Deep learning for survival and competing risk modelling

Gabriel Blumenstock, Stefan Lessmann and Hsin-Vonn Seow

Journal of the Operational Research Society, 2022, vol. 73, issue 1, 26-38

Abstract: The article examines novel machine learning techniques for survival analysis in a credit risk modelling context. Using a large dataset of US mortgages, we evaluate the adequacy of DeepHit, a deep learning-based competing risk model, and random survival forests. The observed results provide strong evidence that both models predict default and prepayment risk more accurately than statistical benchmarks in the form of the Cox proportional hazard model and the Fine and Gray model. The superiority of the machine learning models is robust across different periods including stressed periods. We also find machine learning models do not require larger amounts of training data than the statistical benchmarks. Finally, we extend methods for estimating feature importance scores to deep neural networks for survival analysis and clarify which covariates determine the estimated survival functions of DeepHit. An online companion with additional results is available in Supplementary Information.

Date: 2022
References: Add references at CitEc
Citations: View citations in EconPapers (4)

Downloads: (external link)
http://hdl.handle.net/10.1080/01605682.2020.1838960 (text/html)
Access to full text is restricted to subscribers.

Related works:
This item may be available elsewhere in EconPapers: Search for items with the same title.

Export reference: BibTeX RIS (EndNote, ProCite, RefMan) HTML/Text

Persistent link: https://EconPapers.repec.org/RePEc:taf:tjorxx:v:73:y:2022:i:1:p:26-38

Ordering information: This journal article can be ordered from
http://www.tandfonline.com/pricing/journal/tjor20

DOI: 10.1080/01605682.2020.1838960

Access Statistics for this article

Journal of the Operational Research Society is currently edited by Tom Archibald

More articles in Journal of the Operational Research Society from Taylor & Francis Journals
Bibliographic data for series maintained by Chris Longhurst ().