Credit Scoring Reshape - Credit Research Centre · Advanced Analytics- Credit Scoring to the Rescue...
Transcript of Credit Scoring Reshape - Credit Research Centre · Advanced Analytics- Credit Scoring to the Rescue...
SW-EUB022-2007-03-19-CMD-V7
Credit Scoring Reshape because of and thanks to
the Financial Crisis
Zoom – Focus – Synthesis
Christos Glimidakis, Fotis Tsiamas, Maria Bakogeorgou July 2013
Page 2
STRICTLY CONFIDENTIAL
Macroeconomic Indicators of Greece
Sources: Hellenic Statistical Authority- Bank of Greece
Page 3
STRICTLY CONFIDENTIAL
Financial Indicators of the Bank
From 2009 and onwards (Crisis Period) the 90+
dpd for the Mortgage Portfolio increased to
relatively high levels. The 90+ dpd for the year
2012 was at a considerable higher rate than
2008, though quite less than the average Greek
Market.
The major hit from the crisis was at the
Consumer lending portfolio. Again, from 2009
and onwards (Crisis Period) the 90+ dpd for the
Consumer Portfolio had reached high levels. The
90+ dpd for the year 2012 was at a significant
higher rate than 2008 and at similar rates of the
whole Greek Market.
Sources: Investors’ Presentations
Page 4
STRICTLY CONFIDENTIAL
Information- Before Crisis Vs. After Crisis
Before the crisis:
• The market was at infant stage and most banks main objective was increasing volumes
• The Credit Policy took into consideration only One Dimension information
• No combinations between Existing Score Models (Application & Behavioral) with Policy rules
• Limited information based on competitors
After the crisis:
• Mature Portfolio
• Availability of Information regarding Bureau Score (2010 & onwards)
• Drill down analysis in areas that allowed advanced analytics to take a leading role
Examples Before Vs. After Crisis regarding information in the portfolio
Page 5
STRICTLY CONFIDENTIAL
Bank’s Responsiveness towards the Crisis
It was apparent that the Bank turned towards the judgmental process (Human Underwriting).
From 4% before the crisis 2007 the judgmental contribution to the Total %Approval Rate increased to 20%.
This was technically enlarged because the bank tightened all policy rules and subsequently assigned to the
Credit officers the task of balancing positive and negative characteristics of the applications.
Page 6
STRICTLY CONFIDENTIAL
Zooming in the Bank’s Decision
Analyzing the data it was quite obvious that the most important override reasons from the
judgmental process were categorized into 3 main groups:
• Credit History (Behavioral Scores, Bureau Scores, Buckets of the Bank, Buckets of
Competitors etc.)
• Debt to Income (Balances, Exposures, Limits, Income etc.)
• Demographics & Other (Customer Characteristics, Product etc.)
The increase of the human intervention led to important fluctuations in regards to the handling of
applications with a similar profile.
This was due mainly to the psychological factor which played a crucial role since the country was in
turmoil.
Page 7
STRICTLY CONFIDENTIAL
Advanced Analytics- Credit Scoring to the Rescue
3 Independent Models founded on the specific reason of Rejection
• Credit History Model- based on the Behavior of the Applicant
• Debt to Income Model- based on Financial Situation of the Applicant
• Demographics & Other Model- based on Demographics & Product Related Information
1 Dependent Model
• Synthesis Model (Combining all 3 models into 1)
Demographics & Other
Debt To Income
Credit History
Synthesis Model
Page 8
STRICTLY CONFIDENTIAL
Credit History Related Variables D
elin
que
ncie
s
Recent
White Bureau
Own
Max Ever
Bank’s
Spouse
Me
mb
ers
Sin
ce
Bank’s
White Bureau
Bu
rea
u S
co
res
Own
Spouse
Behavio
ral S
core
s
Bank’s
Page 9
STRICTLY CONFIDENTIAL
DTI Related Variables D
ebts
Consumer
White Bureau
Own Mortgage Bank’s
Spouse Small Business
Unused Consumer Limits
Requested Amount & Product
Incom
e
Own
Family
De
posits
Bank’s
Page 10
STRICTLY CONFIDENTIAL
Demographics & Other Related Variables
Dem
ogra
phic
s
Post Code
Marital Status
Residential Status
# of Children
Years in Job
Years in adress
Age
Profession
Gender
Car
Education Level
Nationality
Available Phones
Real Estate
Pro
du
ct
Product
Channel
Deposit (Auto)
Loan Duration
Guarantor
Add-on
Oth
er
Visa
Master
Diners
Amex
Other Card
Previous Rejection s
Applications in last 1.3.6 m
Group Sales
Staff
Private Banking
Customer/No Customer
Page 11
STRICTLY CONFIDENTIAL
Synthesis: Combine 3 Models into a Single one
Synthesis Risk Class
Demographics & Other
Debt & Income
Credit History
Tree
1
1-2
1-5
1
…
…
10
7
… 8-10
…
… 5-6
…
…
…
… 10
1-4
1-5
…
6-10
…
… 8 - 10
…
…
…
10
Decision Tree
3 Partial Risk Classes ->
1 Final Synthesis
Page 12
STRICTLY CONFIDENTIAL
Models Essentials
Bad Definition: Ever 90+
dpd & Restructuring
Outcome:
12 months
Exclusions:
Rejected for Other
Reasons
Sample:
All Consumer
Applications 2010
Score
card
s C
rite
ria
Credit History 11
Debt To Income 12
Demographics 12
Sco
reca
rds K
-S
Credit History 47,2%
Debt To Income 45,2%
Demographics 41,8% S
ynth
esis
Model
K-S
44.5%
Page 13
STRICTLY CONFIDENTIAL
Reject Inference- Focus
RJ Inference Methodology
Approved/
Rejected Sample
Applications
Rejected
Unknown
Known Performance from Other Products in
Bank
Approved Known Performance
Misalignment Adjustment (Outliers):
If
%Approval Rate Aki << %Approval Rate Ai
then
%Approval Rate Aki = %Approval Rate Aj
i,j = 1,2,…,10 (Risk Classes)
Ak= subgroup of variable A
Parceling Inference
Reject Only
Credit History
Reject Only
Debt To Income
Reject Only
Demographics &
Other
Models
Model
Credit History
Model
Debt To Income
Model
Demographics &
Other
Synthesis Model
Final Model
Page 14
STRICTLY CONFIDENTIAL
%Inferred Bad & Approval Rate per Risk Class
Page 15
STRICTLY CONFIDENTIAL
Synthesis Model – Swap Set
Theoretically, the Synthesis Model should reduce the levels of risk significantly from 1,7% to 1,0% by
increasing the %Approval Rate from 55,7% to 62,5%, hence increasing the overall market share of the Bank.
However, when these models are applied in production the Actual Approval Rate increases by 2,1%.
57,8 %
Actual %Approval Rate
1,0%
Estimated %Bad Rate
Theoretical:
Implementation:
Page 16
STRICTLY CONFIDENTIAL
Challenger Model
Sample:
All Consumer Applications 2010
Exclusions:
Clear Cut Policy Rejected
Outcome:
12 months
Bad definition:
90+ dpd
Challenger Model
∙ Add-on /Guarantor
∙ Age
∙ Area
∙ Family Status
∙ Gender
∙ Group Sales
∙ #of Appl.last 6 m
∙ # of previous rejections
∙ Percent Deposit
∙ Profession
∙ Requested Tenor & Product
∙ Residential
∙ Time in Job
Demographics (13)
∙ Type of Detrimental
∙ Max Bucket In Bank
∙ Max current Bucket in Bank
∙ Max Current Bucket in Bureau
∙ Max Bucket in Bureau
∙ Months in Bureau
∙ Months Since Detrimental
Credit History (7)
∙ Amortized Balance in Bank
∙ Deposits amount
∙ Family Income to Requested
Amount
∙ Total Balance in Bureau to Total
Exposure in Bureau
Debt to Income (4)
K-S
43,3% Va
ria
ble
s
Page 17
STRICTLY CONFIDENTIAL
Synthesis Vs. Challenger Model Validation
Sample Period: 2011 Normal Applications
Bad Definition: 90+dpd or Restructured @12 months
Synthesis Model Validation
Comparison Synthesis to Challenger Model (K-S)
44,5%
53,8%
Synthesis Model
43,3%
50,5%
Challenger Model Development Period
Performance Validation Period
PSI<10%
K-S =53,8%
Page 18
STRICTLY CONFIDENTIAL
Advantages of the Synthesis
•Time in job
More Variables
35 Synthesis Vs.
24 Challenger
Synthesis Model
Reject Inference
Focus on specific Reject Reasons
Crystal
Clear
Adverse
Reasons
Higher K-S
& PSI
stable
throughout
time
Flexible
Monitoring
& Submodels
Re
Development
Macroeconomics
Data
Future Steps of the Models
Inclusion in the
Re Development Process
Page 19
STRICTLY CONFIDENTIAL
Future Next Steps
• Different Modeling Approach for the Synthesis Model e.g. Clustering, Regression etc.
• Inclusion of Macroeconomics Data in the Models
• Further investigation regarding the swap set Analysis
e.g. Why do we still have Rejections in the best Risk Class (RC 1) is it due to Regulatory, Policy restrictions?
• Misalignment between two different periods before crisis and after crisis.
Why for the same Risk Class we have diverse Bad Rates?
What factors have contributed to this change throughout time within each Risk Class?
Page 20
STRICTLY CONFIDENTIAL
THANK YOU !
Christos Glimidakis: [email protected]
Fotis Tsiamas: [email protected]
Maria Bakogeorgou: [email protected]