psychometric testing in credit risk assessment · psychometric testing in credit risk assessment...
TRANSCRIPT
Psychometric testing in credit risk assessment
A digital readiness primer and case study
@eflglobal // [email protected]
Agenda
EFL Intro Where EFL works best Digital readiness primer HCD: Designing for the feature phone Equity Bank case study Q&A
Traditional credit scoring is blind to billions of people 3 billion are unscorable, another 3 billion will enter the middle class in the next 20 years
Covered by credit
bureau, 30%
Not covered,
70%
Not covered • Billions of good applicants rejected • Enormous missed opportunity
70%
Source: http://www.ey.com/gl/en/issues/driving-growth/middle-class-growth-in-emerging-markets; http://data.worldbank.org/indicator/IC.CRD.PRVT.ZS
Copyright © 2017. All Rights Reserved. 3
We are entering a new era of lending
1990’s
Loan officers with deep
relationships
Loan officers +
Traditional credit scoring
2000’s Today
Digital approvals without loan officers
+ Traditional and
alternative scoring
2010’s
Loan officers +
Traditional and alternative scoring
And credit scoring needs to adapt to serve more people with end to end digital processes
Predictive power and availability vary by data source Psychometric is only data source with 100% availability – everyone has a personality
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
0% 5% 10% 15% 20% 25% 30% 35% 40% 45%
Predictive Power (Out of sample gini coefficient with comparable model & sample)
Less
Ava
ilabl
e
Mor
e Av
aila
ble
PSYCHOMETRIC DATA
MOBILE DATA
BUREAU DATA ONLINE DATA
Less Predictive More Predictive
Avai
labi
lity
(Hit
rate
on
sam
ple
of lo
w-in
com
e ap
plic
ants
in e
mer
ging
mar
kets
)
Copyright © 2017. All Rights Reserved. 6
EFL is the proven universal score for today’s lenders
• 4 years R&D at Harvard on psychometric, demographic, social & CDR scoring
• 6 years implementing scores with lenders across Africa, Latin America & Asia
Alternative scoring leader • Score anyone, anywhere
with our psychometric and behavioral science based assessment
• No client history or 3rd party database needed
Universal coverage • Published 3rd party
validations by World Bank and others
• Won multiple awards including G-20 SME Finance Challenge, African Business Award for Innovation, and Mahindra Finance Challenge
Third party vetted
Link
Empowering digital and in-person approvals across $1.5B+ and counting in disbursed loans
Copyright © 2017. All Rights Reserved. 7
EFL assesses risk with a short online loan application
Unscorable Applicant Takes EFL assessment on the web, mobile or SMS
EFL Algorithms Analyze the answers and
determine risk level Score Generation Create 3-digit EFL Score and
Score Confidence rating
Continuous Learning EFL receives monthly loan
repayment used to constantly customize and improve
models
1 3
2
4
1
Attitude
Control
Entrepreneurial potential
Interaction metadata
Social Behavior
How we score anyone The character traits our assessment measures
Copyright © 2017. All Rights Reserved. 8
Agenda
EFL Intro Where EFL works best Digital readiness primer HCD: Designing for the feature phone Equity Bank case study Q&A
Copyright © 2017. All Rights Reserved. 10
EFL works best when there’s a clear need for an alternative credit score, and the FI is challenged by:
Low Acceptance Rates (due to lack of data)
High rejection rates due to lack of data – too
many grey zone or thin file clients
High Levels of Default
Default rates over target; hesitant to dramatically decrease acceptance
rate
Costly & Lengthy Loan Application Processes
Loan application process takes time and is resource-
intensive for lender
1. 2. 3.
Copyright © 2017. All Rights Reserved. 11
Use the EFL score to… Our partners typically use our behavioral psychometric score in one of three ways:
Increase Your Acceptance Rate
Approve more people, without taking on more
risk
Decrease Your Default Rate
Decrease and control your default rate, with limited volume impact
Increase Your Operational Efficiency; Lower Application Cost
Fast track: Cut pre-filters
for high-scoring applications
1. 2. 3.
‘Institutional Readiness’ plays a large role in success
Ability to receive a 3-digit, numerical, credit score and use it in their existing credit decision making
process.
Receiving thousands of credit applications each month.
Extra Plus? Ability to combine EFL
score with other credit scores via scorecard or ensemble model
Strong desire across management & teams to
implement and scale solution quickly.
Motivated team, with
willingness and motivation to change existing process
Willingness to try new solutions and consider a departure from
traditional processes.
Willingness to take calculated risks and move in digital direction.
High Volume, Credit Score Ready Driven to Push Innovation Mindset
Copyright © 2017. All Rights Reserved. 12
We’ve noticed certain traits within organizations that allow the EFL digital credit assessment to thrive
Agenda
EFL Intro Where EFL works best Digital readiness primer HCD: Designing for the feature phone Equity Bank case study Q&A
Copyright © 2017. All Rights Reserved. 14
Financial institutions are going digital globally Digital technologies can cut the cost of providing financial services by 80 to 90% and improve scalability
Annual cost to serve one customer in emerging economies, 2014 ($)
Traditional bank branch
Digital
Cost savings due to digitization
Accounts Cash-in, cash-out Transactions Total
Source: McKinsey Global Payments Map; “Fighting poverty, profitably: Transforming the economics of payments to build sustainable, inclusive financial systems,” Bill and Melinda Gates Foundation, September 2013; McKinsey Global Institute analysis
Copyright © 2017. All Rights Reserved.
Going digital? It’s more than building an app What does it mean for a bank to be “digital”?
15
Affordability Risk KYC / ID Servicing Acquisition Online acquisition is
possible, though usually pricey.
Web scrappers are emerging to support digital affordability, however, limited coverage and
inconsistent data remain an issue
Multiple companies, including EFL are proving that risk
profiles can be built via digital
interactions.
Proving ID online is an emerging field,
however, in EM limited databases
coverage is a factor.
Few companies are focused on are looking to outsource collections.
Being digital means being scalable. If you have to add staff to increase lending, you aren’t digital.
But there are challenges to going digital Moving the credit application process from in person with a loan officer to online comes with its own risks
Increased Risk Digital applications can come with increased risk and fraud. Effective credit scoring is key to keeping a healthy portfolio.
Lose Personal Touch No human interaction can dampen the bank-to-customer relationship. Tools are needed to increase human designed interaction.
EFL is your digital loan officer. Efficiently assess risk on anyone, anywhere, online. Build trust with applicants.
Copyright © 2017. All Rights Reserved. 16
Copyright © 2017. All Rights Reserved. 17
Going digital is a process
No existing digital processes
Hybrid model – some parts digitized
Fully digital process; pushing for scale
Add EFL to partially digital process flow, see results in: • Operational efficiency
Use EFL as a stand alone tool, focused on: • Customer Acquisition • Reject Recovery
Integrate EFL into a digital application: • Control risk while
maintaining efficiency.
• Increase acceptance rates. %
DIG
ITA
L
100%
0%
We work with partners at each stage
Agenda
EFL Intro Where EFL works best Digital readiness primer HCD: Designing for the feature phone Equity Bank case study Q&A
Copyright © 2017. All Rights Reserved.
Product Design: Understand Your “Bones”
19
What framework are you working within?
• SMS = a necessarily text-heavy product • Question length limited to 160 chars –containing both question and answer key • Need to battle potential mistrust of the product – concerns of SMS scams • Open answer responses – need to validate and send appropriate responses
Understand your customers: • Design for the mass-market, and try to incorporate the low-common denominator as
much as possible. People will struggle with:
Technology Literacy Understanding Connectivity
Know What You NEED
Honesty
Intelligence
Attitudes
Stability
General Behavior
Fraud
Hypothesize: Why does someone
default?
• Relatively easy to understand (both the question and desired answer format)
• Predictive even within question
number and character limits • Forgiving if users make mistakes
and need to correct answers • Trust building – Reliable product
that is connected to bank and reward believed
Product design PROCESS
1. Field observation & interviews: on
both question framing and content
2. Experimental design: validate/quantify hypotheses from field via more real-life
channel using flow split-testing and within-flow experiments
3. Data analysis of flows and field
interviews: ensure which key metrics are improving and
which aren’t
Watch! How and when are people interacting with their phones?
Get to know your competition – what mobile services are people
using? What do they like / dislike about them?
Build: Fast and easy prototypes Paper print outs of phones with text questions
Talk with people- get them to use your product, explain as little as possible.
Copyright © 2017. All Rights Reserved. 22
Get OUT! 1
“Hey! Can I ask your opinion?”
“What do you think it means?”
START
Copyright © 2017. All Rights Reserved. 23
2 Create the RIGHT data Test many versions of your product, over wide audiences, through various channels
Output: Well-structured quantitative and qualitative data to analyze
CHANNEL EXPERIMENTAL DESIGN
Parallel Flows
Distance Testing
In Person
Single Flows
Test two similar products at the same time, with different versions of the same content you need to test – isolates the “treatment effects”
Intersperse new iterations of product parts within your current product iteration to see if behavior changes – controls for individual differences
Copyright © 2017. All Rights Reserved. 24
3 Explore Your DATA Define metrics for what makes the product “good” – explore the data to see what's working…
• Answer distribution (extreme values vs. well-distributed • Answer timing (confusing questions take longer) • Drop off rates (do certain questions cause people to disengage?) • % of unusable answers
0%
5%
10%
15%
20%
25%
30%
35%
10-point scale Percent scale
Percent of responses with extreme answers or unusable answers 10 Point Scale
Reply with a NUMBER ONLY on a scale from 0 - 10 (0=never 5=sometimes 10=always). Do you show up late or miss appointments?
Percentages What PERCENT (%) of the time do you show up late or miss appointments?
BEF
OR
E A
FTER
…Do we see an improvement in performance metrics as a result of a new iteration?
Copyright © 2017. All Rights Reserved. 25
Final SMS Content: Three Targeted Core Traits
Locus of Control
Opportunism/ Progress mindset
Integrity
Understanding the user’s concept of
honesty and fairness in the context of her
financial life
Understanding the user’s sense of
personal responsibility for and control over
her life
Understanding the user’s perspective on
opportunities and motivations to take risks
to improve her life
Objective: Identify a few core traits that are most relevant to the population. Focus a majority of psychometric content around these to max predictive power.
10%
15%
20%
25%
Locus of Control, Integrity,
Opportunism
Confidence, Optimism,
GSE
% of responses with no valid answer
Copyright © 2017. All Rights Reserved. 26
Maximizing predictive power 1) 3 core traits focused on 2) Granular answer potential (%s)
… through ensuring understandability 1) Answer in percentages 2) Non-hypotheticals – concrete questions / “mini stories” 3) Data quality check – smartphone? Help on survey?
… and building trust 1) Human approach – make people feel comfortable 2) Start with easy, but non-invasive questions (back load more intrusive
demographic questions) 3) Trust this product – start with the person’s name
Reply with a NUMBER in KSH. If you found 100,000 KSH today, how much would you spend on leisure?
Final SMS Product
Agenda
EFL Intro Where EFL works best Digital readiness primer HCD: Designing for the feature phone Equity Bank case study Q&A
Copyright © 2017 EFL. All Rights Reserved. 28
Overview - Scoring on the feature phone Equity extends credit through chat, proves ability to boost acceptance rates
The Challenge High volume, feature phone mobile lending comes with significant risk. Equity needed an innovative solution to accept more of their customers without taking on additional risk. The Engagement EFL designed a new product – taking the learnings of our traditional psychometric content, and transforming them into a 22 SMS question assessment. Instead of being flat-out rejected, low data customers could take the assessment remotely on their feature phones, and gain an opportunity to be approved.
The Results Collecting 24k completed SMS tests, EFL built a custom SMS model, and saw significant predictive power in the data (psychometric & meta). The strong model results indicate potential SMS uses across markets / channels.
Tine, an Equity customer, applies for a loan via her Equitel SIM card.
She’s new to the bank, so she receives the first EFL SMS asking her to start the test.
Tine’s survey is analyzed, and her credit score sent back to Equity in real time. Tine is disbursed!
How does it work?
Copyright © 2017. All Rights Reserved.
Designing the SMS Pilot – Process & Criteria
29
…TO VALIDATE TWO MAIN HYPOTHESIS
Build & Integrate SMS Platform
Design SMS Test 1
2 Run Pilot;
Gather Results 3 Build Custom
Model 4
HYPOTHESIS MEASUREMENT DESIGN
1. The SMS channel will facilitate a high
conversion rate. UX of SMS
Test
Test Completion Rate (%s in
funnel)
2. SMS psychometric testing will enable Equity
to better segment and predict risk
Psychometric Assessment
Predictive Power of SMS
Model
Out-of-Sample Model
Validation 5
FIVE STEPS UNDERTAKEN TO COMPLETE THE SMS PILOT… Equity & EFL partnered in 2013 to develop a custom
feature phone credit scoring tool
SMS – USSD ~20 psychometric questions
4 test sections Any phone, any location Only need to score once
Whatpercent(%)ofbusinessescheattheir
clients?
Derives predictive power from: 1. Psychometric answers 2. Timing data 3. Free-text input data 4. Self-reported demographic data
Assess willingness to repay through a conversation Through EFL’s chatbot platform, we converse with your clients, one SMS at a time. Through an automated conversation, we ask our traditional psychometric questions – straight to their phone. Predictive, real-time, credit scores available in minutes.
Copyright © 2017. All Rights Reserved. 30
© 2016, EFL Global. All rights reserved. 31
Benson,anewcustomer,appliesforloan.Normally,hewouldberejected.
Instead,heimmidietelyreceivesaninia9onSMSmaskedfromthe
lender
Bensonengages,conversa9onstyle,inthe
SMStest.
CREDITSCORE337
CreditScorereturnedtolender
PsychometricsBehavioral – Text Analysis; Timer Analysis
What time of day are they responding?What are response answer patterns?
Do they understand the question?How is their spelling? Grammar?
What words or phrases do they use?
Seamlessly assess an applicant’s character PR
ODU
CT F
LOW
Copyright © 2017. All Rights Reserved.
EFL SMS test seamlessly fits in with existing UX mobile lending product flow Before rejecting low-data and new applicants, Equity can digitally score and asses risk on borrowers.
32
Benson,acustomerofEquityBank,
appliesforloanviaEquitelSIMSDK/
USSDmenu-
Immidietelyreceivesinia9on
SMSfromdedicatedEquity
shortcode-
BensonrepliestotheSMS(zero-
rated)andbeginsthetest
Bensonengages,conversa9onstyle,intheSMStest.
-Answerstestin4secPons.SMSreminderssenttokeepBenson
focused.
Totaltest)me<15minutes,takenatonceoroverhours/
days
CREDITSCORE337
CreditScore-
Returnedinreal-PmethroughAPIto
Equity’sCU
Copyright © 2017. All Rights Reserved.
Results: Predictive Model; Strong Usability
33
1 Worst 2 3 4 5 Best
Def
ault
Rat
e
EFL Score Quartiles
bad45_m3 bad30_m3 bad60_m3
Customized Model With 24k tests completed, EFL built a customized model for SMS – achieving a predictive power of 63.7% AUC
Usability validated through low drop off rates.
Average completion rate of 78%, with an average test time of 33 minutes.
80% of drop-off occurred within first 4 questions. 56% of tests were completed within 15 minutes.
Out of Sample AUC: 63.7
Low Default
High Default
1 Worst Scoring
4 Best Scoring
3
2
Def
ault
Rat
e
Copyright © 2017. All Rights Reserved.
Ability to increase acceptance rates by 20% when combined with internal behavioral score
EFL & Equity’s Internal Behavioral Score
34
Combining the EFL score with an internal behavioral (transactional) score: boost acceptance rate by 20% without taking on additional risk
Ensemble Experian EFL
CumulatedDefaultRates
20%
Def
ault
Rat
e
Acceptance Rate
1x 1.2x
Low Default
High Default
High Low
Behavioral
Copyright © 2017. All Rights Reserved.
EFL finds different bads, and different goods, than traditional credit bureau scores
35
EFL & Traditional Bureau Score Combining the EFL score with a traditional bureau score: boost acceptance rate by 53% without taking on additional risk
Ensemble Bureu EFL
Cumulated Default Rates
Def
ault
Rat
e
Acceptance Rate
Low Default
High Default
High Acceptance
Low Acceptance
Bureau
53%
Agenda
EFL Intro Where EFL works best Digital readiness primer HCD: Designing for the feature phone Equity Bank case study Q&A