Who s Afraid of Artificial Intelligence? Frank Cuypers

Transcription

1 Who s Afraid of Artificial Intelligence? Frank Cuypers

2 Scenario Solvency II initiates New players flood the market with digital alternatives to insurance Insurance industry flood the market with Fickle Superannuations with little Solvency II capital requirements 30% of the insurance players file for bankruptcy The weaknesses of the Solvency II standard formula become obvious Who should have known? Who should have warned? The actuarial profession is discredited The last Presidents of the SAA, DAV and IFO are burnt at the stake 2

3 Scenario Solvency II initiates New players flood the market with digital alternatives to insurance Insurance industry flood the market with Fickle Superannuations with little Solvency II capital requirements 30% of the insurance players file for bankruptcy The weaknesses of the Solvency II standard formula become obvious Who should have known? Who should have warned? The actuarial profession is discredited The last Presidents of the SAA, DAV and IFO are burnt at the stake 3

4 Scenarios 1995 Chief Actuaries in EB 2035 actuaries bring value again 2015 actuaries tolerated 2035 actuarial profession disappears 4

5 Whom Shall we Still Need in 2035? Specialized jobs Drivers? Nurses? Information intensive jobs Translators? Lawyers? Physicians? Creative jobs Architects? Scientists? The programmers? 5

6 How Shall we Live in 2035? Will our children still learn how to read & write? Shall we be terminated? Shall we become bionically enhanced? Will eventually joint human & machine crowd intelligence supersede artificial intelligence? 6

7 What s Artificial Intelligence? artificial intelligence machine learning data analytics predictive modelling cognitive computing Neural networks Decision trees k-nearest neighbours Support vector machines Bayesian networks Genetic algorithms data mining 7

8 Ubiquitous Neural Networks OCR Higgs search Spam filters Image compression Travelling salesman problems Medical diagnosis Voice recognition & generation Translation translate.google.com Natural languages processing infocodex.com Gaming Face recognition how-dude.me how-old.net Insurance? 8

9 Neural Networks (in a nutshell) 1 h: sigmoid h(a 1 +b 1 x) x 2 y y obs =f(x)+ε h(a 2 +b 2 x) y NN =c 0 +c 1 h 1 +c 2 h 2 +c 3 h 3 h(a 3 +b 3 x) 3 training: minimize Σ (y NN - y obs ) 2 9

10 Neural Networks bye bye models 10

11 Neural Networks (2 hidden layers) 11

12 Applications regression classification data processing control systems robotics 12

13 Applications Supervised Learning Regression Output = e.g. size of claim Each output neuron gives a number, which can take any value Classification Output = e.g. type of claim Each output neuron gives a probability, which all add up to 100% 30'000 30'000 25'000 20'000 15'000 20' linear model neural network 10'000 10' ' '000 5'000-10'000 10'000-15'000 15'000-20' '000 20'000 30'000 goal GLM neural network '000-25'000 25'000-30'000 13

14 Applications Unsupervised Learning 14

15 Applications Unsupervised Learning 15

16 Applications Unsupervised Learning 16

17 Patents Neural network for classifying speech and textural data based on agglomerates in a taxonomy table System and method for automated establishment of experience ratings and/or risk reserves 17

18 Ubiquitous Neural Networks OCR Higgs search Spam filters Image compression Travelling salesman problems Medical diagnosis Voice recognition & generation Translation Natural languages processing Gaming Face recognition Insurance? 18

19 Ubiquitous Neural Networks Insurance? Actuarial engineering Individual claims development Pricing Alternative to replicating portfolios Claims Regulation of attritional claims Fraud detection Underwriting & customer relations Lapse prediction Retention programs Behavioural advice (telematics, health, ) Alternative insurance??? 19

20 Traditional Loss Development DY in in in 2014 AY in in in individual claims 20

21 Traditional Loss Development Aggregate all claims of a given AY into a single aggregate loss DY AY individual claims annual aggregate loss 21

22 Traditional Loss Development Aggregate all claims of a given AY into a single aggregate loss Develop with Chain Ladder Born-Ferg Cape Cod Assume Homogenous portfolio Independent AY 22

23 Individual Claims Development Aggregate all claims of a given AY into a single aggregate loss Use individual claims information individual claims annual aggregate loss 23

24 Individual Claims Development Aggregate all claims of a given AY into a single aggregate loss Use individual claims information cascading DY neural network 24

25 Individual Claims Development Aggregate all claims of a given AY into a single aggregate loss Use individual claims information cascading DY neural network 25

26 ASTIN Working Party on ICDML Didactic implementation 2 types of synthetic claims Excel Cascading DY 1 hidden layer 8 neurons Paids only 26

27 Synthetic Claims Controlled environment 27

28 Synthetic Claims Controlled environment in-sample we know NN knows use for training out-of-sample we know NN knows not use for testing 28

29 ASTIN Working Party on ICDML Didactic implementation 2 types of synthetic claims Excel Cascading DY 1 hidden layer 8 neurons Paids only 1'500'000 1'000' '000 out-of-sample 1'200'000 1'000' ' ' ' '000 1'000'000 1'500'000 goal DY 2 DY 3 DY 4 DY 5 DY 6 DY 7 DY 10 DY ' true 570 in-sample 570 out-of-sample 783 true 783 in-sample 783 out-of-sample 29

30 ASTIN Working Party on ICDML Didactic prototype 2 types of synthetic claims Excel Cascading DY 1 hidden layer 8 neurons Paids only Experimental implementation Several types of synthetic claims R, Python, Cascading DY & AY 1 2 hidden layers 2 many neurons Paids & outstandings Productive roll-out Real data R or Python or SAS Cascading DY or AY? hidden layers? neurons Paids & outstandings Other explanatory variables 30

31 ICD vs ALD Aggregate Loss Development Develop with Chain Ladder Born-Ferg Aggregates all claims of a given AY into a single aggregate loss Works either on paid or incurred losses Assumes Homogeneous portfolio Independent AY Individual Claims Development Develop with DY or AY cascades Convolutional networks Considers all individual claims features, including non monetary inputs Considers simultaneously payments and reserves Works with Heterogeneous portfolios Dependent AY 31

32 DY Cascade vs Chain Ladder 6'000'000 4'000'000 Differences 2'000' '000'000-4'000'000-6'000'000-8'000'000 True - NN True - CL True - NN Total True - CL Total Accident Year 32

33 DY Cascade vs Chain Ladder 20'000'000 0 Differences -20'000'000-40'000'000-60'000'000-80'000' '000' '000'000 True - NN True - CL True - NN Total True - CL Total -140'000' Accident Year 33

34 DY Cascade vs Chain Ladder 20'000'000 0 Differences -20'000'000-40'000'000-60'000'000-80'000' '000'000 True - NN True - CL True - NN Total True - CL Total Accident Year 34

35 Challenges Architecture Data pre-processing Training Cross validation Communication 35

36 Challenges: Architecture Monkey & octopus Can solve similar problems Have completely different brains (octopus has 9 brains ) Dyslexic & autistic humans Have same brain architectures Have completely different skills Neural network? Activation function (sigmoid)? Penalty function? Number of layers? Number of neurons? Training strategy? Fully-connected vs convolutional network? 36

37 Challenges: Data Pre-Processing Humans are good at catching flying objects But less if they are myopic Humans are good at communicating orally But less if they are hearing-impaired Neural network! Pre-process inputs! Scale outputs requires a healthy understanding of the underlying phenomena 37

38 Challenges: Training How do you learn A poem A foreign language A programming language A mathematical method Neural network Minimize penalty function over a high dimensional parameter space Backpropagation Very fast (Python, Matlab) Steepest gradient local minima Simulated annealing? Global minimum? Untested? 38

39 Challenges: Communication You ride a car do you know how your ABS works? your airbag triggers? it will drive on its own? You implement Chain Ladder do you understand why the link factors take these values? you may apply this method? Richard Feynman: Nobody understands Quantum Mechanics! Produce with neural networks illustrate with decision trees 39

40 Challenges: Cross Validatio Important technical issue As important as AvE 40

41 Synthetic vs Real Data Synthetic data Training: ignore known DY Validation: use these DY Real data Training: use all known DY Validation: cross-validate w/in AY 41

42 Claims Generator Generate individual claims with probability distributions of Severity: ultimate UU~LLLL μμ, σσ Development patterns: age-to-ultimate FF tt ~LLLL μμ tt, σσ tt Components Paid PP tt = UU FF PP tt Outstanding OO tt = UU FF OO tt Incurred II tt = PP tt + OO tt Patterns FF PP tt : μμ tt = 1 ee tt ττ λλ FF OO tt : μμ tt = ααee tt ττ λλ 2 αα 140% 120% 100% 80% 60% 40% 20% 0% paid outs incu 100% 50% Dependence Frank Copula FF PP tt FF OO tt for each tt 0% 0% 50% 100% 42

43 Synthetic Claims Generate as many individual claims as needed 1'400'000 1'200'000 1'000' ' '000 1'600'000 1'400'000 1'200'000 1'000' ' ' ' ' ' ' paid outs incu paid outs incu Mix individuals claims from different models 43

44 Cross Validation 44

45 Cross Validation 45

46 Cross Validation training set 1 validation set 1 training set 2 validation set 2 training set 2 training set 3 validation set 3 training set 3 46

47 Cross Validation Residual Analysis YY ii = true values YY ii = predicted values training set 1 validation set 1 yy ii = YY ii YY ii YY ii = residual 2'000'000 YY 100% P YY 100% P yy 80% 80% 60% 60% 1'000'000 40% 40% 20% 20% 0 YY 0 1'000'000 2'000'000 P YY 0% 0% 20% 40% 60% 80% 100% 0% -10% 0% 10% 20% yy goal training validation training validation training validation 47

48 Cross Validation Measures of Fit YY ii = true values YY ii = predicted values 100% 80% 60% 40% 20% validation training CoV = 1 NN YY ii YY ii 2 1 NN YY ii discrepancy = 1 NN yy ii 2 validation training validation ~ training 0% -1.5% -1.0% -0.5% 0.0% 48 CoV discrepancy

49 Advantages Respond very fast but training can take long Can generalize and may get it wrong Are robust most of them Are very flexible with regard to inputs if well pre-processed Can update their knowledge continuously with reinforced learning 49

50 Ubiquitous Neural Networks Insurance? Actuarial engineering Individual claims development Pricing Alternative to replicating portfolios Claims Regulation of attritional claims Fraud detection Underwriting & customer relations Lapse prediction Retention programs Behavioural advice (telematics, health, ) Alternative insurance??? 50

51 Food for Thoughts 1995 Chief Actuaries in EB 2015 actuaries tolerated 2035 actuarial profession disappears 51

52 et Carthago delenda est! Statutory reserving Different models depending on data availability / quality line of business / market processes / products actuarial judgment 1 st moment of a distribution Solvency II Different models depending on data availability / quality line of business / market processes / products actuarial judgment n th moment of a distribution standard reserving model standard solvency formula 52

53 et Carthago delenda est! Internal models Numerical aggregation of realistic distributions SSSSSS risk 1 risk 2 risk 3 Probe the true tail BOF Standard Solvency II formula Analytic linear approximation SSSSSS σσ 2 = ρρ iiii σσ ii σσ jj Probe the tail with 2 nd moments 53

54 et Carthago delenda est! Internal model Numerical aggregation of realistic distributions SSSSSS risk 1 risk 2 risk 3 Probe the true tail BOF Cut along dotted line 54

55 Food for Thoughts 1995 Chief Actuaries in EB 2035 actuaries bring value again 2015 actuaries tolerated It s not quantum field theory! If you can prototype it in Excel, then it can t be difficult 2035 actuarial profession disappears 55

56 Lecturer s Coordinates Frank Cuypers +41 (41) frank.cuypers@prs-zug.com 56