Why Triangle Structure Matters
A loss triangle is not merely a collection of observed numbers. In a typical run-off triangle, each row represents an origin period, such as an accident year, and its entries represent development ages. Entries later in development are observed for older origin periods but are missing for newer ones. In addition, residuals from the same origin period can be related across development ages: a period with unexpectedly high early development may also have unexpectedly high later development.
A bootstrap that samples each cell independently can keep the marginal distribution of residuals while breaking this within-row relationship. A simulated triangle may then contain development patterns that were never represented by the observed experience. Because reserve projections use relationships across development ages, that can alter the distribution of projected reserves, not just the appearance of the simulated data.
The correct resampling unit depends on the dependence being modeled. Whole-row resampling retains the observed pairing of development-age residuals within an origin year. It does not, by itself, preserve dependence between different origin years or shared calendar-period effects. A block or other structure-aware scheme may be needed if those features matter.
What Independent Cell Resampling Changes
Consider two development ages and two possible residual vectors, \((1,1)\) and \((-1,-1)\), each occurring with probability one-half. The two components of each observed vector move together. If a bootstrap draws each development-age residual independently from the observed marginal values, it can also produce \((1,-1)\) or \((-1,1)\), although neither vector appeared in the data.
Under independent marginal draws, the cross-product expectation becomes \(E[e_1]E[e_2]=0\), so the bootstrap covariance is zero. The method has changed the joint behavior even though it has not changed either marginal distribution. In a reserve application, it can understate or overstate uncertainty depending on how the dependence affects the projection.
Row-Vector Resampling and Triangle Constraints
A row-vector bootstrap samples a whole observed residual vector for an origin year and carries its components together. To reconstruct incremental amounts, first fit the selected triangle model and calculate fitted values. Then resample residual vectors and add or otherwise combine them with the fitted values according to the model’s residual definition.
For example, under an additive residual reconstruction, let \(\hat{\mu}_{ij}\) be the fitted incremental value in cell \((i,j)\), and let \(s_{ij}\) be its residual scale. If the standardized residual is \(e_{ij}=(C_{ij}-\hat{\mu}_{ij})/s_{ij}\), a bootstrap increment is reconstructed as \(C^*_{ij}=\hat{\mu}_{ij}+s_{ij}e^*_{ij}\). This reconstruction is specific to that additive setup. A different fitted model may require a different scale or reconstruction, and an additive reconstruction can produce negative increments; it should not be used without checking that it is appropriate.
The triangle is ragged: rows have different numbers of observed development ages. Therefore, row-vector resampling needs a rule for compatible vectors. Possible choices include resampling among rows with the same observed length, using only a defined set of fully developed rows, or using a model-based method that handles the missing cells. Each choice has costs. For instance, restricting to fully developed rows can leave few donor rows and may not represent recent experience. Missing cells must not be treated as observed zero residuals merely to make vector lengths match.
A useful justification for keeping components together follows directly from covariance. For a resampled vector, the empirical covariance between two ages depends on the paired products \(e_{i,j}e_{i,k}\). Resampling the whole vector retains those observed pairs as possible bootstrap outcomes. Independent cell draws instead combine one component from one row with another component from a different row; their expected product is the product of the separate marginal means, not the empirical paired-product average.
The displayed expression uses denominator \(n\), as an empirical covariance for a finite set of rows. Using denominator \(n-1\) gives the usual unbiased sample covariance instead. Either way, preserving row pairings is the key distinction from independent cell resampling.
Worked Example: Seeing Dependence Lost by Cell Resampling
A hypothetical insurer has two observed origin-year residual vectors for two development ages: \(A=(1,1)\) and \(B=(-1,-1)\). Each vector is equally likely to be selected. Compare the possible outcomes under row-vector resampling and independent cell resampling, and calculate the cross-product average for each scheme. Use the average over the four equally likely outcomes in the independent scheme.
Method. Keep the two residual components paired for row-vector resampling. For independent cell resampling, draw each component separately from \(\{-1,1\}\), each value with probability one-half.
Row-vector scheme. The only outcomes are \((1,1)\) and \((-1,-1)\). Their cross-products are \(1(1)=1\) and \((-1)(-1)=1\). Thus the average cross-product is \((1+1)/2=1\).
Independent-cell scheme. The four equally likely outcomes are \((1,1)\), \((1,-1)\), \((-1,1)\), and \((-1,-1)\). Their cross-products are \(1,-1,-1,1\), so the average is \((1-1-1+1)/4=0\).
Both marginal distributions still give probability one-half to each of \(-1\) and \(1\). But the cross-development dependence differs: the row scheme retains the observed pairing, while independent cell resampling removes it in expectation. Common mistake avoided: treating agreement of the two marginal distributions as proof that the bootstrap preserves the triangle’s joint structure. It does not.
Reconstructing a Bootstrap Row
A dependence-preserving bootstrap still needs a coherent path from residuals to simulated increments and then to a reserve statistic. The following example uses an explicitly specified additive reconstruction and a supplied projection factor. It illustrates the mechanics; it does not establish that this residual model is appropriate for every triangle.
Worked Example: Reconstructing Increments and a Reserve
For a hypothetical origin year, fitted increments at development ages 1 and 2 are \((100,50)\), with residual scales \((10,5)\). A row-vector bootstrap selects the standardized residual vector \((1,-1)\). The projected cumulative amount at the end of age 2 is multiplied by a supplied factor of \(1.4\) to estimate ultimate loss. Calculate the bootstrap increments, the projected ultimate, and the reserve beyond age 2.
Method. Use the stated additive formula \(C^*_{j}=\hat{\mu}_{j}+s_j e^*_{j}\), apply the same selected residual vector to both ages, and then apply the given cumulative projection factor.
At age 1, \(C^*_1=100+10(1)=110\). At age 2, \(C^*_2=50+5(-1)=45\). The reconstructed cumulative amount is \(110+45=155\).
The projected ultimate is \(155(1.4)=217\). The reserve beyond the observed cumulative amount is \(217-155=62\).
Check. The projected ultimate equals the observed cumulative amount plus the reserve: \(155+62=217\). The age-1 residual raises its fitted increment by 10, while the age-2 residual lowers its fitted increment by 5. Common mistake avoided: resampling the two residuals separately and accidentally using a pair that was not selected as one row vector. The calculation uses the stated factor as given; estimating that factor and accounting for its uncertainty are separate modeling steps.
Choosing a Resampling Unit
A row-vector scheme is appropriate when the main dependence of interest is across development ages within origin years. It is not a universal solution. If several cells share a calendar-year shock, for example, those cells may lie on a diagonal rather than in one row. Resampling origin-year rows independently can break that diagonal association. Similarly, a row bootstrap does not automatically preserve dependence between adjacent origin years.
A block bootstrap can address some forms of local dependence by resampling groups of related observations together. The block should be defined according to the hypothesized structure: a block of adjacent origin years for serial dependence between origins, or a calendar-period grouping for shared calendar effects. A triangle may not permit every dependence dimension to be preserved simultaneously, especially when it is small or ragged. State which dependence the design aims to retain and what it leaves unmodeled.
| Resampling unit | Can preserve | Important limitation |
|---|---|---|
| Individual cells | Marginal residual behavior by cell or development age, if sampled accordingly | Usually breaks observed pairing across development ages |
| Whole origin-year residual vectors | Observed within-row pairing across development ages | Does not automatically preserve dependence across origin years or calendar periods |
| Blocks of related rows or periods | Selected local dependence represented by the chosen blocks | Requires a defensible block definition and enough data to populate it |
Worked Example: Selecting a Scheme for Calendar-Year Dependence
A hypothetical reserving team finds that residuals from cells sharing a calendar year tend to move together, even when the cells belong to different origin years. It also wants to retain the observed relationship between development ages within each origin year. The triangle has only a few complete rows. Choose a resampling design and identify a tempting but inadequate alternative.
Method. Identify the dependence units before selecting the bootstrap. Here, both within-origin-year development vectors and shared calendar-period behavior are relevant. A simple independent-cell bootstrap cannot retain either pairing, and a plain row bootstrap addresses only the within-row feature.
A defensible design would use a structure-aware resampling or simulation method that represents the relevant calendar-period effects while keeping the residual components associated with an origin-year vector together. One option is to fit an explicit model with origin-year and calendar-period effects, then generate bootstrap outcomes with those effects represented jointly. If a block bootstrap is used instead, its blocks must be defined to retain the calendar-period groupings and the within-row development relationships required by the analysis.
With only a few complete rows, restricting the donor pool to complete rows may make the resampling distribution unstable. The team should assess whether there are enough compatible blocks or vectors and report limitations if the data do not support the intended dependence structure. It should not fill missing triangle cells with zero residuals to manufacture complete vectors.
Conclusion. A row-only bootstrap is insufficient for the stated calendar-year dependence, while independent cell sampling is also inadequate. The choice is not simply “row versus cell”; it is a choice of a resampling unit or fitted structure that represents the dependencies relevant to the reserve question. Common mistake avoided: assuming that preserving one dependence dimension necessarily preserves all of them.
Practical Checks Before Interpreting Results
Before using bootstrap reserves, verify that the resampling scheme matches the data layout and the modeled source of dependence. In particular:
- State whether the sampled units are cells, complete rows, compatible row segments, or blocks.
- Confirm that missing cells remain missing unless the chosen model explicitly generates them.
- Check that residual standardization and reconstruction match the fitted model.
- Inspect simulated triangles for impossible values or development patterns inconsistent with the model.
- Compare the dependence features of the observed and simulated data, not only their marginal residual distributions.
- Describe which dependence the scheme preserves and which dependence it does not claim to preserve.
A bootstrap distribution is conditional on its fitted model and resampling design. Retaining row pairings is a structural improvement over independent cell draws when within-row dependence matters, but it is not proof that the model is correct or that the resulting reserve distribution captures every source of uncertainty. This distinction is aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.
Check Your Understanding
For each problem, identify the relevant resampling structure and calculate the requested quantity.
- A residual vector is selected as \((2,-1)\). Fitted increments are \((80,40)\), and scales are \((5,4)\). Under additive reconstruction, calculate the two bootstrap increments and their cumulative amount. (Answer: increments \(90\) and \(36\); cumulative amount \(126\).)
- Two equally likely residual vectors are \((2,2)\) and \((-2,-2)\). What is their average paired cross-product? What is the expected cross-product if the two components are sampled independently from their marginal distributions? (Answer: paired average \(4\); independent expected cross-product \(0\).)
- A row-vector bootstrap produces increments \(90\) and \(36\), and the supplied factor from the cumulative amount to ultimate is \(1.25\). Calculate the ultimate and reserve beyond the observed cumulative amount. (Answer: ultimate \(157.50\); reserve \(31.50\).)
- Residuals in the same origin-year row are associated across development ages, and there is no evidence of cross-row or calendar-year dependence. Which basic scheme directly retains the observed within-row pairing, and what does independent cell resampling lose? (Answer: resample whole origin-year residual vectors; independent cell resampling loses the observed cross-age pairing and its covariance.)
- Cells on the same calendar-year diagonal appear related, while the analysis also needs within-row development-age dependence. Is a plain row bootstrap guaranteed to preserve both features? State the design implication. (Answer: no; it preserves within-row pairings but does not automatically retain diagonal calendar-year dependence, so use a structure-aware method that represents both or state the limitation.)