Recognizing Loan Losses in Banks: An Examination of Alternative Approaches

Abstract

I investigate the accounting rules for loan loss recognition in banks. In June 2016 the FASB issued a new rule, effective in December 2019, that will replace current GAAP with a model that allows banks to use broader information to estimate loan loss allowances. To empirically examine current GAAP and the new model, I exploit differences in the information sets allowed under the old and the new rules. Using a methodology that combines micro data and machine learning techniques, I provide evidence that it is possible to construct a loan loss recognition model that outperforms the current GAAP without expanding the information set beyond that permitted under the current rule. I find that expanding this model’s information set does not significantly improve its performance. My model’s predicted allowances would have been materially larger at the outset of the financial crisis than actual reported bank estimates. The differences are due to that my model consistently assigns larger weights to certain input variables relative to current GAAP. I also find that weakly capitalized banks under-provision relative to well capitalized banks. My results provide a novel method to examine aspects of the new accounting rule before it comes into effect. The findings suggest that the way information is used, rather than the use of broader information set improves the estimates of loan loss allowance.