After reading our post, Unemployment as a Commercial Loan Loss Predictor, someone asked, “…what would be a better leading indicator than unemployment rate in your opinion?”

At a national level, there are economic variables that lead bank C&I credit losses—they’re available for free at FRED—and we won’t completely discard unemployment, just yet, but the more interesting question is why many banks use the unemployment rate in the first place.

From our experience, banks that use national unemployment as a C&I loss regressor tend also to be firms that don’t segment or partition industries, and, therefore, use it as a one-size-fits-all regressor.¹

While there are a variety of reasons that developers might be attracted to unemployment, we think it’s frequently related to model choice. For example, the data requirements of using the transition matrix method are tough. There need to be sufficient observations to fill cells with meaningful and reliable transition rates, and that makes it nearly impossible for all but a handful of giant banks to treat industries or segments separately.

However, if other regression methods are used, it’s possible to analyze partitions or grouping of industries—either by industry or geography or both—and then it’s possible to use more intuitive variables that are related to those segments. There is no theory, per se, but folks who have been around and have learned from their expertise can easily provide parameters to investigate when the industries are grouped logically. Sometimes they work, sometimes they don’t. An obvious, and usually correct example, is oil prices and the energy industry defaults.

For both types of credit loss forecasts—CECL and stress testing—we would encourage banks to try to use segments similar to how they run the business and manage risks, including C&I portfolio risks and industry or geographic concentrations and limits.

There are instances where it’s not feasible to be at that level of detail, but the goal is to partition industries/geographies into groups that are  homogeneous within and heterogeneous across those groups. Broadly put, when there’s an intuitive reason to combine industries, and those industries behave similarly with respect to the explanatory variables, the segment makes sense. Likewise, when other industries behave differently to those same variables, then they should be in other segments. While we write, “industries,” we really mean the industries’ default rates, LGDs or other borrower or loan characteristics that need to be investigated and modeled.

Why bother? Because such segmentation allows for much more meaningful analysis and discussion at ALLL/CECL committee meetings, as well as much more meaningful and useful sensitivity analysis. This type of development work—by either confirming or denying suspicions of relationships among variables—provides valuable information about the loan portfolio. It’s basic research of the bank’s C&I portfolio.

That being said, it’s likely that developers will need to test many different partitions of industries. For example, we once tested 19 different ways to partition a C&I portfolio with 21 different “industries” or collections of NAICS (as regularly reported by its portfolio management group) into homogeneous groups, each with its own regressors and coefficients.

Of the 19 tried, one was selected as best, but we recommended that the bank continue to run the other 18 schemes as a way to show the robustness (or not) of the chosen one. All modeling requires arbitrary choices, and by running all these different schemes through time, we were able to show that anticipated losses didn’t vary more than 10%; so, even if an executive didn’t like our recommendation of the “best” one, he or she could feel comfortable that his or her preferred scheme would have generated similar results. Of course, you can only do that IF you have knowledgeable developers, and good model governance, including well-documented developmental evidence.

All else equal, we discourage our clients from using the national unemployment rate for C&I loss forecasting, and with partitioning, its use is often unnecessary—except possibly for that last segment that always exists: “everything else.” There are frequently better, more-specific regressors that are simpler to explain and understand and more relevant to the segment, especially for a regional bank’s footprint.

In that regard, we frequently ask the rhetorical question, “if you can’t forecast your losses, how can you possibly manage your portfolio?” If all you can say is that you estimate losses through the unemployment rate, then you haven’t told me much, and based on the longer time series graph in the earlier post, you’re probably late, too.

A few closing points:

1. In the current Corona Virus Crisis, it seems that unemployment will lead losses and, therefore, charge-offs. However, with the government backstopping many businesses via the CARES Act, it’s way too early to estimate charge-offs. (Valeriu: this gets to your first question.)
2. We have a post that shows CO vs. U correlations, and some very interesting “regime shifts” that occurred after 2010.
3. We also have a post about CECL and the Corona Virus that describes some of the modeling issues related to this unprecedented situation.