Does COVID-19 spell "V" or "U" for banks? A study of modelled assumptions and credit losses

Executive Summary

Beyond the profound human consequences of the COVID-19 pandemic, all economic and financial markets have been affected, with nearly all individuals, companies, regions and industries impacted to some degree. In the words of International Monetary Fund (IMF) chief executive Kristalina Georgieva, "we have witnessed the world economy coming to a standstill. We are now in recession. It is way worse than the global financial crisis".

The recent publication of EBA Guidelines on payment moratoria provides unambiguous guidance that we have entered new territory for operational treatments: Liquidity-forbearance treatments under payment moratoria should not on their own trigger distressed restructuring or the Net Present Value (NPV) test for default. Whilst this is likely to guide firms to act in such a way as to dampen, delay or otherwise attenuate the impact of the exogenous shock on credit losses (and hence their own balance sheets, systemic risk, and the real economy), it remains vitally important for banks to have a view on the range of credit losses that might be reasonably expected to occur.

Using public data and standard modelling assumptions, we have forecasted a range of credit cycle and collateral value scenarios, to determine a range of impacts on Lifetime Expected Credit Loss (LECL) with respect to what would have been expected at year-end 2019. We have not sought to condition the analysis on any subjective prior belief about future macro paths, and instead invoke the assumption that the market-driven data already describes all available in information (public and private), including unpublished unemployment data and published medical research.

Our results suggest that under a "V" or sharp rebound scenario, shorter maturities' LECL estimates are impacted by a factor of three, with longer maturities impacted by factors closer to one. But under a longer and more-pessimistic "U" scenario, LECL estimates are impacted by a factor of up to five (unsecured) or 23 (secured). (By way of comparison, the UK mortgage write-off rate has been 1 basis point since 2016, and peaked at 26bps in 1993 when the Bank of England time series commenced). Full results are presented below.

In addition, our analysis serves to highlight that even sophisticated models are conditioned on historical data that was generated under potentially quite a different macro environment; and are likely to have poor goodness-of-fit under stressed conditions. All things being equal, our model's central forecast anticipates a slower return to normal conditions than was experience in the 2008/9 stress. Readers are invited to draw their own conclusions, as to where in the range, a true and fair unbiased expectation lies.

Technical Appendix

The remainder of this article presents details of our analysis, including data selection, model selection, results and discussion.

Introduction

As noted in our recent IFRS in focus article, “the models used by many banks were not developed to accommodate the extreme economic conditions and the levels of government support measures being introduced and so entities may find that additional model development is required.” With a lack of historical data describing credit outcomes for liquidity-forbearance events, modelled assumptions that link observed behaviour to future cash flow performance become all but irrelevant. Further, the recent Dear CEO letter discusses the belief that activity will rebound sharply once social distancing measures are lifted, in a V-shaped scenario that existing stress testing and forecasting assumptions have not generally been devised to describe.

Nevertheless, liquidity events have historically been predictive of credit events. Ultimately, some illiquid customers are likely to be forced, to sell assets for below book value in order to raise cash to settle their liabilities as they fall due, and spiral into insolvency.

In this article, we investigate a range of possible outcomes in terms of Lifetime Expected Credit Loss (LECL), using modelled assumptions and public data. Modelling is of course a subjective discipline. We have sought to minimise subjectivity via application of parsimony and the scientific method. Nevertheless, the range of results is wide. Where in that range, to place the overall expectation is conditional on readers’ subjective prior belief about how the future might unfold. Readers are invited to draw their own conclusions as to the validity of the assumptions, and what a true and fair estimate of the expected impact might be.

Credit Cycle Index Modelling 

This section explores CCI modelling.

Source of CCI

We first seek to describe cyclical changes in credit risk. In line with standard industry assumptions, we have assumed that cyclical changes in credit risk are explained by a single risk factor, the Credit Cycle Index (CCI), which is standardised and referred to as Z. In the real-world, this assumption may not always hold (e.g. would manufacturers of both sunblock and raincoats both experience good trading performance if we were to have a wet summer?) and banks would segment, or decompose portfolios, into the most granular homogeneous segments that can be identified without over-fitting.

Our choice of CCI is a UK index of market-implied real-world Estimated Default Frequencies (EDFs) sourced from the National University of Singapore Risk Management Institute. EDFs have the particular advantage of being market-implied, allowing us to invoke Efficient Market Hypothesis (EMH). EDFs incorporate all available information (public and private) that might influence corporate credit risk including: unpublished unemployment data; the nature and probability of state interventions; advancements in medicine; expectations of corporate behaviour and capital structure; and a complete study of all published literature about the nature and severity of the pandemic. We have avoided traditional macro variables such as GDP and unemployment. These tend to lag and would require us to take a subjective forward-looking view of the macro indices, as well as establish links between the macro-economy and the default risk. In addition, studies suggest that large corporate default frequencies do not correlate well with traditional macro indices, supporting the view that EDF indices should be modelled as latent random variables without covariates. In some geographies, EDF indices also explain cyclical movements in retail credit risk; whereas in other geographies (including the UK), true forward-looking indices are harder to come by, increasing the reliance on subjective assumptions about the future.

CCI Model Selection

Having selected our CCI, we turn our attention to modelling. We performed a model selection procedure, whereby a variety of Seasonal Autoregressive Integrated Moving-Average with eXogenous regressors (SARIMAX) models were fitted. The optimal model was selected using the Akaike Information Criterion (AIC). The selected model has three autoregressive terms for the trend component, and two autoregressive terms for the seasonal component. The three autoregressive terms allow for continuity in level, and the first two derivatives, with eventual reversion to the historical mean (steady state). This construct is frequently used in time series modelling of the macro economy, and assumes that government and market participants’ objective is to steer or otherwise return the economy to a steady state. This assumption would of course be invalidated in the event of a structural break in the time series – a topic we revisit later. Time will tell whether the present discussion of changes to working habits and expectations of state support, in a fundamentally free market system, crystallise into changes to the way the market perceives systematic risk, and hence will result in such a structural break in CCI behaviour.

We also investigated fitting a Markov switching model. The AIC did not out-perform the SARMIAX models, and therefore, we did not explore this avenue further.

CCI Model Discussion

Under a broad range of situations, the CCI model performs, and produces plausible and intuitive forecasts. The intuition in particular has a convenient real-world parallel. Just like a car suspension, the CCI path continues in its current direction but also decays or reverts to the steady state.

However, like all models, it is worth considering whether performance is maintained in more-extreme situations. We therefore investigated qualitative and quantitatively. The figure below shows the CCI model’s prediction at year-end 2008, along with the 95% confidence interval. Whilst not conclusive, this provides comfort that the model produces plausible results for tail events (accepting that this is in-sample), albeit with a wide confidence interval. However, detailed examination of the first few months of the turnaround suggest that the model over-states risk during that initial period, and that the upper bound path would, during that period, be a more-appropriate choice.

For a more-conclusive view of tail performance, we present diagnostic plots. Like all regression models we would expect better goodness-of-fit to be best around the centre of mass, and less-good in the tails. The diagnostic plots below seem to confirm that this is indeed the case. In particular, the Normal Q-Q plot suggests that the model under-estimates severe downwards movements in the CCI.

One could legitimately challenge that the CCI itself does not behave rationally in a stress, and could be more influenced more by market sentiment. However, this is mitigated by the market-consistent nature of EDFs. If EDFs are over-stating risk, then this would present an arbitrage opportunity that would soon disappear.

Thus, we know that the performance of even this reasonably sophisticated time series model will deteriorate, even in the kinds of stresses already seen in the sample. Returning to the year-end 2008 prediction, we see that it does not keep up with the actual sharp turnaround that occurred in the following months. The upper confidence interval would perhaps be a better path to assume, initially.

For completeness we present below two further forecasts.

In the figure below, we show what the model would have forecasted at year-end 2019. Conditions were already close to the steady state, and we see a relatively small movement, returning towards Z=0.

In the figure below, we show what the model is forecasting at the end of March 2020. As is often the case with exogenous shocks, the movement in Z values was rapid, and occurred at such a rate that the model now predicts a continuation before returning towards the steady state. However, as discussed above, we have a prior belief that this kind of model is initially poor at keeping up with the turnaround and may choose not to put all of our faith or belief in the central forecast. In addition, it can reasonably be argued, that there exist several steady state equilibria in the data – including the period before Bank of England independence, and the period of recovery during 2009 and 2012.

CCI and HPI model 

Thus far we have considered only the impact on default risk. Another key determinant of credit losses is the value of collateral. We therefore fitted a second model, to investigate joint movements in the CCI and House Price Index (HPI).

CCI and HPI model selection

For the joint CCI-HPI model, we used vector auto regression (VAR) modelling to establish a relationship between CCI and the change in the natural log of HPI (denoted dlnhpi in the figures below – the month-on-month change in natural log of HPI). In parallel with the CCI model described above, we fitted a vector autoregression model with three lags. Note that the HPI data has already been smoothed by 3 months, to make a basic conversion of historical quarterly data to monthly samples.

The modelled relationship is inconclusive. Whilst there is clearly some degree of correlation between dlnhpi and CCI, the actual statistical relationship is weak. Knowledge of HPI does not seem to enhance the quality of Z forecasts, and vice versa. Although the HPI forecast seems to acknowledge that a Z-shock is associated with a HPI shock, the central prediction is mild with respect to 2008 experience, due the fact that HPI is not already falling (that can be observed.

More-formally, the Impulse Response Function (IRF) plots show that a unit shock to Z generates a <0.006 shock to dlnhpi. The present three-unit shock to Z would therefore on its own not be sufficient to reach the 2008 trough in dlnhpi.

Does this mean that we are to expect only a mild and temporary impact to HPI?

Not necessarily. It is reasonably likely that our fundamental assumption of a precise and random sample of the data generating process is breached: The dynamics of the housing market have changed over time, and may continue to change. Examples include:

• Affordability testing is objectively stricter than 15 years ago;
• The regime for interest on Buy-To-Let (BTL) properties is objectively different to even five years ago;
• One may also consider subjective forward-looking views on whether remote working is the new normal; and
• The effect of inflation, which masks real-terms HPI falls in the early 1990s, could also be considered.

Our model simply does not know, and returns a commensurately wide confidence interval.

Financial Impacts 

Since the adoption of the IFRS 9 accounting standard, the concept of the Expected Credit Loss (ECL) in terms of cash shortfalls has achieved widespread acceptance as one of the de-facto measures of credit risk, alongside Basel Expected Loss (EL) and Risk Weighted Assets (RWAs).

The concept of an expectation implies integration over all possible outcomes, which we recognise is difficult to do, given the subjectivity in identifying what those outcomes might be – either as discrete cases, or by numerical integration over a large number of randomised paths. To simplify comparison, as well as consistency with most banks’ choice to use discrete scenarios, we therefore present only three scenario-dependent ECL estimates, as multiples of the central scenario from year-end 2019.

One key limitation of both the upside and downside scenarios, is that neither reverts to the steady state. In practice this is unrealistic, and we have commented previously in our blogs, on why scenario quantile is not the same as loss severity quantile. All other things being equal, the scenario that maintains the 97.5th percentile of Z is likely to generate a loss severity corresponding to a percentile significantly beyond 97.5th percentile. Nevertheless, we have continued on the basis that the upside and downside scenarios provide a basis for making like-for-like comparisons between estimates without introducing additional, subjective, assumptions.

The results in the tables below were generated using an ECL model that is sensitive to the reporting date Hybrid PD, PITness of the hybrid PD, remaining maturity, CCI (Z), HPI, prepayment risk, migration risk, collateral recovery rate, unsecured recovery rate, amortising amount, customer rate and Effective Interest Rate (EIR). The model uses a monthly time-step.

Unsecured Facilities

The table below presents a range of LECL uplifts with respect to the 2019 year-end central scenario, for an unsecured facility with Through-The-Cycle PD of 1% and Unsecured recovery rate of 20%.

Secured Facilities

The table below presents a range of LECL uplifts with respect to the 2019 year-end central scenario, for a secured facility with Through-The-Cycle PD of 0.6%, Loan To Value (LTV) of 70%, Forced Sale Discount (FSD) of 20%, and FSD standard deviation of 0.2.

Discussion

As discussed above, there is evidence that the CCI model is at-best over-fitted to the previous downturn, and also suffers from poor goodness-of-fit in that downturn.

• If we believe that the activity (and hence the CCI) rebound sharply in a V-shaped scenario, then shorter maturities’ LECL might be expected to double with respect to year-end 2019, with longer maturities reasonably unscathed.
• 1.7x to 3.7x might be expected for unsecured lending, with the greatest impacts occurring at shorter maturities; and 3.6x to 4.6x might be expected for secured lending, again with the greatest impacts occurring at shorter maturities.
• If we are pessimistic then impacts of up to 5.1x (unsecured) or 23.1x (secured) might be expected, with a rank order reversal occurring for secured lending due to HPI deflation occurring.

The 23.1x uplift may seem severe, but in the context of 2019 write-off rates of around 1 basis point, a movement to 23bps would not be out of line with the 26x range of historical mortgage write-off rates in the Bank of England database.