Following on from our blog IFRS 9 | 1Q21 results update: the beginning of the end?, this post looks at seven areas we think lenders should be thinking about in advance of year-end 2021.
As always, we would be delighted to discuss any of these issues in more detail should you wish!
1. PD model monitoring and calibration
In many firms, Probability of Default (PD) model performance monitoring has been triggering red flags during the last year due to models over-predicting defaults. This was caused by lower observed default rates from both changing customer behaviour (spending less) and government support (payment holidays / COVID loans) suppressing default emergence. Some firms froze Point-in-Time calibrations because management assessed the low observed default trends to be giving an artificially benign view of risk.
Although less severe than anticipated in June 2020, meaningful deterioration in the credit environment is still expected as the economic fallout of the crisis emerges. Additionally, there is likely to be some “catch-up” from the normal level of defaults that did not emerge in 2020/21.
Managing model risk in a rapidly changing credit environment is difficult and firms should make sure PD model monitoring is performed on a regular, timely, and sufficiently granular basis, supported by analysis, to understand when and why model performance is deteriorating and what action(s) might need to be taken (e.g. recalibrating to reflect a higher observed default flow). Firms should also consider how to adapt their calibration windows and data to prevent over- or under-estimation of ECL (i.e. managing the potentially distorting impacts of the artificially benign environment to date, payment holidays, repossession moratoria and changes in origination criteria).
2. Macroeconomic response models
Most macroeconomic response models used by banks to incorporate forward economic guidance within the Expected Credit Loss (ECL) include unemployment and/or GDP in some form. GDP has had sharp swings in the last year and its relationship with credit defaults has broken down. In Chart 2 we have inverted the change in UK GDP to illustrate how, counterintuitively, an increase in UK unemployment coincided with the increase in economic activity at the end of the second lockdown (in the red circle). At the same time, the flow to default has been benign and firms have had to use judgement to mitigate volatile model outputs.
Firms need to be alert to the possibility that, even as state support is withdrawn, previous relationships between the economy, credit risk drivers and ECL may remain altered; the predicted recovery is very different from 2008/9 and models may struggle with this. For example, the unemployment spike towards the end of 2020 has not (yet) led to an increase in observed mortgage or consumer lending defaults. Whether this default “bump” will emerge in time or whether it will be suppressed and/or eliminated by the strong labour market (vacancies are approaching pre-pandemic levels) will become clearer in time but the models will be trying to predict defaults based on 2008/9 parameterisation.
As with PD, monitoring of macroeconomic response models (to predict how flow to default changes as economic conditions change) will be increasingly critical over the next few years. Both back-testing, to understand the accuracy of feeder models, and analysis to sense check / benchmark forward-looking guidance to ensure that the models remain credible as the economic situation evolves, are important. As with all model monitoring, we consider setting well defined model KPIs (Key Performance Indicators) and associated thresholds to be good practice. Firms should think about what actions to take if fewer defaults emerge than predicted at year end: should the total level of expected defaults be kept the same and their timing adjusted or should the missing defaults be pushed into future expectations? There will likely be a large degree of expert judgement in these decisions.
3. Sectoral risk
Firms’ models do not tend to segment by industry or sector and the precedent data from 2008/9 did not show the industry sector risk trends that have been present in the current crisis. Firms are increasingly assessing the risk that different sectors present and analysing whether this is appropriately captured by the models.
Firms often address gaps or limitations in model segmentation via Post-Model Adjustments that add risks for specific industries to the ECL. Firms should think about whether the aggregate level of ECL will change or just the allocation (e.g. some sectors are performing well through the crisis and may have lower risk than predicted).
The picture around sectoral risk is not entirely clear. For example, the difficulties faced by the accommodation and food service sector are well known with 48% (933k employees) on furlough at 30 April 2021 and almost 14% of jobs having been lost since the start of the pandemic. But the most recent ONS job vacancy data shows 266% quarterly growth in vacancies which may indicate there is less risk than expected at year-end 2020.
4. Model adjustments
Year end 2020 saw significant use of “In-Model Adjustments” (IMAs) and “Post-Model Adjustments” (PMAs) to address a range of model weaknesses, manage spurious model outputs, and use management judgement to help (credibly) assess the likely credit impact of the crisis given its unprecedented nature. At some firms, these added to an existing stock of adjustments, that addressed a wide variety of issues.
Model adjustments are an important part of model risk management to mitigate known model weaknesses. Ideally, the problems identified via validation and monitoring, issues with data, and risks that can’t be modelled should all be systematically considered to determine if mitigation via a model adjustment is warranted. However, where adjustments relate to issues that are “modellable”, firms should have a plan to build these into the core model; this was reflected in the PRA’s 2020 Dear CEO letter in order to reduce reliance on PMAs.
Model adjustments should be re-assessed at each reporting date with supporting commentary of the specific risks addressed, quantification approach used, conditions for unwind and clear documentation of data sources and controls applied. In-line with the PRA’s 2020 letter, high-quality, well-defined governance and approval of model adjustments is a critical component of the control environment.
While the forward-view of the economy and the expected credit outcome have improved since mid-2020, the level of uncertainty remains high. The range of expectations has narrowed across the industry but is still very wide.
Despite the improvements in inputs (i.e. better economics and payment holidays having largely ended), lenders are concerned about the unpredictability of the crisis and potential cliff-edge effects when state support is withdrawn. As at end April 2021, 3.4m employees were still furloughed, the risk of new variants lurks in the background, and post-Brexit trading processes are causing friction for some businesses.
It is important for firms to think through where they are addressing uncertainty in their IFRS 9 frameworks (e.g. scenario definition / scenario weighting / other sensitivity testing), how it has been quantified, and how the risk is allocated in the loan book. Central overlays based on management judgement offer a lower-quality assessment of the risk compared to more targeted overlays which can be explained quantitatively and unwound once that specific risk is addressed.
6. Stage 2 triggers
So far, at a high level, Stage 2 has done what it was designed to: as the forward-looking view of the economy deteriorated, loans were shifted from Stage 1 (12month losses) into Stage 2 (lifetime losses) and, as the forward-look has improved, they have started migrating back.
Setting quantitative forward-looking lifetime PD thresholds for transfer to Stage 2 is a material judgement, the thresholds should be monitored regularly and (re)validated in the same way as any other part of the model. To do this, firms should have established (or seek to enhance) the monitoring of Significant Increase in Credit Risk with defined tolerances and performance criteria, not least because this can identify potential weaknesses and optimisation opportunities.
Given the benign flow to default, assets are often staying “good” in Stage 2 without flowing through credit stages to hit the 30-day backstop or default. Before any recalibration and threshold adjustment goes ahead, firms need to consider how the recent benign credit performance may have distorted staging performance. At some firms, there have been additional distortions resulting from the tranche of COVID-19 payment holidays granted in 2020 and delayed payments on government-backed loans.
7. Other model changes
With “fire-fighting” and initial responses to the pandemic now abating, firms should think through their strategic IFRS 9 model roadmap for the next 3-5 years.
While it will take some time for the outcome data to emerge to assess and understand the full impacts of the crisis on model design and future remediation, some issues can be identified and addressed now. These could be issues that are not related to the crisis and risks which impact only a segment of the portfolios. For example, some portfolios have shown insufficient modelled economic sensitivity and not all mortgage models account for the end of term risk associated with interest only mortgages.
There is a heavy schedule of upcoming Basel regulatory changes that will have a downstream impact on IFRS 9 models and processes. This does not only affect banks on the Internal Rating Based (IRB) approach as the fundamental changes to the definition of default have already gone live for Standardised Approach lenders. This may mean elements of their IFRS 9 frameworks need remediation to be consistent and appropriately aligned. Firms also often use the same macroeconomic response models for IFRS 9 and stress testing / planning. Therefore, model roadmaps should be holistic across expected loss, unexpected loss and stress loss regimens and reflect data, definitional and model interdependencies.