Find the right technique(s) for each portfolio

The FASB was careful not to prescribe specific approaches to developing CECL forecasts. This is due to several factors including a range of skill sets and the types of data and relevant history. Also, cost was a consideration. As a result, lenders can apply modeling approaches that are appropriate for them. The data they have is a big determinant.

One very effective approach can be to track losses over the life of loans. This requires tracking how losses develop as accounts age to develop cumulative loss curves. This can be done based on the cumulative number of accounts writing off (multiplying by average write-off amounts by age) or by tracking actual losses by account age.

The historical curves are the starting point. Expected future losses are estimated by comparing the cumulative performance curves of newer vintages to the base history curves. Typically differences in curves continue on at a similar magnitude as lifetime losses are estimated.

The final step is to take into account changes in the overall economy's health. Estimating the impacts can be challenging but this is important to accurately applying the cumulative loss curves to younger parts of the portfolio.

We should note that this can be done at the portfolio level which is especially appropriate for smaller portfolios. However, if you are able to segment portfolios into pieces with different risk profiles it helps accounts for changes in portfolio mix that might have occurred over time.

A more 'sophisticated' approach is to estimate the likelihood that an individual account will charge off over its life. This requires a broad set of individual account characteristics. It's also preferable to obtain updated characteristics periodically to ensure that predictions of older accounts are as accurate as possible (versus using only data available from the loan application). Once the probability of default is determined for each account you multiply the risk times the estimated loss given default. The amounts are summed to determine the reserve requirement.

Larger lenders might apply this approach given their larger portfolio sizes. While there is no hard and fast minimum size requirement portfolio size tends to correlate with the scope of data available for modeling. As a result, the individual risk assessment mainly applies to larger lenders. Smaller lenders might benefit from this approach by pooling their performance data with other lenders in order to achieve scale. That would be one advantage of working with CECLNow.

Different modeling techniques can be used to develop the estimated default risk. Many new techniques have been developed in recent years including machine learning algorithms. These take advantage of lower cost computing power and larger data sets. 

The most accurate model may prove to be a combination of several techniques including the 2 mentioned above. Why? Because the short-term predictive window of the individual risk modeling may be more accurate than the cumulative loss approach. But the longer term loss forecast may be more accurate with the cumulative loss model. Combining the 2 may lead to the best estimate. This is especially true for loans with longer terms. 

It's also possible to apply a combination of sophisticated modeling approaches but research into this is required and likely differs for different loan products. This is beyond the scope of this write-up.

Building a model is a waste if it's inaccurate. Once you select a model the final step in development is to validate the model on a holdout set of data. Seeing that the model works as well on the test set as it does on the development data proves its effectiveness.

That validation takes a new form after implementation as you compare model estimates with actual performance. Any significant differences may be due to model errors (time to redo!) or changes in portfolio behavior. If the latter it's important to dig into the data to determine what changes are occurring. Is the change in performance happening across the portfolio or just certain segments? Is the local or national economy experiencing shifts? Are your peers seeing similar patterns? If it seems the economy is performing differently than expected it's time to revisit key assumptions about the economy to assess how to modify them to improve forecast accuracy.