Machine Learning at Zopa 7 December 2017 Vlasios Vasileiou Head of Data and Data Science Vlasios Vasileiou Head of Data Science 

Simple loans. Smart investments. Borrowers Investors Invests Repayments Interest + capital Loans 

A pioneering financial services company World’s 1st peer-to-peer lending platform in 2004 £2.8 billion lent to date. Strong annual growth >50% 246,000 people have taken a Zopa loan 59,000 actively invest through Zopa 

Strong, consistent growth Amount lent and forecast by year 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 £2bn £1bn 

ML & Advanced Analytics have been a strong supporter of our growth ü Credit Risk Assessment ü Onboarding Funnel Simulation ü Pricing Optimization ü Fraud Identification ü Document Forensics ü Optimized Prospect and Existing Customer Marketing ü Income and Rent Estimation for Affordability Evaluation 5 

ML & Advanced Analytics have been a strong supporter of our growth 6 ü Credit Risk Assessment ü Onboarding Funnel Simulation ü Pricing Optimization ü Fraud Identification ü Document Forensics ü Optimized Prospect and Existing Customer Marketing ü Income and Rent Estimation for Affordability Evaluation 

ML & Advanced Analytics have been a strong supporter of our growth 7 How we did get here? What were our challenges and learnings? ü Credit Risk Assessment ü Onboarding Funnel Simulation ü Pricing Optimization ü Fraud Identification ü Document Forensics ü Optimized Prospect and Existing Customer Marketing ü Income and Rent Estimation for Affordability Evaluation 

Analytical infancy (2004 – 2013) • SQL • Excel • Some Python • Externally produced credit models 

Analytical renaissance (2014 – 2015) 2014, £15m investment Board recognized need for data-driven growth ü Creation of Data Science function ü Investment in Data Analytics 

Analytical renaissance (2014 – 2015) Let’s start building some models!! Created by Alekksall - Freepik.com 

Systematizing Machine Learning at Zopa (2014) Wanted to be able to produce ML models that were: § Rapidly generated § Easily vettable § Highly predictive § Easily deployable Several considerations: n Common codebase or personal choice of tools? n Buy or build? n Which language? Which package? Created by Alekksall - Freepik.com 

Wanted to be able to produce ML models that were: § Rapidly generated § Easily vettable § Highly predictive § Easily deployable Several considerations: n Common codebase or personal choice of tools? n Buy or build? n Which language? Which package? Systematizing Machine Learning at Zopa (2014) Common codebase 

Systematizing Machine Learning at Zopa (2014) Common codebase Built in-house Wanted to be able to produce ML models that were: § Rapidly generated § Easily vettable § Highly predictive § Easily deployable Several considerations: n Common codebase or personal choice of tools? n Buy or build? n Which language? Which package? 

Wanted to be able to produce ML models that were: § Rapidly generated § Easily vettable § Highly predictive § Easily deployable Several considerations: n Common codebase or personal choice of tools? n Buy or build? n Python Systematizing Machine Learning at Zopa (2014) Common codebase Built in-house 

Predictor – Zopa’s ML Toolkit (2014) n Streamlined, Automated ML Application n Implements all stages of producing an ML model, requiring minimal user input n Leverages PyData Ecosystem n 9k lines 

How is Predictor Used? n Same exact code used by Data Scientists and in Production √ No “Model Translation” Overhead √ No restrictions on which ML techniques can be used n Training n Needs unprocessed data and a simple config file n Driven via CLI, interacting with Python codebase, or Jupyter GUI n Querying Mode n Driven by all of the above + n rest API (Flask-based microservice) in Production, <1—2 s/ call 

Learnings? n We have full control of how we do ML √ Methodology √ Deployment √ Features √ Competitive advantage, selling point to DSs. √ One day we will open source à Brand boost Ø Need to maintain at least 2 people familiar with the code base Ø Need to keep improving it – otherwise we’ll miss latest developments 

How have we used Predictor? n Was used to credit assess >£2billion worth of loans since 2015 n Production: credit and fraud risk assessment, affordability evaluation n Offline: Pricing and marketing optimization n Future: Operational efficiency for Underwriting, Collections, Customer Services n Techniques we have been using n Neural Networks, n Bagged logistic regressions and Multivariate Adaptive Regression Splines n Gradient Boosted Trees, and Random Forests 

Thank you! Further reading blog.zopa.com/2016/10/21/the-birth-of-predictor/ blog.zopa.com/2016/12/02/data-democratization/ Come work with us! Hiring for Data Scientists, Analysts, and Engineers https://jobs.lever.co/zopa [email protected] 

No content