Why Bootstrap Aggregate Estimates?
An aggregate loss estimate depends on the experience observed in a finite set of periods. A large claim, an unusually quiet year, or a small number of observed periods can materially change the estimated mean or upper tail. A bootstrap uses the observed data as a resampling population: repeatedly draw samples of the same size, with replacement, and recompute the estimate. The resulting replicates describe how sensitive the estimate is to sampling variation under the assumption that the observed units are representative and exchangeable.
The key decision is what one resampled observation represents. If the data contain complete annual aggregate totals, resample whole periods. If they contain period claim counts and a separate list of claim severities, resample those records in a way that preserves the frequency-severity structure. Resampling individual claims alone cannot represent periods with zero claims and generally does not preserve the observed frequency variation.
Choose the Resampling Unit
For complete, consistently defined policy-period aggregate losses, a period is the observation. Resample the observed totals as whole units, including zero totals. This retains the observed mixture of zero-claim and positive-claim periods and directly supports bootstrap estimates of aggregate means and empirical tail summaries.
If records instead provide period counts and individual claim severities, the aggregate mean can be estimated from the product of the mean count and mean severity, provided the frequency and severity samples represent the same target population and the frequency-severity relationship is appropriately treated as independent. Resample the period counts—including zeros—and claim severities separately to quantify uncertainty in that product. For an aggregate distribution or its tail, use the resampled frequency and severity distributions to construct period totals. In each constructed period, draw exactly as many severities as the simulated count.
| Available data | Resampling unit | Important condition |
|---|---|---|
| Complete aggregate totals by period | Whole period total | Periods must be comparable and exchangeable for the target estimate |
| Counts by period and a claim-severity sample | Counts and severities resampled separately | Use the count sample, including zero counts; state the frequency-severity assumptions |
| Incomplete, censored, or changing-coverage records | A simple resampling scheme may not be suitable | Resampling cannot by itself repair censoring, truncation, or inconsistent coverage |
Bootstrap Aggregate Means
Suppose \(m\) comparable periods have observed total losses \(T_1,\ldots,T_m\). The empirical mean aggregate loss per period is \(\bar{T}=m^{-1}\sum_{j=1}^m T_j\). In each bootstrap replicate, draw \(m\) period totals with replacement and calculate their mean. With \(B\) replicates, the bootstrap standard error is the sample standard deviation of the \(B\) replicate means. A percentile interval can be formed from the chosen lower and upper quantiles of those means.
Worked Example: Bootstrap Standard Error from Period Totals
A hypothetical benefit program has complete totals of $0 and $200 for two comparable periods. Enumerate all ordered bootstrap samples of size two, with replacement, and find the bootstrap standard error of the mean total.
Method. Resample the two complete period totals. There are \(2^2=4\) equally likely ordered samples: \((0,0)\), \((0,200)\), \((200,0)\), and \((200,200)\). Their bootstrap means are, respectively, \(0\), \(100\), \(100\), and \(200\) dollars.
The average of the four replicate means is \((0+100+100+200)/4=100\) dollars. Their variance, using the bootstrap distribution probabilities, is \(\{(0-100)^2+(100-100)^2+(100-100)^2+(200-100)^2\}/4=5{,}000\) dollars squared. Thus the bootstrap standard error is \(\sqrt{5{,}000}=\$70.71\), rounded to the nearest cent.
Check. The original sample mean is \((0+200)/2=\$100\), matching the mean of the bootstrap means. Resampling only the positive claim would discard the observed zero-claim period and would not represent the period-total data. The calculation assumes the two periods are comparable; two observations provide a very limited basis for estimating uncertainty.
Frequency and Severity Records
Let \(N_j\) be the claim count in period \(j\), and let \(X\) denote a positive claim severity. For \(m\) observed periods and \(n\) observed claims, the empirical estimates are \(\bar{N}=m^{-1}\sum_{j=1}^m N_j\) and \(\bar{X}=n^{-1}\sum_{k=1}^n X_k\). Under a model in which count and severity are independent, the estimated mean aggregate loss per period is \(\bar{N}\bar{X}\). Bootstrap uncertainty in this estimate can be assessed by independently resampling \(m\) counts and \(n\) severities, then multiplying their respective sample means in each replicate.
For tail estimates, the product of means is not enough: it describes a mean, not an aggregate distribution. In an outer bootstrap replicate, resample the count records and severity records to define bootstrap empirical distributions. Then generate synthetic periods: draw a count from the resampled count distribution, and draw that number of severities, with replacement, from the resampled severity pool. Sum those severities to obtain the synthetic period total. Repeating this process yields an aggregate distribution for that outer replicate. Repeating the outer bootstrap yields a distribution of the resulting aggregate quantile or TVaR estimates. This procedure has two levels of randomness, so the number of synthetic periods should be large enough to make inner simulation noise small relative to the outer-bootstrap variation.
Here \(N^{*}\) is a simulated period count drawn from the bootstrap count distribution. Conditional on \(N^{*}=k\), the sum contains exactly \(k\) severity draws; if \(k=0\), the total is zero. The resampled severity pool used to supply these draws may have a fixed size, but the number of draws for a particular synthetic period is determined by that period’s count.
Worked Example: Resampling Counts and Severities
A hypothetical portfolio has observed period counts \(0,1,2\) and positive claim severities of $50, $100, and $150. In one bootstrap replicate, the resampled counts are \(0,0,2\), and the resampled severity pool is $50, $150, and $150. Calculate the replicate estimate of mean aggregate loss using the product-of-means method. Then construct one synthetic set of period totals using two severity draws for the period with two claims.
Step 1: Resampled count mean. The counts sum to \(0+0+2=2\) across three periods, so their mean is \(2/3\) claims per period.
Step 2: Resampled severity mean. The severities sum to \(50+150+150=\$350\) across three claims, so their mean is \(350/3=\$116.666\ldots\).
Step 3: Replicate aggregate-mean estimate. Multiply the two means: \((2/3)(350/3)=700/9=\$77.777\ldots\), or $77.78 to the nearest cent.
Step 4: Construct period totals for a tail calculation. The resampled counts require \(0+0+2=2\) severity draws in total. Draw exactly two values from the resampled severity pool; for this construction, suppose they are $50 and $150. The three synthetic period totals are therefore \(0\), \(0\), and \(50+150=\$200\).
Check and common mistake. The constructed totals have mean \((0+0+200)/3=\$66.67\). That need not equal the product-of-means estimate in this particular draw: the product uses a fixed-size resampled severity sample to estimate a product of means, whereas the synthetic totals use severity draws in the number required by the resampled counts. It would be incorrect to assign all three resampled severities to these periods: the resampled counts call for only two draws. Conversely, a count resample summing to eight would require eight severity draws for the synthetic totals, not three.
Bootstrap Tail Summaries
A bootstrap tail estimate is calculated by applying the same tail definition to each replicate. For example, define the empirical \(p\)-quantile as the observation at rank \(\lceil mp\rceil\) in the ordered sample of \(m\) totals. For a nonnegative aggregate total \(T\), define \(\operatorname{TVaR}_p(T)\) as the average of the quantile function over the upper \(1-p\) fraction: \(\operatorname{TVaR}_p(T)=(1-p)^{-1}\int_p^1 \operatorname{VaR}_q(T)\,dq\). For an empirical distribution with discrete masses, this integral handles the probability mass at the quantile boundary fractionally. Use the same definitions in every replicate, then examine the replicate distribution or report a percentile interval.
Worked Example: Bootstrap Variation in an Aggregate Tail
Four hypothetical complete period totals, in dollars, are \(0,100,200,300\). For illustration, use the empirical nearest-rank quantile with rank \(\lceil 4p\rceil\), set \(p=0.75\), and consider these five specified bootstrap samples: \((0,0,100,200)\), \((100,200,200,300)\), \((0,100,300,300)\), \((0,200,200,300)\), and \((100,100,200,300)\). Calculate each replicate’s \(\operatorname{VaR}_{0.75}\) and \(\operatorname{TVaR}_{0.75}\).
The nearest-rank position is \(\lceil4(0.75)\rceil=3\), so \(\operatorname{VaR}_{0.75}\) is the third ordered total. For four equally weighted observations, the upper quarter of the empirical quantile function consists of the largest observation, so \(\operatorname{TVaR}_{0.75}\) is the maximum.
The first sample orders as \(0,0,100,200\): its third value is $100 and its maximum is $200. The second orders as \(100,200,200,300\): its third value is $200 and its maximum is $300. The third orders as \(0,100,300,300\): its third value is $300 and its maximum is $300. The fourth orders as \(0,200,200,300\): its third value is $200 and its maximum is $300. The fifth orders as \(100,100,200,300\): its third value is $200 and its maximum is $300.
Thus the five replicate quantiles are \($100,$200,$300,$200,$200\), and the five replicate TVaRs are \($200,$300,$300,$300,$300\). Their averages are $200 and $280, respectively. These five specified replicates illustrate how to calculate the summaries; they are too few to support a reliable interval. With many bootstrap replicates, quantiles of the replicate estimates can summarize uncertainty. The small original sample also means that resampling cannot produce losses above the observed maximum of $300.
Assumptions, Limitations, and Exam Traps
A simple bootstrap is most defensible when the sampled units are comparable and exchangeable for the target population. Period totals should use consistent exposure, coverage, and valuation. The separate count-and-severity approach requires counts that include zero-claim periods and a severity sample representative of the claims in those periods. Treating frequency and severity as independent is an assumption, not a consequence of resampling. If counts and severities are associated, independently resampling them can break that relationship and misstate aggregate uncertainty.
Likewise, a bootstrap does not turn censored claim amounts into complete losses, restore missing zero-claim periods, or make periods with changing coverage comparable. A resampled sample also cannot contain values outside its observed support. Thus estimates of extreme tails can be unstable, especially with few periods or claims; a narrow-looking interval does not establish that the unobserved tail is well estimated.
- Wrong unit: resampling claims when the available observations are period totals loses the period-level structure and can omit zero periods.
- Dropping zeros: resampling only positive counts overstates frequency when zero-count periods belong in the exposure base.
- Mismatched number of severity draws: for synthetic aggregate totals, draw as many severities as the simulated counts require, not a fixed number regardless of count.
- Confusing a mean with a tail: \(\bar{N}\bar{X}\) estimates a mean under its assumptions; it does not by itself determine a quantile or TVaR.
- Overinterpreting a bootstrap interval: resampling describes variation under the observed sample and assumptions; it does not remove bias or create evidence beyond the data.
Check Your Understanding
For each problem, identify the appropriate resampling calculation and check the data structure before interpreting the result.
- A hypothetical insurer records complete totals of $40, $80, and $120 for three comparable periods. One bootstrap sample is $80, $40, and $80. Find its mean. (Answer: $66.67, to the nearest cent.)
- Four periods have counts \(0,1,1,2\). A bootstrap resample of counts is \(0,0,2,2\). Find the resampled mean count and explain why resampling only the four positive claims would not represent this frequency sample. (Answer: Mean count \(=1\); resampling claims alone omits the two zero-count periods.)
- The observed period counts are \(1,0,2\), and the observed severities are $20, $40, and $60. In a replicate, counts are \(0,2,1\), and resampled severities are $20, $60, and $60. Find the product-of-means estimate of aggregate loss per period. (Answer: $46.67, to the nearest cent.)
- For the count and severity resample in question 3, how many severity draws are needed to construct the three synthetic period totals? If the draws for the two-claim period are $20 and $60, and the one-claim period draws $60, what are the totals in the order \(0,2,1\)? (Answer: Three draws; totals are $0, $80, and $60.)
- Four complete period totals are \(0,50,150,250\). Using the nearest-rank definition with rank \(\lceil4p\rceil\), find the empirical \(\operatorname{VaR}_{0.75}\) and \(\operatorname{TVaR}_{0.75}\). (Answer: \(\operatorname{VaR}_{0.75}=\$150\); \(\operatorname{TVaR}_{0.75}=\$250\).)