SlideShare a Scribd company logo
Risk Modeling and Pricing Paradigms:
Past, Present, and Emerging
A Consultant’s View
Keith Shields
SVP, Analytic Services – Loan Science
Chief Analytics Officer – Magnify Analytic Solutions
Introduction and Background
 Who are we and why this topic? And do we know anything?
 Is credit scoring dead?
– Uber?
– P2P?
 In exploring past and present paradigms we remember how
useful credit scoring is.
 We also see how the “Big Data Revolution” allows us to get so
much more out of it.
 But everything boils down to DEPLOYMENT.
A Past and Present Paradigm: Generic Scores
 Generic scores / Bureau Scores
– Have been, and still are, the basis for many credit decisions
– Very useful properties
• Rank order the risk - below left
• Quantifiable relationship with the probability of default - below right
• Facile deployment
• Pervasive…they exist in almost everyone’s vernacular
Worst 40% => 70% of
defaults
A Past and Present Paradigm:
Pricing with an LTV-dependent matrix to control the contract…
 The Bureau Score-LTV matrix
– The pricing problem is more or less solved within the grid
– Important factors like PTI considered during underwriting
– The importance of facile deployment
Can we price for a
22.2% default rate?
If so, then aren’t
these approvable?
Default Rate
Credit Score 80% 90% 100% 110% 120% 130%
540 22.2% 31.9% 36.6% 40.0% 41.5% 43.0%
560 19.1% 25.7% 27.3% 31.6% 32.5% 33.6%
600 17.2% 24.9% 26.4% 28.8% 30.9% 28.4%
640 10.9% 16.7% 19.0% 23.2% 25.2% 27.0%
680 6.3% 6.7% 9.1% 11.8% 14.5% 13.8%
720 2.7% 3.7% 4.7% 5.7% 8.4% 8.5%
760 0.9% 3.4% 3.7% 5.2% 5.5% 5.6%
800 0.5% 1.2% 1.8% 2.9% 5.1% 3.5%
840 0.2% 0.4% 0.8% 1.5% 2.0% 2.6%
LTV
Moving On From The Past…
 Why stay with the generic
scores?
 Can we get a better ranking with
our own data?
 Won’t a better risk ranking
enable more approvals and a
reduction in defaults? See graph
right.
 Things to consider:
– Capability and Scale
– Business and process
nuances
– Deployment
– The Credit Score-LTV
matrix is essentially a
custom model.
– Validation and CFPB
Worst 40% => 75% of
defaults
A Present Paradigm: The Pervasiveness of Custom Scores
 Driving the need to improve over the generic scores…
– “Big Data”: Data retention and scalable data platforms (data warehouses)
• The more we store, and the more we analyze, the more we know
– Modeling exercises begin as an effort to understand and evolve into an
effort to predict.
– Evolution and availability of nuanced and sometimes esoteric credit
attributes (which can improve segmentation as well as models)
• Bureaus have made these easily deployable as well
– Cost of funds disadvantage for some lenders creates a need to find “650s
that perform like 700s”. When there is a cost of funds disadvantage, the
lender can’t afford to make the same decisions as the competition.
– Subprime lending – success very much depends on our ability to get good
ranking between 500 and 600…which results in our having to differentiate
between a “good 550” and a “bad 550”.
The Custom Score and Pricing
 To date custom scores have offered improved risk assessment
largely through the incorporation of contract data.
– A customized score-to-odds:
• ln(p/(1-p)) = A + B*BUREAU_SCORE
– The basis for the Bureau Score-LTV matrix is, as previously
mentioned, a custom model:
• ln(p/(1-p)) = A + B*BUREAU_SCORE + C*LTV +
D*BUREAU_SCORE*LTV
The Custom Score and Pricing
 Our desire to price based on the custom score creates a
deployment problem.
– Pricing decisions are no longer based just on LTV and credit score:
Nuanced credit attributes, PTI, Term, Vehicle Age, or perhaps even
the presence of a SID
– Contracts have to be entered before a score can be generated, and a
score must be generated before a price can be determined, but a
price must be assumed to generate a contract.
– The vicious circle:
Gotta know the
contract to get a
score
Gotta know the
score to get an
APR
Gotta know the
APR to get a
contract
An Emerging Paradigm:
Making the Most of the Custom Score “Math”
 Probability of Default = f(Custom Score)
 Custom Score = f(Credit attributes, Applicant Info, LTV, PTI, Term)
– Credit, applicant and income are all fixed
– Monthly payment = f(Loan Amount, APR, Term)
– What if PD is also fixed? What if we have a “target PD”?
 Target PD = f(Fixed values, Loan Amount, Vehicle Value, APR, Term)
 => APR = F(Target PD, Loan Amount, Vehicle Value, Term, Fixed
Values)
 This question:
– What probability of default is created by the proposed contract?
 Becomes:
– What is the APR, Loan Amount, and Term that creates a desired
probability of default (one that we are priced for)?
 And the best contract needn’t be on the proposed vehicle. Can we marry
the customer to any vehicle on the lot?
 Lenders already do this on a limited scale. But why limit the scale?
An Emerging Paradigm: Matching a Customer to a Vehicle
 LTV = Loan Amount / Vehicle Value
– LTV has a lot of leverage on the probability of default and loss given
default (a level of complexity not discussed on the previous slide)
– It can be reduced through the numerator (additional cash in a deal is
logical and popular condition), but it can also be reduced though the
denominator.
– The range of the denominator is dependent upon the vehicles we can
offer.
 APR = F(Target PD, Loan Amount, Term, Vehicle Value, Fixed
Values)
– The mathematical expression above allows us to solve for a loan
amount and APR for any vehicle-customer combination.
– What about a virtual vehicle lot (like Ebay Motors)? The math allows
us to generate thousands, perhaps millions, of potential contracts on
a given customer.
– This becomes similar to a P2P lending paradigm for autos.
The Auto-Lending “Platform”
 Three questions we force upon ourselves when
evaluating emerging paradigms:
1. Is our industry ripe for disruption?
2. Is there is “Big Data” play driving the disruption?
3. What’s the plan?
Questions
Questions?
Thank you for your time and attention.
www.loanscience.com
www.magnifyas.com

More Related Content

Similar to LoanScience_AFSummit_presentation

The Crisis And The Future Of Valuations
The Crisis And The Future Of ValuationsThe Crisis And The Future Of Valuations
The Crisis And The Future Of Valuations
pswinhoe
 
Credit Risk Analytics
Credit Risk AnalyticsCredit Risk Analytics
Credit Risk Analytics
Senthil Ramanath
 
Creditscore
CreditscoreCreditscore
Creditscore
kevinlan
 
HWZ-Darden Konferenz: Building a Sustainable Analytics Orientation
HWZ-Darden Konferenz: Building a Sustainable Analytics OrientationHWZ-Darden Konferenz: Building a Sustainable Analytics Orientation
HWZ-Darden Konferenz: Building a Sustainable Analytics Orientation
HWZ Hochschule für Wirtschaft
 
Adept Change Management_Panna Visani 2015_1
Adept Change Management_Panna Visani 2015_1Adept Change Management_Panna Visani 2015_1
Adept Change Management_Panna Visani 2015_1
Panna Visani MBCS ACCA
 
La Dove Associates -- CRM/Customer Care Consulting Overview
La Dove Associates --  CRM/Customer Care Consulting Overview La Dove Associates --  CRM/Customer Care Consulting Overview
La Dove Associates -- CRM/Customer Care Consulting Overview
LaDove Associates
 
Helping Our Clients Select Best Quote
Helping Our Clients Select Best QuoteHelping Our Clients Select Best Quote
Helping Our Clients Select Best Quote
Alexander Levine
 
Business modeling for startups part I
Business modeling for startups part IBusiness modeling for startups part I
Business modeling for startups part I
Pierre-Yves Pau
 
Credit Risk Modeling for Alternative Lending
Credit Risk Modeling for Alternative LendingCredit Risk Modeling for Alternative Lending
Credit Risk Modeling for Alternative Lending
Magnify Analytic Solutions
 
ASA_CSP_2016_Shields_Lund
ASA_CSP_2016_Shields_LundASA_CSP_2016_Shields_Lund
ASA_CSP_2016_Shields_Lund
Keith Shields
 
Services Business Unit - Sales training - How to sale services
Services Business Unit - Sales training - How to sale servicesServices Business Unit - Sales training - How to sale services
Services Business Unit - Sales training - How to sale services
Valdir Gomes Silva
 
Applications and Service Offering - Brochure
Applications and Service Offering - BrochureApplications and Service Offering - Brochure
Applications and Service Offering - Brochure
Sohail_farooq
 
Personal Loan Risk Assessment
Personal Loan Risk Assessment Personal Loan Risk Assessment
Personal Loan Risk Assessment
Kunal Kashyap
 
How GetNinjas uses data to make smarter product decisions
How GetNinjas uses data to make smarter product decisionsHow GetNinjas uses data to make smarter product decisions
How GetNinjas uses data to make smarter product decisions
Bernardo Srulzon
 
risk.ppt
risk.pptrisk.ppt
EY + Neo4j: Why graph technology makes sense for fraud detection and customer...
EY + Neo4j: Why graph technology makes sense for fraud detection and customer...EY + Neo4j: Why graph technology makes sense for fraud detection and customer...
EY + Neo4j: Why graph technology makes sense for fraud detection and customer...
Neo4j
 
EY: Why graph technology makes sense for fraud detection and customer 360 pro...
EY: Why graph technology makes sense for fraud detection and customer 360 pro...EY: Why graph technology makes sense for fraud detection and customer 360 pro...
EY: Why graph technology makes sense for fraud detection and customer 360 pro...
Neo4j
 
Telemarketing prediction project
Telemarketing prediction projectTelemarketing prediction project
Telemarketing prediction project
Learnbay Datascience
 
Magnify DMA presentation 2014
Magnify DMA presentation 2014Magnify DMA presentation 2014
Magnify DMA presentation 2014
Magnify Analytic Solutions
 
Neural Network Model
Neural Network ModelNeural Network Model
Neural Network Model
Eric Esajian
 

Similar to LoanScience_AFSummit_presentation (20)

The Crisis And The Future Of Valuations
The Crisis And The Future Of ValuationsThe Crisis And The Future Of Valuations
The Crisis And The Future Of Valuations
 
Credit Risk Analytics
Credit Risk AnalyticsCredit Risk Analytics
Credit Risk Analytics
 
Creditscore
CreditscoreCreditscore
Creditscore
 
HWZ-Darden Konferenz: Building a Sustainable Analytics Orientation
HWZ-Darden Konferenz: Building a Sustainable Analytics OrientationHWZ-Darden Konferenz: Building a Sustainable Analytics Orientation
HWZ-Darden Konferenz: Building a Sustainable Analytics Orientation
 
Adept Change Management_Panna Visani 2015_1
Adept Change Management_Panna Visani 2015_1Adept Change Management_Panna Visani 2015_1
Adept Change Management_Panna Visani 2015_1
 
La Dove Associates -- CRM/Customer Care Consulting Overview
La Dove Associates --  CRM/Customer Care Consulting Overview La Dove Associates --  CRM/Customer Care Consulting Overview
La Dove Associates -- CRM/Customer Care Consulting Overview
 
Helping Our Clients Select Best Quote
Helping Our Clients Select Best QuoteHelping Our Clients Select Best Quote
Helping Our Clients Select Best Quote
 
Business modeling for startups part I
Business modeling for startups part IBusiness modeling for startups part I
Business modeling for startups part I
 
Credit Risk Modeling for Alternative Lending
Credit Risk Modeling for Alternative LendingCredit Risk Modeling for Alternative Lending
Credit Risk Modeling for Alternative Lending
 
ASA_CSP_2016_Shields_Lund
ASA_CSP_2016_Shields_LundASA_CSP_2016_Shields_Lund
ASA_CSP_2016_Shields_Lund
 
Services Business Unit - Sales training - How to sale services
Services Business Unit - Sales training - How to sale servicesServices Business Unit - Sales training - How to sale services
Services Business Unit - Sales training - How to sale services
 
Applications and Service Offering - Brochure
Applications and Service Offering - BrochureApplications and Service Offering - Brochure
Applications and Service Offering - Brochure
 
Personal Loan Risk Assessment
Personal Loan Risk Assessment Personal Loan Risk Assessment
Personal Loan Risk Assessment
 
How GetNinjas uses data to make smarter product decisions
How GetNinjas uses data to make smarter product decisionsHow GetNinjas uses data to make smarter product decisions
How GetNinjas uses data to make smarter product decisions
 
risk.ppt
risk.pptrisk.ppt
risk.ppt
 
EY + Neo4j: Why graph technology makes sense for fraud detection and customer...
EY + Neo4j: Why graph technology makes sense for fraud detection and customer...EY + Neo4j: Why graph technology makes sense for fraud detection and customer...
EY + Neo4j: Why graph technology makes sense for fraud detection and customer...
 
EY: Why graph technology makes sense for fraud detection and customer 360 pro...
EY: Why graph technology makes sense for fraud detection and customer 360 pro...EY: Why graph technology makes sense for fraud detection and customer 360 pro...
EY: Why graph technology makes sense for fraud detection and customer 360 pro...
 
Telemarketing prediction project
Telemarketing prediction projectTelemarketing prediction project
Telemarketing prediction project
 
Magnify DMA presentation 2014
Magnify DMA presentation 2014Magnify DMA presentation 2014
Magnify DMA presentation 2014
 
Neural Network Model
Neural Network ModelNeural Network Model
Neural Network Model
 

LoanScience_AFSummit_presentation

  • 1. Risk Modeling and Pricing Paradigms: Past, Present, and Emerging A Consultant’s View Keith Shields SVP, Analytic Services – Loan Science Chief Analytics Officer – Magnify Analytic Solutions
  • 2. Introduction and Background  Who are we and why this topic? And do we know anything?  Is credit scoring dead? – Uber? – P2P?  In exploring past and present paradigms we remember how useful credit scoring is.  We also see how the “Big Data Revolution” allows us to get so much more out of it.  But everything boils down to DEPLOYMENT.
  • 3. A Past and Present Paradigm: Generic Scores  Generic scores / Bureau Scores – Have been, and still are, the basis for many credit decisions – Very useful properties • Rank order the risk - below left • Quantifiable relationship with the probability of default - below right • Facile deployment • Pervasive…they exist in almost everyone’s vernacular Worst 40% => 70% of defaults
  • 4. A Past and Present Paradigm: Pricing with an LTV-dependent matrix to control the contract…  The Bureau Score-LTV matrix – The pricing problem is more or less solved within the grid – Important factors like PTI considered during underwriting – The importance of facile deployment Can we price for a 22.2% default rate? If so, then aren’t these approvable? Default Rate Credit Score 80% 90% 100% 110% 120% 130% 540 22.2% 31.9% 36.6% 40.0% 41.5% 43.0% 560 19.1% 25.7% 27.3% 31.6% 32.5% 33.6% 600 17.2% 24.9% 26.4% 28.8% 30.9% 28.4% 640 10.9% 16.7% 19.0% 23.2% 25.2% 27.0% 680 6.3% 6.7% 9.1% 11.8% 14.5% 13.8% 720 2.7% 3.7% 4.7% 5.7% 8.4% 8.5% 760 0.9% 3.4% 3.7% 5.2% 5.5% 5.6% 800 0.5% 1.2% 1.8% 2.9% 5.1% 3.5% 840 0.2% 0.4% 0.8% 1.5% 2.0% 2.6% LTV
  • 5. Moving On From The Past…  Why stay with the generic scores?  Can we get a better ranking with our own data?  Won’t a better risk ranking enable more approvals and a reduction in defaults? See graph right.  Things to consider: – Capability and Scale – Business and process nuances – Deployment – The Credit Score-LTV matrix is essentially a custom model. – Validation and CFPB Worst 40% => 75% of defaults
  • 6. A Present Paradigm: The Pervasiveness of Custom Scores  Driving the need to improve over the generic scores… – “Big Data”: Data retention and scalable data platforms (data warehouses) • The more we store, and the more we analyze, the more we know – Modeling exercises begin as an effort to understand and evolve into an effort to predict. – Evolution and availability of nuanced and sometimes esoteric credit attributes (which can improve segmentation as well as models) • Bureaus have made these easily deployable as well – Cost of funds disadvantage for some lenders creates a need to find “650s that perform like 700s”. When there is a cost of funds disadvantage, the lender can’t afford to make the same decisions as the competition. – Subprime lending – success very much depends on our ability to get good ranking between 500 and 600…which results in our having to differentiate between a “good 550” and a “bad 550”.
  • 7. The Custom Score and Pricing  To date custom scores have offered improved risk assessment largely through the incorporation of contract data. – A customized score-to-odds: • ln(p/(1-p)) = A + B*BUREAU_SCORE – The basis for the Bureau Score-LTV matrix is, as previously mentioned, a custom model: • ln(p/(1-p)) = A + B*BUREAU_SCORE + C*LTV + D*BUREAU_SCORE*LTV
  • 8. The Custom Score and Pricing  Our desire to price based on the custom score creates a deployment problem. – Pricing decisions are no longer based just on LTV and credit score: Nuanced credit attributes, PTI, Term, Vehicle Age, or perhaps even the presence of a SID – Contracts have to be entered before a score can be generated, and a score must be generated before a price can be determined, but a price must be assumed to generate a contract. – The vicious circle: Gotta know the contract to get a score Gotta know the score to get an APR Gotta know the APR to get a contract
  • 9. An Emerging Paradigm: Making the Most of the Custom Score “Math”  Probability of Default = f(Custom Score)  Custom Score = f(Credit attributes, Applicant Info, LTV, PTI, Term) – Credit, applicant and income are all fixed – Monthly payment = f(Loan Amount, APR, Term) – What if PD is also fixed? What if we have a “target PD”?  Target PD = f(Fixed values, Loan Amount, Vehicle Value, APR, Term)  => APR = F(Target PD, Loan Amount, Vehicle Value, Term, Fixed Values)  This question: – What probability of default is created by the proposed contract?  Becomes: – What is the APR, Loan Amount, and Term that creates a desired probability of default (one that we are priced for)?  And the best contract needn’t be on the proposed vehicle. Can we marry the customer to any vehicle on the lot?  Lenders already do this on a limited scale. But why limit the scale?
  • 10. An Emerging Paradigm: Matching a Customer to a Vehicle  LTV = Loan Amount / Vehicle Value – LTV has a lot of leverage on the probability of default and loss given default (a level of complexity not discussed on the previous slide) – It can be reduced through the numerator (additional cash in a deal is logical and popular condition), but it can also be reduced though the denominator. – The range of the denominator is dependent upon the vehicles we can offer.  APR = F(Target PD, Loan Amount, Term, Vehicle Value, Fixed Values) – The mathematical expression above allows us to solve for a loan amount and APR for any vehicle-customer combination. – What about a virtual vehicle lot (like Ebay Motors)? The math allows us to generate thousands, perhaps millions, of potential contracts on a given customer. – This becomes similar to a P2P lending paradigm for autos.
  • 11. The Auto-Lending “Platform”  Three questions we force upon ourselves when evaluating emerging paradigms: 1. Is our industry ripe for disruption? 2. Is there is “Big Data” play driving the disruption? 3. What’s the plan?
  • 12. Questions Questions? Thank you for your time and attention. www.loanscience.com www.magnifyas.com