Paper: arXiv 2505.01044
Authors: Arno Botha, Tanja Verster, Bernard Scheepers
Abstract
In the pursuit of modelling a loan’s probability of default (PD) over its lifetime, repeat default events are often ignored when using Cox Proportional Hazard (PH) models. Excluding such events may produce biased and inaccurate PD-estimates, which can compromise financial buffers against future losses. Accordingly, we investigate a few subtypes of Cox-models that can incorporate recurrent default events. Using South African mortgage data, we explore both the Andersen-Gill (AG) and the Prentice-Williams-Peterson (PWP) spell-time models. These models are compared against a baseline that deliberately ignores recurrent events, called the time to first default (TFD) model. Models are evaluated using Harrell’s c-statistic, adjusted Cox-Sell residuals, and a novel extension of time-dependent receiver operating characteristic (ROC) analysis. From these Cox-models, we demonstrate how to derive a portfolio-level term-structure of default risk, which is a series of marginal PD-estimates at each point of the average loan’s lifetime. While the TFD- and PWP-models do not differ significantly across all diagnostics, the AG-model underperformed expectations. Depending on the prevalence of recurrent defaults, one may therefore safely ignore them when estimating lifetime default risk. Accordingly, our work enhances the current practice of using Cox-modelling in producing timeous and accurate PD-estimates under IFRS 9.
Complexity vs Empirical Score
- Math Complexity: 7.0/10
- Empirical Rigor: 8.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper employs advanced survival analysis and recurrent event models (Cox variants) with mathematical derivations, while rigorously testing models on real mortgage data using statistical diagnostics and time-dependent ROC analysis.
Research Flowchart
flowchart TD
A["Research Goal: Model Lifetime PD<br>for loans, including recurrent defaults"] --> B{"Data Source"}
B --> C["Mortgage Data<br>South African Loans"]
C --> D["Modeling Strategy"]
subgraph D ["Methodology: Recurrent Event Cox Models"]
D1["Time to First Default<br>Baseline"]
D2["Anderssen-Gill AG<br>Recurrent Event"]
D3["Prentice-Williams-Peterson PWP<br>Spell-Time"]
end
D --> E["Model Evaluation"]
subgraph E ["Validation Metrics"]
E1["Harrell's c-statistic"]
E2["Adjusted Cox-Sell Residuals"]
E3["Time-Dependent ROC"]
end
E --> F["Key Outcomes"]
subgraph F ["Findings"]
F1["Portfolio-level<br>Term Structure of PD"]
F2["TFD & PWP models<br>perform comparably"]
F3["AG model<br>underperformed"]
F4["Recurrent defaults may be<br>ignored if prevalent"]
end
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes a tutorial comparing recurrent-event Cox‑regression subtypes (Andersen‑Gill, Prentice‑Williams‑Peterson) against a baseline time‑to‑first‑default model for modelling lifetime default risk.
Source sentence
we investigate a few subtypes of Cox-models that can incorporate recurrent default events. We explore both the Andersen-Gill (AG) and the Prentice-Williams-Peterson (PWP) spell-time models using real-world data as an illustration.
Assumes 5
- The baseline hazard remains relatively constant between successive spells for the AG and PWP models to yield similar results. model
Source sentence
they demonstrated that the AG– and PWP-subtypes will resemble each other in results, but only if the baseline hazard remains relatively constant between successive spells.
- The Andersen‑Gill model assumes a common baseline hazard function across all subsequent spells. model
Source sentence
the AG-model cannot easily incorporate spell-specific effects since it assumes a common baseline hazard function across all subsequent spells.
- The study uses the cohort‑end definition to reduce right‑censoring effects. data
Source sentence
We shall therefore restrict our study to the cohort-end definition in the interest of expediency.
- Resolution rates of datasets should be close to zero difference to minimise sampling bias. statistical
Source sentence
should be as close to zero as possible, which would minimise sampling bias and affirm both datasets to be representative of each other.
Show 1 more
- Only a small fraction of loans experience multiple defaults (≈7%). data
Source sentence
only a paltry 7% of loans in our sample have experienced multiple defaults.
Data
- Residential mortgage portfolio macro unknown
Source sentence
Each technique was fit to a data-rich portfolio of residential mortgages from the South African credit market as an illustration.
Costs
Not applicable
Finds 4
- The Andersen‑Gill model underperformed substantially, indicating that assuming a common baseline hazard is a subpar modelling choice. negative
Source sentence
the AG-technique underperformed our expectations quite substantially, which implies that assuming a common baseline hazard is a subpar modelling choice.
- The Prentice‑Williams‑Peterson technique slightly outperforms the Andersen‑Gill technique, suggesting subtle baseline hazard differences across spells. performance
Source sentence
the PWP-technique slightly outperforms the AG-technique, which implies that subtle differences exist in the baseline hazard function across subsequent spells.
- Time‑to‑first‑default and PWP models show negligible differences in MAE, indicating little benefit from including recurrent events in this dataset. statistical
Source sentence
the difference in MAE-values is negligible between the TFD– and PWP-techniques, with the PWP-technique outperforming the former only marginally.
- All three Cox models achieve remarkably similar goodness‑of‑fit and discriminatory power across diagnostics. statistical
Source sentence
the three different Cox-models achieve remarkably similar levels of both goodness-of-fit and discriminatory power, even when measuring the latter over different time horizons.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 20 related papers). Coverage is partial, so this is neither replication nor refutation.
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (61)
- Modelling the term-structure of default risk under IFRS 9 within a multistate regression framework — → 2502.14479
- The TruEnd-procedure: Treating trailing zero-valued balances in credit data — → 2404.17008
- Defining and comparing SICR-events for classifying impaired loans under IFRS 9 — → 2303.03080
- A procedure for loss-optimising the timing of loan recovery under uncertainty (2021) — link
- Simulation-based optimisation of the timing of loan recovery across different portfolios (2021) — link
- Towards modelling lifetime default risk: Exploring different subtypes of recurrent event Cox-regression models [Source Code] (2025) — link
- Credit risk scorecards: developing and implementing intelligent credit scoring (2005)
- Consumer Credit Models: Pricing, Profit and Portfolios (2009) — link
and 53 more
- Credit risk analytics: Measurement techniques, applications, and examples in SAS (2016)
- Credit scoring and its applications (2002) — link
- Credit Risk Management: Basic Concepts (2009) — link
- Survival analysis: a self-learning text (2012) — link
- Multivariate survival analysis and competing risks (2012)
- Survival analysis (2005) — link
- Modeling survival data: extending the Cox model (2000) — link
- An evaluation of the Cox-Snell residuals (2015)
- International Convergence of Capital Measurement and Capital Standards: A Revised Framework, Comprehensive Version (2006) — link
- The Basel Framework (2019) — link
- International Financial Reporting Standard ( IFRS ) 9: Financial Instruments (2014) — link
- Regulation (EU) No 575/2013 of the European Parliament and of the Council of 26 June 2013 on prudential requirements for credit institutions and investment firms and amending Regulation (EU) No 648/2012 Text with EEA relevance (2013) — link
- Statistical classification methods in consumer credit scoring: a review (1997) — link
- Review of the literature on credit risk modeling: development of the past 10 years (2010) — link
- Classification methods applied to credit scoring: Systematic review and overall comparison (2016)
- Time Varying and dynamic models for default risk in consumer loans (2010)
- Impairment of Financial Instruments under IFRS 9 (2018)
- IFRS 9 : Expected Credit Losses (2014)
- Credit risk term-structures for Lifetime Impairment Forecasting: A practical guide (2017) — link
- Estimating Lifetime Expected Credit Losses Under IFRS 9 (2016) — link
- It’s about time: Using discrete-time survival analysis to study duration and the timing of events (1993)
- It’s d 'e j `a vu all over again: Using multiple-spell discrete-time survival analysis (1995) — link
- Survival analysis (2016) — link
- Survival analysis and interpretation of time-to-event data: the tortoise and the hare (2018) — link
- Survival analysis and the credit granting decision (1992)
- Not if but when will borrowers default (1999) — link
- Survival analysis methods for personal loan data (2002) — link
- Credit scoring with macroeconomic variables using survival analysis (2009) — link
- Forecasting and stress testing credit card default using dynamic models (2013)
- Retail credit stress testing using a discrete hazard model with macroeconomic factors (2014) — link
- Time to default in credit scoring using survival analysis: a benchmark study (2017) — link
- Modelling recurrent events: a tutorial for analysis in epidemiology (2015) — link
- A systematic comparison of recurrent event models for application to composite endpoints (2018) — link
- Additional File to the Article ’A Systematic Comparison of Recurrent Event Models for Application to Composite Endpoints’ (2018) — link
- Regression models and life-tables (1972)
- Modelling and forecasting recurrent recovery events on consumer loans (2019) — link
- Applying recurrent event analysis to understand the causes of changes in firm credit ratings (2012) — link
- An overview of statistical methods for multiple failure time data in clinical trials (1997) — link
- Survival analysis for recurrent event data: an application to childhood infectious diseases (2000) — link
- An introduction to ROC analysis (2006) — link
- Time-dependent ROC curves for censored survival data and a diagnostic marker (2000) — link
- A tutorial on evaluating the time-varying discrimination accuracy of survival models used in dynamic decision making (2018) — link
- Nearest neighbor estimation of a bivariate distribution under random censoring (1994) — link
- Variable selection–a review and recommendations for the practicing statistician (2018) — link
- Applied Data Science Using PySpark (2021)
- Building Multiple Regression Models Interactively (1981) — link
- Development of an Impairment Point in Time Probability of Default Model for Revolving Retail Credit Products: South African Case Study (2021) — link
- Too big to fail: large samples and the p-value problem (2013) — link
- On the modelling of prognosis from delinquency to normal performance on retail consumer loans (2016) — link
- External validation of a Cox prognostic model: principles and methods. (2013) — link
- Concordance probability and discriminatory power in proportional hazards regression (2005) — link
- Information Theory and an Extension of the Maximum Likelihood Principle (1998) — link
- A window of opportunity: Assessing behavioural scoring (2013) — link
3 of 61 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.