Answers are grounded in this paper and cite their sources.
Random survival forests for competing risks (CR-RSF) are nonparametric ensemble methods for predicting event-specific cumulative incidence functions when subjects face multiple competing events. However, limited interpretability methods have hindered their adoption in regulated industries such as financial services, where model transparency is essential for risk management and regulatory compliance. This paper extends three model-agnostic global interpretability methods to the competing risks setting: Permutation feature importance (PFI), partial dependence plots (PDP), and accumulated local effects (ALE). These methods have been developed for single-risk random survival forests (RSF) but, to our knowledge, have not previously been adapted to competing risks. Demonstrating them on a car lease dataset, we find that covariate effects are often event-specific and vary across the lease lifecycle, and that the common alternative of fitting separate single-risk models distorts the interpretability outputs themselves for predictors whose signal is shared across competing events. As a supporting contribution, we construct a valid dataset for the joint estimation of default and prepayment incidence by categorising contracts according to their time to maturity, which avoids the bias that arises when non-random maturity events are treated as censored. These results emphasise the importance of dynamic, event-specific risk assessment throughout the lease lifecycle.
No takes yet. Share an insight, caveat, or question.
Sexton et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: