Upload
brandon-stafford
View
213
Download
1
Embed Size (px)
Citation preview
Personal Lines Actuarial Research Department
Generalized Linear ModelsCAGNY
Wednesday, November 28, 2001
Keith D. Holler Ph.D., FCAS, ASA, ARM, MAAA
Personal Lines Actuarial Research Department
2
High Level
e.g. Eye ColorAgeWeight Coffee Size
Given Characteristics:
Predict Response:e.g. Probability someone takes Friday off, given it’s sunny and 70°+e.g. Expected amount spent on lunch
Personal Lines Actuarial Research Department
3
Personal auto or H.O. class plansDeductible or ILF severity models Liability non-economic claim settlement amountHurricane damage curves* Direct mail response and conversion*Policyholder retention*WC transition from M.O. to L.T.*Auto physical damage total loss identification*Claim disposal probabilities*
Insurance Examples
* Logistic Regression
Personal Lines Actuarial Research Department
4
Example – Personal Auto
Log (Loss Cost) = Intercept + Driver + Car Age Size Factor i Factor j
Driver Age Car Size
Intercept Young Older Small Medium Large
6.50 .75 0 .50 .20 0
e.g. Young Driver, Large CarLoss Cost = exp (6.50 + .75 + 0) = $1,408
Parameters
Personal Lines Actuarial Research Department
5
Technical Bits
1. Exponential families – gamma, poisson, normal, binomial2. Fit parameters via maximum likelihood3. Solve MLE by IRLS or Newton-Raphson4. Link Function (e.g. Log Loss Cost)
i. 1-1 functionii. Range Predicted Variable ( - , )iii. LN multiplicative model, id additive model
logit binomial model (yes/no)5. Different means, same scale
Personal Lines Actuarial Research Department
6
Personal Auto Class Plan Issues:
1. Territories or other many level variables2. Deductibles and Limits3. Loss Development4. Trend5. Frequency, Severity or Pure Premium6. Exposure7. Model Selection – penalized likelihood an option
Personal Lines Actuarial Research Department
7
Why GLMS?
1. Multivariate – adjusts for presence of other variables. No overlap.
2. For non-normal data, GLMS better than OLS.3. Preprogrammed – easy to run, flexible model structures.4. Maximum likelihood allows testing importance of variables.5. Linear structure allows balance between amount of data and
number of variables.
Personal Lines Actuarial Research Department
8
Software and References
Software: SAS, GLIM, SPLUS, EMBLEM, GENSTAT, MATLAB, STATA, SPSS
References: Part 9 paper bibliographyGreg Taylor (Recent Astin)Stephen Mildenhall (1999)Hosmer and LemeshowFarrokh Guiahi (June 2000)Karl P. Murphy (Winter 2000)