A Proportional Hazards Model of Bank Failure: An Examination of Its Usefulness as an Early Warning Tool

The number of U.S. bank failures jumped sharply in the mid-1980s and has remained disturbingly high, averaging roughly 170 banks a year over the 1985-1990 period. Furthermore, large-bank failures have become increasingly common. For a variety of reasons, the timing of closures and the resolution techniques used have severely strained the resources of the Federal Deposit Insurance Corporation (FDIC). These developments have stimulated a great deal of debate about the causes of costly bank closures and about alternative ways to prevent them. One focus of this debate has been on the appropriate roles of market versus regulatory discipline. A necessary condition for effective discipline by either force is the ability to identify high-risk banks accurately at a reasonable length of time prior to failure without the use of expensive and time-consuming on-site examinations. This requires the use of some sort of statistical model, conventionally labeled an "early warning model," to translate bank characteristics into estimates of risk. There is considerable debate about whether models of sufficient accuracy can be built using only currently available accounting data. This study examines a particular type of early warning model called a Cox proportional hazards model, which basically produces estimates of the probability that a bank with a given set of characteristics will survive longer than some specified length of time into the future. The sample consists of all banks that failed between January 1, 1987 and October 31, 1990 and a randomly selected group of roughly 1,500 nonfailed banks. Using a relatively small set of publicly available explanatory variables, the model identifies both failed and healthy banks with a high degree of accuracy. Furthermore, a large proportion of banks that subsequently failed are flagged as potential failures in periods prior to their actual demise. The classification accuracy of the model over time is impressive, since the coefficients are based on 1986 data and are not updated over time. In short, the results demonstrate that reasonably accurate early warning models can be built and maintained at relatively low cost.