Stochastic Reserving: Mack's Model and the ODP Bootstrap

Deterministic chain-ladder ultimates (see Reserving Methods) give a point estimate but no sense of how uncertain that estimate is. Two complementary stochastic frameworks — Mack's distribution-free model, which produces an analytic standard error, and the ODP bootstrap, which produces a full simulated distribution — are the standard tools for quantifying reserve variability.

Mack's model: three assumptions

Mack (1993) showed that the chain-ladder point estimates can be justified, and their variance derived, without assuming any particular distribution, using only three assumptions about the cumulative loss process $C_{i,k}$ across origin years $i$ and development ages $k$:

1. Conditional mean: $E[C_{i,k+1}\mid C_{i,1},\dots,C_{i,k}] = f_k\,C_{i,k}$ — the expected next value depends only on the current value times a development factor common across origin years (this is what justifies chain-ladder's core multiplicative structure).
2. Conditional variance: $\mathrm{Var}(C_{i,k+1}\mid C_{i,1},\dots,C_{i,k}) = \sigma_k^2\,C_{i,k}$ — variance is proportional to the current cumulative value (a "square-root" scaling of standard deviation to volume, common in claims data).
3. Independence across origin years: the development of different accident years $\{C_{i,\cdot}\}$ are independent random vectors — no calendar-year (diagonal) effects such as inflation shocks that would correlate origin years within the same evaluation period.

Under these three assumptions, the volume-weighted average ATA factor is the minimum-variance unbiased estimator of $f_k$, which is the formal justification for the ubiquitous choice of volume weighting.

Mean squared error of prediction

Mack derived the mean squared error of prediction (MSEP) of the chain-ladder reserve estimate for a single origin year $i$:

$\mathrm{mse}(\hat R_i) = \hat U_i^2\sum_{k=I-i+1}^{n-1}\frac{\hat\sigma_k^2}{\hat f_k^2}\left(\frac{1}{C_{i,k}}+\frac{1}{\sum_{j=1}^{I-k}C_{j,k}}\right).$

Reading each piece:

• $\hat U_i^2$ scales the relative variance up to the ultimate loss dollar level.
• The sum runs over every remaining development age $k$ from the year's current maturity to the last observed age $n-1$ — uncertainty accumulates additively across every future development step.
• $\hat\sigma_k^2/\hat f_k^2$ is the (squared) coefficient-of-variation-like term for that individual development factor — how uncertain that specific link ratio is.
• $\dfrac1{C_{i,k}}$ is the process variance contribution — the parameter-specific uncertainty from projecting this origin year's own random future development.
• $\dfrac1{\sum_j C_{j,k}}$ is the parameter (estimation) variance contribution — the uncertainty in the estimated $\hat f_k$ itself, which was fit using all origin years' data at that age; it shrinks as the triangle's total volume at that age grows.

Mack also gives a formula for the MSEP of the total reserve across all origin years, which adds a covariance term between origin years (they are correlated through sharing the same estimated $\hat f_k$, even though the underlying loss processes are independent):

$\mathrm{mse}\Big(\sum_i \hat R_i\Big) = \sum_i \mathrm{mse}(\hat R_i) + \hat U_i\left(\sum_{j\ne i}\hat U_j\right)\sum_{k}\frac{2\hat\sigma_k^2/\hat f_k^2}{\sum_l C_{l,k}}.$

Estimating $\sigma_k^2$

$\sigma_k^2$, the process variance parameter at development age $k$, is estimated from the observed scatter of individual link ratios around the selected $\hat f_k$, weighted by volume:

$\hat\sigma_k^2 = \frac{1}{I-k-1}\sum_{i=1}^{I-k} C_{i,k}\left(\frac{C_{i,k+1}}{C_{i,k}}-\hat f_k\right)^2,$

i.e., a volume-weighted sample variance of the individual factors, with $I-k-1$ degrees of freedom (number of origin years available at that age, minus 1 for the estimated mean). Because the last column often has only one or two data points, $\hat\sigma_{n-1}^2$ is frequently extrapolated (e.g., $\hat\sigma_{n-1}^2 = \min(\hat\sigma_{n-2}^4/\hat\sigma_{n-3}^2,\ \min(\hat\sigma_{n-2}^2,\hat\sigma_{n-3}^2))$, Mack's recommended extrapolation) rather than computed directly.

The ODP bootstrap

The over-dispersed Poisson (ODP) bootstrap, developed by England & Verrall, simulates a full predictive distribution of reserves rather than just a standard error, by resampling from the chain-ladder model's residuals. The algorithm:

1. Fit the chain-ladder model to the triangle (equivalently, fit a GLM with a log link, Poisson error, and factors for origin year and development year — this reproduces exactly the chain-ladder fitted values).
2. Compute scaled Pearson residuals for every cell in the upper (observed) triangle: $r_{i,k} = \dfrac{C_{i,k}-\hat C_{i,k}}{\sqrt{\hat\phi\,\hat C_{i,k}}}$, where $\hat\phi$ is the Pearson-based dispersion (scale) parameter, $\hat\phi=\frac1{n-p}\sum r_{i,k}^2$ (with $p$ fitted parameters).
3. Resample residuals with replacement to build a new pseudo-triangle: $C_{i,k}^{*} = \hat C_{i,k} + r_{i,k}^{*}\sqrt{\hat\phi\,\hat C_{i,k}}$.
4. Re-fit the chain-ladder model to the pseudo-triangle, producing a bootstrapped set of development factors and a bootstrapped fitted "expected" future triangle.
5. Add process variance: simulate actual future losses for the lower triangle from a distribution (typically Gamma or ODP) with mean equal to the bootstrap fitted future value and variance $\hat\phi\times$mean, since the bootstrap resampling in steps 3–4 only captures parameter uncertainty — process risk (actual future randomness even with perfectly known parameters) must be layered on separately.
6. Sum the simulated future losses across the lower triangle to get one simulated total reserve; repeat steps 3–5 thousands of times (e.g., 10,000 iterations) to build an empirical distribution of the total reserve.

Percentiles and risk margins

From the simulated distribution of total reserves $\{\hat R^{(1)},\dots,\hat R^{(N)}\}$, any percentile can be read directly (e.g., the 75th percentile reserve is the value below which 75% of simulations fall), and a risk margin for solvency or economic capital purposes is typically set as the difference between a high percentile (e.g., the 75th or 87.5th percentile, common in different jurisdictions' capital regimes) and the mean reserve:

$\text{Risk Margin} = \hat R_{p} - E[\hat R], \qquad p \in \{0.75,\ 0.875,\ 0.995,\dots\}.$

The bootstrap's chief practical advantage over Mack's closed-form formula is that it directly produces the full shape of the distribution (typically right-skewed), not just a standard error, which the actuary would otherwise have to assume a shape for (e.g., lognormal) before deriving percentiles from a Mack standard error alone.

Pitfalls

• Treating Mack's standard error as if it were a full distribution — it is a mean and variance only; assuming normality can understate right-tail risk margins.
• Bootstrapping only parameter uncertainty and forgetting to add process variance in the final simulation step, which understates total reserve variability significantly.
• Ignoring Mack assumption 3 (origin-year independence) when a real calendar-year effect (e.g., a claims-inflation shock or reserve strengthening) is present in the data — both Mack and the standard ODP bootstrap will understate variability if diagonal effects are ignored.
• Using too few bootstrap iterations (e.g., under 1,000), producing unstable tail percentile estimates.

Exam relevance

Mack's model and the ODP bootstrap are core, heavily tested content on CAS Exam 8 (unpaid claim estimate variability).

Further reading

• Mack, Distribution-free calculation of the standard error of chain ladder reserve estimates, ASTIN Bulletin 23(2), 1993.
• England & Verrall, Stochastic Claims Reserving in General Insurance, British Actuarial Journal.
• CAS Working Party on Quantifying Variability in Reserve Estimates.