Modelling the term-structure of default risk under IFRS 9 within a multistate regression framework
Arno Botha, Tanja Verster, Roland Breedt · 2025
Read Modelling the term-structure of default risk under IFRS 9 within a multistate regression framework by Arno Botha, Tanja Verster, Roland Breedt (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-14479.txt.
Introduction
Arno Botha, Tanja Verster, and Roland Breedt study how banks can guess the risk that a loan will fail over its whole life, in a 2025 paper. Banks must set aside money for such losses under a rule called IFRS 9. A poor guess can leave a bank with too little set aside.
The team compares three ways to model how loans move between states, such as paid, late, and default. One is a simple chain model, and the other two are forms of regression that use loan and economy data. They test the models on home loan data. Each richer model did better than the one before it. The team also offers simple checks that other banks can reuse.
Catalog record
| CREATOR | Arno Botha, Tanja Verster, Roland Breedt |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 91 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-14479
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Modelling the term-structure of default risk under IFRS 9 within a multistate regression framework Arno Botha, Tanja Verster, Roland Breedt Abstract The lifetime behaviour of loans is notoriously difficult to model, which can compromise a bank’s financial reserves against future losses, if modelled poorly. Therefore, we present a data-driven comparative study amongst three techniques in modelling a series of default risk estimates over the lifetime of each loan, i.e., its term-structure. The behaviour of loans can be described using a nonstationary and time-dependent semi-Markov model, though we model its elements using a multistate regression-based approach. As such, the transition probabilities are explicitly modelled as a function of a rich set of input variables, including macroeconomic and loan-level inputs. Our modelling techniques are deliberately chosen in ascending order of complexity: 1) a Markov chain; 2) beta regression; and 3) multinomial logistic regression. Using residential mortgage data, our results show that each successive model outperforms the previous, likely as a result of greater sophistication. This finding required devising a novel suite of simple model diagnostics, which can itself be reused in assessing sampling representativeness and the performance of other modelling techniques. These contributions surely advance the current practice within banking when conducting multistate modelling. Consequently, we believe that the estimation of loss reserves will be more timeous and accurate under IFRS 9. Keywords— Multistate modelling; Beta regression; Multinomial logistic regression; IFRS 9; Credit risk. JEL: C32, C52, G21. Word count (excluding front matter and appendix): 8211 Disclosure of interest and declaration of funding This work is financially supported wholly/in part by the National Research Foundation of South Africa (Grant Number 126885), with no known conflicts of interest that may have influenced the outcome of this work. The authors would like to thank all anonymous referees and editors for their extremely valuable contributions that have substantially improved this work. 1 Introduction Lending poses the fundamental risk of capital loss should the borrower fail to repay their loan, which necessitates the accurate prediction of the borrower’s underlying probability of default (PD). This task usually involves finding a statistical relationship between a set of borrower-specific input variables and the binary-valued repayment outcome (i.e., defaulted or not) over some outcome period. The literature on this particular classification task is considerable and spans various forms of supervised statistical learning, including machine learning; see Hand & Henley 1997a , Siddiqi 2005a , Thomas 2009a , Hao et al. 2010a , Baesens et al. 2016a , and Louzada et al. 2016a . However, these credit rating systems focus mostly on producing a conservative PD-estimate that remains static (but stressed) over the lifetime of each loan, often by design, as discussed by Crook & Bellotti 2010a . The broad goal of such systems is to facilitate the estimation of regulatory and economic capital, which should absorb any catastrophic (or unexpected) losses under the Basel framework from the BCBS 2019a . Any temporal effects that might affect the PD during loan life are therefore largely ignored, together with any macroeconomic influences; particularly since the latter is already assumed to be stressed to a recession-like level during PD-estimation. Doing so renders the resulting PD-estimates as through-the-cycle (TTC) in that they should at least approximate the long-run averages of 1-year historical default rates over a full macroeconomic cycle, as required during capital estimation. While these TTC PD-estimates are certainly stable over time by design, they are also typically inaccurate within any other setting besides capital estimation. The alternative, as reviewed and compared by Crook & Bellotti 2010a , is a point-in-time (PIT) rating system, which produces more dynamic PD-estimates that agree more closely with the observed variation in default risk over loan life, as well as incorporate any temporal macroeconomic effects. Such dynamicity is perhaps inappropriate for capital estimation since capital levels should preferably not fluctuate wildly over time. However, these PIT-based PD-estimates are more flexible in that they can be consumed across a greater variety of settings than their TTC-based counterparts. In fact, the introduction of the IFRS 9 accounting standard by the IASB 2014a provided additional impetus for such dynamicity in PD-modelling. Under IFRS 9, a financial asset’s value should be comprehensively adjusted according to a bank’s (evolving) expectation of the asset’s credit risk over time, i.e., the potential loss induced by default. In principle, the bank willingly forfeits a portion of its income at each reporting period into a loss provision, which should ideally offset any amounts that are written-off in future; thereby smoothing away volatility. The provision size is regularly updated based on a statistical model of the asset’s expected credit loss (ECL), wherein the PD is embedded as perhaps the most important risk parameter. The bank dynamically adjusts its loss provision either by raising more from earnings or releasing a portion thereof back into the income statement, defined respectively as an impairment loss or gain according to §5.5.8 in IFRS 9. Any unnecessary impairment variation, perhaps due to an inaccurate ECL-model, would therefore directly affect a bank’s income statement. The ECL may itself be calculated by following a 3-stage approach (§5.5.3, §5.5.5) given the extent of the perceived deterioration (or improvement) in credit risk. In principle, the ECL-estimate of each impairment stage should become increasingly severe, thereby allowing the timeous – and more dynamic – recognition of credit losses; see §B5.5.3 of IFRS 9, PwC 2014a , EY 2018a , and Botha et al. 2025a . In achieving such dynamicity, and especially for Stages 1-2, Skoglund 2017a noted that risk models need to project default risk ideally over various time horizons across loan life and against the changing macroeconomic background. This rather non-trivial task implies the estimation of a marginal (or PIT) PD as a function of a rich set of input variables, including macroeconomic covariates. These inputs are measured at each discrete period t=t_{1},\dots,\mathcal{T} during a loan’s lifetime \mathcal{T} , starting from its time of initial recognition t_{1} . The collection of these PD-estimates over the lifetime of a loan is then called the term-structure of default risk. This term-structure typically manifests as a non-linear and right-skewed curve over loan life. Put differently, the non-linearity of the term-structure speaks to the required dynamicity of ECL-estimates under IFRS 9. However, there are certain modelling challenges to rendering such dynamic and time-sensitive PD-estimates. Chief among them is due to the fact that ‘default’ is not necessarily an absorbing state into which a loan is forever trapped, as discussed by Botha 2021a . This dynamicity is acknowledged in both §36.74 of the Basel framework and in Article 178(5) of the Capital Requirements Regulation (CRR), promulgated by the European Parliament 2013a for EU-markets, which requires banks to rate loans as performing whenever default criteria cease to apply. If ‘default’ is structured as a transient state during PD-estimation, then one can leverage the full credit histories that are otherwise etched with multiple cycles of curing from default and defaulting again. Another major modelling challenge arises from the fact that ‘default’ is not the only failure-inducing event, despite its importance in credit risk modelling. Other events that may ultimately affect the risk of loss under IFRS 9 include prepayments (or early settlement), write-offs, and restructures. These competing risks will preclude the default-event from