320 likes | 329 Views
Learn how to successfully deploy R models into production and overcome the associated challenges. Find out about the historical perspective, underwriting process, next-generation approach, institutional skepticism, and technical challenges.
E N D
From Data to Deployment: overcoming the challenges of embedding R models in Production Gwilym MorrisonHon W Yau
Largest mutual life, pensions and investment company in the UK. Group funds under management of £114 billion. Group businesses provide around 9 million policies and have 1.4 million members. We Employ about 3,500 people. Founded as a Friendly Society in a London coffee shop in 1861. Started out with the aim to help people avoid the stigma of a pauper’s grave.
Intermediated Life Insurance Value Chain Reinsurers Distributors Policyholder Employers Insurers
A Historical Perspective What is this strange-looking device?
Stochastic Neural Analogue Reinforcement Computer (SNARC) The SNARC was a Neural-Network maze solver built by Marvin Minsky in 1951
Minsky’s other Inventions Minsky is also the originator of several modern implementations of “useless machines”.
Deploying Models into Production Today Today we have software implementations of Machine Learning models and so soldering irons are not required to implement these
Underwriting is a Moment of Truth in Life Insurance The traditional approach to Underwriting effective in many respects. However, it can often be uncomfortable and distressing for customers. It is often not looked up favourably by our distribution partners. In addition, it’s relatively expensive.
Life Insurance Underwriting Underwriting is the process of assessing an applicant’s claim risk in order to price their policy as accurately as possible. Weeks/Months Days Minutes Customer completes application form Underwriter triages application Customer Receives Decision Rules engine makes initial assessment Underwriter obtains additional evidence
Effect of Manual Underwriting Manual Underwriting has a number of problems associated with it. Takes a long time Expensive It’s intrusive 01 02 06 05 03 Increases drop-off rates Prone to complications 04 Unpleasant experience
Next Generation Approach to Life Underwriting Online Rules Traditional Predictive
Quote and Apply For Independent Financial Adviser (IFA) Users
Quote and Apply Automated Underwriting Decisions
Quote and Apply Referral Reduction with Scoring Models
Scoring Model & Rules Why AI Models Help Underwriting Decisions
Scoring Model & Rules Smart Automation
Institutional Scepticism and Challenges Following traditional Project delivery processes
Institutional Scepticism and Challenges Data Science Model Delivery
Institutional Scepticism and Challenges How We Think We (Probably) Are
Institutional Scepticism and Challenges How Project Delivery Teams (Probably) Saw Us
Institutional Scepticism and Challenges Language Barrier
Institutional Scepticism and Challenges Development Team Quality Gates
Institutional Scepticism and Challenges Data Science Quality Gates
Scepticism and Challenges Management £
Scepticism and Challenges Not Bankrupting the Company *.Re
Scepticism and Challenges Production Support
Scepticism and Challenges What if we Cause…
Technical Challenges Version Control – For Multi-Year Transactions
Technical Challenges Memory & Response Times
Summary of Challenges And Solutions Solution Challenge The technology is (relatively) new • Early engagement • Spend time building trust • Get sponsorship right • Third party buy-in It has to play nicely with existing processes • Spend time educating colleagues • Agree a streamlined release process • Define quality gates and sign-off process up-front Model release capacity can be limited • Explore “lighter” models • Use cloud deployment