Current Expected Credit Loss, or CECL, is an important upcoming accounting requirement that requires financial institutions to attempt to predict the expected losses on loans and other debt securities over the entire life of the loan. Large retailing banks and credit unions of all sizes can benefit from an accurate CECL model as both entities provide much of the same services to their customers and members, respectively.
The two main metrics you have to consider when choosing the right CECL model should be accuracy and procyclicality. If a loss model lacks accuracy and consistency, what’s the point of spending all that time, money, and effort in a meaningless implementation? A good CECL model will be adequately equipped to better track credit losses. There is a strong correlation between the credit cycle and the economic cycle. Models that account for implied volatility better estimate the timings and severity of economic recessions and manage to do so in a timely manner.
In the webinar, “Which CECL Model Should You Use”, Dr. Joseph Breeden, Chief Scientist and COO, at Deep Future Analytics and Prescient Models LLC, talks about the various types of CECL models. He clarifies the key differences between simple “spreadsheet” models and more advanced statistical models and how they can directly benefit credit unions with improved predictability.
Spreadsheet models for CECL are so simple that anyone could code them in an hour or so in a standard excel spreadsheet document. There are many different types of spreadsheet models, but as Breeden mentions in the webinar, they would not pass normal model validation.
Breeden briefly discusses the recent popularity with historic weighted average models, vintage copy-forward models and loss timing functions. After thorough testing though, he concluded that spreadsheet methods “could only be used as a starting point for making manual adjustments”, and that “statistical models are dramatically better.” The only plausible reason to use spreadsheets is if you have nothing else to use or are simply using them for short-term loans (3-6 months). Beyond that, the predictability on credit losses diminishes greatly and the forecast error increases proportionally with increasing loan term lengths.
The Various Types of Statistical Models
- Time Series & Roll Rates
Time series and roll rate statistical models are similar in complexity, and of the statistical models, are the simplest. The time series model takes historic loss rates monthly and correlates them with various economic factors. The roll rate model takes it a step further and buckets the accounts by how many months delinquent they are. Bucketing accounts allows for a closer look at delinquency changes by using ratios. These ratios compare the delinquency amounts between periods of interest (i.e. 1 mo., 1 yr. etc.) and slightly increase accuracy. Roll rates use delinquency ratios, while time series simply use charge-off rates.
Both methods are better suited for the short-term and do not provide the necessary accuracy for loans greater than one year in length. That being said, they are still generally much more accurate at predicting credit losses than spreadsheet methods. Roll rate and state transition models (explained below) lose long-term accuracy because of their strong focuses on delinquency. Breeden explains, “Delinquency is a lagging indicator of what is coming. Delinquency is often driven by economic factors which causes them to miss some of the changing credit factors.”
- State Transition
A state transition model, like mentioned above, is very similar to the roll rate model as they both focus on delinquency. However, the state transition model is a more universal approach to tackling CECL. In addition, there is increased complexity involved with this method as there is the option to personally assign different grades of delinquency states. Breeden discusses the difference between a state transition and roll rate model:
Roll rate is just aggregate time series. State transition is account by account looking at all possible transitions. If I’m in bucket two today: I could miss another payment and go to bucket three, make a one-month payment and stay in bucket two, make a two-month payment and go back to bucket one or I could I cure all the way back to current. I could also charge-off or pay off because I’ve gone out and refinanced or sold my home, etc.
State transition models are much more sophisticated because there are many different states to model, but this really improves accuracy and flexibility for CECL.
- Vintage Models
The statistical models discussed thus far are relatively accurate models but “lack precise estimates of the lifestyle,” as Breeden puts it. This is where the two types of vintage models come in. Vintage models are the most accurate in terms of forecast error and are significantly better than the other models at predicting losses on long-term loans. These types of models include survival models (i.e. multi-horizon discrete time survival) and age-period-cohort models.
Survival models are essentially a lower level version of vintage. The multi-horizon model is a combination of vintage modeling and behavior scoring that puts together a loan-level CECL answer. Finally, we have the age-period-cohort model. This is a model that is often already written, meaning all that is required is to drop in vintage aggregate data. Once accomplished, the model will systematically measure life cycle, credit quality, and the economic environment. The data can then be utilized for queries and forecasts.
The downside to these vintage models is that they do not pay attention to delinquency. This makes their projections and forecasts for the short-term less accurate. The strength of the low-end statistical models (time series, roll rate, and state transition) are reflected as weaknesses for the vintage models and vice versa.
How the Credit Cycle and Economic Cycle Fit into This?
There is a strong correlation between the credit cycle and the economic cycle. Why is this relevant or important to CECL models? It’s important because these behaviors try to predict and explain market conditions. Volatility and economic recessions are a part of any healthy economy and are altogether unavoidable. All good CECL models will not fail to consider procyclicality and market sentiment. Where the credit/economic cycles lie greatly depends on loan demand from the market.
According to SLOOS (Senior Loan Officer Opinion Survey), the loan demand has been slowly declining over the past few years due to slightly lowering interest rates and increasing asset prices. When loan demand is high, the credit risk is low, and the opposite is true when loan demand is low. In the webinar, Breeden compares the credit cycle with unemployment and finds seemingly counterintuitive results. After graphing the two variables throughout history, he concludes that bad loans are always booked in good economic conditions and vice versa. He continues, “As unemployment gets better, credit quality is actually getting worse. Sometime in the next year to 18 months is probably when we face the greatest risk of the worst point in the credit cycle.”
Name that Price: How Much is CECL Worth in your Organization?
The spreadsheet and statistical methods discussed above only begin to crack the surface of the CECL capabilities. The “Which CECL Model Should You Use” webinar discusses the various strengths and weaknesses for each CECL model. There is a lot of flexibility in creating a good CECL model, but every model must be able to accurately predict one thing, and that is the overall credit quality in the market.
The purpose of utilizing a CECL model is to accurately predict the timing and extent of losses. How much early warning is desired as an organization? Breeden puts it, “The amount of investment you make in your model directly correlates with how much early warning you get.” If you want plenty of early warning from market conditions, you will have to pay additional monetary value for the better CECL model.
Which model is best? Breeden closes, “The vintage models aren’t the winner because of the accuracy measure, it’s the winner here because it looks at credit quality. It’s able to pick up early that the credit quality being booked is not the same as the recent past.”