As we all well know, our beliefs about the world are shaped by our experiences.
Like many of you, we have read numerous white papers on CECL and participated in many presentations on the subject. Nearly all have been done by external auditing firms. As a result, the emphasis is on the accounting implications with far less guidance on the financial modeling.
We have been modeling and forecasting lifetime expected credit losses for more than a decade. Our experience has led us to believe:
- Smaller institutions should consider using less complex models such as incurred loss or a simple vintage analysis. You do not need to purchase expensive software and should be able to derive reasonable estimates using Excel.
- Larger and more complex institutions should select credit loss models based on how they intend to use the CECL calculation results. If the goal is to generate ALLL estimates for accounting and regulatory reporting, then aggregate models such as vintage analysis will likely suffice. If the goal is to better manage the balance sheet, then we believe the institution should utilize more robust and less aggregated models.
- Larger and more complex financial institutions should consider using more than one model and select credit loss models depending on the type of loan or investment. For example, we believe that loss estimates on commercial real estate loans should be undertaken at the loan level using probability of default and loss-given default estimates. On the other hand, we believe that relatively more homogeneous loans, such as residential real estate loans, are best modeled using statistical techniques.
- The more granular the modeling, the more reliable the result. For example, a prediction of credit losses for a portfolio of residential real estate loans that is based on individual loan attributes and modeled at the loan level using updated credit indicators such as FICO and loan-to-value ratio will be more accurate than estimates derived by inferring losses from past aggregate performance. This is especially true if an institution has made changes to underwriting standards and/or the forecasted macroeconomic conditions are different than those recently experienced.
- The more granular the modeling, the less likely an individual financial institution will have past performance data in sufficient quantity to make it “statistically credible” and the greater the need to incorporate industry data. For example, an institution might not have sufficient experience with below prime, long-term auto loans to be able to make accurate predictions for this cohort.
- The use of robust, granular, discounted cash flow models which incorporate estimates of prepayment (“conditional repayment rate” or “CRR”) and default (“conditional default rate” or “CDR”) along with the loss to be incurred on a default (“loss severity”) have several advantages over other statistical modeling techniques and are best used if the goal is to better manage balance sheet risk.
- The technique explicitly incorporates voluntary prepayments which must be considered as part of a CECL calculation and are a key driver of interest rate risk.
- The total loss estimate is decoupled using an incidence of default forecast and a loss should a default occur. This means you can have a default without having a loss. For example, let’s assume we are modeling a residential real estate loan with a FICO score of 620 and a loan-to-value ratio of 125%. Let’s further assume an increase in unemployment is forecast. The loan has a relatively high risk defaulting at a loss because the loan is 25% “underwater” before foreclosure and liquidation costs. Now let’s instead consider a loan with the same FICO but a loan-to-value ratio of 50%. Not only does the loan have less chance of defaulting, it is likely the lender will not incur a loss if the loan did default.
- Robust DCF models can be run at the loan level. To further our example, let’s say that you are modeling these loans in bullet 2 at the portfolio level. For the sake of simplicity, both loans equal $100,000. In this case, you have a portfolio of $200,000 with a FICO of 620 and a loan-to-value ratio of 87.5%. Given these aggregate attributes, you would expect little to no credit losses for the portfolio while in reality you have significant risk of loss for the 125% LTV loan.
- The DCF models are prospective in nature and changes in forecasted macroeconomic conditions can be relatively easily incorporated into them. For example, our default rate assumptions for consumer and residential real estate loans are based on the existing and forecasted unemployment rate and we can easily incorporate expected changes in the unemployment rate into our models. Similarly, our loss severity assumption for residential real estate loans is based on the combined loan-to-value ratio of the loan. Our DCF models dynamically vector the combined LTV based on expected changes in housing prices. We believe this is much more straightforward, and thus more easily verified, than making environmental and qualitative adjustments to aggregate analyses.
- Modeling using the same primary indicators an institution’s lenders use to make loans such as FICO and loan-to-value ratio for residential real estate loans can facilitate better risk management across the organization because everyone is “speaking the same language.” Those responsible for credit risk can communicate in ways that are directly actionable – “we are facing an uptick in unemployment, and let’s reduce our exposure to low FICO long-term auto loans” as opposed to “based on our vintage analysis we believe we could see an increase in losses due to the forecasted change in unemployment and we would like you to take less risk.”
- We believe that robust DCF models have another advantage:
- Discounted cash flow models using the SIFMA standard financial mathematics we described above are widely used by valuation experts. This facilitates communication and review because the resulting cash flows are in a standard protocol used in nearly all public offerings of financial instruments. The focus can then be on better understanding the drivers of loss in the model – the input assumptions.
In short, we believe that the use of dynamic, robust, granular prospective DCF models to calculate expected lifetime credit losses based on the credit attributes an institution’s lenders use to make loans can lead to better, more informed management decisions and better allocations of capital.