From cox to DeepHit: A comprehensive analysis of aircraft longevity using time-varying and machine learning models

Estimating the economic useful life (EUL) of commercial aircraft is governed by extreme right-censoring: with over 70% of 35,212 aircraft still in service, retired-only analysis underestimates longevity through exclusion bias, while full-fleet analysis must work from relatively few observed retirements. We argue that this censoring, rather than the choice of model, sets the limit of what can be estimated, and we address it with a staged strategy comprising nonparametric Kaplan-Meier estimation, Cox models with time-varying coefficients, and a DeepHit neural survival network, each introduced to overcome a specific limitation of the previous stage. In individual aircraft-model × usage cells, estimated median EULs range from roughly 25–30 years for passenger jets to 29–42 years for their cargo counterparts, and cargo aircraft outlive passenger aircraft within every estimable model pair (87.7% vs. 33.2% 30-year survival across the pooled fleet). Leasing is associated with later retirement (24.2% lower estimated hazards for passenger aircraft and 42.6% lower estimated hazards for cargo aircraft), and, in pooled passenger-aircraft models, the estimated Airbus–Boeing contrast reverses over the lifecycle, with Airbus aircraft facing higher early but lower late retirement risk; this contrast may partly reflect differences in the manufacturers’ model composition. The censoring constraint becomes especially apparent at the finest resolution: many recently introduced types and modern freighter cells have too few retirements for reliable model-level estimation, motivating a pooled model that shares information across aircraft types. DeepHit, using a learned type embedding and lagged oil prices as leading predictors, attains the strongest predictive accuracy (C-index 0.89 vs. 0.64–0.77 for the Cox-family models), an indicative gain rather than a strict like-for-like ranking because the models are scored on different samples and representations. Reliable EUL estimation thus depends less on any single method than on matching the method to what extreme censoring leaves observable.