Sie sind auf Seite 1von 10

Bosch Presentation

BHANUTEJA ARYASOMAYAJULA
About me

• Graduated in Mechanical Engineering from IIT Kharagpur in 2016

• Joined American Express as a Data Scientist in Aug ‘16

• Got promoted to Assistant Manager in May ‘18

• 3 years of experience in building machine learning models related to credit risk

• Enjoy travelling, playing badminton & watching movies in free time


What do I do at AmEx?

• Build Machine Learning models to predict

Business Partners in EMEA & APAC


 Default risk of AmEx credit card holders

 Credit Limit of new applicants

 Profitability of customers
Data Sources dealt with

• Information from credit card application

• Credit Bureau Information

• Digital attributes & 3rd party information

• Amex intelligence
What is default risk?

• Probability that the person doesn’t make his credit card payment (defaults)
post the due date (P)

• Statement Balance – 2000 $

• Due Date – May 1st 2019

• P is the dependent variable in the default risk model


Drivers of default risk – How much is P?

• Has a good bureau score • Has a poor bureau score

• Has limited debt • Has a lot of debt

• High income & never misses loan payments • Limited earnings, doesn’t make
payments on time

P = 0.01 P = 0.5
Very low chance of defaulting High chance of defaulting
Default risk model – Impact & Achievements

• Built & implemented 6 GBM models from scratch in a span of 6 months

• Achieved a GINI of 60%, a 30% lift from the previous version of models

• Net profit of 5 MM USD over a span of 2 years due to loss savings created
from better prediction of default risk
Modelling credit limit

• Income = 100,000 USD • Income = 20,000 USD

• Has good bureau score • Has a poor bureau score

• Low P value from default model • High P value from default model

Credit Limit = 10000 USD Credit Limit = 1500 USD


Credit limit model – Impact & Achievements

• Built & implemented 2 kNN models to assign credit limits to more than half a
million accounts per year

• Net profit of 10 MM USD in 2 years as a result of accurate credit limit


assignment
Tools & Techniques used

TOOLS TECHNIQUES

• Python • Linear Logistic & Non-Linear Regressions


• SAS
 pandas • Decision Trees
• SQL
 numpy
• Random Forests
 scikit-learn
• Hive
• Nearest Neighbors
 seaborn • MS Excel
 matplotlib • Clustering

• Neural Networks

Das könnte Ihnen auch gefallen