Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Incorporating data drift to perform survival analysis on credit risk

Jianwei Peng, Stefan Lessmann

TL;DR

This paper tackles non-stationarity in mortgage default prediction by introducing a dynamic LMISO framework that combines a balance-deviation longitudinal marker with a discrete-time hazard, augmented by landmarking and isotonic calibration to handle data drift. The approach links endogenously evolving borrower behaviour to time-to-default through a shared longitudinal signal, while accounting for temporal heterogeneity and calibration drift via landmark-specific baselines and a monotone calibrator. Empirical results across sudden, incremental, and recurring drift scenarios show LMISO outperforms classical survival models, drift-adaptive learners, and gradient boosting methods in discrimination and calibration, highlighting the practical value for risk monitoring and stress testing in non-stationary portfolios. The work contributes a transparent, scalable framework suitable for large mortgage portfolios and points to extensions toward multi-trajectory modelling, macroeconomic integration, and competing risks.

Abstract

Survival analysis has become a standard approach for modelling time to default by time-varying covariates in credit risk. Unlike most existing methods that implicitly assume a stationary data-generating process, in practise, mortgage portfolios are exposed to various forms of data drift caused by changing borrower behaviour, macroeconomic conditions, policy regimes and so on. This study investigates the impact of data drift on survival-based credit risk models and proposes a dynamic joint modelling framework to improve robustness under non-stationary environments. The proposed model integrates a longitudinal behavioural marker derived from balance dynamics with a discrete-time hazard formulation, combined with landmark one-hot encoding and isotonic calibration. Three types of data drift (sudden, incremental and recurring) are simulated and analysed on mortgage loan datasets from Freddie Mac. Experiments and corresponding evidence show that the proposed landmark-based joint model consistently outperforms classical survival models, tree-based drift-adaptive learners and gradient boosting methods in terms of discrimination and calibration across all drift scenarios, which confirms the superiority of our model design.

Incorporating data drift to perform survival analysis on credit risk

TL;DR

This paper tackles non-stationarity in mortgage default prediction by introducing a dynamic LMISO framework that combines a balance-deviation longitudinal marker with a discrete-time hazard, augmented by landmarking and isotonic calibration to handle data drift. The approach links endogenously evolving borrower behaviour to time-to-default through a shared longitudinal signal, while accounting for temporal heterogeneity and calibration drift via landmark-specific baselines and a monotone calibrator. Empirical results across sudden, incremental, and recurring drift scenarios show LMISO outperforms classical survival models, drift-adaptive learners, and gradient boosting methods in discrimination and calibration, highlighting the practical value for risk monitoring and stress testing in non-stationary portfolios. The work contributes a transparent, scalable framework suitable for large mortgage portfolios and points to extensions toward multi-trajectory modelling, macroeconomic integration, and competing risks.

Abstract

Survival analysis has become a standard approach for modelling time to default by time-varying covariates in credit risk. Unlike most existing methods that implicitly assume a stationary data-generating process, in practise, mortgage portfolios are exposed to various forms of data drift caused by changing borrower behaviour, macroeconomic conditions, policy regimes and so on. This study investigates the impact of data drift on survival-based credit risk models and proposes a dynamic joint modelling framework to improve robustness under non-stationary environments. The proposed model integrates a longitudinal behavioural marker derived from balance dynamics with a discrete-time hazard formulation, combined with landmark one-hot encoding and isotonic calibration. Three types of data drift (sudden, incremental and recurring) are simulated and analysed on mortgage loan datasets from Freddie Mac. Experiments and corresponding evidence show that the proposed landmark-based joint model consistently outperforms classical survival models, tree-based drift-adaptive learners and gradient boosting methods in terms of discrimination and calibration across all drift scenarios, which confirms the superiority of our model design.
Paper Structure (19 sections, 18 equations, 2 figures, 18 tables)

This paper contains 19 sections, 18 equations, 2 figures, 18 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Longitudinal Marker
  5. Behavioural marker: scheduled vs actual repayment
  6. Loan-specific longitudinal trajectory (per-loan regression)
  7. Landmarking
  8. Discrete-Time Hazard (Logistic survival model)
  9. Landmark One-Hot Encoding (LM): adjusting for temporal non-stationarity
  10. Isotonic Calibration (ISO): making predicted probabilities well-calibrated
  11. Joint LMISO model summary
  12. Data and Experimental Settings
  13. Data Description and Preprocessing
  14. Data Drift Simulation
  15. Experimental Settings
  16. ...and 4 more sections

Figures (2)

  • Figure 1: Parametric drift schedules used for simulation (schematic). The month index spans the first 60 months for illustration; $t_s=20$ and $t_e=40$ correspond to one-third and two-thirds of the span, respectively.
  • Figure 2: Drift level comparison 2020.