a new approach to consumer credit
TRANSCRIPT
A New Approach to Consumer Credit Analytics
We have created a better consumer credit analytics paradigm
Focused on developing a deep understanding of who the borrower is – their character
• Priorities and goals
• Attitudes and concerns
• Competencies
Anchored by a groundbreaking analytical engine
• Proprietary machine-learning framework
• Coupled with a nuanced understanding of risk drivers…
• …To identify deep character traits such as• Spending priorities: more long term to more impulsive
• Attitude to risk: more risk averse to more risk tolerant
• Planning competence: more proactive to more reactive
Our approach generates outstanding, demonstrable results
• Calibrated on Lending Club data – the best publicly available credit data source…
• ...It outperforms both Lending Club’s (sub-)grades and FICO scores in predicting default (see Appendix for details)
• Benefits are greatest in market segments with the most untapped opportunities: sub-prime and thin credit file
Lending Club (sub)grade
IRR - Garden methodology
Downside IRR - Garden methodology
Lending Club IRRDownside Lending Club
IRR
C3 10.38% 7.08% 7.63% 1.59%
D3 12.38% 6.07% 9.51% 0.92%
E3 12.16% 2.69% 8.92% -3.51%
F3 13.42% 0.72% 7.69% -10.06%
G3 12.23% -5.45% 8.68% -12.09%
Applicable across a broad range of credit-linked asset classes
• Consumer unsecured installment loans
• Credit card receivables
• Auto loans
• Residential mortgages
• SMB loans
Current credit methodologies predict future outcomes by scoring past outcomes
• But the relationship between past outcomes and future outcomes is one of correlation, not causality
Past outcomes• Delinquencies• Inquiries• …
Future outcomes• Delinquencies• Inquiries• …
Causation
Future outcomes = score(Past outcomes, Context)
Context• Income, wealth• Life events• …
Correlation
Past outcomes are merely imperfect proxies for deeper personal character traits
• The fundamental question is: Who is the borrower?
Past outcomes• Delinquencies• Inquiries• …
Character• Risk aversion• Impulsivity• …
Future outcomes• Delinquencies• Inquiries• …
CausationContext• Income, wealth• Life events• …
Deduction
Deduction
We have developed a methodology to identify character from traditional or alternative data
• On a foundation of…• Structural modeling of drivers of risk
• Deductive machine learning algorithms
• …We have built a better credit model
• Our approach enables a far greater leveraging of both traditional and alternative data
Benefits vs status quo Bases
Lower prediction error More efficient use of data
Less overfitting bias Less sensitive to data noise
More forward-looking predictions Causation not correlation
The benefits go beyond credit underwriting
• Customer marketing and acquisition
• Fraud detection
• Risk management
• Customer engagement and retention
• Customer development
Appendix: Performance
We have developed a differentiated underwriting framework that is best in class
• Calibrated on Lending Club data, the best publicly available data source of actual credit histories
• All loans issued in 2014 (approx. 235,000) so that we could track defaults 1 year out (i.e. in 2015)
• Default rates indicated in this appendix are for the 12 months from loan issuance
• Model trained on 70% of the data (approx. 165,000)
• 30% of the data held out for testing (approx. 70,000)
• All results generated on the held out data: no model overfitting bias
Our framework outperforms both Lending Club and FICO-based credit assessments
• Lending Club assigns each loan to 1 of 7 credit grades, A through G, based on their proprietary risk assessment• ‘A’ loans are deemed the least risky and ‘G’ loans the most
• In addition all loans are assigned 1 of 5 sub-grades within each grade to further differentiate risk levels• E.g. among C grade loans, C1 loans are the least risky and C5 the most
• We have compared our results to Lending Club results based on grade and sub-grade
Moreover, our machine learning methodology should consistently outperform even the most advanced alternatives
• In particular, it should outperform conventional implementations of Random Forest and Neural Network models
• Consequence of the fact that character is a complex, latent, constellation of borrower features
• Quantity of data needed to train alternative models simply does not exist
Interpretation of results graphs
• In the following graphs, the horizontal axis indicates the % of loans that are “accepted” for underwriting. Only the best x% are accepted• In the case of Lending Club grading (“LC” line), the selection is based on sub-grades: e.g.
first the B1 sub-grades are accepted, then the B2 sub-grades and so forth
• Our proprietary model determines the acceptance criterion reflected in the “Proprietary” line
• The vertical axis indicates the reduction in default rates relative to random acceptance of x% of loans• Higher is better
• As the graphs suggest, the performance benefits of our approach are greatest for the riskiest loans
• All results were calculated on held out data
Our approach would correctly reclassify many sub-prime loans based on risk
37%
18%
44%
Reclassification of Sub-Prime1 Borrowers by Whole Lending Club Grade
Upgrade Downgrade No change
1Sub-prime equates to a FICO score of 680 or below