Covarity’s recent webinar, Essentials of Commercial Portfolio Stress Testing garnered wide interest from nearly 300 banking executives. Co-sponsored by the RMA, we were very fortunate to have Will Calendar from First Manhattan Consulting Group lend his expertise to the subject. If you missed the event, you can download the slides or watch the recording.
While it was a highly informative session on a topic that is clearly close to many, we found the Q&A component – which continued for a few days following the live session – to be equally informative.
For the benefit of all that attended – and in many cases asked questions – here is a summary of the Q&A. All responses were provided courtesy of Will Calendar.
Q: Absent extensive “actual” loss data, what is the best methodology to develop assumptions for default probabilities and loan losses.
In most cases we find that data may exist, but it is arguably difficult to extract or manipulate, which complicates the analyses.
In the absence of meaningful historic data, other sources of interest could include:
- Peer data organized by product portfolio and adjusted using management interpretation to account for perceived differences in underwriting characteristics, vintaging, business strategy/focus, geography\
- Other vended sources of default or loss data, including:
- Real Estate related: CMBS market data, data from proprietary providers (e.g., CBRE, PPR)
- C&I related: Moody’s published and subscription based default studies
Q: What do you mean by “relationship matters” when examining losses? Do you have examples?
Relationship can be defined broadly in this context, but there are some specific examples we can point to where historical loss experience by sub-portfolio shows differing performance.
Lower default levels on residential portfolios for customers with deposit relationships have been viewed at multiple clients. In some cases we see default averages that are 1.5 to 2.0X for non-deposit relationship customers. In the commercial space we have seen directionally similar influences for deposit vs. non-deposit clients, although the data may be more cumbersome to analyze and the outcomes more volatile.
Additionally, relationship can be defined by how the client came into the bank. In some cases we have found different default and loss results for similar types of products based on their origination characteristics, for example whether the business unit originating that type of credit was doing so as a focus of its business or for accommodation reasons.
Q: I have been trying to find loss and default data specific to CRE property types such as multi-family, retail, office, etc. Where can I find more granular default and loss data on CRE?
CMBS market data is a publicly available option but will likely pose complications getting to the level of sub-segment definition you would like over a reasonable historical period. Granular loan level data is largely proprietary. You may want to investigate availability of CBRE or PPR databases as they likely have highly granular data collected over the years.
Q: Is the Expected Loss (PD x LGD x EAD) the primary intent of stress testing? If we want to have insight on the economic capital under the stressed scenario, how would you associate the stressed EL with a stressed loss distribution?
The EL used in stress testing should be, by design, point in time oriented. That is to say, you are trying to parameter the loss over a pre-defined time horizon based on known and expected conditions consistent with that time period.
In comparison, the EL used for capital purposes (i.e. Basel) is meant to be over-the-cycle in nature. It should not represent a specific bias towards the future state of macro or other loss-driving conditions. The stress for economic capital/regulatory capital is applied through a confidence interval formula that adjusts over the cycle to a defined level.
Q: For historically-based models, how do you feel about including or excluding data points during the great recession as it may skew either the inputs or the outputs of the model?
The answer here depends upon the context of what you are looking to accomplish. In a stress/scenario analysis, it may be very legitimate to include these points if you are trying to forecast a period where losses could resemble a historical period.
If you are looking to calculate PDs and LGDs for capital estimation purposes, including recent data points becomes problematic. For Basel purposes, the PD should represent an over the cycle view, and most would consider inclusion of recent history as unduly influencing a true over-the-cycle view.
The approaches around this include different options, some of which are more defensible than others:
- Defining a specific set of years as the cycle based on overlaying a view of macro factors. This approach becomes difficult to maintain over time.
- Weighting observations differently through the process, although there needs to be a very coherent rationale applied to avoid views that this becomes ambiguous.
- Separately identifying macro vs. underwriting loss drivers and then applying a probabilistic view of macro scenarios. This approach is more common for consumer type portfolios and difficult to model for many commercial books
Q: How do you recommend stressing a construction loan portfolio?
A construction loan book can be approached in a similar fashion to CRE-Investor approach that was discussed [David: in the webinar] with some additional caveats that need to be factors in.
You need to pay particular attention to exposure at default as the project likely has multiple draw events when additional funding will be released. Depending on how this is structured and the lender’s ability to mitigate them based on contractual obligations, your loan balance could be higher in the future than currently recorded.
Additionally, in stress testing C&D portfolios you need to consider the appraisal standard (e.g., “as constructed, “as-is”, “as-stabilized”) as this can vary meaningfully in an LGD context.
Finally, C&D exposures need to be considered with respect to any interest reserves that exist and their current impact on default likelihood.
Q: Could you also discuss how future expected loss can be distributed across time horizons to create a loss distribution over time?
Distributing losses across the time horizon over which you are forecasting stress losses is a beneficial but difficult activity. I use the term “losses” here to differentiate between EL. EL as a common definition would align to “over the cycle” PD times over the cycle LGD. In this generic context you can use both of these variables for calculating capital (with some adjustments) in a Basel context. Losses in a stress test case need to be point-in-time over a defined horizon. While the PD x LGD construct still holds the inputs would not be the same as what you would use for capital estimation.
First you would of course start with developing a total quantum of point in time losses over the defined horizon. Once that is agreed you can use other overlay factors to apply an expectation of loss timing to that total quantum. The factors you would want to consider here (informed by supplemental ex-post analysis) could include: maturity timing, potential loan modifications, type of loan, location, type of property, borrower industry, etc.
Q: Wouldn’t LTV determine LGD map and DSCR determine PD?
That is predominately the driving force in CRE-Investor. However, in some cases LTV can be a real influence on PD too. For example, there may be properties that have acceptable “current” DSCRs but where the LTV is so high that the borrower will walk away from the property because the likelihood of recovery is so low or long-dated that continuing to service the loan is perceived as non-economical.
Any other questions or feedback? Let us know.
While the results are by no means scientific and the nature of the subject would naturally introduce bias, in our poll 67% of respondents were either considering a pivot to working capital lending, were currently pivoting or had themselves recently pivoted.
The presentation followed with some use case examples of “the good, the bad and the ugly” before closing with some practical best practice considerations.