Harmonising RWA Estimates: Striking the Right Chord for IRB Unsecured Retail Portfolios

In early 2016, regulators embarked on a policy objective of reducing undue variability between Risk Weighted Assets (RWAs) amongst firms using the Internal Ratings Based (IRB) approach. RWA harmonisation is desirable not only from a supervisory perspective, but also internally and externally in the market. Internally, a “unit” of RWAs should carry the same meaning across products and portfolios. And externally, investors typically want to buy the same number of units of credit risk per unit of equity, across firms.

As firms reach the end of their mortgages internal ratings based (IRB) credit model redevelopment journeys, our thoughts turn to unsecured retail lending. It would seem sensible to adopt consistent assumptions, and apply learnings, from the mortgages journey. However, in practice, similarities between retail mortgages and unsecured lending are limited – as such a didactic harmonisation becomes unlikely.

To achieve an output of RWA harmonisation requires an input of harmonised assumptions. In this article we discuss four areas of potential divergence that are relevant to retail unsecured: Data representativeness, Probability of Default (PD), Exposure at Default (EAD) and Loss Given Default (LGD).

We suggest that in the short to medium term, the wide range of plausible, yet divergent, modelling options should motivate firms to build and impact-assess multiple prototypes. When planned from the outset, this can be achieved efficiently and without harming deadlines. This will allow firms to obtain a deeper understanding of the model impacts across the organisation, prior to submitting the finalised models for internal governance and approval, and ultimately for regulatory submission.

Data representativeness

The unsecured lending market today is significantly different to that which existed during the 2007/2008 crisis period and is unrecognisable with respect to the early 1990s. Sector-wide examples include new product innovations, the withdrawal of “teaser” credit card balance transfer deals, and significant changes in consumer legislation including treatment of persistent debt, repeat users and vulnerable customers. Firm-specific changes include changes to customer-level default flagging, and customer-level collections and recoveries strategies that consider a customer’s aggregate borrowings.

Whilst a complete treatise on changes in the market is beyond the scope of this article, it is reasonable to conclude that recent conduct rule changes render historical loss data somewhat less than fully representative of today’s (and the likely future) data generating process. Regulation requires firms to make “appropriate adjustments” for data representativeness, before resorting to margins of conservatism.

As an example, firms shifting from facility-level to customer-level collections and recoveries face similar choices. Such a material change in policy/process may motivate firms to re-state loss data as if gathered under today’s policies and processes. However, this would introduce a degree of subjectivity. It may be more straightforward for developers to model the observed data, apply conservatism for lack of representativeness, and document the limitation so that it is transparent to users.

Within wholesale IRB, firms adopted divergent approaches to historical loss data restatement under today’s obligor, default and loss definitions. This divergence in approaches to loss data restatement seems likely to continue in retail unsecured, thus driving RWA variability between firms.

Probability of Default Calibration

For mortgages, whole market arrears data, as well as many firms’ internal default histories, support a reasonably strong consensus for defining a complete economic cycle and therefore the appropriate calibration level for Long Run Average (LRA) PDs.

To harmonise unsecured lending with mortgages calibration assumptions, firms would need to extrapolate macroeconomic indices sufficiently far back into (or beyond) the early 1990s. However, unsecured data does not typically yield an obvious economic cycle such as “peak to peak” or “trough to trough”. Deloitte analysis of market write-off data suggests that different yet reasonable assumptions about the complete economic cycle can lead to a factor of three difference in calibration level.

Faced with such subjectivity, use of a point-in-time (PIT) calibration philosophy may seem appealing when the world is viewed solely through a model developer’s lens, with focus on completing model development with minimal subjectivity. Indeed, the fact that SS11/13 only explicitly disallows the approach for mortgages could, at a stretch, be extrapolated as meaning that the approach is permissible in other portfolios.

Yet it is difficult to see how PIT calibrations would be compliant with the CRR or EBA Guidelines, or indeed CP16/22 (which reaffirms a commitment to discrete bins for PD calibration). The issue of completing the self-assessment notwithstanding, when the world is viewed through the finance or business lens, the financial impact of a 100% PIT calibration may be undesirable due to greater peak-stress RWAs than under an LRA calibration.

Considering that existing unsecured hybrid PD approvals have LRA calibrations, any move towards PIT calibrations is likely to introduce RWA variability between firms.

Exposure At Default

The Basel 3.1 reforms (currently in consultation) harmonise firms into using the 12-month fixed horizon method for estimating conversion factor (CF) models. This represents a departure from the current consensus of using the cohort method and treating Exposure at Default (EAD) as the dependent variable, as well as an inconsistency with the way that models are applied on reporting date. Furthermore, the 12-month fixed horizon and CF models are generally perceived as less stable statistically.

Firms may choose to prioritise Use Test principles and continue to model EAD using the cohort method. Such an approach is likely to require conservatism to account for the methodology uncertainty with respect to 12-month fixed horizon CF models and may therefore lead to RWA variability between firms.

Loss Given Default

For mortgages LGD there exists a near universal consensus to apply two-stage workout models. Within secured lending, such approaches substantially explain the observed bimodality in realised losses. However, losses on unsecured products associated with the “non cure” outcome typically retain a significant degree of bimodality, meaning that a probability of cure model, P(Cure), may not be a statistically optimal approach to modelling realised losses. In such situations, it may be preferrable to select a broader dependent variable and model P(low loss); or introduce a two-step decision tree with P(Cure) followed by P(Low Loss).

All things being equal, total realised loss is fixed, and all approaches would receive the same calibration level. But all things are not equal: SS11/13 permits applying a lower discount rate for a proportion of the “synthetic” cash flows on cure, to the extent that a cure model may be more attractive in terms of calibration level than direct estimation, even after excluding low-loss non-cure events. However, multi-step workout trees come with the risk that components’ downturn calibrations occur at different times, resulting in an overall calibration that exceeds the downturn realised loss. Choices in LGD model construct and calibration therefore stand out as a potential source of RWA variability between firms.

Discussion

In the short to medium term, the range of plausible yet divergent modelling options should motivate firms to design and populate their reference datasets and estimation code in such a way as to perform detailed impact analysis under multiple combinations of options. When planned from the outset, code can be developed to run combinations of choices as configurations options, without a last-minute scramble to re-engineer or duplicate existing code.

The final set of design choices should be made in the context of wider impacts across the entire gamut of areas envisaged under the Use Test. Particularly important is for firms to understand financial impacts under stress, as well as ensure the alignment of modelled assumption to the underlying business processes.