Mining Intelligent Insights: AI/ML for Financial Services

© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.

Américo de Paula

Solutions Architecture Manager - LATAM

Mining intelligent insights:

AI/ML for Financial Services

Breakthrough

advances

Optimization and

automation

AI and ML enable innovation at scale…

New features

for existing products

“After decades of false starts, artificial intelligence is on the verge of a

breakthrough, with the latest progress propelled by machine learning.”McKinsey Global Institute, Artificial Intelligence The Next Digital Frontier? June 2017

…and could revolutionize Financial Services

The Economist, May 25, 2017

“AI could contribute up to $15.7

trillion to the global economy in

2030…. Healthcare, automotive

and financial services are the

sectors with the greatest potential

for product enhancement and

disruption due to AI.”

Sizing the prize: What’s the real value of AI for your

business and how can you capitalise? PwC report,

June 2017

Immense opportunities… …but huge risks of disruption

The potential impact of AI/ML is enterprise-wide

Compliance,

Surveillance, and

Fraud Detection

Pricing and

Product

Recommendation

Document

Processing

Trading Customer

Experience

• Credit card/account fraud

detection

• Anti-money laundering/

Sanctions

• Investigations optimization

• Sales practices/

transaction surveillance

• Compliance processes

optimization

• Regulatory mapping

• Enhanced customer

service through voice

services and chatbots

• Call center

optimization

• Personal financial

management

• Loan/Insurance

underwriting

• Sales/recommendations of

financial products

• Credit assessments

• Contract ingestion and

analytics

• Financial information

extraction

• Common financial

instrument taxonomy

• Corporate actions

• Portfolio management/

robo-advising

• Algorithmic trading

• Sentiment/news analysis

• Geospatial image analysis

• Predictive grid computing

capacity management

AI/ML use cases are gaining traction in Financial Services

But overall the industry has been slow to invest

Source: McKinsey Global Institute, Artificial

Intelligence The Next Digital Frontier?

An ambivalent response to AI

• Strong overall appetite for adopting AI

• History of digital investment and

strong foundation for integrating AI

technologies

• Large volumes of data to support

model training and development

• Comparatively low investment in AI

What is preventing the industry from moving ahead?

AI/ML expertise is

rare

Building and

scaling AI/ML

technology is hard

Deploying and operating

models in production is

time-consuming and

expensive

A lack of cost-effective,

easy-to-use, and

scalable AI/ML services

AWS offers a range of solutions to make AI/ML more accessible

PollyLex Rekognition

Deep Learning FrameworksMachine Learning Platforms Amazon AI/ML Services

Usability/simplicity:

leverages AWS AI/ML expertise

Greater control:

customer-specific models

Amazon ML

Spark & EMR

Kinesis

Batch

ECS

Customization of

offerings at scale

More personal and

efficient customer

interactions

Operational

efficiencies

Novel investment/

trading opportunities

Benefits for Financial Services Institutions

and others...

Our deep experience with AI/ML differentiates our services

Product

recommendation

engine

Robot-enabled

fulfillment

centers

New

product

categories

Amazon has invested in AI/ML since our

inception, and we share our knowledge and

capabilities with our customers20171995

Natural language

processing-supported

contact centers

ML-driven supply

chain and

capacity planning

Checkout-free

shopping

using deep learning


And today, enterprises across industries run AI/ML on AWS



AWS Machine Learn ing Stack

Frameworks &

Infrastructure

AWS Deep Learning AMI

GPU(P3 Instances)

MobileCPU

(C5 Instances)

IoT

(Greengrass)

Vision:

Rekognition Image

Rekognition Video

Speech:

Polly

Transcribe

Language:

Lex Translate

Comprehend

Apache

MXNetPyTorch

Cognitive

ToolkitKeras

Caffe2

& CaffeTensorFlow Gluon

Application

Services

Platform

Services

Amazon Machine

Learning

Mechanical

Turk

Spark &

EMR

Amazon

SageMakerAWS

DeepLens

Fraud.net is running AI/ML on AWS to predict financial crime

“Amazon Machine Learning helps us reduce complexity and make sense of emerging fraud patterns.

We can see correlations we wouldn’t have been able to see otherwise and answer questions it would

have taken us way too long to answer ourselves.

”Fraud.net is the world’s leading

crowdsourced fraud prevention

platform, aggregating and

analyzing large amounts of

fraud data from thousands of

online merchants in real time.

The platform protects more

than 2 percent of all U.S. e-

commerce.

- Oliver Clark, CTO, Fraud.net• To address its scalability needs, Fraud.net chose AWS to host its customer

platform, relying on services including DynamoDB, Lambda, S3, and Redshift

• Recently, Fraud.net started using Amazon Machine Learning, which helps its

developers build models and enables the use of APIs to get predictions for

applications without having to deploy prediction generation code

• Fraud.net can now easily launch and train new machine-learning models to target

evolving forms of fraud

• Using AWS, Fraud.net can maintain its fast application response times of under

200 milliseconds and save its customers about $1 million a week through fraud

detection and prevention

BuildFax uses Amazon ML to help insurers avoid losses

“Amazon Machine Learning democratizes the process of building predictive models. It’s easy and fast

to use and has machine-learning best practices encapsulated in the product, which lets us deliver

results significantly faster than in the past.

”BuildFax aggregates dispersed

building permit data from across

the United States and provides it

to other businesses, especially

insurance companies, and

economic analysts. The

company also tracks trends like

housing remodels and new

commercial construction.

- Joe Emison, Founder & Chief Technology Officer, BuildFax

• BuildFax’s core customer base is insurance companies, which spend billions of

dollars annually on roof losses

• The company initially built predictive models based on ZIP codes and other general

data, but building the models was complex and the results did not provide enough

differentiators

• BuildFax now uses Amazon Machine Learning to provide roof-age and job-cost

estimations for insurers and builders, with property-specific values that don’t need

to rely on broad, ZIP code-level estimate

• Models that previously took six months or longer to create are now complete in four

weeks or fewer

@dmbanga @apwenchel© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.

Adam Wenchel, VP of AI


Remember the infrastructure Built on AWS ML takes time, and technical debt

Democratize AI, responsibly Maximize scarce experts’ productivity There is a lot more than the ML model

Machine Learning @ Capita l One


The road to production is long and arduous

Prepare for the journey


Model selection and Hyper-

parameter Tuning

GPU Optimization Self-service

Production is a hard place Continuous Monitoring Rapid ML model refit/deploy

Capital One AI: The Road to Production


Machine Learning Technology @ Capital One

Built on AWS


Amazon AI AI Platforms AI Engines & InfraAI Services

Capital One AI: Democratizing ML in a Well-Managed Way

Experiment

ManagementLogging Versioning Reproducibility Monitoring

Optimize &

ScaleModel Optimization GPU saturation

Data/Compute

coordination

Capital O

ne


Organizing for Success

You want a Data Science Team, then what ?


People

Machine Learning is inherently Interdisciplinary

Physical

ScientistsAnalysts

Architects/

Integrators

Computer

Scientists


People

Physical

ScientistsAnalysts

Architects/

Integrators

Computer

Scientists

Used to solving

problems by

applying ML to

sensors data



People


Physical

ScientistsAnalysts

Architects/

Integrators

Computer

Scientists

Used to solving

problems by

applying ML to

sensors data

Understand

Business

Problem

Framing


People


Physical

ScientistsAnalysts

Architects/

Integrators

Computer

Scientists

Used to solving

problems by

applying ML to

sensors data

Understand

Business

Problem

Framing

Understand

Computer

Logic and ML

Science


People


Physical

ScientistsAnalysts

Architects/

Integrators

Computer

Scientists

Used to solving

problems by

applying ML to

sensors data

Understand

Business

Problem

Framing

Understand

Computer

Logic and ML

Science

Can put

everything

together


Capital One AI: How We Organize for Success

Create a positive work environment for ML

talent

Centralize new technologiesAttract and motivate the right talent



Opportunity: Transform Financial Services

Increase efficiency, creating value for our

customers

New ways of empowering customers to

take control of their financial livesCreate amazing customer experiences


Nima Najafi

Scotiabank Senior Manager

Data Science & Model Innovation

Opt imiz ing Payments

Col lect ions wi th Conta iners

and Machine Learn ing


Who is Scot iabank?

• Third-largest bank in Canada

• On Top 10 List Of World's Strongest Banks

• Providing services in North America, Latin America, the

Caribbean and Central America, and Asia-Pacific

• 88,000 employees

• $907B assets (as at July 31, 2016)

• Dedicated to helping its 23 million customers become better

off through a broad range of advice, products, and services,

including personal and commercial banking, wealth

management and private banking, corporate and investment

banking, and capital markets


Business problem:

Credit card portfolio has been growing in the past

years, requiring more effective collections to handle

the growing volume.

Bank strategy:

Grow digital presence in order to better serve our

customers and play a leading role in transforming the

banking sector for the digital age.

What were we t ry ing to so lve?

Acquisition

Account management

Collections


Why star t wi th payments co l lect ions?

Quarter All Credit Cards Other

2017 Q2 2.21 2.47 1.96

2017 Q1 2.18 2.42 1.97

2016 Q4 2.16 2.37 1.97

2016 Q3 2.08 2.29 1.91

2016 Q2 2.04 2.20 1.89

2016 Q1 1.99 2.16 1.83

US delinquency rates

In Canada, the average consumer

balance for credit cards grew by

2.07% from Q2 2016 to Q2 2017.

Canadian delinquency rates


What came before deep learn ing?

• Models were more basic (logistic regression/linear

regression)

• Process was manual and time consuming

• Complete data sources weren’t readily available

• Less accurate predictions

• Additional opportunities for financial savings


Deep learning model

Written in Python

Takes ~300 credit bureau and internal variables as inputs

Hyper-parameters (learning rate, batch size, #layers, #npl, #epochs, activation

functions, ...) were selected via grid search

10-fold cross validation with stratification

Percentage of charge-offs captured in bottom 10%

Out-of-time

datasetDeepLearni.ng model

Incumbent (in-house)

modelLift

Nov-14 54.5% 45.3% 20.3%

Jan-15 55.0% 46.2% 19.1%

May-15 55.4% 46.8% 18.2%

Aug-15 50.1% 42.1% 19.0%


Results!• Create/orchestrate/terminate the application from Amazon EC2

Container Registry (Amazon ECR) to Amazon EC2 Container Service

(Amazon ECS)

• Synchronize the data between Amazon Simple Storage Service

(Amazon S3) and GitHub LFS

• Secure the cluster through IAM roles (cluster only accepts traffic from

Scotiabank IPs)

• Application is fully containerized and versioned

• Changes to the cluster infrastructure are managed by pull requests only

• Rapid development and traceability through “everything through code”

paradigm


Success storyamzn.to/FSV305-WSJ

http://amzn.to/FSV305-WSJ

Alexa...

25,000+ skills

The Alexa Family


A Li t t le About Al ly…

…A digitally disruptive, diversified financial services firm

with a full suite of financial products and services

including banking, auto finance, and mortgage offerings.

Beyond its services, Ally is known for its unique culture,

straight forward approach and customer-centric business

philosophy.


Al ly Business L ines

• Full-spectrum provider

• New, used, leasing

• Floor plan and

commercial services

• Dealer online services

• Auto-focused insurance

business

• Direct banking platform

• Deposit products:

savings, checking, CDs,

IRAs

• Mobile banking

• Credit card

• Mortgage

• Financing for mid-market

companies in technology,

healthcare, retail, and

automotive

• Digital portfolio

management platform

• Ally Invest

Auto Finance Ally Bank Corporate FinanceWealth Management

& Online Brokerage


The Design Journey



* Solved through fine-tuning utterances


Technology @ Work

Amazon

Alexa

Amazon Developer Portal

Voice Interaction Model

Ally Interactive API

(AWS Lambda)

Transactions

Accounts

Welcome

Transfers

Rates

CurrenSee

Amazon

CloudWatch

AWS X-Ray

AWS IAM

AWS KMS

Shared

Services

AuthN

Token Secure

Gateway

Ally

Microservices

Enrollment APIs

Step Up Auth

AuthN APIs

Banking APIs

Notify Services


The Al ly Ski l l

"ALEXA, OPEN ALLY”

“GET MY BALANCE” “HOW MUCH DID I SPEND LAST MONDAY?”

“ALEXA, TRANSFER $10 FROM MY

ALLY BANK [PRODUCT NAME] ACCOUNT

TO MY [ELIGIBLE EXTERNAL ACCOUNT]”

TO VERIFY IT'S YOU, TELL ME YOUR

6-DIGIT PASSCODE

“WHAT'S TODAY’S RATE FOR A 12 MONTH HIGH YIELD CD”

“CONVERT $300 INTO CURRENSEE”“ALEXA, ASK ALLY FOR HELP.”

Go to amzn.to/takeselfie

Américo de PaulaSolutions Architecture Manager

[email protected]

Worldwide | N. America | LATAM | UK/IR | EMEA | APAC | Japan | China