Data Science in Financial Services Michel Kamel Senior

Data Science in Financial Services Michel Kamel Senior Risk Associate at Group Risk Analytics – Bank Audi Data Scientist – Harvard Extension School

Mining • Data mining

Outline • Evolution of data & Analytics • Machine learning overview • Case studies

Data Evolution • Banks have data from their first day of 1770 1950 1980 -1990 2005 -2010 2012 -2018 2050

Outline • Evolution of data & Analytics • Machine learning Models overview • Case studies

Machine Learning Models • Machine Learning is the process of building models to help computers learn from data (experience). • Start by defining a business objective, find your data to fit the appropriate model which will go live

Machine Learning Models • Clustering & Segmentation • Regression • Decision tree • Neural Network

Outline • Evolution of data & Analytics • Machine learning Models overview • Case studies

Case Study 1: Application Scorecard - Turkey Data (Internal & KKB) 3, 897, 584 rows x 680 Columns Bureau and internal data Accept Reject Expert Model PD Model Income Model

Model development • Comparison of models (logistics, neural network…etc) • Maximize the predictive power of the model

Application Scorecard Before After Approval Rate 57% Bad Rate X% Bad Rate 65%*X

Scorecard Tracking

Behavioral scorecards • Business Objective: Predict the risk of default of each existing customer 497, 382 rows x 4, 767 Columns • Collect 360 degree view customer data (static, transactional, market, …) Bureau and internal data • Keep the bank up-to-date with the customer • Less time constraints

Behavioral Scorecard – Predictive Characteristics (example) • Default rate by Age

Collection Scorecard • Issue: Thousands of applications monthly • Objective: What is the probability a client will repay his past-due loans once he becomes bad? • Data: • 360 degree view of the customer • Trails history of collection department

Collection Scorecard • Action strategy based on collection score X, 000 of > delinquent loans (per month) Low Probability Medium Prob. High Prob.

Clustering & Segmentation: Branches Segmentation • Issue: Set yearly targets objectively • Objective: which branches are similar in terms of target? • Data: 1. Number of employees 2. Market share 3. Total Assets 4. Total Liabilities 5. Number of customers 6. Average number of transactions

Clustering & Segmentation: Branches Segmentation

Attrition Scorecard • Objective: Which customers will leave the Bank? • Data: • 360 degree view of the customer • Transactional historical data • • # of transactions via branch/ATM Average limit utilization # of loyalty cards products Others

Attrition Scorecard Trend of Assets # of Prod Jan-2015 Jan-2016 Customer Complaints # of web/Mobile Transactions Social Media ? Jan-2017 Who stopped its relationship with the bank?

ML for Anti-Money Laundering • Issue: AML requirements demand analysis of multiple data: Web channel s Sancti on lists Client data KYC Structur ed data ü The need for highly effective & fast engine Non. Struct ured Data

ATM Cash optimization • Issue: Excess or shortage of cash during days. • Predict the amount of cash need for each ATM each day • A regression (panel data) model is used • Cost optimization following model confidence

Sentiment Analysis – Text Mining • Twitter

Data science challenges • Proper data governance framework • Too many “silos” – data is not pooled for the benefit of the entire organization • Time taken to i) extract and ii) analyze large data sets • High cost of storing and analyzing large data • Limited skilled people for data science • Buy-in from the top