Document type
Date
CONFIDENTIAL AND PROPRIETARY
Any use of this material without specific permission of McKinsey & Company is strictly prohibited
Remarks on “Measuring the Quality of Credit Scoring Models”
October 22, 2010
Discussant remarks

1 Footnote
SOURCE: Source
Title
Unit of measure
‹#›
|
1
Summary and contribution of the paper
Summary
▪Credit scoring models are widely used by financial institutions to predict the probability of
default of their clients
▪The paper reviews the methods used to measure quality of credit scoring models, based on both
distribution functions and density functions, extends existing formulas and illustrates their
properties in simulation, as well as on real financial data
▪
Contribution
▪Useful review of the methods for quality measurement, largely based on existing literature, which
may be appreciated by practitioners
▪Pointing out that the default (bad/good) client definition, which is crucial for correct
functioning of models, and suggesting that intermediate categories should not be used
▪Extension of existing indices (KS, Gini, Lift statistics) under general case (i.e., good and bad
clients have different normal distributions)
▪Graphical simulation and real-data illustration of these indices and suggestions of computational
approaches

1 Footnote
SOURCE: Source
Title
Unit of measure
‹#›
|
2
Potential points to develop further
Suggestions for further developing the paper
▪Clarify what is the audience for this paper before developing the paper further – is the main aim
to provide new methodology or assist practitioners in their work?
▪Depending on the focus, develop and clearly stress the “so what” (main implications) in the
abstract/introduction/conclusions – for instance, in case the focus is to help practitioners
–Which methods for quality measurement are more suitable under different circumstances, given their
behavior?
–What are the thresholds for model performance that should be used in practice?
–What if scores are not normally distributed? What is the most likely deviation from normality and
what implications does it have for quality measurement?
–Can authors share programs they developed to make the practitioners’ job easier?

1 Footnote
SOURCE: Source
Title
Unit of measure
‹#›
|
3
7/1/04
For discussion: What could be wrong with models estimating probability of default?
Jan 06
Jan 07
Jan 04
Jan 05
Jan 08
Jan 09
Jan 10
Jan 11
6x
Percent
SOURCE: S&P/Experian Consumer Credit Default Indices, available at www.standardandpoors.com
First mortgage defaults (US)