Why Bootstrap a Loss Development Triangle?
A reserve estimate from a loss development triangle is a point estimate. It does not, by itself, show how much the estimate might vary if the observed experience had been different. A bootstrap uses the observed triangle, or deviations from a fitted triangle, to generate alternative triangles. Refitting and projecting each alternative gives a collection of reserve estimates that describes variability under the bootstrap assumptions.
The central design decision is what to resample. Resampling individual cells treats them as separate observations. Resampling accident-period blocks keeps the selected cells from each period together. These choices imply different dependence assumptions, and therefore can produce different reserve variability. The resampling scheme must match the structure the analyst intends to preserve.
Let \(X_{ij}\) denote the incremental loss for accident year \(i\) and development age \(j\). The cumulative loss through age \(j\) is \(C_{ij}=\sum_{k=1}^{j}X_{ik}\). In a conventional chain-ladder projection, an age-to-age factor is calculated from available cumulative observations, then applied to the latest cumulative amount for each accident year. A bootstrap adds a resampling step and repeats that fit-and-project process.
Two Resampling Choices
A residual bootstrap first fits a model to the observed triangle. It then resamples deviations between observed increments and fitted increments, reconstructing pseudo-increments around the fitted values. This keeps the fitted development pattern as the starting point while introducing variation based on observed deviations.
For example, under an additive model with fitted increment \(\hat X_{ij}\), an ordinary residual is \(e_{ij}=X_{ij}-\hat X_{ij}\). If the residuals are resampled without centering, the resampled values may have a nonzero mean and systematically shift the fitted level. A common preparation is to subtract their sample mean before resampling. If the reconstructed increment \(\hat X_{ij}+e^*_{ij}\) would be negative, the analyst needs a stated treatment, such as using a suitable positive-valued model or a resampling rule designed to prevent negative increments. Silently retaining impossible negative losses is not appropriate.
When variability scales with the fitted mean, a standardized residual may be more suitable. Under the variance assumption \(\operatorname{Var}(X_{ij}\mid\text{fitted information})=\phi\hat X_{ij}\), where \(\phi\) is a dispersion parameter, the Pearson residual is:
A resampled standardized residual \(r^*_{ij}\) is converted back to the increment scale using \(X^*_{ij}=\hat X_{ij}+\sqrt{\phi\hat X_{ij}}r^*_{ij}\). The dispersion term belongs in the denominator and in the reconstruction. This model-based standardization is not a claim that every triangle follows a particular probability distribution; the variance model and residual treatment must be justified for the application.
An incremental-loss bootstrap instead resamples observed increments, or suitable groups of increments, directly. This can preserve the empirical shape of observed losses, but raw increments from different development ages need not be exchangeable: their means and distributions can vary by age. Resampling them indiscriminately can erase the development pattern or place an early-age amount in a context where it is not comparable.
| Approach | What is resampled | Key modeling concern |
|---|---|---|
| Residual bootstrap | Deviations from fitted increments, often standardized | Fit quality, residual scaling, centering, and treatment of negative reconstructions |
| Incremental-loss bootstrap | Observed increments, individually or in groups | Whether the resampled increments are comparable across the ages or periods being combined |
| Accident-period block bootstrap | A vector or block of experience belonging to one accident period | Whether blocks are comparable and whether their dependence structure is reasonably preserved |
Worked Example: A Chain-Ladder Reserve Before Resampling
Worked Example: A Chain-Ladder Reserve Before Resampling
A hypothetical insurer has the following incremental losses, in thousands of dollars. The triangle is assumed to have a stable development pattern for this illustration. Calculate chain-ladder age-to-age factors and the reserve for the two immature accident years.
The observed rows are: AY 1: \(100,60,40\); AY 2: \(120,70\); AY 3: \(90\). The cumulative rows are AY 1: \(100,160,200\); AY 2: \(120,190\); AY 3: \(90\).
Step 1: Calculate the age 1 to age 2 factor. The volume-weighted factor is the sum of observed cumulative losses at age 2 divided by the sum at age 1:
Step 2: Calculate the age 2 to age 3 factor. Only AY 1 has an observed cumulative amount at age 3, so:
Step 3: Project AY 2. Its age 2 cumulative loss is \(190\), so its projected ultimate is \(190(1.25)=237.5\). Its reserve is \(237.5-190=47.5\) thousand dollars.
Step 4: Project AY 3. Its age 1 cumulative loss is \(90\). Its projected ultimate is \(90(1.590909)(1.25)=178.9773\) thousand dollars, so its reserve is \(178.9773-90=88.9773\) thousand dollars.
The total projected reserve is \(47.5+88.9773=136.4773\) thousand dollars, or $136,477.27 to the nearest cent. The calculation uses no interest or mortality assumption; neither is relevant to this loss-development projection. The result is checked by summing each immature period’s projected ultimate less its latest cumulative amount.
Reconstructing a Residual Bootstrap Replicate
After fitting the original triangle, a residual bootstrap samples prepared residuals and reconstructs pseudo-increments. The model must specify how fitted means and residual scales are obtained. The reconstructed triangle is then refitted: keeping the original development factors fixed would omit the variation caused by refitting the development pattern.
Choose and fit a development model, and calculate its fitted increments and residuals. State any variance model used to standardize residuals.
Decide whether residuals are resampled within development age, across comparable cells, or in accident-period blocks. The choice states which features of the observed structure are treated as exchangeable.
Resample the prepared residuals and add them back on the appropriate fitted scale, or resample the specified incremental-loss units. Apply the stated rule for inadmissible values.
Recalculate the development selections from the pseudo-triangle, project its outstanding losses, and record the resulting reserve.
Repeat the complete procedure to obtain a bootstrap distribution of reserve estimates. Its spread describes variability under the selected model and resampling assumptions.
Worked Example: One Residual Bootstrap Replicate
Worked Example: One Residual Bootstrap Replicate
Use the triangle from the preceding example and an additive residual bootstrap in thousands of dollars. For this small illustration, use ordinary residuals and resample from the centered observed residuals across the available cells. This assumes that these additive residuals are suitable for pooling; a real application would need to assess that assumption and the risk of negative reconstructed increments.
Step 1: Calculate fitted increments and residuals. At age 2, fitted increments are the age 1 cumulative amounts multiplied by \(\hat f_{12}-1=0.590909\). Thus AY 1 has fitted increment \(100(0.590909)=59.0909\) and residual \(60-59.0909=0.9091\). AY 2 has fitted increment \(120(0.590909)=70.9091\) and residual \(70-70.9091=-0.9091\). At age 3, the fitted AY 1 increment is \(160(1.25-1)=40\), giving residual \(40-40=0\).
The three residuals sum to zero, so their sample mean is zero. For this replicate, resample the residual \(+0.9091\) for both observed age 2 cells and the residual \(0\) for the observed age 3 cell.
Step 2: Reconstruct the observed cells. The pseudo-increments are AY 1: \(100,60,40\), and AY 2: \(120,70.9091\). AY 3 remains \(90\). The pseudo cumulative values used for refitting are AY 1: \(100,160,200\); AY 2: \(120,190.9091\); AY 3: \(90\).
Step 3: Refit the factors. The pseudo age 1 to age 2 factor is:
The pseudo age 2 to age 3 factor remains \(200/160=1.25\), because AY 1 is the only observed period at that transition.
Step 4: Project and calculate the replicate reserve. AY 2 has reserve \(190.9091(1.25-1)=47.7273\) thousand dollars. AY 3 has projected ultimate \(90(1.595041)(1.25)=179.4421\) thousand dollars and reserve \(179.4421-90=89.4421\) thousand dollars. The replicate total is \(47.7273+89.4421=137.1694\) thousand dollars, or $137,169.42 to the nearest cent.
The replicate exceeds the original point estimate of $136,477.27 by $692.15. That difference is one bootstrap outcome, not a reserve adjustment by itself. The common mistake this example avoids is resampling residuals and then leaving the original development factors unchanged; the factors are refitted to the reconstructed triangle.
Why the Resampling Unit Matters
Accident periods can share conditions that affect several development ages at once. If such within-period dependence is relevant, sampling each cell independently can break the observed association between that period’s increments. A block bootstrap resamples a vector of related residuals together. This does not guarantee that the blocks are representative, but it preserves the pairing within each sampled block.
The reason development factors are refitted after resampling can also be seen directly. For a transition from age \(j\) to age \(j+1\), the volume-weighted chain-ladder factor is the ratio of the sum of observed cumulative amounts at the later age to the sum at the earlier age. It is a weighted average of the individual ratios \(C_{i,j+1}/C_{ij}\), with weights proportional to \(C_{ij}\):
Thus, changing the pseudo cumulative amounts changes the weights, the individual ratios, or both. Recalculating the factor lets the bootstrap replicate reflect this fitting variation. This identity assumes the denominators \(C_{ij}\) are positive for the included observations.
Worked Example: Resampling Accident-Period Residual Blocks
A hypothetical portfolio has three comparable accident periods. For each, a fitted model gives two future incremental estimates, in thousands of dollars, and the observed residual vector for those two development ages. The fitted vectors are AY A: \((30,20)\), AY B: \((40,25)\), and AY C: \((50,30)\). Their residual vectors are respectively \((2,-1)\), \((-1,2)\), and \((-1,-1)\). All values are in thousands of dollars, and the residuals are assumed suitable for additive reconstruction.
For one block-bootstrap replicate, draw three period blocks with replacement: A, C, and C. Apply each selected block’s two residuals to the corresponding fitted future vector.
For the first target period, the reconstructed future amounts are \((30+2,20-1)=(32,19)\), totaling \(51\). For the second, they are \((40-1,25-1)=(39,24)\), totaling \(63\). For the third, they are \((50-1,30-1)=(49,29)\), totaling \(78\).
The replicate total is \(51+63+78=192\) thousand dollars, or $192,000. As a check, the unadjusted fitted total is \((30+20)+(40+25)+(50+30)=195\) thousand dollars. The selected residual-block totals are \(2-1=1\), \(-1-1=-2\), and \(-1-1=-2\), so the net adjustment is \(1-2-2=-3\) thousand dollars and the reconstructed total is \(195-3=192\) thousand dollars.
The block is the pair of development-age residuals for a period, not a single residual. This preserves the observed pairing in each drawn block. It does not establish that all accident periods are interchangeable; that comparability is an assumption. The example also illustrates the common trap of pairing residuals from different periods independently when the intended method is a period-block bootstrap.
Assumptions, Interpretation, and Exam Traps
A bootstrap distribution is conditional on its construction choices. It does not automatically include every source of uncertainty. For example, the fitted development model, selected tail treatment, calendar effects, and residual dependence may all affect the results. A bootstrap that holds a feature fixed cannot measure uncertainty in that feature. The process of refitting captures some estimation variability, but only for the parts of the model that are actually refitted.
- Check comparability. Residuals or increments should not be pooled across development ages or periods without a reason to treat them as comparable.
- Preserve intended dependence. If the method assumes accident-period blocks, keep the block together. If it assumes independent cells, state that assumption rather than implying the data prove independence.
- Keep scales consistent. A standardized residual must be reconstructed using the same fitted standard deviation used to define it. Under the variance model \(\phi\hat X_{ij}\), this is \(\sqrt{\phi\hat X_{ij}}\), not merely \(\sqrt{\hat X_{ij}}\).
- Refit each replicate. Reusing the original development factors after changing the triangle misses variation from factor estimation.
- Handle sparse development ages cautiously. A transition represented by only one observed accident period provides little empirical information for resampling that age-specific residual distribution.
- Distinguish a replicate from a final summary. One simulated reserve is one outcome. Measures such as the bootstrap mean or a percentile require a collection of replicates and do not remove model risk.
For a professional application, the analyst should state the fitted model, residual definition, centering or scaling adjustment, resampling unit, treatment of inadmissible pseudo-increments, refitting procedure, and which features are held fixed. These details determine what reserve variability the bootstrap can reasonably represent. This treatment is aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.
Check Your Understanding
Work each problem through the stated resampling or projection steps before checking the final answer.
- A triangle has AY 1 cumulative amounts \(80,120,150\) and AY 2 cumulative amounts \(100,?\) at ages 1 and 2. The observed AY 2 age 2 amount is \(150\). Calculate the volume-weighted age 1 to age 2 factor and the fitted AY 2 age 2 cumulative amount under that factor. (Answer: factor \(270/180=1.5000\); fitted cumulative amount \(150\).)
- A fitted increment is \(64\), an observed increment is \(70\), and the variance model is \(\operatorname{Var}(X\mid\text{fit})=\phi\hat X\) with \(\phi=2.25\). Calculate the Pearson residual. (Answer: \(6/\sqrt{144}=0.5000\).)
- A fitted future reserve is \(420\) thousand dollars. A resampled additive residual block has components \(+8\) and \(-5\) thousand dollars. Reconstruct the reserve using this block and state the result in dollars. (Answer: \(420+8-5=423\) thousand dollars, or \($423,000\).)
- Two accident-period residual blocks are \((3,-2)\) and \((-1,4)\), in thousands of dollars. A replicate samples the first block twice. What is the total adjustment across the two sampled periods, and why should the two residuals within each draw remain paired in a block bootstrap? (Answer: each block totals \(1\), so the adjustment is \(2\) thousand dollars; pairing preserves the within-period association represented by each vector.)
- A bootstrap reconstructs a pseudo-triangle using resampled residuals but retains the original age-to-age factors for every projection. Identify the omitted source of variation. (Answer: variation from refitting the development factors to the pseudo-triangle.)