Download - 5 Lyn Thomas Risk Stream Ccri2011
-
7/31/2019 5 Lyn Thomas Risk Stream Ccri2011
1/13
1
Lyn C Thomas
Quantitative Financial Risk Management Centre,
School of Management
University of Southampton, UK
ICCR London Oct 4 2011
Modelling credit risk in portfolios of consumer loans:
how to uncrunch credit
Why and how to put economicfactors into risk management
systems
-
7/31/2019 5 Lyn Thomas Risk Stream Ccri2011
2/13
2
Outline
Scoring as a way of modelling retail credit risk
Problems with scoring because no economics involved
Incorporating macro economic impacts into scorecards
Case study from invoice discounting
-
7/31/2019 5 Lyn Thomas Risk Stream Ccri2011
3/13
3
Scoring as a way of modellingretail credit risk
Retail credit risk models traditionally based on scorecards Application scorecards and behavioural scorecards
Use sample from 2-3 years ago to relate borrowercharacteristics (application, performance, bureau) to
default status 1 year later. Objective is to get ranking accurate
Measured using KS, Gini Cut-off chosen using business measures not default probability
Assumption is relative ranking of credit worthinessconstant over time
No need to include economic variables
-
7/31/2019 5 Lyn Thomas Risk Stream Ccri2011
4/13
4
reas where lack ofeconomic factors affected scorecards
Problems with bureau scores in US subprime mortgage crisis
Other issue- fraud, disconnect between time periods of model and originator risk SEC highlighted scores inability to respond to changes in economy
Problems with credit rating agency PD estimates for portfolios of mortgages ( RMBS) Models did not integrate application scores and economics correctly so ratings were
downgraded ll
Basel required score to translate to long run average probability of default (PD)
Score to PD does vary over time because of economic changes Without economics in scorecards how to estimate score to long run average PD How to build models for stress testing
Vantage ScoreDetails of Real estate scores2003-2008
-
7/31/2019 5 Lyn Thomas Risk Stream Ccri2011
5/13
5
Decomposition of log odds score
Log odds score ( logistic regression gives log odds scores; linear
regression gives transformed log odds )
Use Bayes theorem to split into population odds plus weights ofevidence ( adjustment due to individual characteristics)
If pG pB proportion of Goods ( Bads in population)n
( | )( | ) ( | )( ) ln ln ln ln ln ln ( ) ( )
( | | ( | ) ( | )
G G
pop Pop
B B
p p G pp G p Gs o I s woe
p B p p B p p B
xx x
x x x
x x x
( )
( ) ( )
( | )( ) ln ( | ) ( | ) 1
( | )
1( | )
1 1
s
s s
p Gs p G p B
p B
ep G
e e
x
x x
x
x x x x
x
x-
+ =
= =+ +
X
x
@
-
7/31/2019 5 Lyn Thomas Risk Stream Ccri2011
6/13
6
hy introduce economic andmarket variables into score?
Normally scores thought of as static but they are really dynamic : want score
at time t to be
What scorecard gives iswhere to is when scorecard built
Solution: put economic conditions, e(t), into scorecard
Obviously spop(e) depends on e
spop(e) is transformation of population default rate; must change over time
Does woe(x,e) depend on e: if so need interaction terms between economicvariables and borrower characteristics
( , ) ( , ( )) ( ( )) ( , ( ))Pop
s t s t s t woe t x x e e x e
( , ) ( ) ( , )Pops t s t woe t x x
0 0 0( , ) ( ) ( , )
Pops t s t woe t x x
1 1( ( )) ( ) ( )pop m ms e t c e t c e t
1 1
( , ) ( ( ) ( ))m n
ij j i
i j
woe t c IndicatorFunction x e t
x
-
7/31/2019 5 Lyn Thomas Risk Stream Ccri2011
7/13
7
Which economic variables?
Little published literature on this
Some work on which variables impact on corporate defaults Some work on impact of economic conditions on mortgage defaults Really nothing on unsecured consumer credit
Possible variables
General economy:
GDP Libor interest rates Production index FTSE
Impact of economy on households Unemployment rate
Price indices Consumer confidence House price index
Lending environment Net lending Mortgage lending
-
7/31/2019 5 Lyn Thomas Risk Stream Ccri2011
8/13
8
Case Study:Invoice Discounting Example
Scorecard built to estimate default risk of small firms, wherebank is invoice discounter (like factoring)
Give loan using firms invoices to customers as collateral
Scorecard built circa 2005/6 continued to discriminate wellthrough 2009
But estimate of number of firms defaulting grossly
underestimated in 2008/9.
-
7/31/2019 5 Lyn Thomas Risk Stream Ccri2011
9/13
9
Scorecard without Economic Variables
Training In time Test Out of time Test
Gini 62 63 60
KS 46.46 48.83 46.34
HLtest (Chi
square)43.16 26.93 1470.94
Actual defaults 4666 2247 1409Expected defaults 4666 2201 605
-
7/31/2019 5 Lyn Thomas Risk Stream Ccri2011
10/13
10
Scorecard with Confidence Index andFTSE as economics estimate s pop(e)
Version 1 Training In time Test Out of timeTest
Gini 63 63 59
KS 47.12 49.00 48.80
HLtest (Chi
square)32.51 29.95 63.81
Actual defaults 4666 2247 1409Expected defaults 4666 2202 1306
-
7/31/2019 5 Lyn Thomas Risk Stream Ccri2011
11/13
11
Model with interactions(Confidence and FTSE):economics estimate of spop(e) and woe (e)
Training In time Test Out of timeTest
Gini 63 63 57
KS 47.03 49.27 44.17
HL(Chi square) 31.75 21.49 81.69
Actual defaults 4666 2247 1409
Expected
defaults4666 2204 1383
-
7/31/2019 5 Lyn Thomas Risk Stream Ccri2011
12/13
-
7/31/2019 5 Lyn Thomas Risk Stream Ccri2011
13/13
13
Conclusions Economic variables can be added to scorecards
Need longer time periods in samples to get varying economic conditions
Adding straight variables estimates spop
No improvement in discrimination Impressive improvement in probability of default prediction
Adding interaction variables estimates woe(x,e)
Not clear what improvement this gives Which variables are affected by economics Will segmentation work better ?
Need to ensure not all time dependent changes have to be explainedby economics
Alternative approach is to keep scorecard fixed and to use economics inscore to PD transformation
Reinterpret hazard function approach in this way Roll rates/markov chains can be functions of economics