Uncertain times and early predictions of bank failure

• Author: Cullen F. Goenner
• DOI: http://doi.org/10.1111/fire.12213
• Published date: 01 November 2020
• Date: 01 November 2020

DOI: 10.1111/fire.12213

ORIGINAL ARTICLE

Uncertain times and early predictions of bank

failure

Cullen F.Goenner

Department of Economics and Finance,

University of North Dakota,Grand Forks, North

Dakota

Correspondence

CullenF. Goenner,Department of Economics

andFinance, University of North Dakota, 293

CentennialDrive, Grand Forks, ND 58202.

Email:cullen.goenner@und.edu

Abstract

The Great Financial Crisis shows that bank failure in the United

States, while rare, is a concern during uncertain times. Interest here

is in the ability to predict future failures at the start of a crisis, when

the recent past has few events on which to base inferences. I show

that policy makers using estimates based on the Savings and Loans

crisis would identify in early 2009 that 2.0% of banks were in critical

condition and 7.0% were unhealthy. This is comparable to the 1.7%

of banks that failed within a year and the 3.9% of banks that would

fail during the crisis.

KEYWORDS

bank failure, banking crisis, Bayesian model averaging,prediction

JEL CLASSIFICATIONS

G17, G21, G28

1INTRODUCTION

When Lehman Brothers filed for bankruptcy on September 15, 2008, it became quite clear to everyonethat the finan-

cial sector was again in crisis and that commercial banks in the United States and elsewhere were at a heightened risk

of failure. The magnitude of the risk to banks was unclear at the time, as the previous 15 years had seen, on average,

less than eight failures a year in the United States and 2 years (2005, 2006) with none. Traditionalearly warning mod-

els of bank failure rely on the recent pattern of previous failures to base their predictions. With few recent failures to

drawfrom, it would seem natural to look at a past period of crisis to guide U.S. policy makers’ predictions in late 2008 of

the banks that would fail subsequent to the Great Financial Crisis. Previous research (Cole, Cornyn, & Gunther,1995;

Cole & Gunther, 1998) shows that the statistical models used bythe Federal Reserve were quite accurate in predict-

ing failures during the last major banking crisis of 1985–1992, referred to as the Savings and Loans (S&L) crisis. These

models’ predictions are based on the financial conditions of banks that are captured in their call reports and reflect

measures of banks’ capital adequacy, asset quality,management quality, earnings, liquidity, and sensitivity to market

risk (CAMELS). Cole et al. (1995) and Cole and Gunther (1998) find that using model estimates based on data from

an earlier year during the S&L crisis enables them to accurately predict failures out-of-sample later during the same

Financial Review.2020;55:583–601. wileyonlinelibrary.com/journal/fire c

2019 The Eastern Finance Association 583

584 GOENNER

crisis. I consider whether failure patterns established during the S&L crisis are also useful in predicting failures at the

start of the Great Financial Crisis.

Several studies (Cleary & Hebb, 2016; Cole & White, 2012; DeYoung& Torna, 2013; Jin, Kanagaretnam, & Lobo,

2011; Ng & Roychowdhury,2014) examine bank failures during the Great Financial Crisis. Similar to Cole et al. (1995)

and Cole and Gunther (1998), these studies use data exclusiveto a particular crisis period for both their modeling and

evaluation purposes. The contributions of these studies provide policy makersimportant insights into the factors that

influence failure during the financial crisis. Forexample, Jin et al. (2011) find that a bank’s choice of auditor plays a role

in failures during the crisis, as does treatment of loan loss reserves as regulatory capital (Ng & Roychowdhury,2014).

DeYoung and Torna (2013) find that a banks’ exposure to nontraditional banking activities (insurance underwriting,

securitization, investmentbanking, and venture capital) puts a bank at a higher risk of failure during the Great Financial

Crisis. The issue from the policy maker’s perspectiveis that the usefulness of these models’ estimates is retrospective.

In other words, they are useful in helping to understand the conditions that contribute to failure after the crisis period

examined has passed, while their relevanceto failures in a future c risis is yetunclear.

This paper follows the observation made by Cole and White (2012) that the same financial conditions influencing

bank failures in 2009 also affected failures during the 1980s (Cole & Gunther,1998; Lane, Looney, & Wansley, 1986;

Thomson, 1992; Whalen, 1991). My focus differs from Cole and White (2012) and others in that I explicitly use the

previousfailure experience of banks during the S&L crisis to build a prediction model, which is applied to data observed

by policy makers at the start of 2009 for out-of-sample predictions of failures in the period 2009–2015. In a sense, I

attempt to identify initial conditions that serve asc ommonr isk factors of bank failuresacross different crises episodes,

whichallow for the creation of risk scores at the start of a subsequent crisis. If these risk scores are accurate, then policy

makersin early 2009 would have a means of assessing risks to banks early at the start of the crisis. It would also suggest

that lessons learned from the Great Financial Crisis may help predict risk exposureduring the next crisis.

I assess the risk of bank failure using both logit and survival models. Others (Cole & Wu, 2010; Mayes & Stremmel,

2014) focus on the relative accuracy of the two models, whereas I focus on the different information that each model

provides policy makers at the start of a crisis. The logit model’s estimates using data from the S&L crisis can correctly

identify 106 of the 129 (82%) banks that actually fail within a year using year-end financial data from 2008 and a cut-

off established in the earlier period. The Cox proportional hazards model uses estimates from the S&L crisis and year-

end data from only 2008 to predict risk scores and the survival experience of banks throughout the period 2009–

2015. I show that the model’s prediction of risk of failure is as accurate later in the crisis period as it is early in the

crisis period, which indicates banks’ initial financial conditions are good for assessing the risk to banks throughout a

crisis period. I measure accuracywith a time-dependent version of the area under the receiver-operating characteristic

(ROC)curve (AUC; Heagerty & Zheng, 2005). Policy makers, using these Cox estimates and classification of risk scores,

would identify in early 2009 that 2.0% of banks were in critical condition, 7.0% were unhealthy,and 91% were healthy.

This is comparableto the 1.7% of banks that failed within a year, the 3.9% of banks that would fail later during the crisis,

and the 94.3% that remained healthy throughout.

This paper contributes to the literature, as I demonstrate models of bank failure are subject to the uncertainty

of which variables to include in the model’s specification. Cole et al. (1995) note that the Federal Reserve identifies

approximately 30 financial variables most likely to affect the probability of bank failure.1From this list, Cole et al.

(1995) use stepwise selection to determine the subset of these variables relevantto failures during the S&L crisis. Bank

failures, even during crises, are relatively rare events, and in such cases where there are manypotential risk factors,

predictions based on a single model specification are likelysensitive to variable selection (Volinsky,Madigan, Raftery, &

Kronmal, 1997). I use techniques of Bayesianmodel averaging (BMA) to base my inferences on estimates that explicitly

account for the uncertainty of the model’s specification by averaging overthe estimates from several different speci-

fications. By accounting for model uncertainty,I improve, relative to stepwise selection, the out-of-sample predictions

of the logit and Cox models.

1However,Lane et al. (1986, p. 516) note there is little agreement by regulators on which of these factors are most important.