2017 Predictive Analytics Symposium

Transcription

1 2017 Predictive Analytics Symposium Session 2, General Session: Panel of Predictive and Data Analytics Heads for Financial Services Companies Moderator: David L. Snell ASA, MAAA Presenters: Peter Banthorpe James C. Guszcza, FSA, FCAS SOA Antitrust Compliance Guidelines SOA Presentation Disclaimer

2 James Guszcza US Chief Data Scientist, Deloitte Consulting James Guszcza is Deloitte Consulting's Chief Data Scientist, and a pioneering member of Deloitte s original Data Science practice. Jim has applied statistical and machine learning methods to such diverse business problems as healthcare utilization, customer and employee retention, talent management, customer segmentation, insurance pricing and underwriting, credit scoring, child support enforcement, patient safety, claims management, and fraud detection. He has also spearheaded Deloitte s use of behavioral nudge tactics to more effectively act on model indications. Deloitte Consulting LLP 555 West 5th Street, Suite 2700 Los Angeles, CA James Guszcza, PhD, FCAS Chief Data Scientist Deloitte Consulting US Tel: jguszcza@deloitte.com Member of Deloitte Touche Tohmatsu A frequent author and conference speaker, Jim designed and teaches hands-on business analytics training seminars for both the Casualty Actuarial Society and the Society of Actuaries. Jim is a former professor at the University of Wisconsin-Madison business school, and has a PhD in the Philosophy of Science from The University of Chicago. He is a Fellow of the Casualty Actuarial Society, and serves on its board of directors.

3 Peter Banthorpe Head of Global Research and Data Analytics Peter Banthorpe is Senior Vice President and Head of Global Research and Development with RGA Reinsurance Company. In this role, he leads RGA s advanced analytics capabilities, International experience study unit, and oversees RGA s biometric research and related strategic research programme. Peter is also the Lead Risk Management Office for Mortality at RGA. Peter Banthorpe MA FIA Senior Vice President Head of Global Research and Data Analytics Reinsurance Group of America 16th Floot, 5 Aldermanbury Square London, EC2V 7HR T: pbanthorpe@rgare.com Prior to joining RGA in 2010, Peter was Head of Pricing and Reinsurance for AIG Life UK (formerly Fortis). He was one of its first employees, a member of its senior management team, and led the development of its actuarial, management information and reinsurance capabilities from initial creation until several years after its launch. Peter is a graduate of Cambridge University and qualified as an actuary in He has been a member of the Institute and Faculty of Actuaries (IFoA) CMI Executive Committee since 2008 and was formally a founding Director of CMI Limited. He has previously chaired IFoA working parties on Critical Illness pricing and Genetics and the IFoA s Mortality Research Steering Committee. Peter is also a regular presenter at industry events on a diverse range of subjects.

4 Panel of Predictive Analytics/Data Analytics/Similar heads for Financial Services Companies Peter Banthorpe SVP, Head of Global Research and Data Analytics September 14, 2017

5 Getting Maximum Value From Data and Analytics 01 Business Need No shortage of demand - need to prioritize 02 Data Limited touchpoints, could use, should use 03 Talent and Hard Work Data engineering, data manipulation, data science 04 Algorithm Correlation vs. Causation 05 More Talent A collection of distinct roles 06 Change Programs A need for agility 07 Iteration No model lasts forever 2

6 Getting Maximum Value From Data and Analytics 01 Business Need No shortage of demand - need to prioritize 02 Data Limited touchpoints, external data, use 03 Talent and Hard Work Data engineering and data manipulation 04 Strategy A Model Range of Models and Correlation vs. Causation 05 More Talents A collection of distinct roles 06 Change Programs A need for agility 07 Iteration No model lasts forever 3

7 A Quick Poll what role do you most closely align to? Actuary Data Engineer Data Scientist Data Strategist Data Analytics Leader Actuary and Data Scientist Something else entirely 4

8 Business Needs: Insurance Value Chain and Applications 01 Pre-Sale Underwriting In-force Management Claims Price Elasticity (re-pricing strategy) Propensity to Complete Purchase Cross-sell / Upsell Claims Triage Distributor Quality Management Underwriting Triage More granular value recognition Non-disclosure / Fraud Prediction Moment of Truth Determine Underwriting Ratings Proactive Lapse management Preferred Risk Selection (new rating factors) Non-disclosure / Fraud Prediction Customer Lifetime value Propensity to Buy Underwriting Risk Segmentations Level of demand that RGA has experienced High Medium Low 5

9 What data could we use? Estimated average number of customer interactions per year Life Non-Life Motor Health Retail Telco Banking Technology Companies Social Media Source: McKinsey & Company, presented in Transforming Life Insurance with Design Thinking 6

10 What data should we use? Regulations Increasing focus EU General Data Protection Regulation (GDPR) Anti-discrimination legislation Profiling versus Decisions Risks Reputational damage Financial Penalties Theft / Mis-use / Loss Data governance procedures and culture Reasonable Expectations Fairness Country Specific Industry and data specific What have you told the client Does it benefit the consumer? Rewards Consumer Company (Red vs Blue Ocean) Employee Regulator 7

11 Modelling Toolkit Machine Learning (and into Deep learning) Data Mining Not intended to be exhaustive Operational Research Simulation Optimisation Simulated Annealing Linear, Logistic Regression, GLMs Neural Networks Genetic Algorithms Time-series Analysis K-Means Clustering Bayesian Networks Decision Trees Link Analysis Graph Theory Querying OLAP Cross-tabs Business Intelligence Exponential Smoothing Winter s methods State Space Models Recurrent Neural Networks Deep Boltzmann Machines Deep Belief Networks Random Forest Convolutional Neural Nets Association Rule Learning Principle Components Support Vector Machines Lasso / Ridge penalties Gradient Boosting Visualisation Monte Carlo Fourier Transforms Wavelets Harmonic Analysis Forecasting Reliability/Survival Analysis ANOVA MANOVA Correlation Factor Analysis Statistics 8

12 Domain Knowledge for Life and Health Basic Demographics Age Gender Education Occupation Insurance purchasing Social Media Usage Marital status Number of Children Avocations Driving Behaviour Credit Behaviour Personality Well-being Stress Benefit amount Where you live Investment choices Car ownership Smoker Status Care provider Sleep Previous claims Activity history Diet Behavioural Not intended to be exhaustive Socio- Economic Assets Salary House price Health and Biometrics Disease history Family History Prescription History Heart Rate / HRV / etc Genetic Data Omic Data 9

13 The parable of the blind men and an elephant 04 10

14 What model to apply when? 04 Unit Cost and Cost of Error Causation necessary Correlation sufficient Number of Transactions Concept gratefully borrowed from Michael Natusch from Pru. 11

15

16 Getting Maximum Value From Data and Analytics Business Need No shortage of demand - need to prioritize 01 Data Limited touchpoints, could use, should use Talent and Hard Work Data engineering, data manipulation, data science Algorithm Correlation vs. Causation More Talent A collection of distinct roles Change Programs A need for agility Iteration No model lasts forever Strategy 13

17 2017 RGA. All rights reserved.

18 Data science from idea to impact Jim Guszcza, PhD, FCAS SOA Predictive Analytics Symposium Chicago September 14, 2017

19 What is data science? bringing it to life

20 You can t make this stuff up

21

22 In the center of it all Drew Conway s data science Venn diagram Jeff Hammerbacher and DJ Patil coin the term data scientist circa

23 In the center of it all Where is the actuarial profession?

24 In the center of it all Where is the actuarial profession? Here?

25 In the center of it all Where is the actuarial profession? Or here?

26 Getting it built Data science is neither engineering, nor is it traditional statistics

27 Getting it built Data science is neither engineering, nor is it traditional statistics Substantive domain / institutional knowledge is necessary to effectively decide: What questions to address Where to look for the data Nature of the data-generating process Limitations of the data How to adjust the data How the analysis will be used

28 Getting it built Data science is neither engineering, nor is it traditional statistics Substantive domain / institutional knowledge is necessary to effectively decide: What questions to address Where to look for the data Nature of the data-generating process Limitations of the data How to adjust the data How the analysis will be used Not all people on a data science team will be equally strong in each area. Effective collaboration is necessary.

29 The original data scientist An approximate answer to the right question is worth a good deal more than an exact answer to an approximate problem. The combination of some data and an aching desire for an answer does not ensure that a reasonable answer can be extracted from a given body of data. -- John Tukey

30 1993

31 1993

32 2001 (Now you know where data science came from)

33 2017

34 2017

35 2017

36 2017

37 Data science a narrow frame The tech-centric view Amenable to Kaggle-style crowdsourcing DATA MODEL

38 The answer is 42 STRATEGY DATA MODEL

39 The answer is 42 Remember the Tukey quote STRATEGY DATA MODEL

40 The answer is 42 Remember the Tukey quote STRATEGY DATA MODEL And Blei / Smyth:

41 Making it real Good models are necessary, but not sufficient for positive ROI. Models can point us to the cases in need of attention, but do not tell us how to act on these insights. Effective model operationalization is crucial. STRATEGY DATA MODEL Modeling might only be 10-20% of a predictive modeling / data science project!

42 Big data neither necessary nor sufficient for positive ROI

43

44 Big data

45 Big data and its discontents

46 Big data and its discontents

47 A parable about data science and inefficient markets Superior management could run circles around taller piles of cash. Michael Lewis, Moneyball

48 NYC does actuarial prediction big data

49 Sometimes big data is not enough Data: Digital: Design: Statistical fraud detection using web click data Interactions mediated by web site Optimize behavioral nudge pop-up messages (use A/B testing)

50 Uberizing insurance?

51 Prosocial applications of big data

52 It s in the news