Loss Reserving: Stochastic Methods and Bootstrap
Apply stochastic reserving methods including bootstrap techniques for Exam MAS-II.
Why Stochastic Reserving Matters
Traditional deterministic reserving methods like the chain ladder produce point estimates of reserves but say nothing about the uncertainty around those estimates. In practice, regulators, rating agencies, and management need to know not just the best estimate of reserves but also the range of plausible outcomes. Stochastic reserving methods provide a distribution of possible reserve amounts, enabling the calculation of reserve ranges, risk margins, and the probability of reserve adequacy at various confidence levels.
The two main approaches to stochastic reserving are analytical methods (such as Mack's model, which derives closed-form expressions for prediction error) and simulation methods (such as the bootstrap, which generates a reserve distribution through resampling). Both approaches build on the deterministic chain ladder framework, adding a probabilistic structure that quantifies uncertainty.
Mack's Model
The Mack model provides a distribution-free stochastic framework for the chain ladder. It makes three assumptions. First, accident years are independent. Second, the conditional expectation of cumulative losses at age j+1 given losses at age j satisfies E[C_{i,j+1} | C_{i,j}] = f_j * C_{i,j}, which is the standard chain ladder assumption. Third, the conditional variance satisfies Var(C_{i,j+1} | C_{i,j}) = sigma_j^2 * C_{i,j}, which specifies how the variability of development depends on the current loss level.
The development factors f_j are estimated exactly as in the standard chain ladder (weighted average of individual link ratios). The variance parameters sigma_j^2 are estimated from the residuals: sigma_j^2_hat = (1/(n minus j minus 1)) * sum over i of C_{i,j} * (C_{i,j+1}/C_{i,j} minus f_j)^2. For the last development period, where only one data point is available, sigma is typically estimated by extrapolation from the preceding periods.
Mean Squared Error of Prediction
The mean squared error of prediction (MSEP) for Mack's model decomposes into two components: process variance (irreducible randomness in how losses develop, even with known parameters) and estimation variance (additional uncertainty because the development factors are estimated from data rather than known exactly). The MSEP for accident year i's reserve is:
MSEP(R_i) = C_{i,n}^2 * sum from j=J_i to n-1 of (sigma_j^2 / f_j^2) * (1/C_{i,j} + 1/sum_k C_{k,j}), where J_i is the current development age for accident year i, C_{i,j} is the projected cumulative loss at age j, and the sum over k runs over all accident years with data at age j. The first term in parentheses (1/C_{i,j}) captures process variance, and the second term (1/sum_k C_{k,j}) captures estimation variance.
The total reserve MSEP accounts for correlation between accident year estimates. The correlation arises because all accident years use the same estimated development factors. Mack's formula for the total MSEP includes cross-product terms that reflect this shared estimation uncertainty.
The Over-Dispersed Poisson Model
The over-dispersed Poisson (ODP) model provides a GLM framework for the chain ladder, which facilitates bootstrap implementation. The model assumes that incremental losses Y_{i,j} follow an over-dispersed Poisson distribution with E[Y_{i,j}] = mu_{i,j} = alpha_i * beta_j, where alpha_i is an accident year parameter and beta_j is a development period parameter. The over-dispersion parameter phi scales the variance: Var(Y_{i,j}) = phi * mu_{i,j}.
Under the ODP model, the maximum likelihood estimates of the development factors are identical to the standard chain ladder estimates, and the projected reserves match those from the deterministic chain ladder. The key advantage is that the model provides a probabilistic framework for the residuals, enabling bootstrap resampling and simulation of the predictive reserve distribution.
Bootstrap Reserving Procedure
The bootstrap approach generates a distribution of reserves through the following steps. First, fit the chain ladder model (equivalently, the ODP model) to the observed triangle and compute Pearson residuals: r_{i,j} = (Y_{i,j} minus mu_hat_{i,j}) / sqrt(mu_hat_{i,j}). Second, resample the residuals with replacement to create a set of pseudo-residuals. Third, construct a pseudo-triangle by reversing the residual formula: Y*_{i,j} = mu_hat_{i,j} + r*_{i,j} * sqrt(mu_hat_{i,j}), where r*_{i,j} is the resampled residual for cell (i,j).
Fourth, re-estimate the chain ladder development factors from the pseudo-triangle and project reserves. Fifth, add process variance by simulating the future incremental losses from an appropriate distribution (for example, a Gamma distribution with the mean from the projected reserves and variance proportional to the mean, matching the ODP assumption). Repeating steps 2 through 5 many thousands of times (typically 10,000 or more) generates a predictive distribution of reserves.
Interpreting Bootstrap Output
The bootstrap reserve distribution provides several useful quantities. The mean of the distribution serves as the point estimate (which should be close to the deterministic chain ladder reserve). The standard deviation estimates the prediction error. Percentiles provide reserve ranges: for example, the 75th percentile represents the reserve level that is adequate with 75% probability. The difference between the 75th percentile and the mean is sometimes used as a risk margin.
The coefficient of variation (standard deviation divided by mean) summarizes the relative uncertainty. Skewness of the reserve distribution indicates how asymmetric the uncertainty is (reserve distributions are typically right-skewed, meaning there is more upside risk than downside). Comparing the bootstrap distribution to a normal approximation shows whether the normal assumption is adequate for setting reserve ranges or whether the skewness is material.
Common Pitfalls
A frequent error in bootstrap reserving is forgetting to add process variance in step 5. Without process variance, the bootstrap captures only estimation uncertainty, underestimating total prediction error. Another common mistake is using raw residuals instead of Pearson residuals. Pearson residuals are standardized to have approximately constant variance, which is necessary for the resampling to be valid.
Students also sometimes confuse the Mack model (which is distribution-free and provides analytical formulas for MSEP) with the ODP bootstrap model (which assumes a specific distributional form and uses simulation). While both methods quantify reserve uncertainty around the same point estimate, they make different assumptions and can produce different uncertainty estimates. The Mack model is simpler to implement but provides only the first two moments of the prediction error, not a full distribution. The bootstrap provides a full distribution but requires more computational effort and distributional assumptions.