Time-to-Event Prediction with Neural Networks

In the last decades the analytical value of data has really become apparent and the amount of data collected has vastly increased. This enables us to approach problems in more data driven manners. In the thesis, I have combined recent developments in machine learning with statistical methods to better answer the question: “When in the future will a given event occur?” The first part of the thesis was done in collaboration with the Norwegian bank DNB. We created new methods for predicting when in the future customers will default on their mortgage loans. By investigating the historical balances of the customers’ checking accounts, savings accounts and credit cards, we found that we could improve on existing methods for predicting mortgage defaults. In the second part of the thesis, our attention was directed toward more general methodology that may be applied to a number of problems. Our proposed improvements were illustrated using a selection of available datasets, ranging from how gene and protein expression profiles affect the mortality of breast cancer patients, to how customer information can help determine if customers are likely to continue to subscribe to a music streaming service.