Improving Customer Retention by Identifying Characteristics of Auto Loan Prepayers using MARS
John Trimble Wells Fargo
An Appeal of Data Mining
Efficient identification of:
– Most important predictors – Local relationships
Offers ability to target different strategies:
– Within and … – Without local segments
MARS and other DM techniques have these capabilities
Topic of This Presentation
—Prepayments in 1st 18 Months—
Loans paid off prior to their maturity date Imposes costs on the lender
– Increases administrative burden
More dealer discounts, marketing costs, general admin burden
– Imposes reinvestment risks:
May be at lower rate if rates falling May take longer to match with borrowers if rates rising May have to accept lower credit quality if occurs in midst of market share battle
If can predict prepayments, can devise mitigating strategies
Rate of Voluntary Prepayment
Approximately 2-3% of Loans Per Month 30% Prepay Over 18 Months 60% Prepay Over Life of Loan
– Includes insignificant prepayments Payoff on Next to Last Payment Payoff With Small Balance
About Wells Fargo
Auto Finance
Auto Loan Originations
—Automated Approval—
Car Dealer
Wells Fargo
Lender #2
Lender #3
Wells Fargo Portfolio
$14+ billion in loans and leases (retail and wholesale) 840,000+ Customers Approve new loans with credit scoring models to large degree
– Model assesses risk of approx. 1.8 million
individuals per year (800,000 applications)
Characteristics of Prepayers
Using MARS to Understand Who Prepays Auto Loans
⎯Potential Causes⎯
Natural Hazards
– Road hazards – Fire, Flood, Theft
Prepayment
Voluntary Causes
– – – –
Advantageous refinancing opportunities Vehicle replacement Interest rate levels Manufacturer incentives
Retention
– Could be lack of attractive financing alternatives
Insurance/Hazard Prepayment
⎯Likely Difficult to Predict⎯ By Nature Due to Random Events
– Likely Difficult to Predict
Predicting Payoff More Difficult
– The Greater the % of Payoffs Due to Natural
Hazards – Ins Payoffs May Mask Payoffs Motivated by Personal Interests
Following three slides seem to bear this out
Predicting Prepayments
—Available Attributes—
Over 900 Credit Bureau and Application Attributes on 100,000+ Records Available Examples include:
– Credit Inquiries – Loan Maturity – Recent Delinquencies – Credit Utilization – # Accounts – Debt Burden – Down Payment
MARS Prepayer Profile
- Loan Term + Credit Union Accts + Recent Account Openings
– - Carries Revolving Balances
- Low Revolving Debt + ACH payment + Poor or Thin Delinquency History + High Percent Down Basis Functions=30, Interactions=2
+(-) Indicates Positive (Negative) Impact on Prob Payoff
⎯MARS vs Descending FICO Score⎯
100 90 80 70
Prepayer Logistic Model
Lorenz Curves FICO MARS
% of PAYOFFS
60 50 40 30 20 10 0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100 % of OBS
Barely Exceeds Random Ranking
Insurance Prepayers
Using MARS to Help Identify Them in the Sample
Identifying Ins Prepayers
Information Contained in Customer Service Comments Files Need to Write Script to Search These Files No Standardization
– Difficult to Write Scripts
C/S Reps Do Tend To Follow Personal Habits Exploit This To Identify Ins Prepays Involves Visual Inspection of Comment Files
Basis Functions=30, Interactions=2 +(-) Indicates Positive (Negative) Impact on Prob Payoff
⎯Need to Classify Prepays from Info Like This⎯
cust c/i to say she got coupon book in mail today and will mail pmt today, will call back if she decides to do ach. CCI WE REFINANCED VEH CUST STATES THERE IS STILL A BAL REMAINING ASKED CUST IF SHE CANCELLED HER WARR W/ THEM CUST REALIZE WHY SHE WILL CANCEL WARR. W/ GEICO. DAGWOOD says this is a loan that she recieved by mail. She was w/ Wachovia Bank before and says we haven't paid off her loan in full, transfd to on-line services # since this is considered internet loan for further assistance. man @ poe sttd no DAGWOOD Bumstead wrks there deleting (xxx) xxx-xxxx tt DAGWOOD, sd she is changing jobs, starts new job 2/18, does not have new bp#, cust auth aps 2/15 for jan iao 341.55 +5 fee, included lc, check #176, con # 111win010101, ver demos, sd next payment will probably be late but will get it in asap 40L to 30i all late fees paid eff 3/28/02, then 36L all late fees outstanding, late fee code should be '0' per ARSI late fee project CS ACTIVITY ONML REVIEW DATE 06-04-1922 BALANCE 14687.72 HP CS ACTIVITY OBNC REVIEW DATE 06-05-1922 BALANCE 14687.72 TT DAGWOOD SD CAN’T MAKE PYMT TIL NEXT TWO WEEKS RFUSED POST DATED CK ADVS FUR COLL CS ACTIVITY OBWP REVIEW DATE 06-06-1922 BALANCE 14687.72 BP (xxx) xxx-xxxx TT FEMALE DAGWOOD NO LONG EMP AT cUST CI INQ ON ach DAGWOOD ci, wanted to make pymnt via phone, adv her that slc is handling gave #(xxx)-xxx-xxxx.... CS ACTIVITY IBPA PROMISE 06-13-1922 325.29 REVIEW DATE 06-14-1922 BALANCE 14687.72 **SPEED PAY/chk#193 / $325.29 + 6 FEE /06 10 02 *** IB TT DAGWOOD VI CS ACTIVITY OBPA PROMISE 06-14-1922 325.29 REVIEW DATE 06-17-1922 BALANCE 14687.72 ***APS/325.29/6/12/22*** HP TT DAGWOOD VI CS ACTIVITY OBNC REVIEW DATE 07-03-1922 BALANCE 14503.46 HP T DAGWOOD SAYS SHE IS RUNNING BEHIND HASNT DONE BILLS WILL DO IT AND THAN SHE CAN GET US OUR MONEY. V/I CS ACTIVITY IBPA PROMISE 07-18-1922 325.29 REVIEW DATE 07-19-1922 BALANCE 14503.46 ***APS/325.29/071602/CK#111/*** TT DAGWOOD VI CS ACTIVITY OBPA PROMISE 07-17-1922 325.29 REVIEW DATE 07-18-1922 BALANCE 14503.46 ***APS/325.29/071622/CHK#111/*** HP TT DAGWOOD V/I Good Till 7/20/22 franklin from ford dealership IL Data as of 07/20/22 Daily Dollar Rate: 2.537 Payoff Quoted: $14,281.07, Balance: $14,265.85 CS ACTIVITY ONML REVIEW DATE 08-12-1922 BALANCE 14265.85 P/O RECVD FROM: DAFFY DUCK AUTO PK ADDRESS: 1111 DRYLAKE BLVD, USELESS PARK FL zzzzz MAKER: SUNTRUST #zzzzz $14288.68 \P/O $ 14288.68 O/P $ N CS ACTIVITY ONML REVIEW DATE 08-14-1922 BALANCE 14265.85 CS ACTIVITY IBPO REVIEW DATE 08-20-1922 BALANCE 14265.85 IB TT DAGWOOD SD THAT DEALERSHIP IS SENDING IN THE PAYOFF VI WAAS TRADED ON 072122 TRANSFER From CLT001 to TPPP01 PAID FILE PROCESSED
Customer Service Comment
Names, dates, locations and identifying numbers have been altered to preserve account holder privacy
Strategy to Identify Ins Cases
Visually Inspect a Sample of Comments
– Classify into Ins Payoff and Non-Ins Payoff
Build MARS Model of Non-Ins Payoff
– Classify Records – Search for Ins Company Names
Helpful Strategy
– But Still Need Substantial Visual Inspection to Build
Viable Script to Accurately Classify Records
Insurance Figures in Most Prepayments
92% of prepays involve insurance payment
– Ranges from “total loss” to – Partial loss—buyer completes the payoff – A very high % likely involve vehicle replacement
Remaining 8% most likely related to:
– Vehicle replacement and … – New financing
Cheaper rate (lower financing cost) Extension of term (lower monthly payment)
Insurance Prepayers
Using MARS to Predict Insurance Prepayers
MARS Ins Prepayer Profile
- Term of Loan Longer Term
– - Carries Revolving Balances – - Credit Utilization – - Low Percent Down Pmt
+ High Percent Down Pmt + Co-applicant’s delinquency record
– - Co-applicant has thin credit file
+ Credit Seeking Activity - Carries Revolving Balances - Tenure of trades
+(-) Indicates Positive (Negative) Impact on Prob Ins Payoff
Basis Functions=30, Interactions=2
Ins Prepay Logistic Model
⎯MARS vs Descending FICO⎯
MARS
100 90 80 70
Lorenz Curves FICO
% of INS PAYOFFS
60 50 40 30 20 10 0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100 % of OBS
Models Barely Exceed Random Ranking
Non-Insurance Prepayers
Using MARS to Predict NonInsurance Prepayers
MARS Non-Ins Prepayer Profile
Upper 25% Credit Quality
– + Credit Quality
Bottom 75% Credit Quality
– + Credit Utilization – + Credit Seeking Activity – + Monthly Pmt for Shelter
Co-applicant has weak credit record
– – – –
+ Credit Seeking Activity - Credit Utilization + Applicant’s Age - Down Pmt
Basis Functions=30, Interactions=2
+(-) Indicates Positive (Negative) Impact on Prob Non-Ins Payoff
Non-Ins Prepay Logistic Model
⎯MARS vs Ascending FICO⎯
100 90 80
Lorenz Curves MARS FICO
% of NON-INS PAYOFFS
70 60 50 40 30 20 10 0 5
• MARS Predicts Best • FICO Rank Orders, But Overpredicts
10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100 % of OBS
Insurance vs Non-Insurance Prepayers
Using MARS to Predict Which
MARS Prepayer Profile
⎯ Non-Ins vs. Ins⎯
Upper 35% credit quality
– + Credit Quality
Bottom 65% Credit Quality
– – – –
+ Credit Quality + Early in Life of Loan - Later in Life of Loan + Loan-to-value ratio > 130%
Co-applicant Has Weak Credit Record
– - Credit Utilization – - High Down Pmt – + Low Down Pmt
Basis Functions=30, Interactions=2
+(-) Indicates Positive (Negative) Impact on Prob that a Payoff will be Non-Ins vs Ins
MARS Prepayer Profile
⎯ Non-Ins vs. Ins (continued)⎯ Co-applicant Has Weak Credit Record
– + Early in Life of Loan – + Age of Applicant – + Credit Seeking Activity – + Carries Revolving Balances – + Longer Term of Loan
Basis Functions=30, Interactions=2
+(-) Indicates Positive (Negative) Impact on Prob that a Payoff will be Non-Ins vs Ins
Non-Ins vs Ins Logistic Model
⎯PAYOFFS: MARS vs Ascend. FICO⎯
100 90 80
Lorenz Curves MARS FICO
% of NON-INS PAYOFFS
70 60 50 40 30 20 10 0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100 % of PAYOFF OBS
• MARS Predicts Best, incl. Timing • FICO Rank Orders, But Overpredicts
What Have We Learned?
About the Motivations for Prepayment
Motivations for Prepayment
Vehicle Replacement is #1 Reason
– Vast Majority Involve Substantial Ins Payment – Remainder of replacements are likely vehicle upgrades
Avoid rising maintenance costs or Upgrade vehicle status
Other Motivations Not Easily Determined
– Reduce Size of Payment – Switch to Alternative Financing (e.g. Home Equity) – Desire and capability to own outright
What Have We Learned?
About Who Prepays
Who Prepays
Virtually Anyone if Insurance Payment Involved For Non-Insurance Prepayment
– Bottom 75% of Credit Quality Distribution most likely
Otherwise Ascending NOT Descending FICO Would Have Rank-Ordered Non-Ins Prepays
– Top 25% Credit Quality Distribution is small segment – When Co-Applicant Has Weak Credit
Older owners more likely to pay off Those in weaker financial condition more likely to pay off
– Higher debt utilizers – Those who sought easier terms (e.g. longer term, lower down pmt)
What Have We Learned
About When Prepayment Occurs
When Prepayment Occurs
Early in Life of Loan More Likely
– If Borrower(s) in Lower 75% of Credity Quality – If Co-Applicant Has Weak Credit
Later in Life of Loan Becomes Less Likely
– If Borrower(s) in Lower 75% of Credity Quality and – If Co-Applicant Has Weak Credit – And if Life of Loan Has Passed Early Payoff Peak
Don’t Know About Top 25% Credit Quality
– Would Expect A Pattern⎯Either Early or Late
Conclusions
About Possible Mitigation Strategies
⎯About Possible Mitigation Strategies⎯
Predicting Payoff (All Reasons) at Origination
– Predictive Capability Too Weak – Strategy Based on This Likely Unsuccessful
Conclusions
Predicting Non-Ins Payoff at Origination
– Significant Capability To Predict – Strategy Likely to Be Beneficial – Further Identifying Reason For Payoff in Sample
Likely Enhance Ability To Build More Predictive Model