Taylor, Verrall, Marshall Flashcards
Exponential Dispersion Family
Theta: Location Parameter
Phi: Scale/Dispersion Parameter
b: Shape function
c: Normalizing function (unit total mass)
V: Variance function
Tweedie Sub-Family
Restricts V = mu^p
Where
p=0 Normal
p=1 ODP
p=2 Gamma
p=3 Inverse Gaussian
GLM Specifications
Error dist.
Explanatory variables
Link function
Standardized Pearson vs Standardized Deviance Residuals
SPR: Reproduces any dist. in original data
SDR: Normal, more useful for model assessment
GLM Value Correction
Outliers: 0 weight
Large variance: Less weight
Parametric Mack Model
Remove variance assumption
Define Next Incr.| Cuml. as an EDF dist.
Theorem: Parametric Mack Model
- If original variance assumption holds: MLEs of parametric Mack are CL LDFs
- If dist. is restricted to ODP AND dispersion params only vary by column:
-CL LDFs are MVUEs
-Cuml. loss and reserve estimates are MVUEs
This is stronger than non-parametric Mack model (unbiased but min variance of only linear combos of LDFs)
Cross-Classified Model Assumptions
- Incr. loss stochastically independent
- Incr. loss dist. in EDF
- Expected Incr. = Alpha_AY * Beta_Dev
- Sum of beta’s = 1
Theorems: Cross-Classified Model
- If ODP AND dispersion parameters all equal:
-MLEs same as CL LDFs - If ODP AND dispersion all equal AND fitted values corrected for bias:
-they are also MVUEs
Implies:
-Forecasts from parametric Mack and ODP Cross-Classified and Chain Ladder all the same
-Can forecast from ODP Cross-Classified while working as if ODP Mack
When to incorporate expert opinion
- Change in payment pattern
- Change in benefits due to legislation
Ways to incorporate expert opinion
- Override LDFs in particular row
- Use only n-year averages for LDFs
- Choose a different a priori estimate for BF
Types of Stochastic Models
- Mack
- ODP
- Over-Dispersed Negative Binomial
- Normal Approx. to Negative Binomial
Stochastic Mack Pros and Cons
Pro
-Easy
Cons
-No predictive dist. specified
-Need to estimate parameters for variance separate from LDFs
Stochastic ODP Pros and Cons
Pros
-Can handle some negative incr.
-Same reserve estimate as CL
-More stable than lognormal
Con
-Difficult to explain (Less connected to CL)
Stochastic OD-Neg-Bin Pros and Cons
Pros
-Can handle some negative incr.
-Same reserve estimate as CL
-Same form as CL
Con
-Can’t handle negative column sums (negative variance)