Using Machine Learning for Default Prediction
Name
Martti Praks
Abstract
Default prediction is one of the key activities for a financial institution when estimating credit risks. The likelihood of default is an important indicator to decide if and with what conditions credit can be given and how the whole credit portfolio is performing. In general, used models are divided into two domains: statistical approaches and machine learning techniques. The main result of the thesis is a comparison between models created with logistic regression and other main machine learning techniques using actual data from AS LHV Group.
The thesis displays different default prediction models and includes discussions over benefits provided by different machine learning techniques. Best results are achieved with the random forest method. Different methods are used to explain the decision-making mechanisms of the random forest model to support using it in practice. Considering recent successes of decision tree-based models, the results are not surprising. Random forest results are explained by feature influences and concrete examples of why some probabilities were provided. Whether these methods are enough to use the random forest in practice, is left up to decide by the end-user.
The thesis displays different default prediction models and includes discussions over benefits provided by different machine learning techniques. Best results are achieved with the random forest method. Different methods are used to explain the decision-making mechanisms of the random forest model to support using it in practice. Considering recent successes of decision tree-based models, the results are not surprising. Random forest results are explained by feature influences and concrete examples of why some probabilities were provided. Whether these methods are enough to use the random forest in practice, is left up to decide by the end-user.
Graduation Thesis language
Estonian
Graduation Thesis type
Master - Data Science
Supervisor(s)
Markus Kängsepp, Meelis Kull, Kuldar Kõiv
Defence year
2022