Why Summarize Aggregate Loss Experience?
An observed aggregate total combines all covered losses in a defined period. A collection of such totals can describe how much aggregate loss was typical in the observed periods and how variable those totals were. These summaries can also serve as estimates for a comparable future period—but only if the periods being compared represent the same kind of risk and are reasonably modeled as independent observations from a common distribution.
The unit of observation matters. If each value is one calendar year's total covered loss, then the sample mean estimates the expected aggregate loss for a year with the same coverage and exposure basis. The data do not automatically estimate a policy-level amount, a claim-level severity, or a total for a different exposure base.
The two variance formulas answer related but distinct questions. The empirical distribution variance describes the spread of the finite set of observations when that set itself is treated as a distribution. The sample variance uses \(n-1\) in the denominator and, under independent and identically distributed observations with finite variance, is an unbiased estimator of the variance of the underlying period-total distribution.
Calculating and Interpreting Sample Moments
For an appropriate model, suppose the period totals \(S_1,\ldots,S_n\) are independent and identically distributed, with \(E[S_j]=\mu\) and \(\operatorname{Var}(S_j)=\sigma^2<\infty\). Then \(\bar S\) estimates \(\mu\), and \(s^2\) estimates \(\sigma^2\). Independence and a common distribution are substantive assumptions: they are not guaranteed merely because the observations appear in a table.
To calculate a sample variance reliably, first find the mean, then calculate every deviation from that mean, square each deviation, and sum the squares. The units are squared as well: if losses are recorded in thousands of dollars, the variance is in thousands of dollars squared. The standard deviation, if needed, is the square root of the variance and returns to the original units.
Worked Example: Mean and Variance of Annual Aggregate Totals
A hypothetical portfolio has the following five annual aggregate losses, recorded in thousands of dollars: \(120, 80, 100, 140, 60\). Treat the years as independent observations from a common annual-total distribution. Calculate the sample mean, sample variance, and empirical-distribution variance, then interpret them.
Method. Calculate the center first, then the squared deviations. Keeping the values in thousands of dollars makes the units of the variance explicit.
The total is \(120+80+100+140+60=500\), so \(n=5\) and \[ \bar S=\frac{500}{5}=100 \] thousand dollars.
The deviations from the mean are \(20,-20,0,40,-40\). Their squares are \(400,400,0,1{,}600,1{,}600\), with sum \(4{,}000\). Therefore, \[ s^2=\frac{4{,}000}{5-1}=1{,}000 \] in thousands of dollars squared. The empirical-distribution variance is \[ \frac{4{,}000}{5}=800 \] in thousands of dollars squared.
Interpretation and check. The observed mean is \(\$100{,}000\) per year; under the stated common-distribution model, it estimates the expected annual aggregate loss for a comparable year. The sample variance is \(1{,}000\) thousand-dollars squared. The empirical variance is smaller because its denominator is 5 rather than 4. As a check, the five deviations sum to zero, as deviations from their arithmetic mean must.
Common mistake avoided. Dividing the squared-deviation total by \(n\) and calling the result the sample variance confuses the empirical-distribution variance with the usual unbiased sample-variance estimator.
Empirical Summaries of the Upper Tail
The mean and variance do not show the entire shape of observed aggregate experience. A high percentile or a summary of the largest observations can help describe the upper end of a finite data set. These are descriptive summaries of the observed sample; they do not, by themselves, establish how often a similarly large total will occur in the future.
A sample upper-tail average can be defined directly for a chosen number \(k\): sort the totals from largest to smallest and average the largest \(k\). This answers a finite-sample question such as “What was the average of the two largest observed totals?” It is not automatically an estimate of a conditional expectation above a population quantile. The chosen \(k\), the sample size, and ties all matter when interpreting the result.
Worked Example: An Empirical Quantile and the Largest-Observation Average
A hypothetical insurer observes eight annual aggregate totals, in thousands of dollars: \(20,30,30,40,50,60,90,160\). Calculate the empirical 75th percentile and the average of the two largest totals. Explain what each result describes.
Method. Use the stated inverse empirical distribution convention for the percentile. For the upper-tail average, select exactly the two largest observations.
The data are already sorted, and \(n=8\). The percentile position is \[ \lceil np\rceil=\lceil 8(0.75)\rceil=\lceil 6\rceil=6. \] The sixth observation is \(60\), so \(Q_n(0.75)=60\) thousand dollars. The two largest observations are \(90\) and \(160\), so their average is \[ \frac{90+160}{2}=\frac{250}{2}=125 \] thousand dollars.
Interpretation and check. Under this quantile convention, at least 75% of the observed totals are at or below \(\$60{,}000\); six of the eight are. The average of the two largest observed totals is \(\$125{,}000\). The latter is a description of these two observations, not a claim that future totals above a particular threshold will average that amount.
Common mistake avoided. A percentile convention must be specified, particularly with a small sample. Here the inverse empirical distribution selects the sixth ordered value; interpolating between values would use a different convention.
Why the Sample Variance Uses \(n-1\)
The sample mean is calculated from the same observations whose deviations are being squared. This uses one degree of freedom: once \(n-1\) deviations from the sample mean are known, the last one is determined because all \(n\) deviations sum to zero. More formally, under the independent, identically distributed model with finite variance \(\sigma^2\), the denominator \(n-1\) makes the sample variance unbiased.
Write \(S_j-\bar S=(S_j-\mu)-(\bar S-\mu)\). Summing the squared deviations and expanding gives \(\sum_{j=1}^n(S_j-\bar S)^2=\sum_{j=1}^n(S_j-\mu)^2-n(\bar S-\mu)^2\), since \(\sum_{j=1}^n(S_j-\mu)=n(\bar S-\mu)\).
The first sum has expectation \(n\sigma^2\). Independence gives \(\operatorname{Var}(\bar S)=\sigma^2/n\), and \(E[\bar S]=\mu\), so \(E[n(\bar S-\mu)^2]=n\operatorname{Var}(\bar S)=\sigma^2\).
Thus \(E[\sum_{j=1}^n(S_j-\bar S)^2]=n\sigma^2-\sigma^2=(n-1)\sigma^2\), and consequently \(E[s^2]=\sigma^2\). The argument requires independent observations from a common distribution with finite variance; it does not establish unbiasedness when those assumptions fail.
This derivation explains the purpose of the denominator; it does not make a small historical sample precise. A sample variance can be strongly affected by a single unusually large aggregate total, especially when only a few periods are observed.
Zeros, Extreme Totals, and Comparability
A zero-loss period is still an observed period if the data cover the stated portfolio and time interval completely. Removing zeroes changes the question from the unconditional aggregate loss per period to an amount conditional on a positive total. Similarly, deleting a large total because it seems unusual will generally lower the observed mean and variance. An unusual observation should first be checked for data or coverage errors, not discarded simply because it is influential.
Historical totals are informative about a future period only when the basis is comparable. The covered population, period length, policy terms, and exposure should be consistent with the interpretation being made. If those conditions change, the raw sample mean is not automatically an estimate for the changed situation. The calculations themselves cannot correct for changes in exposure, inflation, coverage, or data recording.
Worked Example: Including Zero-Loss Years in the Estimate
A hypothetical small book records annual aggregate losses, in thousands of dollars, of \(0,0,0,0,500\) over five complete years. Assume the years are comparable independent observations from a common annual-total distribution. Calculate the sample mean and sample variance, and explain why the zeroes remain in the calculation.
Method. Use all five annual totals because the target is aggregate loss per year, including years with no loss.
The sum is \(500\), so the sample mean is \(500/5=100\) thousand dollars. The deviations from 100 are \(-100,-100,-100,-100,400\). Their squared values sum to \[ 4(100^2)+400^2=4(10{,}000)+160{,}000=200{,}000 \] in thousands of dollars squared. Therefore, \[ s^2=\frac{200{,}000}{5-1}=50{,}000 \] in thousands of dollars squared.
Interpretation and check. The observed average is \(\$100{,}000\) per year, despite four years having no loss; that is the relevant sample mean for an unconditional annual aggregate total. The deviations sum to \(-400+400=0\), and the squared-deviation total is \(200{,}000\), confirming the variance calculation.
Common mistake avoided. Averaging only the positive year would give \(\$500{,}000\), which describes the average among these positive-loss years—not the average annual aggregate loss across all five observed years.
Practical Interpretation
Empirical moments and tail summaries are useful because they reduce a set of aggregate observations to interpretable quantities. The sample mean is a natural estimate of the expected total for a comparable period under a common-distribution model. The sample variance describes observed period-to-period dispersion and estimates the underlying variance under the assumptions stated above. A percentile and a largest-observation average add information about the upper part of the observed sample.
None of these summaries guarantees that a future period will resemble the sample average, or that the largest observed amount is a reliable estimate of an extreme future total. Aggregate losses can be skewed, and a short record may contain very few large-loss years. Keep the period, currency scale, denominator, and interpretation attached to every reported result. In particular, state whether a variance is a sample variance or an empirical-distribution variance.
Check Your Understanding
For each problem, treat the listed totals as observations in the units stated. Use the sample-variance denominator \(n-1\), and use the inverse empirical distribution convention for empirical quantiles.
- Four comparable annual aggregate totals, in thousands of dollars, are \(25,35,40,60\). Find the sample mean and sample variance. (Answer: mean \(40\) thousand dollars; sample variance \(216.67\) thousand-dollars squared.)
- Six annual totals, in thousands of dollars, are \(10,20,20,30,50,110\). Find the empirical median and the average of the two largest observations. (Answer: empirical median \(20\) thousand dollars; largest-two average \(80\) thousand dollars.)
- Four comparable annual totals, in thousands of dollars, are \(0,40,80,120\). Calculate the mean and sample variance, including the zero-loss year. (Answer: mean \(60\) thousand dollars; sample variance \(2{,}666.67\) thousand-dollars squared.)
- A data set contains annual totals for five years, but the fifth year covers twice as many policies as each other year. Can the unadjusted mean of the five totals automatically be interpreted as an estimate for a future year with the smaller exposure? (Answer: No; the totals are not on a comparable exposure basis, so the unadjusted mean does not automatically estimate the smaller-exposure year's expected total.)
- Five comparable annual totals, in thousands of dollars, are \(10,10,10,10,210\). Find both the empirical-distribution variance and the sample variance. (Answer: empirical-distribution variance \(6{,}400\) thousand-dollars squared; sample variance \(8{,}000\) thousand-dollars squared.)