credit risk modeling for alternative lending

Keith Shields: Chief Analytics Officer - MagnifyBruce Lund, Ph.D.: Director, Thought Leadership - Magnify

1

Credit Risk Modeling for Alternative Lending

Copyright ©2012 Marketing Associates LLC. All rights reserved. Magnify is a division of Marketing Associates, LLC.

magnifyAnalytic Solutions magnify • simplify • amplify

2


Speaker Intro

Keith Shields 20+ years in financial services, developing and deploying statistical models to predict

credit risk across multiple asset classes

CAO Magnify & SVP Loan Science: 7 years of Analytic Services and Solutions

14 years at Ford Motor Credit, Director of Global Analytics

2+ years at Credit Acceptance Corp, VP Portfolio Management

Dr. Bruce Lund 22 years at Ford Marketing: created the Customer Knowledge System, a DW that

houses customer data, predicts customer behavior, and steers 1-1 marketing efforts for 30+ million Ford customers

Former Math professor

Analytics “Godfather” of Magnify




Tying to the Abstract…

Key Elements: Credit Risk ModelingScorecards and Cutoff ScoresAlternative Lending, Marketplace Lending, and FintechCommon predictive modeling conventionsRegulatory PressureThe Demand for Creativity and DisruptionWhat Makes “Alternative Lending Analytics” Different from “Banking Analytics”?The Continued Relevance of Logistic Regression“It is not the model fitting techniques that need to change, as much as it is the treatment of samples, potential predictors, and model refits”.

3Copyright ©2012 Marketing Associates LLC. All rights reserved. Magnify is a division of Marketing Associates, LLC.


4


Definition and Function A credit risk model is, usually, a linear statistical model that uses combination

of credit, contract, and personal attributes to predict the likelihood that a loan applicant will default (fail to pay back the loan).

“Attributes” = “Characteristics” = Independent variables Credit Score Loan-to-Value Ratio (LTV)…insert diatribe on 2008 mortgage disaster Payment-to-Income Ratio (PTI) Revolving Utilization Recent Credit Inquiries

Integral part of how lending is done. PD’s, Basel III, SR 11-7 Is lending ever done without models? The Big Short…


Credit Risk Modeling


5


History Bill Fair, Earl Isaac…”FICO” Credit Scoring System: Credit Risk Models are converted to

“scorecards” by creating a point system whereby the parameter estimates of the credit risk model are multiplied by the possible values of the independent variables to create points.

Example: Y = b0 + b1*x1. Suppose x1 can take on a value of 1, 2, or 3. Scorecard:


Credit Risk Modeling

X1 Points1 B1 * 1 * 1002 B1 * 2 * 1003 B1 * 3 * 100

Note: Typically we transform X1 first


7


Example: Assume we make $1,000 profit on “good” loans and suffer a $10,000 loss on “bad” loans:


Scorecards and Cutoff Scores

Credit Score% of

Applicants Good Rate Loss (Bad) RateExpectedNet Profit

Cumulative Approval

RateCumulative Loss Rate

500 40% 99.5% 0.5% $945 40% 0.5%400 20% 96.0% 4.0% $560 60% 1.7%300 15% 92.0% 8.0% $120 75% 2.9%200 15% 87.0% 13.0% -$430 90% 4.6%100 10% 75.0% 25.0% -$1,750 100% 6.7%

• Cutoff score = 300. We lose money on borrowers who score below 300.• We will decline roughly 25% of our applicants.• Our loss rate will be 2.9% (as opposed to 6.7% if we approved everyone).• Our profit per contract is $677 (as opposed to $268 if we approved everything).

Some simple math on each credit tranche leads us to very useful conclusions:

(40%*.5%+20%*4%+15%*8%)/75%





Key Elements:

Credit Risk ModelingScorecards and Cutoff ScoresAlternative Lending, Marketplace Lending, and FintechCommon predictive modeling conventionsRegulatory PressureThe Demand for Creativity and DisruptionWhat Makes “Alternative Lending Analytics” Different from “Banking Analytics”?The Continued Relevance of Logistic Regression“It is not the model fitting techniques that need to change, as much as it is the treatment of samples, potential predictors, and model refits”.

8


9


Definition and Players The simplest definition is that an alternative lender is any lender that

isn’t a bank (or captive, in my opinion) . Historically the customers of alternative lenders are people, or small

businesses, who are unable to secure a bank loan. This is still largely the case, but not nearly to the extent it once was.


Alternative Lending


10


Modern Nuances and New Entrants Recently many start-up alternative lenders have distinguished themselves

from banks based on heavier use of technology, removal of traditional loan structures, and tapping into overlooked sources of collateral and non-traditional funding sources: Technology: Fintech platforms Removal of traditional repayment plans: Merchant Cash Advance, Payday

Loans, and Uber Overlooked sources of collateral: merchant invoices Non-traditional funding sources: Peer-to-Peer (P2P) or “Marketplace”

Lending


Alternative Lending


11


Is there a place for us (statisticians, data scientists, analysts)? In a big way yes:

Huge expenditures on alternative data (Yodlee, Clarity, LendDo, Yelp) 100% technology-enabled, model-driven, loan adjudication and pricing Loan acquisition strategies driven completely from customer behavior and

“in-market” models

But there are also minefields: Small samples and the “overfitting contest” Biased samples Relentless pressure to grow (before regulation) and strong appetite from

secondary markets. The Big Short?


Alternative Lending






12


13


Logistic regression offers the right construct for estimating risk… Let PD = “Probability of Default”. Let Y=1 be default and Y=0 be

paid as agreed PD is related to attributes Xk’s by taking a weighted sum of the

attributes called “xbeta” where xbeta = β0 + β1X1 + … + βnXn and substituting xbeta into the logistic function

PD = P(Y=1|Xk) = … the logistic function

We need to find the ’s which best connect the Xk’s to the Y’s If Y=1, want PD big If Y=0, want 1-PD big Done by maximizing the likelihood function “L” as a function

of ’s

Max [ L( X) = (1 - )1- ]Copyright ©2012 Marketing Associates LLC. All rights reserved. Magnify is a division of Marketing Associates, LLC.

The Continued Relevance of Logistic Regression


14


Logistic regression models are at the root of most credit scores Per Wiki: Although logistic (or non-linear) probability

modelling is still the most popular means by which to develop scorecards, various other methods offer powerful alternatives, including MARS, CART, CHAID, and random forests.

Machine learning has become really popular lately, particularly among alternative lenders. The overfitting contest continues (next slide)…


The Continued Relevance of Logistic Regression … But

http://en.wikipedia.org/wiki/Logistic_function

http://en.wikipedia.org/wiki/Non-linear

http://en.wikipedia.org/wiki/Multivariate_adaptive_regression_splines

http://en.wikipedia.org/wiki/Predictive_analytics#Classification_and_regression_trees

http://en.wikipedia.org/wiki/CHAID

http://en.wikipedia.org/wiki/Random_forest




Comparing a Logistic Regression Model with SVM ModelThe sample was split 50/50 (sample1 / sample2).The following steps were performed for both the logistic regression model and SVM model: Sample1 was used to fit a model that was subsequently used to score sample2 Sample2 was used to fit a model that was subsequently used to score sample1

00%-37%

37%-45%

45%-52%

52%-60%

60%-67%

67%-72%

72%+0%

10%20%30%40%50%60%70%80%90%

100%

0%

5%

10%

15%

20%

25%

30%

38.1% 38.5% 40.3%

58.3%66.7%

73.1% 70.3%

Validation of Machine Learning Model: Test Samples

% of Funded Actual Predicted

SVM Prediction

Rene

wal

Like

lihoo

d

% o

f Fun

ded

Cont

ract

s

00%-37%

37%-45%

45%-52%

52%-60%

60%-67%

67%-72%

72%+0%

10%20%30%40%50%60%70%80%90%

0%

5%

10%

15%

20%

25%

37.3% 38.1%

66.7%

47.4%

68.9% 67.9% 68.5%

Validation of Logistic Regression Model: Test Samples

% of Funded Actual Predicted

Logistic Regression Prediction

Rene

wal

Like

lihoo

d

% o

f Acti

ve C

ontr

acts

KS=28.0

KS=24.4






16


17


Building predictive models has become a canned process for many banks – why?

Reg B: Empirically-derived, statistically sound ECOA OCC 11-7: Supervisory Guidelines on Model Risk Management CFPB and lenders’ lack of transparency Credit Ratings and Secondary Markets Banks shrinking desire to do retail banking. The alternative?

Commercial lines-of-credit for alternative lenders Millennials, millennials, millennials…


Regulatory and Societal Pressures


18


But we still have to find ways to accurately assess risk and price loans…

Big Data can’t be just a buzzword

Unstructured data has to be structured in an intelligent way

Personally-verifiable, or volunteered, data

Technology devoted to collecting unstructured data and structuring it


The Demand For Creativity and Disruption


19



Analytics: Banking vs. Alternative LendingIssue Banking Alternative

LendersHow do alternative lenders handle the modeling problems?

Sample Sizes

Large Small Strong priorsFix parameter estimates“Wider” samples, not longer

History Spans multiple economic cycles

1-3 years, these are mostly start-up companies

“Look-alike samples”: BUY DATARefit and relearnIntroduce products that control risk without having to model it (e.g. SoFi)Change the target event to something that happens quickly (e.g. Loan Science)

Regulation Onerous Light Actively explore social media (Yelp)Huge investments in data that banks and captives don’t think they can deploy

Legacy code

Lots of old SAS processes

No legacy processes

Technology “free for all”: R processes are daisy-chained with SAS, then connected to Python. “The IT nightmare”: 100% deployment flexibility


20


“It is not the model fitting techniques that need to change, as much as it is the treatment of samples, potential predictors, and model refits”.

The dependent variable can be redefined to something that is more frequently occurring and ultimately worth predicting.

Unstructured data can be structured as x’s and multiplied by betas to create expanded credit scores. For example: Yelp: # times the word “terrible” is mentioned Student lending: # times “broke” is in the same tweet as “student

loans”

Rapid refit and relearn processes are relatively easy to run routinely when data are appropriately structured.


How Does Alternative Lending Change What We Do?


21


Thank you for your time.Follow-up questions and requests: [email protected]; [email protected] [email protected]


Questions

mailto:[email protected]



credit risk modeling for alternative lending

Data & Analytics