EBA RTS on the specification of the nature, severity and duration of an economic downturn, and the business case for credit cycle modelling
The EBA’s consultation postbag appears to have been heavier than usual, perhaps evidenced by the second public consultation on economic downturn EBA/CP/2018/07, the consultation on guidelines for downturn LGD CP/2018/08, the quantity and nature of feedback referenced in the feedback to the public consultation, and perhaps most importantly the significant change in methodology since the original consultation CP/2017/02. We summarise the changes in each draft below:
Between the second and final drafts, there are relatively few changes in the articles of the RTS. Changes include:
- Definition of the downturn period in terms of peaks and troughs is superseded by the period relating to the most severe values observed.
- Definition of the downturn severity in terms of the preceding 20 years is clarified as the 20 years preceding the downturn observation.
- The downturn duration may exceed 12 months, in the case that multiple consecutive severe index values are observed (which may include severe values across several indices).
- It is clarified that the downturn definition should be reviewed at least annually.
- The regulation applies from 1 January 2021, replacing 31 December 2019.
Between the first and second drafts, the changes were more substantial. An approach that depended on the underlying CCF and LGD methodologies was entirely replaced with a simpler approach that decouples downturn identification from the underlying model design.
The objective of IRB credit modelling (or, indeed, any credit modelling) should be to build models with sufficient goodness-of-fit at an appropriate level of complexity to be useful for internal applications, in a manner that is compliant. The concept of including an exogenous component in IRB model estimation is not only sensible, but presents an opportunity for firms to further harmonise the input assumptions in their IFRS 9 and IRB modelling approaches’ PIT calibrations.
One aspect of PIT modelling that continues to divide opinion is the choice of macroeconomic index in explaining the exogenous component of credit risk performance (i.e. default rates, conversion factors, and loss given default), with broadly two schools of thought:
- We observe many firms that have used regression techniques to link the exogenous component of credit risk performance with familiar macro indices such as GDP and unemployment rate. Indeed, in a previous article we showed how a threshold model linked to output gap can improve the goodness-of-fit in an unsecured retail model. Whilst such approaches pass the “reasonable” and “intuitive” tests, they are generally non-parsimonious and may also rely on Gauss-Markov assumptions that are not met in practice. Further, the goodness-of-fit often leaves a lot to be desired (we have seen firms accept OLS models with R-square values in the 50%-60% range). Nevertheless, the approach generally delivers a materially-correct result that is explainable (or adjustable) via base case macro path assumptions, with shortcomings generally addressable using an overlay or post model adjustment.
- An alternative approach, whereby the exogenous component of credit risk is modelled as a latent random variable has widespread acceptance at larger and more-sophisticated market participants whose lending more-closely resembles publicly traded corporates. Here, market-implied Estimated Default Frequency indices (as well as other credit cycle indices) invoke Efficient Market Hypothesis to argue that all information is already “priced into” the EDFs to deliver highly-accurate PIT PDs. The introduction of additional covariates in the system (such as GDP or unemployment) is rejected as spurious. Published literature (including  and ) observe that cyclical variations in historical default frequencies and recovery rates are well-explained by the latent random credit cycle index, with macroeconomic variables becoming insignificant after the CCI is introduced.
Towards convergence of IRB and IFRS 9?
For a credit modelling community that has embraced PIT modelling, the task of calibrating a downturn model is conceptually relatively straightforward: Find the macro index that explains the exogenous component of realised CCFs or LGDs, and then set that index to its downturn value at model execution.
There is, however, a snag. What if the macro index doesn’t fit particularly well? The index whose inclusion passed the “reasonable” test at financial statement materiality could easily fall foul of sharp-eyed model validators for the reasons outlined above. Where Basel encounters uncertainty, a margin of conservatism (MoC) is required.
Goodness-of-fit is so critical to model selection that it receives a category all of its own (“category C” for “general estimation error”) in EBA GL/2017/16. The impact of historical macro conditions and lending practices can be debated, but general estimation error is directly measurable and impossible to hide. The resulting MoC will directly impact the CET1 ratio, and is likely to eventually influence pricing and returns.
The MoC for estimation error in the downturn calibration is, however, avoidable to some degree. The key is to recognise that Article 2.3 of the RTS requires firms to “consider other economic factors as relevant, where these are explanatory variables for, or indicators of, the economic cycle specific to the type of exposures under consideration”.
In this article we have explored a specific example of where diligent application of the science of model selection and goodness-of-fit is in-effect rewarded in Pillar 1. We have suggested that more-widespread adoption of CCI modelling will lead to greater model accuracy, and lower conservatism under IRB. Firms’ gap assessments and options analysis should consider not only the RWA and EL impacts, but the MoC contribution to those impacts.