining the count may do little to improve the total estimate. Conversely, a volatile count estimate may dominate even when average severity is well supported.

The central calculation is simple multiplication. Interpreting its uncertainty requires more care: the size of each component, its variability, and any relationship between the component estimates all matter. This tutorial develops a practical framework for measuring those effects and for communicating what a sensitivity result does—and does not—say.

Definition: Let $C$ be the expected number of claims in a specified portfolio and period, and let $M=E[X]$ be the expected severity per claim, where $X$ is the loss amount for one claim. The frequency-severity estimate of expected aggregate loss is $\hat L=\hat C\hat M$, where $\hat C$ and $\hat M$ are estimates of $C$ and $M$. The target period, exposure basis, loss definition, and dollar basis must be consistent between the components.

The product represents expected aggregate loss under the usual frequency-severity model: claim severities are identically distributed with mean $M$, and the expected claim count is $C$. Conditional on the count, expected aggregate loss is the count times $M$; averaging over the count gives $E[\text{aggregate loss}]=E[N]M=CM$. This expectation identity does not require count and severity to be independent if each claim has the same expected severity conditional on the count. Independence is needed for some variance results below.

An important distinction is between uncertainty in the estimate and process variability in future losses. Estimate uncertainty concerns how much $\hat C$ and $\hat M$ might change if the experience or assumptions used to estimate them changed. Process variability concerns the randomness of future claim counts and claim amounts even if the underlying parameters were known. A range for the estimated expected loss is not automatically a prediction interval for actual losses.

How Uncertainty Propagates Through the Product

Treat $\hat C$ and $\hat M$ as random quantities representing estimation uncertainty. If their means are $C$ and $M$, their variances are $\sigma_C^2$ and $\sigma_M^2$, and the estimates are independent, then the product has an exact variance that includes a small interaction term.

\operatorname{Var}(\hat C\hat M) = M^2\sigma_C^2+C^2\sigma_M^2+\sigma_C^2\sigma_M^2

The first term is the contribution from count uncertainty, scaled by the square of average severity. The second is the contribution from severity uncertainty, scaled by the square of expected count. The final term arises because both estimates vary at once. If the component estimates are reasonably precise, this interaction term is often small, and a first-order approximation is useful.

\operatorname{Var}(\hat C\hat M) \mathrel{\approx} M^2\sigma_C^2+C^2\sigma_M^2

When the estimates may be related, a first-order approximation includes their covariance:

\operatorname{Var}(\hat C\hat M) \mathrel{\approx} M^2\sigma_C^2+C^2\sigma_M^2+2CM\operatorname{Cov}(\hat C,\hat M)

Positive covariance increases the approximate variance: scenarios that raise the count estimate also tend to raise severity. Negative covariance reduces it. Shared data, common development choices, or related trend assumptions can create dependence, so independence should be justified rather than presumed.

Why the Formula Holds

For independent estimates, $E[\hat C\hat M]=E[\hat C]E[\hat M]=CM$. Independence also gives $E[(\hat C\hat M)^2]=E[\hat C^2]E[\hat M^2]$. Using $E[\hat C^2]=C^2+\sigma_C^2$ and $E[\hat M^2]=M^2+\sigma_M^2$, subtract the squared mean $C^2M^2$:

\begin{aligned} \operatorname{Var}(\hat C\hat M) &=(C^2+\sigma_C^2)(M^2+\sigma_M^2)-C^2M^2\\ &=M^2\sigma_C^2+C^2\sigma_M^2+\sigma_C^2\sigma_M^2. \end{aligned}

This exact result applies to independent estimates with finite variances. It does not establish that the product itself is normally distributed. A normal interval based on the product’s standard deviation is an approximation, most defensible when relative uncertainties are moderate and the product distribution is not strongly skewed.

Worked Examples

Worked Example: Sensitivity to Separate Component Changes

A hypothetical specialty insurer estimates 500 claims for a target period and an average severity of $2,400 per claim. Assess the effect of an 8% change in count and a 5% change in severity, separately and together. This is a deterministic sensitivity calculation: no probability distribution is assigned. There is no discounting or interest adjustment, and mortality is not applicable.

Base estimate. The estimated aggregate loss is $500\times \$2{,}400=\$1{,}200{,}000$.

Count change only. An 8% increase in count gives $500(1.08)=540$ claims. Holding severity fixed, the estimate is $540\times \$2{,}400=\$1{,}296{,}000$, an increase of $96,000. An 8% decrease gives $500(0.92)=460$ claims and $460\times \$2{,}400=\$1{,}104{,}000$, a decrease of $96,000.

Severity change only. A 5% increase gives $\$2{,}400(1.05)=\$2{,}520$. The estimate becomes $500\times \$2{,}520=\$1{,}260{,}000$, an increase of $60,000. A 5% decrease gives $\$2{,}400(0.95)=\$2{,}280$, producing $500\times \$2{,}280=\$1{,}140{,}000$, a decrease of $60,000.

Both increase. The combined estimate is $540\times \$2{,}520=\$1{,}360{,}800$, or $160,800 above the base. The combined percentage increase is $1.08\times1.05-1=0.134$, or 13.4%, not 13%. The additional 0.4 percentage point is the interaction between the changes.

Check. Both decreases give $460\times \$2{,}280=\$1{,}048{,}800$, 12.6% below the base, consistent with $0.92\times0.95=0.874$. The count change has the larger dollar effect in this scenario: $96,000 versus $60,000. This is a sensitivity comparison, not a statement about the probability of either change.

Method and trap avoided. This is one-at-a-time and combined sensitivity analysis. Adding 8% and 5% to claim a 13% joint increase misses the multiplicative interaction; multiplying the changed components gives the correct 13.4%.

Worked Example: Estimating Uncertainty in the Product

For a hypothetical portfolio, suppose the count estimate and severity estimate are independent, with $\hat C\sim N(100,10^2)$ claims and $\hat M\sim N(3.0,0.3^2)$ thousand dollars per claim. These are normal approximations to estimation uncertainty; their standard deviations are 10 claims and $0.3 thousand, respectively. There is no discounting or interest adjustment, and mortality is not applicable.

Point estimate. The estimated aggregate loss is $100\times3.0=300$ thousand dollars, or $300,000.

Variance contributions. In squared thousands of dollars, the count contribution is $3.0^2(10^2)=9(100)=900$. The severity contribution is $100^2(0.3^2)=10{,}000(0.09)=900$. The interaction term is $10^2(0.3^2)=100(0.09)=9$. Thus the exact product variance under the independent-estimate model is $900+900+9=1{,}809$ squared thousands of dollars.

Standard deviation and interval. The product standard deviation is $\sqrt{1{,}809}=42.532$ thousand dollars. A normal approximation using this standard deviation gives a 95% interval of $300\pm1.96(42.532)$ thousand dollars. Since $1.96(42.532)=83.36$, the approximate interval is $216,640 to $383,360, rounded to the nearest $10.

Check. The independent-estimate product has expected value $100(3.0)=300$ thousand dollars. The interval is centered on that value, and its width reflects uncertainty in both inputs. It is an approximation for the estimate, not a 95% prediction interval for actual claims.

Method and trap avoided. This uses the exact variance of a product of independent estimates, followed by a normal approximation. Adding the two standard deviations directly would be incorrect; variances are combined in the formula, and the interaction term is included.

Worked Example: Identifying the Main Uncertainty Driver

A hypothetical program has an estimated 200 claims and mean severity of $1,800. The independent estimate distributions are $\hat C\sim N(200,16^2)$ and $\hat M\sim N(1{,}800,270^2)$. Treat these as approximations to estimation uncertainty. There is no discounting or interest adjustment, and mortality is not applicable.

Point estimate and relative uncertainty. The estimated loss is $200\times\$1{,}800=\$360{,}000$. The count coefficient of variation is $16/200=8\%$; the severity coefficient of variation is $270/1{,}800=15\%$.

One-standard-deviation sensitivities. A 16-claim change alters the product by approximately $16\times\$1{,}800=\$28{,}800$. A $270 severity change alters it by approximately $200\times\$270=\$54{,}000$. The severity uncertainty therefore has the larger first-order dollar effect.

Variance check. The count contribution is $1{,}800^2(16^2)=3{,}240{,}000(256)=829{,}440{,}000$ dollars squared. The severity contribution is $200^2(270^2)=40{,}000(72{,}900)=2{,}916{,}000{,}000$ dollars squared. The interaction is $256(72{,}900)=18{,}662{,}400$ dollars squared. Total variance is $3,764,102,400 squared, giving a standard deviation of about $61,352 and a coefficient of variation of $61{,}352/360{,}000=17.04\%$.

Among the two first-order contributions, severity accounts for $2{,}916/(829.44+2{,}916)=77.85\%$. The interaction term is additional, rather than a contribution attributable solely to one component.

Check. The approximate relative standard deviation, $\sqrt{0.08^2+0.15^2}=17.00\%$, is close to the exact result of 17.04%; the difference comes from the interaction term. The conclusion is that severity is the larger uncertainty driver in this hypothetical estimate, not that severity is always more uncertain than frequency.

Method and trap avoided. Compare component effects on the product, using either dollar sensitivities or variance contributions. Comparing the raw standard deviations, 16 claims and $270, would be meaningless because they have different units.

Applying the Results Carefully

Sensitivity analysis is useful because it translates changes in component assumptions into a change in the estimated total. For small changes, the product’s approximate dollar change is:

\Delta L\mathrel{\approx}M\Delta C+C\Delta M

This follows by expanding $(C+\Delta C)(M+\Delta M)$: the exact change is $M\Delta C+C\Delta M+\Delta C\Delta M$. The approximation omits the last term, which is small when both changes are small. For larger or simultaneous changes, calculate the full product rather than relying on the approximation.

A useful way to compare uncertainty is by relative, rather than absolute, variation. Under the independent-estimate model, define $CV_C=\sigma_C/C$ and $CV_M=\sigma_M/M$. The exact relative variance is:

\frac{\operatorname{Var}(\hat C\hat M)}{(CM)^2} =CV_C^2+CV_M^2+CV_C^2CV_M^2

For small coefficients of variation, the product’s relative variance is approximately the sum of the two component relative variances. This makes a comparison of component contributions straightforward. It also highlights why a component with a modest absolute standard deviation can matter greatly if its mean is small, and why a large standard deviation in one unit cannot be compared directly with another component’s standard deviation.

Key takeaway: The frequency-severity point estimate is a product, but its uncertainty depends on both components’ relative precision and on their dependence. Use dollar sensitivity to describe the effect of plausible changes, and use variance propagation to quantify estimation uncertainty under stated assumptions.

Common Mistakes and Exam Traps

Adding percentage changes instead of multiplying. An 8% count increase and a 5% severity increase produce a 13.4% product increase, not 13%.
Adding standard deviations. Independent variances—not standard deviations—are combined in the product-variance formula. The covariance term must be considered if estimates are related.
Ignoring units or mismatched bases. Count and severity must refer to the same period, portfolio, claim definition, and dollar basis. A severity trend already reflected in one component should not be applied again to that same component when calculating sensitivity.
Confusing estimate uncertainty with actual-loss variability. A confidence-style interval for an estimated expected loss does not describe the full range of possible aggregate claims.
Overstating a normal interval. The product of two normal quantities is not generally normal. State when a normal interval is an approximation, and be cautious when relative uncertainty is large or estimates can be near zero.

Check Your Understanding

For each problem, identify the relevant uncertainty or sensitivity calculation before checking the answer.

A hypothetical portfolio has an estimated 80 claims and a mean severity of $1,250. What is the estimated aggregate loss, and how many dollars does a 5% increase in the count change the estimate if severity is unchanged? (Answer: $100,000; the increase is $5,000.)
Independent estimates satisfy $\hat C\sim N(50,5^2)$ and $\hat M\sim N(4{,}000,400^2)$. Find the product’s mean and exact variance under the independent-estimate model. (Answer: $200,000; variance $804{,}000{,}000$ dollars squared.)
Independent count and severity estimates have coefficients of variation 10% and 20%. Using the exact independent-product formula, what is the coefficient of variation of their product? (Answer: $\sqrt{0.0504}=22.45\%$, to two decimal places.)
Suppose $C=100$, $M=\$2{,}000$, $\operatorname{Var}(\hat C)=64$, $\operatorname{Var}(\hat M)=40{,}000$, and $\operatorname{Cov}(\hat C,\hat M)=-400$. Using the first-order formula, estimate the product variance. (Answer: $496{,}000{,}000$ dollars squared.)
For actual aggregate losses, suppose claim count $N$ is Poisson with mean 40 and independent severities $X$ have mean $2,000 and standard deviation $3,000. What is the variance of aggregate loss if these parameters are known? (Answer: $520{,}000{,}000$ dollars squared; this is process variance, not uncertainty in estimating the expected loss.)

What You'll Learn