Welcome to a Post-FICO World! Consumer credit modeling relies on - - PowerPoint PPT Presentation

Welcome to a Post-FICO World!

Consumer credit modeling relies on data and analytics that haven’t changed in decades

A smarter prime lender could approve almost twice as many borrowers and yet have fewer defaults

Traditional Underwriting Modern Data Science

0% 20% 40% 60% 80% 100%

Average lender approval rates*

Defaults

Percent in US with loans but have never defaulted**

* Source: Prosper, Lending Club **Source: Upstart data study with TransUnion

So why doesn’t everyone do it?

Real data science is hard Regulatory risk is daunting

So you want to add a new variable?

• Broadly available
• Decade+ of training data
• Easily verifiable
• Unbiased and legal

Hint: Facebook is not the answer!

Some helpful attributes

3-Year Student Loan Default Rate (%) School ranking

15 10 5 800 1000 1200 1400 1600

We’ve assembled a collection of variables that are more predictive than the entire credit bureau file

20

Default rate of “best 40%” from sample population

Default Rate (%) 3 6 9 12 15

Random Financial variables Financial variables 
 Obtained a degree Financial variables 
 Obtained a degree 
 School ranking 
 Major Financial variables Obtained a degree School ranking Major SAT/GPA Data from NCES National Education Longitudinal Study

And by layering all of these variables together, we can make smarter credit decisions instantly

Data that is predictive in a recession is even more valuable

Unemployment rate by level of education

A disruptive credit model requires unique predictive data, better math, and faster learning

Traditional

Upstart

Variables

Credit file • Income Credit file • Income • Occupation • Employer • Work Experience • Degrees • Schools • GPA • Test Scores • Job Offers • Cost of Living • etc.

Methods

Black/white decision logic, simple regression Continuous decision logic, cross-validated logistic regression, higher-order variables, random forest, monte carlo methods, ensemble learning

Learning Speed

Lenders 2-3x per year, FICO 2-3x per decade Automated training, daily updates

When you’re building a disruptive credit model, verification of inputs is essential

Upstart

Borrower income verified 100% Borrower education verified 100% Borrower savings verified 100% Verification phone call 100%

0% 5% 10% 15% 20% 25% 30%

M A Y 2 1 4 J U N 2 1 4 J U L 2 1 4 A U G 2 1 4 S E P 2 1 4 O C T 2 1 4 N O V 2 1 4 D E C 2 1 4 J A N 2 1 5 F E B 2 1 5 M A R 2 1 5 A P R 2 1 5 M A Y 2 1 5 J U N 2 1 5 J U L 2 1 5 A U G 2 1 5 S E P 2 1 5 O C T 2 1 5 N O V 2 1 5 D E C 2 1 5 J A N 2 1 6

Approval Rate of Control Group IRR by Origination Month

Proof in the pudding: steadily increasing approval rates and consistent investor returns

Our model has learned quickly, with each cohort performing better than the prior

Cohort # Originated % DQ121+ Q3 2014 852 5.40% Q4 2014 1559 4.49% Q1 2015 2365 2.88% Q2 2015 3356 2.68% Q3 2015 5109 1.23% Q4 2015 7163 0.06%

Our delinquencies by loan grade also provide evidence that we’re accurately pricing our loans

Loan Grade # Originated Average Age (Months) % DQ121+ Modeled % DQ121+ AAA 21 12.6 0.00% 0.02% AA 1391 10.7 0.14% 0.15% A 5052 9.8 0.61% 0.46% B 4639 10.4 2.00% 1.31% C 2578 9.7 2.48% 2.22% D 3795 9.1 3.98% 3.70% E 639 5.4 0.94% 0.94%

“Sounds great, but my lawyers say no!”

• You

So you give loans to wealthy grads from elite schools?

• No. Less than 2% of Upstart borrowers come from elite
• schools. And wealthy people don’t need our loans.

Q: A:

Your average borrower is 28 years old - are you biased against older borrowers?

• No. In fact, all else being equal, an applicant with

longer credit history will get a lower rate on Upstart.

Q: A:

Does your system discriminate against people based on race, gender, or other protected classes?

• No. Using a tool provided by the CFPB, we were able to

demonstrate that our model demonstrates no statistical bias with respect to race or gender.

Q: A:

X

Financial Capacity to Repay Propensity to Repay

( (

=

All successful credit models are based on the same tried & true concepts

f

Income

• Earning potential
• Unemployment potential

Expenses

• Debt obligations
• Living expenses
• Spending habits

Assets

• Available to service debt

Personal Characteristics

• Credit history
• Personal responsibility
• Awareness of credit score

Support Network

• Network connectedness
• Backstop financial support

… but modern data science can make these concepts better

Success in our case means reducing the price of credit to 65M underserved borrowers

Percent of borrowers Borrower age Upstart Lending Club

Thank you!

• dave@upstart.com

@davegirouard