Rogelio Andrade Mancisidor

«Deep Generative Models in Credit Scoring»

 

Popular scientific abstract:

Banks need to develop effective credit scoring models to better understand the relationship between customer information and the customer's ability to repay a loan. The output of such a model is called the default probability and is used to rank loan applications in terms of their creditworthiness. The focus of this thesis is to develop novel credit scoring methodologies that solve well-known problems in the field and that bridge the gap between simple neural networks and advanced methodologies in deep learning applied to credit scoring. Deep learning is a system built of a cascade of trainable modules, where we train all modules simultaneously and each of the modules adjust itself to produce the right answer, which in the case of credit scoring is an accurate estimate of creditworthiness.

We propose a new methodology to learn a useful way to transform customer information into a new data representation that is capable to capture customer creditworthiness in a well-defined clustering structure. These clusters are unknown a priori and are impossible to identified using the original customer information. Further, the clusters that we identified with our proposed method are suitable for marketing campaigns and credit risk assessment.

Banks do not know the creditworthiness of people who apply for loans and are rejected. Therefore, we develop models that can infer the unknown customer’s creditworthiness for rejected applications using probabilistic theory. Our experiments show that adding rejected applications improve the creditworthiness estimation for all loan applications.

Banks have access to multiple sources of information about their customers. For example, historical information obtained on application forms and behavior data collected during the loan period. These different sources of information are called data modalities. We develop a novel methodology that can learn a shared data representation for these two data modalities. Furthermore, our proposed method is capable of generating a missing modality using the information of the available modality. That is, we can generate the data of the future behavior of a certain person using the information collected during the application process.

The thesis is published in Munin and is available at: https://hdl.handle.net/10037/20407

 

Supervisors:

• Professor Robert Jenssen, Department of Physics and Technology, UiT (main supervisor)
• Associate Professor Michael Kampffmeyer, Department of Physics and Technology, UiT
• Researcher Kjersti Aas, Norwegian Computing Center

 

 

Evaluation committee:

• Senior Lecturer Raffaella Calabrese, University of Edinburgh, United Kingdom (1. Opponent)
• Professor Thomas Dyhre Nielsen, Universitetet i Aalborg, Denmark (2. Opponent)
• Professor Fred Godtliebsen, Department of Mathematics and Statistics, UiT (internal member og leader of the committee)

All participants will participate remotely to the defence.

 

Leader of the public defense:
The leader of the public defense is Professor Alfred Hanssen , Vice dean for Innovation, Faculty of Science and Technology, UiT.

 

Opposition ex auditorio:
If you have any questions for the candidate during the public defence, please send an e-mail to the leader of the public defence. They will announce the questions during the defence.

 

Trial lecture:

The trial lecture is held Friday February 5th at 10:15 AM in the same auditorium.

Title of the trial lecture: «Binary classifiers»

 

Streaming:

The defense and trial lecture will be streamed via Mediasite: 

https://mediasite.uit.no/Mediasite/Catalog/Full/97e119178b004b728e80bd7832f093ce21 

 

Audience:

UiT follows the national guidelines regarding infection control. A maximum of 20 people are allowed in the auditorium during the defence, as long as everybody keeps a distance of 1 meter at all times.