Four Key Questions When Estimating Current Expected Credit Losses (CECL)

Under the FASB current expected credit loss (CECL) accounting standard, public entities are required to estimate losses over the contractual term of the financial asset or group of financial assets. However, this requirement is eased to allow entities to make or obtain “reasonable and supportable” forecasts of expected credit losses before reverting to historical information to estimate loss over the full life of the instrument. FASB guidance is clear that measuring expected credit losses for a long-dated financial asset may consist of two components, a “reasonable and supportable” forecast component, and a component based on historical data. The entity should then use a rational and systematic mean-reversion method to link the reasonable and supportable component with the historical component.
In this blog we offer answers to four key questions raised by this requirement.
How long should our “reasonable and supportable” period be?
It is important to understand that greater precision and accuracy in your loss estimate will likely result in greater volatility. Extending the length of the reasonable and supportable forecast period would mean that the forecast is driving more of the loss estimate, and depending on the variability of economic conditions, will introduce more volatility to the loss estimate. Shortening the length of the forecast period would mean that the mean reversion and historical loss information is driving more of the loss estimate, which will be less volatile than the forecast period. However, the gains in lower volatility will produce a loss in accuracy.
We recommend two techniques that will help you determine a time period that will hold up to audit and provide a strong business rationale:

  1. Back-testing. The closeness of fit of a back-test will provide strong support to the choice of the forecast length. When building CECL models, entities should ensure that they have the flexibility to adjust the forecast length and measure its impact on actuals vs. estimates.
  2. Confidence Intervals. Entities can use confidence intervals to justify the length of their forecast period. The confidence intervals of economic forecasts widen at a rate that is proportional to time. An entity may decide to not use an economic forecast once the confidence interval exceeds a certain threshold.

The “right” length of a reasonable and supportable forecast period will balance accuracy vs. volatility, provide a close fit between actuals and estimates when back-tested, and not produce confidence intervals that exceed a threshold of tolerance.
How is the historical loss information calculated?
The length of the historical time period used to calculate the historical loss information should cover, at a minimum, one business cycle. A historical period of two to five years is consistent with regulatory standards for CCAR and Basel models. The time period used to calculate historical loss information will likely differ depending on the portfolio type. While five years of historical data might be sufficient to calculate historical losses for some portfolios, for low-default portfolios, or portfolios with little recent default history, five years might not be sufficient. Data relevance should also be a consideration when calculating the historical loss information. Just because an entity possesses 15+ years of historical data doesn’t mean it should necessarily be used, as the older information might not be relevant to the current state of the portfolio.
Does it matter which mean-reversion method I use?
The choice of mean-reversion method used can significantly impact the loss estimate. FASB guidance does not prescribe a mean-reversion method, instead saying that entities may revert to historical loss information immediately, on a straight-line basis, or use another rational and systematic basis. We believe that entities will want to avoid any jump discontinuities between the forecast period and the historical loss period. Therefore, a straight-line method of mean reversion might be used, which may consume several quarters and constitute a significant portion of the loan’s average life. The speed of the mean-reversion, from forecast to historical, is another parameter that must be calibrated. Entities should test different mean-reversion methods (e.g. straight-line, weighted) keeping in mind the considerations of accuracy and volatility and the closeness of fit between actuals and estimates.

For USDA’s $120.1B Single Family Mortgage Guarantee Program we developed a custom House Price Index (HPI) forecast for rural counties, where third-party HPI forecasts for lightly populated rural counties do not exist. We developed time-series based forecasts that mean-reverted to a long-run national HPI trend based on historical data. Our model produced HPI forecasts for eight quarters, before gradually mean-reverting to the historical trend. Testing showed that loss estimates derived from this mean-reversion method closely matched historical loss estimates collected under similar macroeconomic conditions.

Should I mean-revert inputs to my credit loss models, or only the model output?
Banks that are re-purposing a loan-level CCAR/DFAST model for CECL may choose to mean-revert at the input level, using a bottom-up approach. However, a frequent problem with the bottom-up approach is that the model output from mean-reverted inputs might not align with the mean-reversion of the entire estimate. However, if you have the ability to mean-revert at the input level and produce an accurate expected credit loss, it is recommended over mean-reverting the model output. Mean-reversion at the input level provides additional benefits such as more opportunities for scenario testing and other types of ‘what-if’ analysis. It is unclear how favorably or unfavorably regulators will judge inconsistency in model frameworks within an institution. For this reason, entities may choose to mean-revert at the model output level, as it is easier to implement across all portfolios.
In summary, the FASB guidance for forecasting expected credit losses is not prescriptive, and the “right answer” for your model framework will depend on the size, complexity, and other unique characteristics of your portfolios. Here we summarize key considerations when developing a CECL model framework in the following table.