Early warning models in real time

Each quarter, banks file a call report, or Report of Condition and Income, containing hundreds of accounting items pertaining to their financial condition. Because call reports are filed quarterly, whereas banks are typically examined about once every twelve to eighteen months, statistical early warning models using call report data potentially provide a more up-to-date picture of a bank's condition than on-site exams alone. Often neglected, however, is the fact that call report data are subject to revision. We find evidence of a strong relationship between on-site exams and call report revisions. In addition, we evaluate a major class of early warning models using both originally published and revised data to assess whether model accuracy in real time is appreciably lower than accuracy measured using revised data. The findings indicate revised data overstate the accuracy of early warning models. The substantial effect of revisions on the accuracy of early warning models, coupled with the finding of a relationship between revisions and exams, points to a substantial auditing role for on-site exams. More generally, our findings point to the need for care in the use of call report data for research in which the real-time flow of financial information is of some concern.