Beyond GLMs. Xavier Conort & Colin Priest

Transcription

1 Beyond GLMs Xavier Conort & Colin Priest 1

2 Agenda 1. GLMs and Actuaries 2. Extensions to GLMs 3. Automating GLM model building 4. Best practice predictive modelling 5. Conclusion 2

3 1) GLMs Linear models for statistical distributions that aren t Normal Taught in the actuarial education process Widely used by actuaries for Pricing Understanding lapse rates Marketing Claims reserving 3

4 GLMs are old Developed in 1972 in an era of small data and no PCs Used by actuaries for decades There are newer techniques to choose from 4

5 2) Extensions to GLMs GAMs GAMLSS GLMMs Regularized GLMs 5

6 GAMs Generalized additive models Automatically fits nonlinear relationships Works a bit like rolling averages 6

7 GAMs Good For Designing data transformations for GLMs Unsuitable For Pricing, or whenever a formula is required Data containing only categorical variables 7

8 GAMLSS Generalized additive models with location, scale and shape Allows you to model how the variance and skewness varies 8

9 GAMLSS Good For understanding variability simulating risk e.g. internal models data exhibiting heteroskedasticity Unsuitable For when you only want the prediction data exhibiting homoskedasticity 9

10 GLMMs Generalized linear mixed models Allows for automatic credibility weighting 10

11 GLMMs Good For Categorical variables and interactions between categorical and numeric Sparse data Hierarchical relationships e.g. vehicle make and model Unsuitable For Large amounts of highly credible data Data without categorical variables 11

12 Regularized GLMs Regularisation is any method that penalises overfitting or complexity in models Automatically chooses predictors Automatically allows for credibility 12

13 Regularized GLMs Good For Collinearity in data Making GLMs more reliable Lots of input variables Sparse data Unsuitable For Complex interactions between variables 13

14 3) Automating GLM Building Why automate? Variable selection (feature selection) Linearising (feature engineering) Dimensionality reduction 14

15 Why Automate? Building GLMs requires time and people and both of these are expensive! Most of the resource intensive work is ruledriven without much complex judgement required It s just like when books were copied by hand! 15

16 Variable Selection Variable Importance using machine learning 16

17 Variable Selection Using lasso regularized GLMs 17

18 Variable Selection Using genetic algorithms 18

19 Linearising Using GAMs 19

20 Linearising Using GBMs 20

21 Dimensionality Reduction variable importance for reducing number of categories 21

22 Dimensionality Reduction Text mining for grouping categories together and reducing the number of categories 22

23 4) Best practices observed in 23

24 My own learning curve Previously... now! 24

25 What is Kaggle? A social fight club for data geeks In 2010, Anthony Goldbloom took the SIGKDD and Netflix s model And attracted 371,397 data geeks as of Sept 17, 2015! Kaggle worked with more than 20 Fortune 500 companies including 3 leading insurance companies + 1 Australian insurer represented by Deloitte 25

26 Machine Learning works for insurance too! Won by my colleague Owen Zhang and me! 26

27 Why Geeks like to fight? My motivation has been to learn new things. 27

28 What did I learn? The Machine works much faster and harder than me Feature engineering is key to success. And actuaries are good in this. They can however learn more by being exposed to problems and datasets outside the insurance industry Top Kagglers use actuarial tricks such as credibility estimates Most popular and powerful Machine Learning algorithms used by the data science community are open source algorithms 28

29 Competitions that boosted my learning curve The Machine seems much smarter than I am at capturing complexity in the data even for simple datasets! Humans can help the Machine too! But don t oversimplify and discard any data. Don t be impatient. My best GBM had 24,500 trees with learning rate = 0.01! SVM and feature selection matter too! 29

30 Competitions that boosted my learning curve Word n-grams and character n-grams can make a big difference Parallel processing and big servers can help with complex feature engineering! Glmnet can do a great job! Sklearn in Python is cool too! 30

31 & Machine Learning algos to know to automatically capture complexity in the data Gradient Boosting Machine packages 1. R gbm 2. R xgboost 3. Sklearn GradientBoostingClassifier and GradientBoostingRegressor Forest packages 1. R randomforest 2. Sklearn RandomForestClassifier and RandomForestRegressor 3. R extratrees 4. Sklearn ExtraTreesClassifier and ExtraTreesRegressor Support Vector Machine packages 1. R e Sklearn svc and svr 3. Sklearn Nystroem 31

32 & Machine Learning algos to know. to take advantage of high cardinality categorical features or text data Regularized generalized linear models 1. R glmnet 2. Sklearn Ridge 3. Sklearn LogisticRegression Feature Extraction for categorical features or text data 1. R Matrix 2. Sklearn OneHotEncoder and DictVectorizer 3. R tau 4. Sklearn TfidfVectorizer 32

33 & tools to know to make your code efficient Data manipulation at faster speed 1. R data.table 2. Python pandas Parallel computing 1. R foreach / domc 2. Python joblib 33

34 Adapt feature engineering to ML algo Machine Learning (ML) algo Categorical variables support Features subsampling Sparse support Insensitive to scale & uniform transf Automated nonlinear and interactions modelling Handle missing value R Random Forest Yes. up to 32 levels Yes No Yes Yes No Sklearn Random Forest No Yes Yes but slow Yes Yes No R Gradient Boosting Machine Yes. up to 1024 levels No No Yes Yes Yes Sklearn Gradient Boosted Regression Trees No Yes Yes but slow Yes Yes No extreme Gradient Boosting No Yes Yes Yes Yes Yes Regularized GLMs No No Yes No No No Support Vector Machine No No Yes No Yes No 34

35 Other popular algorithms popular in 35

36 Most important. Don t forget to use your actuarial intuition to help the machine! Always consider simple feature engineering that makes sense for your business such differences / ratios of features Be creative, feature engineering is often key to success. Don t trust features that are too good They can make the Machine lazy! An example: GE Flight Quest or they are likely to be caused by a bug or a leak! 36

37 Hans Bu hlmann ) Conclusion It s time to become actuaries of the 5 th kind Paul Embrechts 2005 Big Data Working Party Actuaries of the First Kind Actuaries of the Second Kind Actuaries of the Third Kind Actuaries of the Fourth Kind Actuaries of the Fifth Kind 17 th century: Life insurance, Deterministic methods Early 20 th century: General insurance, Probabilistic methods 1980s: Assets/derivatives, Contingencies Stochastic processes Early 21 st century: ERM Second decade of 21 st century: Big Data 37

38 Conclusion So that we aren t replaced by robots (or data scientists) 38