Why the Resampling Basis Matters
A reserve estimate is calculated from data and assumptions, so it can vary if the underlying experience or the sample used to estimate it were different. A bootstrap estimates this variability by repeatedly generating a new data set, applying the reserve method to that data set, and examining the resulting reserve estimates. The key choice is how to generate each new data set.
A parametric bootstrap generates observations from a specified probability model, usually using parameters estimated from the observed experience. A nonparametric bootstrap samples observed data units with replacement, without imposing a particular parametric distribution. In reserving, those units might be individual losses, comparable development observations, or appropriately constructed residuals. The two approaches can give different variability estimates because they use different resampling bases.
The reserve calculation must be repeated consistently in every replicate. If the original procedure estimates a parameter and then uses that parameter to calculate a reserve, the replicate procedure ordinarily estimates the parameter again from the bootstrap data before calculating \(R^*\). Holding the estimate fixed answers a different, conditional question and can omit parameter-estimation variability.
Two Bootstrap Approaches
| Feature | Parametric bootstrap | Nonparametric bootstrap |
|---|---|---|
| Resampling basis | Observations generated from a specified fitted distribution | Observed data units sampled with replacement |
| What shapes the results? | The assumed distribution and fitted or specified parameters | The observed values and their empirical frequencies |
| Potential advantage | Can generate outcomes beyond the observed values, consistent with the model | Requires fewer distributional assumptions |
| Potential limitation | Results can be sensitive to model misspecification | Cannot directly generate values outside the empirical support |
For a development triangle, a nonparametric method often resamples suitable experience units or model residuals and reconstructs the data needed for projection. A parametric method instead simulates from a fitted model for the relevant experience, then refits and projects. The precise resampling unit matters: resampling individual cells as if independent may be inappropriate when the data structure induces dependence. The resampling scheme must match the data and the uncertainty question.
Bootstrap variability is not a guarantee that the assumed model is correct. Nor does a large number of replicates repair an unrepresentative sample or a poor resampling design. More replicates reduce simulation noise in the estimated bootstrap summaries; they do not remove uncertainty arising from the choice of model or data basis.
Why the Nonparametric Bootstrap Standard Error Has This Form
For \(n\) observed values \(x_1,\ldots,x_n\), a nonparametric bootstrap draw selects each value with probability \(1/n\). The conditional distribution of one draw therefore has mean \(\bar x\) and variance \(n^{-1}\sum_{j=1}^{n}(x_j-\bar x)^2\). A bootstrap sample mean averages \(n\) independent draws from this empirical distribution, so its variance is the variance of one draw divided by \(n\).
This uses the variance of the empirical distribution, whose denominator is \(n\), not the usual unbiased sample-variance denominator \(n-1\). Equivalently, if \(s^2=(n-1)^{-1}\sum(x_j-\bar x)^2\), then the conditional bootstrap variance of the mean is \((n-1)s^2/n^2\). For a reserve that is a linear transformation of the mean, the variance scales by the square of the multiplier.
Worked Example: A Parametric Bootstrap for a Bernoulli Estimate
A hypothetical reserve estimate is \(8\widehat p\) thousand dollars, where \(\widehat p\) is the success proportion from five independent Bernoulli observations. Three observations are successes. Under a parametric bootstrap, generate five new observations independently from a Bernoulli distribution with fitted probability \(\widehat p\). Find the mean and standard error of the bootstrap reserve. Assume no discounting; the reserve is expressed in thousands of dollars.
Method. The fitted probability is \(\widehat p=3/5=0.6\). Thus the number of bootstrap successes satisfies \(K^*\sim\operatorname{Binomial}(5,0.6)\), and the replicate success proportion is \(K^*/5\). Therefore \(R^*=8K^*/5=1.6K^*\) thousand dollars.
Mean. Since \(E[K^*]=5(0.6)=3\),
The bootstrap mean reserve is \(4.8\) thousand dollars.
Standard error. Since \(\operatorname{Var}(K^*)=5(0.6)(0.4)=1.2\),
To three decimal places, the standard error is \(1.753\) thousand dollars. As a check, the original plug-in estimate is \(8(0.6)=4.8\) thousand dollars, equal to the bootstrap mean. A common mistake is to use \(8\) as the multiplier on the count \(K^*\); the reserve is \(8\) times the proportion, so the multiplier on the count is \(8/5\).
Worked Example: Comparing Poisson and Empirical Resampling
Four hypothetical comparable cells have observed claim counts \(0,1,1,2\). For a simple one-period projection, estimate the future count per comparable cell by the average count and multiply it by a fixed claim cost of \(3\) thousand dollars. Find the bootstrap mean and standard error of the reserve under (a) a parametric Poisson bootstrap with fitted rate and (b) a nonparametric bootstrap of the four observed counts. Treat cells as independent and equally weighted; there is no discounting.
The observed average count is \((0+1+1+2)/4=1\). The reserve estimator is \(R=3\bar X\) thousand dollars, so the original estimate is \(3(1)=3\) thousand dollars.
(a) Parametric method. Fit \(X\sim\operatorname{Poisson}(\lambda)\) by the observed mean, giving \(\widehat\lambda=1\). Generate four independent counts \(X_1^*,\ldots,X_4^*\sim\operatorname{Poisson}(1)\), and calculate \(R^*=3\bar X^*\). The replicate mean count has expectation \(1\) and variance \(1/4\), because each count has variance \(1\). Hence
The parametric bootstrap mean is \(3.00\) thousand dollars and its standard error is \(1.500\) thousand dollars.
(b) Nonparametric method. Each bootstrap draw selects one of \(0,1,1,2\) with equal probability. The empirical mean is \(1\), and the variance of one draw from this empirical distribution is
The mean of four independent draws therefore has variance \(0.5/4=0.125\). Thus
The nonparametric bootstrap mean is \(3.00\) thousand dollars and its standard error is \(1.061\) thousand dollars. Both methods preserve the estimated mean in this example, but they do not give the same variability: the fitted Poisson model permits counts beyond the observed range, while the empirical distribution only resamples \(0,1,2\). The conclusion is specific to these assumptions and data, not a general rule that one method always produces greater variability.
Worked Example: Bootstrap Variability in a Development-Factor Projection
A hypothetical portfolio has a cumulative amount of \(200\) thousand dollars at an earlier development age. Three comparable observed age-to-age factors are \(1.10,1.30,1.50\). For this simplified projection, use the mean factor \(\bar F\) and define the reserve estimate as \(R=200(\bar F-1)\) thousand dollars. Compare the nonparametric bootstrap standard error with a parametric bootstrap that generates three independent factors from \(N(1.30,0.04)\), then re-estimates the mean factor in each replicate. Treat the normal model as an explicit approximation that permits factors below one; assume no discounting.
The observed mean factor is \((1.10+1.30+1.50)/3=1.30\), giving \(R=200(1.30-1)=60\) thousand dollars. The common factor \(200\) is the earlier cumulative amount, so a change in the estimated factor changes the reserve by \(200\) times that change.
Nonparametric bootstrap. The empirical variance of one resampled factor is
A bootstrap mean of three factors has variance \(0.0266667/3=0.0088889\). Multiplying by \(200\) to convert the mean factor to reserve gives
The nonparametric bootstrap standard error is \(18.856\) thousand dollars.
Parametric bootstrap. Under the specified normal model, a replicate mean of three factors has mean \(1.30\) and variance \(0.04/3\). Thus the reserve mean is \(200(1.30-1)=60\) thousand dollars, and
The two means agree because both procedures center the factor estimate at \(1.30\); the standard errors differ because the assumed parametric variance, \(0.04\), exceeds the empirical variance, \(0.0266667\). A common mistake is to multiply the factor’s standard error by \(200\) but forget that the reserve uses \(\bar F-1\); subtracting the fixed value \(1\) changes the mean, not the variance. The calculations also show why the parametric assumption must be stated and assessed, rather than treated as a neutral computational choice.
Interpreting Bootstrap Results in Reserving
A bootstrap standard error summarizes variability under a particular model, data set, and resampling rule. A parametric bootstrap can reflect variation in estimated parameters if each replicate is generated and refitted as specified. Its results remain conditional on the chosen model family and any fixed assumptions, such as a selected development structure. A nonparametric bootstrap reflects the observed empirical distribution; with sparse experience, that distribution may represent only a few distinct outcomes.
In practice, reserve uncertainty is often summarized by the bootstrap mean, standard deviation, and selected quantiles of \(R^*\). These are summaries of the bootstrap distribution, not statements that the true reserve must fall within a particular range. When only a finite number \(B\) of replicates is simulated, the reported summaries themselves have Monte Carlo error. Increasing \(B\) can stabilize those summaries, but it does not change the underlying resampling basis or resolve model misspecification.
Before comparing results, verify that the methods use the same reserve definition, projection horizon, exposure basis, and treatment of fixed assumptions. Otherwise, a difference in bootstrap variability may reflect inconsistent calculations rather than the distinction between parametric and nonparametric resampling.
- Use parametric resampling when a defensible fitted distribution is available and generating outcomes beyond the observed values is important.
- Use nonparametric resampling when observed experience is a suitable empirical basis and the data contain enough relevant variation.
- Compare the resampling bases when assumptions materially affect reserve variability; do not select a method solely because its standard error is smaller.
Check Your Understanding
For each problem, state the resampling basis and calculate the requested reserve summary. Unless stated otherwise, reserves are undiscounted and reported in thousands of dollars.
- A reserve estimate is \(12\widehat p\), where \(\widehat p\) is based on ten independent Bernoulli observations, four of which are successes. A parametric bootstrap uses the fitted probability. Find the bootstrap mean and standard error. (Answer: Mean \(4.800\); standard error \(\sqrt{3.456}\approx1.859\) thousand dollars.)
- Three comparable observations are counts \(1,2,3\). A nonparametric bootstrap resamples three counts and estimates reserve as twice the bootstrap sample mean. Find the bootstrap mean and standard error. (Answer: Mean \(4.000\); standard error \(\sqrt{8/9}\approx0.943\) thousand dollars.)
- Five independent cells are modeled as Poisson with fitted rate \(1.5\). In each parametric bootstrap replicate, estimate the average count and multiply it by \(4\). Find the bootstrap reserve mean and standard error. (Answer: Mean \(6.000\); standard error \(\sqrt{4.8}\approx2.191\) thousand dollars.)
- Four observed development factors are \(0.9,1.1,1.2,1.2\). A nonparametric bootstrap resamples four factors, and reserve is \(100(\bar F^*-1)\). Find the bootstrap mean and standard error. (Answer: Mean \(10.000\); standard error \(100\sqrt{0.00375}\approx6.124\) thousand dollars.)
- Two methods give the same bootstrap reserve mean, but one uses a fitted Poisson model and the other resamples observed counts \(0,1,1,2\). Explain why their standard errors can differ and name one limitation of the nonparametric method. (Answer: The Poisson model and empirical distribution can have different variances and ranges; the empirical bootstrap cannot generate counts outside the observed values.)