Greining á fjölvíddarúmi með sjálfbeinandi aðferðum fyrir lánshæfismat fyrirtækja

Credit risk analysis is a widely researched topic and forms the foundation that aids in decision making for numerous businesses around the world. However, with the increasing data availability and complexity, it has become more difficult to identify key variables that best serve to distinguish between financially healthy and unhealthy firms. This thesis presents an approach to classify firms into homogeneous groups based on their characteristics by applying unsupervised machine learning techniques. Using financial information, we train an autoencoder to perform dimensionality reduction and then proceed to apply and compare different clustering techniques on the reduced feature space. In addition, we compare the performance of our autoencoder with a more traditional approach. Our results show that by performing cluster analysis on a reduced space constructed by an autoencoder we can extract valuable relationships from relatively large data sets by partitioning the data objects based on their similarities. However, our results also indicate that unsupervised techniques perform very poorly in assessing the defaulting probability of firms