Shapland and Leong Flashcards
Alphas, betas in row 3, age 4 for mean of incremental losses
a3 + b2 + b3 + b4
Alphas, betas in row 2, age 2 for mean of incremental losses
a2 + b2
Variance of incremental losses for (w,d) in ODP bootstrap
Var(q(w,d)) = Om
O is dispersion factor
Residual for ODP bootstrap
r = (q - m)/ sqrt(m^z)
Error distributions that result from variance assumption (GLM)
Var(q(w,d) = Om^z z=0, Normal z=1, Poisson z=2, Gamma z=3, Inverse Gaussian
Assumptions needed so GLM result equals CL result
Variance proportional to mean
Poisson errors
Parameter for each row and column (except first column)
5 Bootstrap Steps
- Create sampled triangle from residuals and means
- Determine alphas and betas
- Calculate mean and variance of each cell
- Draw losses from gamma distribution to add process variance
- Sum bottom half of triangle to get simulated unpaid loss
ODP Bootstrap - Dispersion factor
O = sum(r^2)/(n-p) n = # data points p = 2m - 1
Standardized residual
Residual times hat matrix adjustment
Verrall adjustment to residuals
Multiplied by sqrt[n/(n-p)] for degrees of freedom
n = # data points
p = 2m - 1
Three options to deal with negative incrementals in bootstrap
- Model -ln[-q(w,d)]
- Add constant to each loss so sum of losses in each column is positive, then subtract constant from results
- If using simplified GLM, can use CL
Simulating loss when mean is negative
Simulate positive value, then shift the mean Gamma(-m,-Om) + 2m; keeps distribution skewed to right
Stratified sampling (for heteroskedasticity)
Split triangle into groups
When sampling, only sample from residuals in same group
Hetero-Adjustment
Find SD of residuals in each group
Scale residuals by h, so that the SD is same for each group [h = max(SD of groups)/SD of the group]
For whole triangle, sample from these residuals
When applying residual, divide by h to recognize size of residuals in group
Steps to take if negative values lead to extreme results
- Remove outliers
- Recalibrate model
- Limit incremental losses below at zero
- Understand driver of negative development
Parametric bootstrapping
Once residuals have been calculated, fit a distribution to these residuals and sample from distribution
Heteroecthesious Data
Data where exposures are not same:
- Valuation of losses before year is fully earned - model will still forecast full year
- Last diagonal is valued between normal schedule - we don’t have a model for it
Correlating multiple lines of business in bootstrap
Location mapping – take sampled residual from same location in triangle from each business
Re-sorting
Bootstrap box and whisker
Box is between 25th and 75th percentiles (IQ range)
Any residuals exceeding 75th percentile by 1.5x IQ range should be investigated
Bootstrap diagnostics
Mean, standard error, COV, %iles, Min/Max
Standard error should be higher for more recent years
CoV should be highest for older years
Check min/max for reasonability
Future research possibilities for bootstrap model
Test ODP bootstrap against additional data sets
Expand model to include Munich chain ladder, claim counts and severity
Research other risk measures
Use in Solvency II
Research correlation matrix