Integrated Empirical Estimation from Claims Data

Why Claims Data Must Be Read at More Than One Level

A claims file may contain one row per policy, one row per claim, or one row per policy-year. Those rows represent different observational units, so a calculation that is correct for one table can be wrong for another. For example, a claim count is not a count of policies, and the average of recorded claims is not automatically the average annual cost per policy.

An integrated empirical estimate keeps the units connected: exposure measures how much coverage was observed, claim records describe the frequency and size of losses, and policy-year aggregate amounts describe total loss over a defined coverage period. The estimates are descriptive of the experience recorded; they do not, by themselves, establish that future experience will be identical.

Definition: For a set of observations with total claim count $N$ and total exposure $E$, the empirical exposure-adjusted frequency is $N/E$ claims per exposure unit. If the complete claim amounts are $x_1,\ldots,x_N$, the empirical claim severity distribution assigns probability $1/N$ to each recorded claim amount, and its mean is $\bar{x}=(x_1+\cdots+x_N)/N$. The empirical aggregate loss per exposure unit is total recorded loss divided by $E$.

Exposure must be expressed in the intended unit, such as policy-years. A policy observed for half a year contributes $0.5$ policy-year, not one full policy-year. Severity is ordinarily calculated over claims, so each claim contributes one observation unless a different weighting scheme is explicitly specified. This makes it a claim-weighted summary: policies with more claims contribute more claim amounts.

From Claim Records to an Aggregate Estimate

Suppose the claim data are complete for the specified coverage and time period, losses are measured on a consistent basis, and each claim belongs to the exposure being summarized. The empirical estimate of aggregate loss per exposure unit can be written either as a direct total or as a product of frequency and mean severity.

\frac{\sum_{j=1}^{N}x_j}{E} = \frac{N}{E}\left(\frac{\sum_{j=1}^{N}x_j}{N}\right)

The equality follows because $N$ cancels. The left side is total loss divided by exposure. On the right, $N/E$ is claims per exposure unit and $\sum x_j/N$ is dollars per claim, so their product is dollars per exposure unit. This identity is an accounting decomposition, not a claim that frequency and severity are independent or that a particular probability model has been fitted.

Establish the exposure unit.
Sum the exposure for the eligible policies or policy-years, using consistent coverage dates and units.

Count and total the claims.
Use the claim records to obtain $N$ and the total loss $\sum x_j$. Check that each included claim is within the defined scope.

Calculate and reconcile.
Compute $N/E$, $\bar{x}$, and their product. Confirm that it equals total loss divided by exposure.

Worked Example: Reconciling Frequency, Severity, and Aggregate Loss

A hypothetical small portfolio has four policy-years of exposure. Its policy-year records show: Policy-year A has $0.5$ exposure and one claim of \$800; B has $1.0$ exposure and two claims of \$1,200 and \$2,000; C has $1.5$ exposure and no claims; and D has $1.0$ exposure and one claim of \$4,000. Calculate exposure-adjusted frequency, empirical mean severity, and aggregate loss per policy-year.

Method. First summarize exposure and claim count, then summarize the four claim amounts. Including the zero-claim policy-year in exposure is essential even though it contributes no claim severity.

Total exposure is $0.5+1.0+1.5+1.0=4.0$ policy-years. The claim count is $N=1+2+0+1=4$, so empirical frequency is $4/4.0=1.0$ claim per policy-year.

The total claim amount is $\$800+\$1,200+\$2,000+\$4,000=\$8,000$. The empirical mean severity is $\$8,000/4=\$2,000$ per claim. Thus, frequency times mean severity is $1.0\times \$2,000=\$2,000$ per policy-year.

As a check, direct aggregate loss per exposure is $\$8,000/4.0=\$2,000$ per policy-year, the same result. The empirical severity distribution puts probability $1/4$ on each of \$800, \$1,200, \$2,000, and \$4,000.

Common mistake avoided. Dividing \$8,000 by the three policy-years that had at least one claim would produce \$2,666.67 per exposed policy-year with a claim. That is a different conditional quantity, not the portfolio aggregate loss per policy-year. The denominator for the latter is all four policy-years of exposure.

Coverage-Modified Payments

An empirical estimate for gross losses is not automatically an estimate of insurer payments. For a per-payment deductible $d$ and a payment limit $u$ that caps the insurer’s payment after the deductible, the payment for a loss $X$ is

Y^P=\min\bigl((X-d)_+,u\bigr), \qquad (z)_+=\max(z,0).

This definition means the insured absorbs losses up to the deductible, and the insurer pays at most $u$ on any one claim. It is important to establish whether a stated limit caps the insurer’s payment or the total covered loss. The formula above uses a limit on the insurer’s payment; another contract definition can lead to a different calculation.

Worked Example: Applying a Deductible and Payment Limit

A hypothetical portfolio has four policy-years of exposure and four complete gross claims: \$400, \$900, \$1,800, and \$3,500. Each claim has a \$500 deductible and a \$2,000 limit on the insurer’s payment after the deductible. Find the empirical gross and modified aggregate loss per policy-year.

Method. Apply the contract formula to each claim separately. Then total the resulting payments and divide by the full exposure.

For the \$400 loss, $(400-500)_+=0$, so payment is $\min(0,2000)=\$0$. For \$900, payment is $\min(900-500,2000)=\$400$. For \$1,800, payment is $\min(1800-500,2000)=\$1,300$. For \$3,500, the amount above the deductible is \$3,000, so payment is capped at \$2,000.

The gross total is $\$400+\$900+\$1,800+\$3,500=\$6,600$. Gross loss per policy-year is $\$6,600/4=\$1,650$. Modified payments total $\$0+\$400+\$1,300+\$2,000=\$3,700$, giving $\$3,700/4=\$925$ per policy-year.

As a second check, three claims produce a positive payment. Their mean payment per payment is $\$3,700/3=\$1,233.33$ to the nearest cent. The modified payment frequency is $3/4=0.75$ payments per policy-year, and $0.75\times \$1,233.33$ is approximately \$925 per policy-year; using the unrounded mean gives exactly \$925.

Common mistake avoided. Subtracting the deductible from the portfolio total and then applying one limit would treat several separate claims as a single loss. The deductible and per-payment limit apply claim by claim, so the \$2,000 cap is applied separately to each claim.

Sampling Uncertainty and the Choice of Observation Unit

An empirical mean calculated from a sample is subject to sampling variation. For $n$ independent observations from a common population with finite variance, the estimated standard error of their sample mean is $s/\sqrt{n}$, where $s$ is the sample standard deviation computed with divisor $n-1$. This formula describes the variability of a mean under the stated sampling setup; it does not correct for missing claims, changing coverage, or a nonrepresentative sample.

The observation unit matters. If the target is an average policy-year aggregate, use policy-year aggregate amounts as the observations. If the target is average claim severity, use claim amounts, while recognizing that several claims from one policy or year may be dependent. Treating clustered claims as independent can understate uncertainty. Earlier material on standard errors and clustered resampling addresses those designs; the essential check here is to match the sampling unit to the quantity being estimated.

Worked Example: Estimating Uncertainty in Annual Aggregate Loss

A hypothetical set of five independent, equally exposed policy-years has aggregate losses of \$0, \$500, \$1,000, \$1,500, and \$2,000. Estimate the mean loss per policy-year and its standard error from these observations.

Method. Treat each policy-year aggregate as one observation. The five observations are in dollars per policy-year, so the sample mean and its standard error retain those units.

The mean is $(0+500+1,000+1,500+2,000)/5=\$1,000$. The deviations from the mean are $-1,000,-500,0,500,1,000$ dollars. Their squared sum is $1,000,000+250,000+0+250,000+1,000,000=2,500,000$ dollars squared.

The sample variance is $2,500,000/(5-1)=625,000$ dollars squared, and the sample standard deviation is $\sqrt{625,000}=\$790.57$ to the nearest cent. The estimated standard error of the mean is $\$790.5694/\sqrt{5}=\$353.55$, rounded to the nearest cent. The estimated mean is therefore \$1,000 per policy-year, with an estimated standard error of \$353.55 per policy-year.

Check. The standard error is less than the sample standard deviation because it divides the latter by $\sqrt{5}$. This small sample gives a potentially unstable estimate of uncertainty; the calculation is not, by itself, a guarantee that a normal-based interval will have accurate coverage.

Data Checks Before Interpreting the Result

The arithmetic is only as reliable as the definitions and records behind it. Before interpreting an estimate, specify the coverage, period, currency basis, and exposure unit. If claims from different valuation dates are combined, amounts may need consistent inflation treatment. If coverage changed, a claim observed under one deductible or limit may not represent payments under another. A missing or incomplete claim file can also change both the claim count and the observed severity distribution.

Several checks help catch errors before a result is reported:

Reconcile totals. Claim-level amounts should sum to the aggregate total for the same scope, subject to any explicitly documented adjustments.
Keep denominators distinct. Frequency uses exposure; severity uses claim count; aggregate loss per exposure uses exposure.
Check payment definitions. Do not mix gross losses, insurer payments, and amounts after recoveries in one severity calculation.
Inspect zero-claim exposures. They remain part of the denominator for an all-policy exposure-adjusted estimate.
State the scope of inference. A historical empirical average describes the observed data. Generalizing beyond those records requires assumptions about representativeness and stability.

Key takeaway: An integrated empirical estimate is credible only when claim counts, claim amounts, aggregate losses, and exposure refer to compatible populations and periods. Reconcile the direct aggregate-per-exposure calculation with frequency times mean severity, apply coverage terms at the claim level, and estimate uncertainty using an observation unit that matches the target.

Check Your Understanding

For each problem, identify the relevant units before calculating.

A portfolio has $0.5$, $1.5$, and $2.0$ policy-years of exposure and records 1, 3, and 2 claims, respectively. What is the empirical frequency per policy-year? (Answer: $6/4=1.5$ claims per policy-year.)
Four complete claim amounts are \$600, \$900, \$1,500, and \$3,000. Calculate empirical mean severity. (Answer: $\$6,000/4=\$1,500$ per claim.)
Four claims are \$300, \$700, \$2,000, and \$3,000. Each has a \$400 deductible and a \$1,600 limit on insurer payment after the deductible. What is the total modified payment? (Answer: $0+300+1,600+1,600=\$3,500$.)
Five independent one-policy-year observations have aggregate losses \$0, \$500, \$1,000, \$1,500, and \$2,000. Give the sample mean and estimated standard error of the mean, rounded to cents. (Answer: mean $\$1,000.00$; standard error $\$353.55$ per policy-year.)
There are 8 claims over 10 policy-years, with mean claim severity \$2,500. What is the empirical aggregate loss estimate per policy-year, and how can it be checked? (Answer: $(8/10)\times \$2,500=\$2,000$ per policy-year; equivalently, total claims of $\$20,000$ divided by 10 policy-years.)

What You'll Learn