Alessia Ciccaglione
An Equitable Framework for Dynamic Credit Risk: Fairness Penalization in Discrete-Time Survival Models.
Rel. Eliana Pastor, Rosa Elvira Lillo Rodriguez, Sandra Benítez-Peña. Politecnico di Torino, Corso di laurea magistrale in Data Science And Engineering, 2026
Abstract
Credit risk assessment governs access to essential services, with social consequences that make fairness a necessary concern. Because this risk is not fixed but evolves as the loan progresses, fairness too must be understood as changing over time. This thesis proposes a framework for equitable dynamic credit risk that embeds an in-processing fairness penalty into a discrete-time survival model built on a multi-layer perceptron and compares it against an equivalent static classifier. The penalty equalizes error rates across groups defined by sensitive attributes and is applied under several aggregation modes and temporal weighting schedules to reflect different disparity patterns. The dynamic and static models are trained and tested on both a simulated and a constructed real-world dataset that combines, from separate but linkable sources, time-varying loan trajectories and demographic attributes (race, sex and age).
To assess the fairness-accuracy trade-off, the models are evaluated on both dimensions, including a new metric that captures the temporal evolution of group disparities
Relatori
Anno Accademico
Tipo di pubblicazione
Numero di pagine
Informazioni aggiuntive
Corso di laurea
Classe di laurea
Ente in cotutela
Aziende collaboratrici
URI
![]() |
Modifica (riservato agli operatori) |
