Beyond the Cox Model: Applying Machine Learning Techniques with Time-to-Event Data

The enhancement of data storage capacity over the years and the advancement of various technologies that enable data acquisition and support big data have increased interest in collecting information and its variation over time. In many different disciplines, the aim of monitoring this kind of observations is to analyse and predict the time until the occurrence of an event of interest. For 50 years, Cox’s semi-parametric proportional hazards regression model has been the statistical methodology most widely used for survival analysis (the analysis of time-to-event data). However, nowadays, machine learning techniques are becoming more popular and several studies assert their supremacy from the predictive point of view. The purpose of the present research is the application of new methodologies related to the Cox model and their comparison with the classic survival model in the field of corporate default risk prediction. In particular, scikit survival[1], a Python module for survival analysis designed to be compatible with scikit-learn[2] Application Program Interface (API), will be evaluated as a machine learning technique and the results compared with those obtained by applying classical survival analysis techniques implemented with the SAS® PHREG procedure.