Why Bootstrap a Capital Quantile?
Operational-loss capital is often expressed as a high quantile of an annual aggregate loss distribution. That quantile is estimated from finite data and a fitted model, so the reported capital estimate is itself uncertain. A bootstrap estimates how much the estimate might vary across repeated samples generated under an empirical or fitted representation of the data.
The object of interest is the uncertainty in the estimated quantile, not the range of possible annual losses. For example, a 99th-percentile annual loss quantile is a point on a modeled loss distribution. A bootstrap interval around that estimate describes variation in the estimated point across resampled data sets. It is not an interval that will contain 99% of future annual losses.
The choice of what counts as an observation matters. If the data consist of annual aggregate losses, resampling years preserves each year's total as the sampling unit. If the model separately represents annual event counts and event severities, a parametric bootstrap instead simulates the components specified by that model, then refits its parameters and recalculates the annual-loss quantile. Resampling individual events as if they were annual totals would change the question being answered.
Percentile Intervals and Their Interpretation
Let \(\widehat{q}_p\) be the estimated \(p\)-quantile from the original data. After generating \(B\) bootstrap data sets and recalculating the estimate in each one, write the results as \(\widehat{q}_p^{*(1)},\ldots,\widehat{q}_p^{*(B)}\). A percentile interval uses quantiles of these \(B\) bootstrap estimates as its endpoints. A nominal 95% percentile interval uses the 2.5th and 97.5th percentiles.
For all hand-calculated examples below, use the nearest-rank convention: sort \(B\) values, and define the empirical \(r\)-quantile to be the observation at rank \(\lceil Br\rceil\), with ranks starting at 1. This convention makes the calculations reproducible for small examples. In practice, software may use an interpolated quantile convention; state the convention and use it consistently.
The bootstrap distribution is an approximation, not a guarantee of exact coverage. In particular, a small number of replicates gives coarse, unstable interval endpoints. Tail quantiles also depend heavily on the tail information supplied by the data or model. A large bootstrap count reduces Monte Carlo noise in the estimated endpoints, but it cannot create information absent from the input data or correct an unsuitable model.
Nonparametric Bootstrap of Observed Annual Losses
In a nonparametric bootstrap, the observed values define the empirical distribution from which samples are drawn with replacement. This approach avoids specifying a parametric distribution, but it cannot generate losses beyond the observed values. It is therefore especially important to interpret a nonparametric interval cautiously when the target is an extreme quantile and the observed sample is short.
Worked Example: Resampling Annual Aggregate Losses
A hypothetical insurer has observed annual aggregate operational losses of 2, 4, and 8 thousand dollars in three years. Estimate the 80th-percentile annual loss using the nearest-rank sample quantile, then find a 50% percentile bootstrap interval by enumerating all size-three nonparametric bootstrap samples.
Method. Treat each complete annual total as one observation. The original ordered data are \(2,4,8\). The rank for the 80th percentile is \(\lceil 3(0.80)\rceil=\lceil2.4\rceil=3\), so the estimate is the third observation: \(\widehat{q}_{0.80}=8\) thousand dollars.
Each of the three draws in a bootstrap sample can be any of the three observed annual totals. There are \(3^3=27\) equally likely ordered samples. The bootstrap 80th-percentile estimate is the maximum of the three sampled values, because its nearest-rank index is again 3.
The maximum is 2 only when all three draws are 2, which occurs once. The maximum is 4 when all draws are from \(\{2,4\}\), except the all-2 sample; there are \(2^3-1^3=7\) such samples. The remaining \(27-1-7=19\) samples have maximum 8. Thus the bootstrap estimate distribution is:
For a 50% percentile interval, use the 25th and 75th percentiles. In the ordered list of 27 estimates, ranks 1 through 1 contain 2, ranks 2 through 8 contain 4, and ranks 9 through 27 contain 8. The endpoint ranks are \(\lceil27(0.25)\rceil=7\) and \(\lceil27(0.75)\rceil=21\), giving endpoints 4 and 8 thousand dollars.
Result and check. The estimate is 8 thousand dollars, and the 50% percentile bootstrap interval is \([4,8]\) thousand dollars. The upper endpoint cannot exceed the largest observed loss because every bootstrap observation was drawn from the three observed annual totals. That is a property of this nonparametric procedure, not evidence that larger losses are impossible. The small data set also makes this interval a teaching calculation, not a reliable capital uncertainty assessment.
Parametric Bootstrap: Simulate, Refit, Recalculate
A parametric bootstrap starts with an assumed distribution and parameter estimates fitted to the observed data. Each replicate is simulated from that fitted model. The model is then refitted to the simulated data, and the target quantile is recalculated using the replicate's fitted parameters. This sequence approximates the sampling variation in the estimation procedure, conditional on the assumed model being appropriate.
Estimate the model parameters from the observed operational-loss data and calculate the original quantile estimate.
Generate a data set of the same relevant size from the fitted model, using the model's stated observation unit.
Estimate parameters again from that replicate and calculate the same target quantile using the refitted model.
Repeat the procedure \(B\) times and use the empirical distribution of the replicate quantiles to form an interval.
Refitting is essential if the interval is intended to reflect uncertainty from estimating parameters. If the original fitted parameters are held fixed in every replicate, the procedure instead measures variation under a fixed model, or simulation noise in the calculation, depending on what is simulated. It does not capture how parameter estimates would change with new data.
Worked Example: Parametric Bootstrap for Exponential Severity
A hypothetical operational-loss severity sample, measured in thousands of dollars, is \(1,2,3,4\). Assume independent observations from an exponential distribution with mean parameter \(\theta\), so \(F(x;\theta)=1-e^{-x/\theta}\) for \(x\geq0\). Estimate the 95th-percentile severity and form a 50% percentile interval from four specified parametric bootstrap replicates.
Original fit. For an exponential model, the maximum likelihood estimate of the mean is the sample mean. Thus \(\widehat{\theta}=(1+2+3+4)/4=10/4=2.5\) thousand dollars. Solving \(1-e^{-q/\theta}=0.95\) gives \(e^{-q/\theta}=0.05\), so \(q=-\theta\ln(0.05)\). Since \(-\ln(0.05)=2.995732\), the original estimate is \(2.5(2.995732)=7.4893\) thousand dollars.
Bootstrap replicates. Simulate four severities from the fitted exponential model for each replicate, then refit the mean parameter as the replicate sample mean. For a compact hand calculation, suppose the four simulated data sets are \((1,2,3,4)\), \((1,1,2,2)\), \((2,3,4,5)\), and \((1,3,3,5)\), all in thousands of dollars. These are illustrative simulated outcomes under the fitted model.
Their refitted mean parameters are, respectively, \((1+2+3+4)/4=2.5\), \((1+1+2+2)/4=1.5\), \((2+3+4+5)/4=3.5\), and \((1+3+3+5)/4=3.0\). In each replicate, multiply the refitted mean by 2.995732 to obtain its 95th-percentile severity estimate. The four results are \(7.4893\), \(4.4936\), \(10.4851\), and \(8.9872\) thousand dollars.
Interval and check. In sorted order, the estimates are \(4.4936, 7.4893, 8.9872, 10.4851\). With \(B=4\), the nearest-rank 25th and 75th percentiles are at ranks \(\lceil4(0.25)\rceil=1\) and \(\lceil4(0.75)\rceil=3\). The 50% percentile interval is therefore \([4.4936,8.9872]\) thousand dollars. The original estimate, 7.4893, lies within this interval. Four replicates give only a coarse illustration; an applied analysis needs many more replicates and a suitable interval level.
Bootstrapping an Annual Capital Quantile
For an annual aggregate model, the bootstrap must follow the model's frequency and severity structure. A parametric replicate typically simulates a year's event count and then the associated event severities, as specified by the fitted model. Refitting the frequency and severity parameters to each replicate allows their estimation variability to affect the recalculated annual aggregate quantile. The quantile calculation can use an appropriate compound-distribution or simulation method, but the bootstrap question remains: what is simulated, what is refitted, and what estimate is recorded?
Worked Example: Frequency-Parameter Uncertainty in Annual Capital
A hypothetical operational-loss model has annual event count \(N\sim\operatorname{Poisson}(\lambda)\), and every event causes a fixed loss of 10 thousand dollars. The observed counts over four years are \(0,1,1,2\). Estimate the 95th-percentile annual aggregate loss, then calculate a 50% percentile bootstrap interval from four parametric frequency replicates. Treat the fixed event severity as known, so only the frequency parameter is refitted.
Original fit. The Poisson maximum likelihood estimate is the mean count: \(\widehat{\lambda}=(0+1+1+2)/4=1\). For a discrete distribution, define the 95th percentile as the smallest count \(k\) with \(P(N\leq k)\geq0.95\). At \(\lambda=1\), \(P(N\leq2)=0.9197\), while \(P(N\leq3)=0.9810\). Thus the 95th-percentile count is 3, and annual aggregate capital is \(3(10)=30\) thousand dollars.
Replicate data and refits. Suppose four data sets simulated from \(\operatorname{Poisson}(1)\), each with four annual counts, are \((0,1,1,1)\), \((0,1,1,2)\), \((0,0,1,1)\), and \((1,2,2,3)\). Their refitted Poisson means are \(0.75,1,0.5,\) and \(2\), respectively.
For \(\lambda=0.75\), \(P(N\leq1)=0.8266\) and \(P(N\leq2)=0.9595\), so the 95th-percentile count is 2 and capital is 20 thousand dollars. For \(\lambda=1\), the count quantile is 3, giving 30 thousand dollars. For \(\lambda=0.5\), \(P(N\leq1)=0.9098\) and \(P(N\leq2)=0.9856\), so capital is 20 thousand dollars. For \(\lambda=2\), \(P(N\leq4)=0.9473\) and \(P(N\leq5)=0.9834\), so the quantile count is 5 and capital is 50 thousand dollars.
Interval and check. The sorted replicate capital estimates are \(20,20,30,50\) thousand dollars. The nearest-rank 25th percentile is the first value, 20, and the 75th percentile is the third value, 30. The 50% percentile interval is \([20,30]\) thousand dollars. Each endpoint is a multiple of the fixed 10-thousand-dollar event severity, as required by this model. The interval reflects frequency-parameter estimation uncertainty under the assumed Poisson model; it does not include uncertainty about the fixed severity or the model form.
Common Traps and Limits
A common mistake is to simulate bootstrap data sets but keep the original fitted parameter in every replicate, then describe the resulting spread as parameter-estimation uncertainty. For the exponential example, that incorrect procedure would assign every replicate the same 95th-percentile estimate, \(2.5(2.995732)=7.4893\) thousand dollars, if it simply recalculated the quantile from the fixed original mean. The corrected procedure refits the mean in each replicate; the four resulting estimates were \(7.4893,4.4936,10.4851,\) and \(8.9872\) thousand dollars. Holding parameters fixed omits the variation in the fitted mean.
Other traps include using the wrong sampling unit, treating bootstrap replicates as independent real-world observations, and interpreting a percentile interval as a guarantee about future losses. A bootstrap interval also does not automatically account for model selection, scenario uncertainty, or structural uncertainty. If the fitted model is wrong, a parametric bootstrap can give a precise-looking interval around an unreliable estimate. These qualifications are particularly consequential for high operational-loss quantiles, where limited tail data can make estimates sensitive to modeling choices.
Check Your Understanding
Use the nearest-rank convention unless a question specifies otherwise. Amounts are in thousands of dollars where indicated.
- Observed annual aggregate losses are \(3,5,9\). For the empirical 75th-percentile estimate, bootstrap all size-three samples with replacement. What is the 50% percentile interval? (Answer: \([5,9]\) thousand dollars.)
- An exponential severity model has fitted mean 4 thousand dollars. Using \(q_p=-\theta\ln(1-p)\), calculate its 90th-percentile severity estimate. (Answer: \(9.2103\) thousand dollars.)
- Annual event counts over four years are \(1,0,2,1\). Assume a Poisson frequency model and a known fixed severity of 5 thousand dollars per event. Find the fitted 90th-percentile annual aggregate loss. (Answer: \(10\) thousand dollars.)
- A data set contains one annual total for each observed year. For a nonparametric bootstrap intended to estimate uncertainty in an annual-loss quantile, what is resampled, and why is resampling individual event records instead a different procedure? (Answer: Resample complete annual totals, because the year is the observed unit for that data set; resampling event records changes the unit and does not preserve the observed annual aggregates.)
- A parametric bootstrap simulates new data from a fitted severity model. Which additional step is needed in each replicate to reflect parameter-estimation uncertainty in the severity quantile? (Answer: Refit the severity-model parameters to that replicate, then recalculate the target quantile using the refitted parameters.)