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 100% 80% Percent in US with 60% loans but have never defaulted** 40% Average lender approval rates* 20% Defaults 0% Traditional Underwriting Modern Data Science * Source: Prosper, Lending Club **Source: Upstart data study with TransUnion

So why doesn’t everyone do it? Real data Regulatory risk is science is hard daunting

So you want to add a new variable? Some helpful attributes • Broadly available • Decade+ of training data • Easily verifiable • Unbiased and legal Hint: Facebook is not the answer!

We’ve assembled a collection of variables that are more predictive than the entire credit bureau file 20 3-Year Student Loan Default Rate (%) 15 10 5 800 1000 1200 1400 1600 School ranking

And by layering all of these variables together, we can make smarter credit decisions instantly Default rate of “best 40%” from sample population 15 12 Default Rate (%) 9 6 3 0 Random Financial variables Financial variables 
 Financial variables 
 Financial variables Obtained a degree Obtained a degree 
 Obtained a degree School ranking 
 School ranking Major Major SAT/GPA Data from NCES National Education Longitudinal Study

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 Upstart Traditional Credit file • Income • Occupation • Employer • Work Experience • Degrees • Schools • GPA • Test Scores • Credit file • Income Variables Job Offers • Cost of Living • etc. Continuous decision logic, cross-validated logistic Black/white decision logic, regression, higher-order variables, random forest, Methods simple regression monte carlo methods, ensemble learning Lenders 2-3x per year, Automated training, Learning FICO 2-3x per decade daily updates Speed

When you’re building a disruptive credit model, verification of inputs is essential Upstart 100% Borrower income verified 100% Borrower education verified 100% Borrower savings verified 100% Verification phone call

Proof in the pudding: steadily increasing approval rates and consistent investor returns Approval Rate of Control Group IRR by Origination Month 30% 25% 20% 15% 10% 5% 0% 4 4 4 4 4 4 4 4 5 5 5 5 5 5 5 5 5 5 5 5 6 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 Y N L G P T V C N B R R Y N L G P T V C N A U C A U C E O E E A P E O E U U U A U A M J S O F M J S O D A D A M A N J N J J J

Our model has learned quickly, with each cohort performing better than the prior Cohort # Originated % DQ121+ 852 5.40% Q3 2014 1559 4.49% Q4 2014 2365 2.88% Q1 2015 3356 2.68% Q2 2015 5109 1.23% Q3 2015 7163 0.06% Q4 2015

Our delinquencies by loan grade also provide evidence that we’re accurately pricing our loans Average Age Modeled % Loan Grade # Originated % DQ121+ (Months) DQ121+ 21 12.6 0.00% 0.02% AAA 1391 10.7 0.14% 0.15% AA 5052 9.8 0.61% 0.46% A 4639 10.4 2.00% 1.31% B 2578 9.7 2.48% 2.22% C 3795 9.1 3.98% 3.70% D 639 5.4 0.94% 0.94% E

“Sounds great, but my lawyers say no!” - You

So you give loans to wealthy grads from elite Q: schools? No. Less than 2% of Upstart borrowers come from elite A: schools. And wealthy people don’t need our loans.

Your average borrower is 28 years old - are you Q: biased against older borrowers? No. In fact, all else being equal, an applicant with A: longer credit history will get a lower rate on Upstart.

Does your system discriminate against people Q: based on race, gender, or other protected classes? No. Using a tool provided by the CFPB, we were able to A: demonstrate that our model demonstrates no statistical bias with respect to race or gender.

All successful credit models are based on the same tried & true concepts f ( ( Financial Capacity X Propensity = to Repay to Repay Income Personal Characteristics Earning potential Credit history • • Unemployment potential Personal responsibility • • Awareness of credit score • Expenses Support Network Debt obligations • Network connectedness • Living expenses • Backstop financial support • Spending habits • Assets Available to service debt • … but modern data science can make these concepts better

Success in our case means reducing the price of credit to 65M underserved borrowers Upstart Lending Club Percent of borrowers Borrower age

Thank you! dave@upstart.com � @davegirouard �