Machine Learning and the Insurance Industry Prof. John D. Kelleher

Transcription

1 Machine Learning and the Insurance Industry Prof. John D. Kelleher ADAPT Centre, Dublin Institute of Technology The ADAPT Centre is funded under the SFI Research Centres Programme (Grant 13/RC/2106) and is co-funded under the European Regional Development Fund. 1/37

2 ADAPT Research Programme 2/37

3 ADAPT Spokes Programme 3/37

4 ADAPT Value Proposition Having the ability to access and analyse data at a time of our choosing was a great help to us when reporting during the General Election. Will Goodbody RTE 4/37

5 What is Machine Learning? 5/37

6 Data-Driven Decisions Figure: Moving from data to insights to decisions. 6/37

7 Clustering Association Mining Machine Learning Prediction Anomaly Detection

8 What is Machine Learning? Figure: Using machine learning to induce a prediction model from a training dataset. 8/37

9 What is Machine Learning? Figure: Using the model to make predictions for new query instances. 9/37

10 Loan-Salary ID Occupation Age Ratio Outcome 1 industrial repaid 2 professional default 3 professional default 4 professional default 5 industrial default 6 industrial repaid 7 professional repaid 8 professional repaid 9 industrial default 10 industrial default What is the relationship between the descriptive features (Occupation, Age, Loan-Salary Ratio) and the target feature (Outcome)?

11 A Simple Predictive Model for Loan Default if Loan-Salary Ratio > 3 then Outcome= default else Outcome= repay end if 11/37

12 Machine Learning and Insurance Case Study 12/37

13 ML and Insurance Underwriting Insurance underwriting is an clear target for ML A life insurer will typically spend approximately one month and several hundred euros/dollars underwritting each applicant. 13/37

14 Predictive ML Prerequisites 1. A clearly defined target variable, i.e. what the model is trying to predict 2. The availability of a suitably rich data set, in which at least some predictive variables correlated with the target can be identified 3. A large number of observations upon which to build the model, allowing the abiding relationships to surface and be separated from random noise 4. An application by which model results are translated into business actions 14/37

15 Predictive ML Prerequisites Fulfilling these requirements may be difficult for some insurers. 15/37

16 Car Insurance Underwriting Target Variable Claims over 6-month contract Modelling Data Underwriting requirements supplemented by third-part data Frequency of Loss Approx. 10 percent of drivers make claims annually Business Action Underwriting Decision 16/37

17 Life Insurance Underwriting Target Variable Mortality experience over life of product (10,20+ years)* Modelling Data Underwriting requirements supplemented by third-part data Frequency of Loss Fewer than 1 first year death per 1,000 new policies issued* Business Action Underwriting Decision 17/37

18 Target Variables Because life insurance is sold through long duration contracts the target variable also has a very long duration Accessing mortality date over such a long-time span is often a significant challenge (particularly when one needs to connect the data at the time of insurance underwriting on a policy to the mortality event) 18/37

19 Frequency of Loss In order to statistically model variation in insurance claims we need a large sample of loss events Because of the frequency of car insurance claims car insurers can build robust models using loss data from the most recent years Life insurers often have to go back over a much large time horizon (again connecting the data at the time of underwiting to the later events can be very challenging) 19/37

20 The Majority of Effort in ML Projects Focuses on Data Considerations Task % Time Understanding Business Problem 20% Accessing & Preparing Data 36% Generating Models 20% Writing Reports or Presentations 15% Scoring and Deploying 9% 20/37

21 Artificial Intelligence Machine Learning Deep Learning

22 What is a function? A function maps a set of inputs (numbers) to an output (number) sum(2, 5, 4) 11 22/37

23 What is a weightedsum function? weightedsum([x 1, x 2,..., x m ], [w }{{} 1, w 2,..., w m ]) }{{} Input Numbers Weights = (x 1 w 1 ) + (x 2 w 2 ) + + (x m w m ) weightedsum([3, 9], [ 3, 1]) = (3 3) + (9 1) = = 0 23/37

24 What is an activation function? An activation function takes the output of our weightedsum function and applies another mapping to it. logistic(act) Activation 24/37

25 What is an activation function? activation = logistic(weightedsum(([x 1, x 2,..., x m ], [w }{{} 1, w 2,..., w m ])) }{{} Input Numbers Weights logistic(weightedsum([3, 9], [ 3, 1])) = logistic((3 3) + (9 1)) = logistic( 9 + 9) = logistic(0) = /37

26 What is a Neuron? The simple list of operations that we have just described defines the fundamental building block of a neural network: the Neuron. Neuron = activation(weightedsum(([x 1, x 2,..., x m ], [w }{{} 1, w 2,..., w m ])) }{{} Input Numbers Weights 26/37

27 What is a Neuron? x 0 x 1 x 2 w 2 w 0 w 1 Σ ϕ Activation x 3 w 3. w m x m 27/37

28 What is a Neural Network? Input Layer I1 I2 Hidden Layer H1 H2 H3 Output Layer O1 O2 28/37

29 Deep Learning 29/37

30 Deep Learning Output[1] Output[2] Figure: A Network is Deep if it has Multiple Hidden Layers 30/37

31 Deep Learning DATA HAND- ENGINEERED FEATURES MODEL Figure: Standard ML DATA LEARNED FEATURES MODEL Inspired by slide from Kevin Duh (Deep Learning Tutorial) Figure: Deep Learning 31/37

32 Face [Y/N] Input Image Edge Detectors Object Parts Object Models

33

34 Deep Learning Deep Learning Models: Require a very large amount of data to train Most appropriate with high-dimensional data (e.g., video, image, text, time-series) Take a long time to train Black box (GDPR) 34/37

35 Bringing Machine Learning into Your Business Focus: target clearly defined business problems Simple: deep learning is not the only type of machine learning Stable: the best models are the ones that fit into your existing processes Build the human capacity to do data science 35/37

36 ADAPT Industry Showcase - Croke Park - Tomorrow! 36/37

37 Thank you for your time and KELLEHER AND TIERNEY computer science DATA SCIENCE DATA SCIENCE DATA SCIENCE JOHN D. KELLEHER AND BRENDAN TIERNEY KELLEHER AND TIERNEY A short briefing on intellectual property strategy, describing how a flexible and creative approach can help an organization achieve both short- and long-term benefits. THE MIT PRESS MASSACHUSETTS INSTITUTE OF TECHNOLOGY CAMBRIDGE, MASSACHUSETTS THE MIT PRESS ESSENTIAL KNOWLEDGE SERIES 37/37