نبذة مختصرة : A well functioning economy requires a stable credit market. Computational intelligence methods could provide a method to reduce the amount of uncertainty in the markets. This report examines four different methods for predicting the probability for defaults of credit card clients in Taiwan. The four selected methods were Linear Discriminant Analysis, Support Vector Machines, Artificial Neural Networks and Deep Neural Networks. The models were then evaluated with regards to five different methods: Area Under the Curve for the Receiver Operating Characteristic, accuracy, precision, sensitivity and specificity. The results showed that all models performed better than random with similar results, except for the Support Vector Machine which in our testing configuration incorrectly classified almost all debtors that defaulted on their debt. Although there was no clearly superior model the results showed that the Deep Neural Networks and Linear Discriminant Analysis were the two most promising methods. ; En välfungerande ekonomi behöver en stabil kreditmarknad. Maskininlärningsmetoder har potential att reducera osäkerheten på marknaden. Rapporten undersöker fyra olika metoder för att beräkna sannolikheten att en låntagare återbetalar sin kreditkortsskuld baserat på kreditkortsdata från Taiwan. Metoderna som valdes var Linear Discriminant Analysis, Support Vector Machines, Arti- ficiella Neurala Nätverk och Djupa Neurala Nätverk. Modellerna utvärderades med avseende på fem olika metoder: Area Under the Curve for the Receiver Operating Characteristic, nogrannhet, precision, sensitivitet och specificitet. Resultaten visade att alla modeller presterade bättre än slump med liknande resultat utom för Support Vector Machines som i vår testkonfiguration felaktigt klassificerade nästintill alla låntagare som inte skulle återbetala. Även om ingen modell var tydligt bättre än de andra visade resultaten att Djupa Neurala Nätverk och Linear Discriminant Analysis är metoderna som visar mest potential.
No Comments.