Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Using GEV-regression to improve accuracy of probability of default in low default portfolios

Alsin, Fredrik (2014) FMS820 20141
Mathematical Statistics
Abstract (Swedish)
Calculating probability of default (PD) for low default portfolios in a
statistically sound way can be a daunting task. For groups of coun-
terparties with no or few defaults, e.g. large corporates and nancial
institutions, standard approaches like logistic regression fail due to the
low number of events. In this thesis, a dierent approach is used. The
logit function in logistic regression is replaced with the inverse of the
Generalized Extreme Value (GEV) distribution function, called GEV-
regression.
The LASSO regression technique provides a way to automatically
nd the attributes that contain the most information in the data by
including all attributes, and adding a penalty for complexity in the
maximization of the likelihood... (More)
Calculating probability of default (PD) for low default portfolios in a
statistically sound way can be a daunting task. For groups of coun-
terparties with no or few defaults, e.g. large corporates and nancial
institutions, standard approaches like logistic regression fail due to the
low number of events. In this thesis, a dierent approach is used. The
logit function in logistic regression is replaced with the inverse of the
Generalized Extreme Value (GEV) distribution function, called GEV-
regression.
The LASSO regression technique provides a way to automatically
nd the attributes that contain the most information in the data by
including all attributes, and adding a penalty for complexity in the
maximization of the likelihood function for the parameters. This ap-
proach is used to make a strong predictive model. Data was provided
by Danske Bank consisting of two portfolios of anonymized counterpar-
ties. PD was calculated for each observation, and the accuracy of the
predictions using GEV-regression and logistic regression was compared.
The purpose of this thesis is twofold. First, to use the data provided
to create a model for probability of default, using LASSO regression.
Second, to investigate if GEV-regression is suitable for calculating PD
for low default portfolios.
The analysis suggests that it is possible to make a good predictive
model using LASSO regression. It also supports that GEV-regression
is a viable option to logistic regression, as it had a similar accuracy. It
does not however support that GEV-regression is better than logistic
regression in low default portfolios. Further studies with more data is
needed to conclude which methodology is the optimal one. (Less)
Please use this url to cite or link to this publication:
author
Alsin, Fredrik
supervisor
organization
course
FMS820 20141
year
type
H2 - Master's Degree (Two Years)
subject
language
English
id
4611244
date added to LUP
2014-08-26 15:11:39
date last changed
2014-08-26 15:11:39
@misc{4611244,
  abstract     = {{Calculating probability of default (PD) for low default portfolios in a
statistically sound way can be a daunting task. For groups of coun-
terparties with no or few defaults, e.g. large corporates and nancial
institutions, standard approaches like logistic regression fail due to the
low number of events. In this thesis, a dierent approach is used. The
logit function in logistic regression is replaced with the inverse of the
Generalized Extreme Value (GEV) distribution function, called GEV-
regression.
The LASSO regression technique provides a way to automatically
nd the attributes that contain the most information in the data by
including all attributes, and adding a penalty for complexity in the
maximization of the likelihood function for the parameters. This ap-
proach is used to make a strong predictive model. Data was provided
by Danske Bank consisting of two portfolios of anonymized counterpar-
ties. PD was calculated for each observation, and the accuracy of the
predictions using GEV-regression and logistic regression was compared.
The purpose of this thesis is twofold. First, to use the data provided
to create a model for probability of default, using LASSO regression.
Second, to investigate if GEV-regression is suitable for calculating PD
for low default portfolios.
The analysis suggests that it is possible to make a good predictive
model using LASSO regression. It also supports that GEV-regression
is a viable option to logistic regression, as it had a similar accuracy. It
does not however support that GEV-regression is better than logistic
regression in low default portfolios. Further studies with more data is
needed to conclude which methodology is the optimal one.}},
  author       = {{Alsin, Fredrik}},
  language     = {{eng}},
  note         = {{Student Paper}},
  title        = {{Using GEV-regression to improve accuracy of probability of default in low default portfolios}},
  year         = {{2014}},
}