Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
Machine Learning
and
the Future of the Actuary
Frank Cuypers
2
Scenario
2016
Solvency II initiates
2020
New players flood the market with “digital” alternatives to insurance
2025
Insurance industry flood the market with “Fickle Superannuations” with little Solvency II capital
requirements
2030
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?
2035
The actuarial profession is discredited
The last Presidents of the SAA, DAV and IFO are burnt at the stake
3
Scenario
4
Scenarios
2015
actuaries
tolerated
2035
actuarial
profession
disappears
1995
Chief
Actuaries
in EB
2035
actuaries
bring value
again
5
Whom Shall we Still Need in 2035?
Specialized jobs
Drivers?
Nurses?
…
Information intensive jobs
Translators?
Lawyers?
Physicians?
…
Creative jobs
Architects?
Scientists?
…
The programmers?
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?
7
What’s Artificial Intelligence?
artificial intelligence machine learning
predictive modelling data analytics
data mining cognitive computing
…
decision trees
support vector machines
genetic algorithms Bayesian networks …
k-nearest neighbours
neural networks
8
Ubiquitous Neural Networks
OCR
Higgs search
Spam filters
Travelling salesman problems
Medical diagnosis
Voice recognition & generation
Translation deepl.com
Natural languages processing infocodex.com
Gaming
Image & face recognition how-old.net
…
9
𝑦𝑁𝑁 = 𝑐0 + 𝑐1ℎ1 + 𝑐2ℎ2 + 𝑐3ℎ3
Neural Networks in a Nutshell
𝑥 𝑦
1
2
3
𝑦𝑜𝑏𝑠 = 𝑓 𝑥 + 𝜀
Training: minimize 𝑦𝑜𝑏𝑠 − 𝑦𝑁𝑁2
ℎ 𝑎1 + 𝑏1𝑥
ℎ 𝑎2 + 𝑏2𝑥
ℎ 𝑎3 + 𝑏3𝑥
10
Neural Networks
bye bye models…
11
Neural Networks – 2 Hidden Layers
12
Applications
regression classification
data processing control systems
robotics
…
13
-100,000
0
100,000
200,000
300,000
400,000
500,000
600,000
700,000
800,000
900,000
1,000,000
1,100,000
1,200,000
0
500,000
1,000,000
1,500,000
0 500,000 1,000,000 1,500,000
DATA GLM NN
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%
neural network
-1.0
-0.5
0.0
0.5
1.0
-1.0 -0.5 0.0 0.5 1.0
-1.0
-0.5
0.0
0.5
1.0
-1.0 -0.5 0.0 0.5 1.0
-1.0
-0.5
0.0
0.5
1.0
-1.0 -0.5 0.0 0.5 1.0
linear model
14
Applications – Unsupervised Learning
15
Applications – Unsupervised Learning
16
Applications – Unsupervised Learning
17
Ubiquitous Neural Networks
OCR
Higgs search
Spam filters
Travelling salesman problems
Medical diagnosis
Voice recognition & generation
Translation deepl.com
Natural languages processing infocodex.com
Gaming
Image & face recognition how-old.net
…
18
Ubiquitous Neural Networks and Disrupters
OCR
Higgs search
Spam filters
Travelling salesman problems
Medical diagnosis
Voice recognition & generation
Translation
Natural languages processing
Gaming
Image & face recognition
…
Insurance?
Actuarial engineering
• Individual claims development
• Pricing
• Alternative to replicating portfolios
• …
Claims
• Regulation of attritional claims
• Fraud detection
• …
Underwriting & customer relations
• Medical underwriting
• Lapse prediction
• Retention programs
• Behavioural advice (telematics, health,…)
• …
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
• Medical underwriting
• Lapse prediction
• Retention programs
• Behavioural advice (telematics, health,…)
• …
OCR
Higgs search
Spam filters
Travelling salesman problems
Medical diagnosis
Voice recognition & generation
Translation
Natural languages processing
Gaming
Image & face recognition
…
21
Traditional Loss Development
DY
AY
individual claims
2011 2011
in 2012
2014
2013
2012
2011 in
2013
2011 in
2014
2012 in
2013
2012 in
2014
2013 in
2014
22
Traditional Loss Development
Aggregate all claims of a given AY into a single aggregate loss
DY
AY
individual claims annual aggregate loss
23
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
Stationary portfolio
24
Individual Claims Development
Aggregate all claims of a given AY into a single aggregate loss
Use individual claims information
individual claims annual aggregate loss
25
Individual Claims Development
Aggregate all claims of a given AY into a single aggregate loss
Use individual claims information cascading DY neural network
26
Individual Claims Development
Aggregate all claims of a given AY into a single aggregate loss
Use individual claims information cascading DY neural network
…
27
0
500,000
1,000,000
1,500,000
0 500,000 1,000,000 1,500,000
goal DY 2 DY 3 DY 4 DY 5
DY 6 DY 7 DY 10 DY 12
0
200,000
400,000
600,000
800,000
1,000,000
1,200,000
0 5 10 15 20
570 570 570 783 783 783
Simple Test
2 types of claims
Cascading DY
1 hidden layer
8 neurons
Paids only
out-of-sample
28
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
Stationary portfolio
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
Dynamic portfolio
29
DY Cascade vs Chain Ladder
-8,000,000
-6,000,000
-4,000,000
-2,000,000
0
2,000,000
4,000,000
6,000,000
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Dif
fere
nc
es
Accident Year
True - NN
True - CL
30
-120,000,000
-100,000,000
-80,000,000
-60,000,000
-40,000,000
-20,000,000
0
20,000,000
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Dif
fere
nc
es
Accident Year
True - NN
True - CL
DY Cascade vs Chain Ladder
31
-160,000,000
-140,000,000
-120,000,000
-100,000,000
-80,000,000
-60,000,000
-40,000,000
-20,000,000
0
20,000,000
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Dif
fere
nc
es
Accident Year
True - NN
True - CL
DY Cascade vs Chain Ladder
32
Ubiquitous Neural Networks
Insurance?
Actuarial engineering
• Individual claims development
• Pricing
• Alternative to replicating portfolios
• …
Claims
• Regulation of attritional claims
• Fraud detection
• …
Underwriting & customer relations
• Medical underwriting
• Lapse prediction
• Retention programs
• Behavioural advice (telematics, health,…)
• …
OCR
Higgs search
Spam filters
Travelling salesman problems
Medical diagnosis
Voice recognition & generation
Translation
Natural languages processing
Gaming
Image & face recognition
…
33
Insurance AI 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
34
Challenges
Architecture
Data pre-processing
Training
Cross validation
Communication
35
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
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
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
Challenges: Cross Validation
Important technical issue
As important as AvE
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
40
Facts & Alternative Facts
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
Fallacies
Needs big data
… a few hundred claims suffice
Needs an IBM Watson
… AI is more than «deep learning»
… most is much easier than translating
Why change proven processes?
… because they can be misleading
… because they can be improved
… because others could do it
Let’s implement a blockchain
…
41
Food for Thoughts
2035
actuarial
profession
disappears
If PRS can prototype it in Excel,
it’s not quantum field theory!
2015
actuaries
tolerated
2035
actuaries
bring value
again 1995
Chief
Actuaries
in EB
42
Statutory reserving
Different models depending on
data availability / quality
line of business / market
processes / products
…
actuarial judgment
→ 1st moment of a distribution
standard reserving model
Solvency II
Different models depending on
data availability / quality
line of business / market
processes / products
…
actuarial judgment
→ nth moment of a distribution
standard solvency formula
Different models depending on
data availability / quality
line of business / market
processes / products
…
actuarial judgment
… et Carthago delenda est!
43
BOF
Standard Solvency II formula
Analytic linear approximation
𝑆𝐶𝑅 ← 𝜎2 = 𝜌𝑖𝑗𝜎𝑖𝜎𝑗
Probe the tail with 2nd moments
Internal models
Numerical aggregation of realistic distributions
𝑆𝐶𝑅 ← risk1⊗ risk2⊗ risk3⊗⋯
Probe the true tail
… et Carthago delenda est!
44
BOF
… et Carthago delenda est!
Cut along
dotted line
Internal models
Numerical aggregation of realistic distributions
𝑆𝐶𝑅 ← risk1⊗ risk2⊗ risk3⊗⋯
Probe the true tail
46
The IAA
47
ASTIN what it is
Focuses on non-life insurance
Bridges the gap between practitioners and academics
with meaningful dialogues and research that:
stimulate academics to explore frontiers that are relevant in practice
catalyze the transfer of SoA techniques from academia to the industry
Creates an environment that promotes continuing professional development
48
ASTIN what it does
Colloquia
Regional meetings
Webinars
Working parties
ASTIN Bulletin
Training in countries with a developing actuarial profession
Bursaries for young actuaries & researchers from developing economies
Online library of technical and practical documents
Experts helpline (work in progress)
49
ASTIN Working Parties
Goals
Foster applied research in general insurance
Extract more value from the intellectual potential of the ASTIN membership
Topics
Focused and short-lived
Practical and useful
Organisation
Bottom-up initiatives: Champions submit short ToR
ASTIN Board endorses ToR and findings
ASTIN provides funding