Full text of Financial Industry Studies : August 1992

View original document

The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.

FINANCIAL INDUSTRY
August 1 9 9 2

STUDIES
Federal Reserve Bank of Dallas

What Do Early Warning
Models Tell Us A b o u t
Asset Risk and Bank Failures?
Linda M. Hooks

Economist

Banking Difficulties
and Discount W i n d o w Operations:
Is Monetary

Policy

Affected?

Kenneth J. Robinson
Senior

Economist

This publication was digitized and made available by the Federal Reserve Bank of Dallas' Historical Library (FedHistory@dal.frb.org)

Financial Industry Studies
Federal Reserve Bank of Dallas
August 1992

President and Chief Executive Officer
Robert D. McTeer, Jr.
First Vice President and Chief Operating Officer
Tony J. Salvaggio
Senior Vice President
Robert D. Hankins
Vice President
Genie D. Short
Senior Economists
Jeffery W. Gunther
Robert R. Moore
Kenneth J. Robinson
Economist
Linda M. Hooks
Financial Analyst
Kelly Klemme

Industry Studies is published by the Federal Reserve Bank
of Dallas. The views expressed are those of the authors and
do not necessarily reflect the position of the Federal Reserve Bank
of Dallas or the Federal Reserve System.

Financial

Subscriptions are available free of charge.
Please send requests for single-copy and multiple-copy subscriptions,
back issues, and address changes to the Public Affairs Department,
Federal Reserve Bank of Dallas, Station K, Dallas, Texas 75222, (214) 922-5254.
Articles may be reprinted on the condition that the source
is credited and the Financial Industry Studies Department is provided
a copy of the publication containing the reprinted material.

What Do Early
Warning Models Tell
Us About Asset Risk
and Bank Failures?
Linda M. Hooks
Economist
Financial Industry Studies Department
Federal Reserve Bank of Dallas

The findings highlight the importance of
risk-taking in explaining recent bank failures.
The results indicate that asset risk measures
were more important for predicting bank
failures in the mid-1980s than in the late
1980s. Risk-taking was reflected in ex ante
measures of asset risk in 1985, before the
most severe regional economic difficulties.
In 1987 and 1989, however, the effects of
this risk-taking were manifested in ex post
measures, such as low equity-to-asset ratios.
This shift occurred as adverse regional
economic conditions led banks to write
off problem loans.

Eleventh District Banking Conditions

anking difficulties rose dramatically
during the 1980s, as the rate of U.S.
bank failures reached levels not observed
since the Great Depression. While Eleventh
District banks showed signs of recovery in
1991, banking difficulties have persisted in
some regions of the country. The financial
turmoil experienced during the past decade
has prompted renewed interest in predicting
bank failures. Early warning models are
econometric tools designed to estimate the
impact of a set of relevant explanatory
variables on the probability of bank failure.
These models can help regulators identify
potential problem banks and thereby help
to limit potential losses to the Federal
Deposit Insurance Corporation (FDIC)
resulting from bank failures. Bank managers
could also use similar techniques to provide
information for managing portfolio risk.
An important element of early warning
models of bank failure is the riskiness of
bank assets. The accurate measurement of
a bank's risk posture remains a complex
issue, however. Many analysts have focused
on specific aspects of a bank's portfolio
that identify some portion of asset riskiness,
but few have examined a more comprehensive picture of a bank's overall risk exposure.
Using data on banks in Eleventh District
states, this article investigates how information on the contents of a bank's asset portfolio can be used to evaluate asset risk in
the context of an early warning model.1

The profitability cycle experienced by
Eleventh District banks in the 1980s is one
of the most pronounced on record. Chart 1
shows movements in the return on assets
over the past decade for banks in each
state in the District. The dramatic decline in
the return on assets at Texas banks through
1988 is echoed, to a lesser degree, by
Louisiana banks over the same period.
New Mexico banks experienced less drastic
changes in profitability. These swings in
profitability were primarily attributable to
the interaction between changes in the
economic environment in which District
banks operated and changes in the riskiness
of bank portfolios.
Changes that occurred in the District's
economic environment are illustrated in
Chart 2, which shows growth in the annual
averages of nonagricultural employment
levels for the District states compared with
the United States over the past decade.
Employment in Texas and Louisiana
decreased in the early 1980s, then grew
again through 1985. Employment then
declined more severely through 1987. These
movements mirror the changes in bank

1 The Eleventh District comprises Texas, northern
Louisiana, and southern New Mexico. This article considers the banking industries of these three District
states.

profitability shown in Chart 1. Because the
years 1985 and 1987 mark high and low
points in District economic activity, the
early warning models examined later are
estimated for both years, and for 1989.
These estimates facilitate comparisons of
the importance of measures of asset risk in
predicting bank failures under different
local economic conditions.
In addition to reflecting changes in
economic conditions, movements in bank
profitability also reflect the underlying composition of banks' portfolios. Chart 3 shows
a breakdown of the average loan-to-asset
ratio at year-end 1985, before the strong
downturn in the regional economy. District
banks held a higher average concentration
of commercial and industrial loans than
banks elsewhere in the nation, together
with a greater share of assets in both construction loans and loans backed by commercial real estate. Texas banks, in particular,
also held a relatively low proportion of
assets in residential real estate loans. The
average loan-to-asset ratio was higher for
each of the District states than for the rest
of the nation.
The increase in District states' bank failures that followed the abrupt changes in

Chart 1

Return on Assets for Insured Commercial
Banks, 1980-91

Percent

SOURCE: Report of Condition and Income.

Chart 2

Nonagricultural Employment, 1980-91

Index, 1980= 100

SOURCE: CITIBASE, Citibank Economic Database.

District bank profitability is shown in Chart
4. Bank failures in Texas and Louisiana
mounted rapidly over the decade, peaking
in 1989, while bank failures in New Mexico
were spread over the mid-1980s and early
1990s. The increase in bank failures, both
in the District states and in many other
regions in the country, has heightened
interest in developing accurate early warning models of bank failures. Early warning
models can identify potential problem
banks and, thus, help avoid bank failures.
By 1991, District banks' loan portfolios
were more evenly divided among the various
loan categories than earlier in the decade.
While Chart 3 showed that District banks
held fairly concentrated loan portfolios in
1985, Chart 5 shows a more diverse composition of the average loan portfolio for
banks in the District by the end of 1991- A
comparison of Charts 3 and 5 also shows
that banks in Texas and Louisiana held a
substantially smaller proportion of their
assets in loans in 1991 compared with the
rest of the nation. The proportion of loans
relative to total assets decreased because
banks were more cautious in lending, and
fewer borrowers sought bank loans. The
cautious lending positions reflected additional loan write-downs taken during the

period of financial difficulties, and the
declining demand for loans reflected the
regional economic downturn.

Shortcomings of Existing Asset Risk Measures

Chart 4

Bank Failures, 1982-91

Number of failures
160

Measures of asset risk are particularly
important components of an early warning
model, since asset risk can profoundly
affect bank survivability. Asset risk measures
generally are more forward-looking than
other types of explanatory variables and,
therefore, may significantly aid in the identification of potential bank failures. A more
accurate measure of asset risk also would
help bank managers to control risk exposure
more effectively.
Bank analysts sometimes quantify asset
risk by examining the share of loans or
assets in a specific loan category. Users of
this method presume that the chosen loan
share category contributes substantially to
portfolio risk. For example, commercial and
industrial loans may be perceived as a particularly risky loan category. Because this
method fails to take into account components of the portfolio other than the chosen

Chart 3
Average Lending Concentration, 1985
Loan-to-asset ratio, percent

1982 1983 1984 1985 1986 1987 1988 1989 1990 1991

NOTE: There were no bank failures in Texas, Louisiana, or
New Mexico in 1980 or 1981.
SOURCE: Federal Deposit Insurance Corporation.

loan category, it produces an inaccurate
assessment of risk.2
A second approach to measuring bank
asset risk by evaluating loan concentrations
is known as the Herfindahl Index. This
index measures asset risk by identifying
appropriate loan categories, forming the
ratio of each loan categoiy to total assets,
and summing the squares of those ratios.3
The Herfindahl Index method fails to account

For example, a particular loan category may b e risky
in the sense that it generates volatile, or unstable,
returns, but the movement in those returns may be
offset by movements in the returns of another component of the portfolio. In such cases, the return to the
overall portfolio may be relatively stable, even though
the returns to each component of the portfolio are
quite volatile when viewed in isolation.
2

Louisiana

Rest of the
United States

New Mexico

I Residential Real Estate

•

I Commercial & Industrial

• I Consumer

I Commercial Real Estate

E3 Other

Construction

SOURCE: Report of Condition and Income.

The formula for this measure is (real estate loans to

total assets)2 + (commercial and industrial loans to
total assets)2 + (.consumer loans to total assets)2 +
(commercial and industrial loans to total assets)2 +
(foreign loans to total assets)2 + (agricultural loans to
total assetsY + (depository institution loans to total
assets)2. Thomson (1991) demonstrates the use of this
measure in an early warning model.

Chart 5
Average Lending Concentration, 1991
Loan-to-asset ratio, percent
6 0 -i

50 •

Louisiana
I Residential Real Estate

I Commercial & Industrial

EJ Consumer

Construction

I Commercial Real Estate

Other

SOURCE: Report of Condition and Income.

for differing risk levels among the loan
share categories. Each loan category is
given the same weight, even though risk
may differ across categories.4

Finance Theory and Measures of Asset Risk
While no perfectly accurate measure of
asset risk exists, finance theory can address
some problems. Modern finance theory

For example, a bank with a high concentration of
loans in a risky loan category would b e identified as
having the same level of asset risk as a bank with a
high concentration of loans in a less risky loan category.
4

5 Formally, beta for investment i is defined by the
formula covariance (R B )/variance (Rm), where R.
denotes the returns to investment i and Rm denotes
the returns to the market overall. Sharpe (1964) provides a formal derivation of the capital asset pricing
model; Copeland and "Weston (1988) give a useful
exposition of the beta measure.

Beta measures systematic risk, the reaction of an
asset to general market movements. A measure incorporating systematic and nonsystematic asset risk, the
variance of the portfolio, might b e used instead.
6

shows that a portfolio's riskiness depends
on both the loan shares in a portfolio and
the covariation among the returns to the
different shares. Financial analysts considering the riskiness of investments in the stock
market often calcLilate a risk measure called
beta, derived from finance theory, that
accounts for these factors. Of course, any
measure of asset risk will suffer from some
degree of error, but an understanding of
the beta risk measure offers some methods
for improving the measurement of asset risk
in early warning models of bank failures.
Beta summarizes information about the
riskiness of a particular stock that is
conveyed by movements in the marketdetennined price of that stock. Beta measures
the covariance of the returns to a particular
stock with the returns to the stock market
overalls This covariance is relevant because
it measures the portion of risk that arises
from an individual stock, or the contribution of a particular stock to portfolio risk.6
For example, consider an investor who
holds a portfolio of two stocks. If the returns
to both stocks move in tandem with the
returns to the stock market overall, then
the returns to the portfolio will be volatile,
or the portfolio will be risky, because
returns to both components of the portfolio
will move up or down simultaneously. If,
instead, the returns to only one of the
stocks parallel the returns to the stock
market overall, while the returns to the
other stock move inversely with the market,
then the returns to the portfolio will be less
volatile and the portfolio less risky because
changes in the returns offset each other.
A portfolio composed of several stocks has
a beta that is simply the sum of the betas
for the individual investments, weighted by
the proportion of the portfolio in each
investment.
For bank assets, a beta-type measure can
be constructed using beta measures for each
asset category. A bank's asset risk is then
measured as the sum of these proxy betas,
scaled by the corresponding loan shares.
Two methods of incorporating these beta
risk weights are the nonsample information

method and the in-sample information
method.7
The nonsample weighting method calculates risk weights from data external to the
sample. Each loan share category is matched
with a similar industry from the stock market;
the beta for this stock market industry
serves as a proxy for the loan category's
beta. The beta of a bank's loan portfolio is
the sum of these individual proxy betas
weighted by the proportion of the asset
portfolio invested in each loan category. In
other words, a bank's portfolio beta is a
risk-weighted sum of the loan shares in the
portfolio, where the risk weights are betas
from related industries.
The different weights on loan shares
reflecting their differing riskiness can alternatively be calculated from data within the
sample under consideration. This technique
requires that each loan share be included
as an explanatory variable when the early
warning model is estimated. The risk
weights will be the coefficients on each
loan share in the model after it has been
estimated with the sample data.
If the additional information used in the
nonsample weighting method is relevant to
bank portfolio risk, then an early warning
model using the nonsample method should
yield better predictions of bank failures
than one using the in-sample method. If
the additional nonsample information is
irrelevant, then this erroneous information
biases the results of an early warning model,
and the nonsample weighting method
would produce less accurate predictions
of bank failures than the in-sample method
would. For example, the nonsample information may not be helpful if the stock
market proxies chosen for the loan categories are inadequate representations of
the returns on the actual loans.
While the in-sample weighting measure
and the nonsample weighting measure, in
theory, improve the measurement of risk,
in practice, these measures are subject to
errors in implementation just as other measures are. It is important, then, to examine
the extent to which these alternate measures

actually improve the predictions of estimated early warning models. The next
section examines this issue by comparing
estimates of early warning models using
the in-sample weighting and nonsample
weighting asset risk measures with estimates
of early warning models using two other
measures, the Herfindahl Index and the
loan-to-asset ratio.

Early Warning Model Accuracy
Estimates of the early warning model
measure asset risk in four alternate ways to
assess their impact on the accuracy of the
model. These four methods are the insample risk weighting measure, the nonsample risk weighting measure, the Herfindahl Index measure, and the loan-to-asset
ratio. For comparison, a fifth version of the
model that does not include an explicit
measure of asset risk is also estimated; this
serves as a benchmark from which die other
models can be judged. The box titled "Overview of Early Warning Models " explains
the remaining variables used in the early
warning models.
The usefulness of an estimated early
warning model can be evaluated by
examining the model's ability to predict
whether a given bank will fail. Two types
of errors may arise when an estimated
model is used to predict bank failure. First,
a bank that actually fails during the time
period under consideration may be incorrectly predicted to survive; this is referred
to as a Type I error. Secondly, a bank that
actually survives during the specified time
period may be incorrectly predicted to fail;
this is a Type II error. A Type I error means
that the model has faltered in its early
warning capacity, since it misclassified a
bank that actually failed as one that would
not fail; for this reason, a Type I error is
often considered the more serious error.
Type I and Type II errors vary inversely

Schaefer (1987) proposes a related risk measure
based on finance theory.

Overview of Early Warning Models
Early warning models provide a framework for analyzing a bank's financial condition. T h e s e e c o n o m e t r i c m o d e l s
estimate the relationship between a set of
relevant explanatory variables, such as
bank capital, and the likelihood of bank
failure. The estimated relationship can
then be used to forecast future bank failures. An important measure of the accuracy of an early warning model is the
proportion of failed banks that are correctly identified by the model as potential
failures.
Most early warning models use a similar set of explanatory variables. 1 The
model in this article includes among its
explanatory variables proxies for the components of the CAMEL rating used by
examiners to evaluate the financial condition of a bank. 2 For the model estimated
below, capital adequacy is measured by
the ratio of equity capital to assets. Asset
quality is measured by the ratio of net
charge-offs to total loans. 3 Management
expertise is approximated by the ratio of
overhead expenses (noninterest expenses) to assets and by the ratio of loans
extended to bank officers to assets. Earnings are represented by the ratio of net
income to assets, while liquidity is measured by the ratio of cash and securities to
assets. Asset risk measures are incorporated in several different ways, as explained in the main article.

Finally, the early warning model includes several additional explanatory
variables to control for other influences
on bank failures. A dummy variable controlling for the impact of a bank's corporate structure takes on the value of one for
a bank holding company, and zero otherwise. The model includes total assets of
the bank to account for the possible impact of bank size on failure. Because a
bank's access to low-cost funds may be
important to its financial condition, the
ratio of core deposits to total deposits also
is included. The model allows for the
impact of local economic conditions by
including the percentage change in the
county-level annual average per-capita
personal income and employment levels.

1 Avery and Hanweck (1984), Pantalone and Piatt
(1987), Thomson (1991), and Whalen (1991) estimate early warning models using various combinations of the variables mentioned. For a survey of the
results of several different models, see DemirgucKunt (1989).

CAMEL stands for capital, asset quality, management, earnings, and liquidity.

Estimates of early warning models sometimes use
the ratio of nonperforming assets to total assets
instead of net charge-offs. Regressions not reported here indicate that this substitution does not
qualitatively alter the results.
3

for an estimated early warning model. Type
1 error, the proportion of failed banks classified as survivors, falls as Type II error, the
proportion of surviving banks classified as
failures, rises. Curves depicting this tradeoff between the proportions of the two
types of errors can be used to compare the
accuracy of competing early warning models.
Given an acceptable level of Type I error,
the model with the lowest Type II error
would be preferred. That is, analysts desire
early signals about potentially excessive
risk exposure in banks, so they want an
early warning model with a low Type I
error (that correctly identifies a large proportion of the banks that fail); but for a
given level of Type I error, analysts prefer
a model that correctly classifies more
surviving banks than other models do.
Generally, one model performs better than
another if it produces a curve that fits closer
to the graph's axes at each point of the
curve; the closer fit of the curve means that
one model has lower Type II errors for
each level of Type I errors compared with
the other model.
Accuracy of the 1985 early warning
model. The first sample data set consists
of all banks in Louisiana, Texas, and New
Mexico in 1985. The model estimates the
relationship between the 1985 values of the
explanatory variables for the banks and the
failure or survival of these banks over the
period 1986-87. The risk posture of banks
in 1985, as they headed into the economic
downturn, would be expected to affect
their profitability during the downturn.
Chart 6 shows the trade-off curve
between Type I errors and Type II errors
associated with different early warning
models for the period 1986-87. Because
the model's most important function is to
predict bank failures correctly, the acceptable Type I error level chosen should be
relatively low. Choosing the 5-percent level
of Type I error, for example, means that
each model correctly identifies 95 percent
of the banks that fail. When each model
misclassifies only 5 percent of the failed

Chart 6

1985 Model Prediction Errors
for Bank Failures, 1986-87
Type II
Misclassified surviving banks

Type I
Misclassified failed banks

banks, which proportion of the surviving
banks does it misclassify? Chart 6 shows
that, at the 5-percent level of Type I error,
the model using in-sample risk weighting
misclassifies 24 percent of the surviving
banks, while the model using the Herfindahl Index misclassifies 32 percent of the
surviving banks. In comparison, the benchmark model that uses none of the asset risk
measures misclassifies 50 percent of the
surviving banks. The remaining models
produce Type II misclassification rates
between 24 percent and 50 percent.8
While the in-sample and nonsample risk
weight models generate fewer errors than
the other models with risk measures, all

For clarity, each figure depicts the error trade-off
curves only for the benchmark model, the most accurate asset risk model, and the least accurate risk
model. The Appendix titled "Regression Results"
contains tables with prediction errors for each model.

four models that incorporate a risk measure
yield fewer prediction errors than the
benchmark model. Especially at lower
levels of Type I errors, the models for each
of the risk measures outperform the benchmark. These results indicate that asset risk
measures were important in predicting
banking difficulties in the 1986-87 period;
errors in predicting bank failures occur
more frequently if asset risk is not included
as an explanatory variable in the early
warning model.
Accuracy of the 1987 early warning
model. Chart 7 illustrates the classification
errors of three models estimated on 1987
data for the period 1988-89, after the
strong downturn in regional economic
conditions. Among the different measures
of asset risk, the four methods now display
similar misclassification rates, although the
in-sample risk weighting method produces
fewer Type II errors at some levels of Type
I errors. The other misclassification rates

Chart 7

1987 Model Prediction Errors
for Bank Failures, 1988-89
Type II
Misclassified surviving banks

Type I
Misclassified failed banks

are close in value to those of the benchmark model that does not explicitly
incorporate asset risk as an explanatory
variable. While asset risk adds substantially
to the predictive power of early warning
models for 1985, the magnitude of this
difference fades when the early warning
model is reestimated with 1987 data.
The decrease in the predictive power of
asset risk variables indicates that later in the
downturn, bank capital levels reflected
asset quality problems more completely
than earlier. During this later period, past
asset risk difficulties were resolved through
loan write-downs, so that bank capital ratios
incorporated past asset quality problems
more fully than measures of current asset
risk. The relatively good performance of
the benchmark model with the 1987 data
indicates that a bank's capital position and
other variables are as effective as its asset
risk position in explaining a bank's condition during the period 1988-89Accuracy of the 1989 early warning
model. The decline in the predictive
power of asset risk measures persists for
the 1989 model estimates. Chart 8 shows
the error trade-off curve for early warning
models estimated on 1989 values for banks
that survived or failed in 1990 and 1991.
The benchmark model that does not
explicitly include a measure of asset risk
performs about as well as the models
incorporating asset risk measures. At low
levels of Type I error, the Type II errors of
the benchmark model are smaller than some
of the models with measures of asset risk.
At low levels of Type I error, each 1989
model has notably lower Type II errors
than the corresponding 1985 and 1987
models, which indicates that each model
generally made more accurate predictions
in 1989 than it did in 1985 or 1987. By
1989, the variables included in the early
warning models more precisely captured
past risk positions than they did in earlier
models. Thus, the early warning models
based on this information have greater
predictive power than the earlier models.

Chart 8

1989 Model Prediction Errors
for Bank Failures, 1990-91
Type II
Misclassified surviving banks

reflective of past risk postures. The early
warning models estimated confirm that
asset risk measures served as early indicators of banking difficulties, but as asset
quality problems were resolved, measures
of capital became more important indicators
of banking difficulties.

Conclusion

Type I
Misclassified failed banks

Ideally, an early warning model would
provide a truly early warning of potential
problems. Instead, the more accurate
predictions of the 1989 models relative to
the 1985 and 1987 models imply that these
models identify potential bank failures
most accurately after the onset of financial
difficulties.
In sum, results for the 1985 early warning
models confirm that an accurate measure
of asset risk is an important component for
predicting a bank's financial condition
before the significant deterioration in local
economic conditions. The better performance of the in-sample and nonsample
risk weights for loan shares demonstrates
that riskiness should be viewed in terms of
the entire bank portfolio, and not in terms
of only a few components of the portfolio.
In the late 1980s and early 1990s, District
banks resolved many of their earlier asset
quality problems. As banks recognized
troubled assets and wrote off problem
loans, measures of capital became more

The results of the early warning models
estimated above show that financial difficulties were first reflected in measures of
ex ante asset risk, and later in measures
of ex post asset risk. In 1985, before the
severe regional economic downturn, the
best predictors of future bank failures in
Eleventh District states include ex ante
measures of a bank's asset riskiness. After
the downturn had occurred, however,
measures of current bank asset risk become
less important, and the bank's asset quality
and capital position, reflecting past risk
positions, become more important.
The lower prediction errors of the 1989
models relative to the 1985 and 1987 models
indicate that ex ante measures of asset risk
used in the 1985 and 1987 models foreshadow potential bank problems imprecisely.
Ex post measures of bank risk-taking
improve the predictive power for the 1989
models, but, by their nature, they were
reflective of declining banking conditions
only after financial difficulties had arisen,
and hence do not contribute significantly
to early warnings. These results highlight
the difficulty early warning models have
in providing accurate signals of problems
that have not yet emerged.
The imprecision associated with risk
predictions based on call report data highlights concerns about the assignment of
fixed weights to asset categories for the
risk-based capital guidelines. Although it is
desirable to assign different weights to
different asset categories, the estimates of
the early warning models in this article
suggest that it is difficult to do so accurately
over long periods of time. As the relationship between a given asset category and
the probability of bank failure fluctuates

over time, the fixed weight on that asset
category will at times be appropriate, too
low, or too high.9 Moreover, the asset categories themselves provide only a rough
basis for characterizing risk, since a single
category can contain assets with differing
risk characteristics. Hence, for both bank
managers and regulators, the risk weights
should be viewed as imperfect guidelines to
the degree of bank risk exposure. This technique is only one of many elements used

« Bradley, Wambeke, and Whidbee (1991) address
some of the difficulties of assigning accurate risk
weights under the new risk-based capital guidelines
for thrifts.

to detennine the overall riskiness of a bank.
Fortunately, most banks in the Eleventh
District states now are well-positioned to
meet the recently imposed risk-based capital
guidelines. The return on District banking
assets rebounded over the past two years,
and the District's bank failure rate subsided.
These developments suggest that the severe
problems of the late 1980s have been
resolved and that banks in the District states
are poised for profitable growth.

Appendix
Regression Results
This appendix summarizes the estimates of early warning models used in
the text to generate the models' prediction error rates. 1 Because the dependent
variable for the model is one if the bank
fails and zero if it does not fail, the models are estimated using the probit technique instead of ordinary least squares.
The models are estimated for 1985,1987,
and 1989.
The estimates of the early warning
models indicate that the coefficients generally have their predicted signs. The
coefficients on the capital-to-asset ratio
and the net-income-to-asset ratio, for
example, are both significantly negatively
related to the probability of failure. The
coefficients on the asset risk variables
are positive and sometimes statistically
significant. The coefficient on the nonsample risk weights measure is significantly positive in the 1985 model, while
the coefficient on the Herfindahl Index
measure is not significant in any model.
The results of the regressions incorporating the in-sample risk weights show that
the coefficients on the separate loan categories vary in sign and magnitude across
the 1985, 1987, and 1989 models.

One unexpected coefficient sign is the
occasionally significantly negative sign
on the ratio of noninterest expenses to
assets. Supplementary regressions indicate that the coefficient on the salary-toasset component of noninterest expenses
has a negative sign, while the coefficient
on the premise-expenses-to-asset component has the expected positive sign. A
second counterintuitive relationship is the
significantly negative sign on the ratio of
net charge-offs to loans in the 1989 model
estimates. This negative relationship is
reported in similar estimates of early warning models, like Thomson (1991), and
may arise from mismatched timing of
these charges with financial difficulties.
Tables A1, A2, and A3 summarize
prediction errors for each model; these
tables correspond to Charts 6, 7, and 8 in
the main text. Each table lists the Type II
errors for the models at a given level of
Type I errors.

1 Complete regression results are available from the
author upon request.

(Continued on the next page)

Regression Results-—Continued
Table A1
1985 Model Prediction Errors for 1986-87 Bank Failures
Type II errors (Percentage of misclassified surviving banks)
Type I errors
(Percentage of misclassified failed banks)

In
sample
weights

Non
sample
weights

Herfindahl
Index

Loans
to
assets

Benchmark

5
10
30
50
70
90

24
19
6
2
1
0

30
24
6
3
1
0

32
28
7
3
1
0

29
22
7
2
1
0

50
35
10
2
1
0

Table A2
1987 Model Prediction Errors for 1988-89 Bank Failures
Type II errors (Percentage of misclassified surviving banks)
Type I errors
(Percentage of misclassified failed banks)

In
sample
weights

Non
sample
weights

Herfindahl
Index

Loans
to
assets

Benchmark

5
10
30
50
70
90

33
22
9
4
2
0

33
24
10
5
2
0

32
24
10
5
2
0

30
25
10
4
2
0

39
29
11
5
2
0

Table A3
1989 Model Prediction Errors for 1 990-91 Bank Failures
Type II errors (Percentage of misclassified surviving banks)
Type I errors
(Percentage of misclassified failed banks)
5
10
30
50
70
90

In
sample
weights
9
7
2
0
0
0

Non
sample
weights

Herfindahl
Index

Loans
to
assets

Benchmark

13
7
2
1
0
0

12
8
2
0
0
0

11
8
2
0
0
0

12
6
3
0
0
0

References
Avery, Robert B., and Gerald A. Hanweck
(1984), "A Dynamic Analysis of Bank
Failures," Proceedings of a Conference on
Bank, Structure and Competition, Federal
Reserve Bank of Chicago, 380-95.

Schaefer, Stephen M. (1987), "The Design
of Bank Regulation and Supervision: Some
Lessons from the Theory of Finance," in
Threats to International Financial
Stability,
eds. Richard Pones and Alexander Swoboda
(Cambridge: Cambridge University Press).

Bradley, Michael G., Carol A. Wambeke,
and David A. Whidbee (1991), "Risk
Weights, Risk-based Capital and Deposit
Insurance," Journal of Banking
and
Finance 15 (September): 875-93.

Sharpe, W. F. (1964), "Capital Asset Prices:
A Theory of Market Equilibrium under
Conditions of Risk," Journal of Finance 19
(September): 425^12.

Copeland, Thomas E., and J. Fred Weston
(1988), Financial Theory and
Corporate
Policy (New York: Addison-Wesley Publishing Company), 193-98.

Thomson, James B. (1991), "Predicting
Bank Failures in the 1980's," Federal
Reserve Bank of Cleveland Economic
Review Quarter 1, 9-20.

Demirguc—Kunt, Asli (1989), "DepositInstitution Failures: A Review of Empirical
Literature," Federal Reserve Bank of Cleveland Economic Review, Quarter 4, 2-18.

Whalen, Gary (1991), "A Proportional
Hazards Model of Bank Failure: An Examination of Its Usefulness as an Early Warning
Tool," Federal Reserve Bank of Cleveland
Economic Revieiv Quarter 1, 21-31.

Pantalone, Coleen C., and Marjorie B. Piatt
(1987), "Predicting Commercial Bank
Failure Since Deregulation," New England
Economic Review July/August, 37-47.

Banking Difficulties
and Discount Window
Operations:
Is Monetary Policy Affected?
Kenneth J. Robinson
Senior Economist
Financial Industry Studies Department
Federal Reserve Bank of Dallas

n important function of the Federal
Reserve at its inception in 1913 was to
serve as lender of last resort. This role
relates to the question of how a central
bank should respond to financial crises. In
particular, under what conditions should
the Fed provide liquidity to the financial
system to avert widespread financial panics?
The classic treatment of this lender-of-lastresort function can be traced to the writings
of Henry Thornton in An Enquiry into the
Nature and Effects of the Paper Credit of
Great Britain, published in the early
nineteenth century, and Walter Bagehot's
definitive work, Lombard Street, published
in 1873. Bagehot summed up what is
required of a lender of last resort: "Theory
suggests, and experience proves, that in a
panic the holders of the ultimate Bank
reserves should lend to all that bring good
securities quickly, freely, and readily."
The Federal Reserve System has carried
out its lender-of-last-resort responsibilities
primarily through its discount window
operations.1 Federal Reserve advances to
depository institutions are grouped into
three broad categories. Adjustment
credit
represents short-term loans extended to
depository institutions when other sources
of funds "...are not reasonably available and
when the need for credit is appropriate.
Guidelines for administering adjustment
credit are fairly general, in recognition of
the wide range of circumstances that may

give rise to borrowings by institutions that
differ in size, in the nature of their business, and in the economic environments in
which they operate." 2
The second category of discount window
advances is seasonal credit, which is available to depository institutions of relatively
small size that generate a clear pattern of
recurring swings in funding needs. The
Federal Reserve established the seasonal
program in the early 1970s because of
some small banks' lack of access to national
money markets. Under this program, institutions "may obtain longer-term hinds from
the discount window during periods of
seasonal need so they can carry fewer liquid
assets during the rest of the year and can
make more hinds available for local
lending." 3
The third category of discount window
loans is known as extended credit. Extended
credit is issued to depository institutions
experiencing "...special difficulties arising
from exceptional circumstances or practices
involving individual institutions or from
liquidity strains affecting a broad range of
depository institutions."4 The extended
credit category of discount window advances
most closely approximates central bank
operations under the traditional concept of
lender of last resort.
While the lender-of-last-resort function
remains vital, the most essential role of
the central bank today can be found in its
monetary policy objectives. That is, the

1 The Federal Reserve could also carry out its lenderof-last-resort function through open market operations. Before 1980, discount window advances were
available only to banks that were members of the
Federal Reserve System. Since passage of the Depository Institutions Deregulation and Monetary Control
Act of 1980, all depository institutions are eligible for
discount window loans.

Federal Reserve System, 1990.

Chart 1

promptly as there is with traditional shortterm adjustment credit," the Federal Reserve
Bulletin states, "the money market impact
of extended credit is similar to that of nonborrowed reserves."' Thus, as Bagehot
stressed, any liquidity assistance by the
central bank should be short-temi in nature,
mainly to ensure that the effect on monetary policy goals is inconsequential. 6 Humphrey (1975, 4) points out, though, that a
conflict between lender-of-last-resort and
monetary policy goals is not inevitable:

Extended Credit
Billions of dollars

T h e r e n e e d b e n o conflict b e t w e e n the
m o n e t a r y control a n d lender o f last resort
functions, h o w e v e r , s i n c e the first refers to
the long run a n d the s e c o n d to temporary
SOURCE: CITIBASE, Citibank Economic Database.

periods o f e m e r g e n c y . If the central bank,
in its role as l e n d e r o f last resort, r e s p o n d s
appropriately to the threat o f a liquidity
crisis, the panic will b e averted quickly.

central bank is also charged with the responsibility of controlling the growth of the
nation's money supply. The exercise of
lender-of-last-resort responsibilities to cope
with liquidity needs could lead to a temporary abrogation of these monetary policy
responsibilities. Repeated bank and thrift
failures in the 1980s resulted in the more
frequent use of extended credit. Extended
credit, in effect, adds to bank reserves, and,
thus, can potentially alter the course of
monetary policy. "Because there is not the
same need to repay such borrowing

See Federal

Reserve Bulletin,

notes to RESERVES

AND BORROWINGS, Depository Institutions Table.
6

This article focuses on the narrow category of ex-

tended credit. For more on the role and functions of
lender of last resort, see Humphrey (1975) and Garcia
and Plautz (1988).
Before the early 1980s, extended credit was, for the
most part, not issued. The exception was in 1974,
w h e n extended credit increased sharply. While individual institutions are not identified as recipients of
extended credit, the timing of this spike in extended
credit coincided with the difficulties of the Franklin
National Bank.
7

C o n s e q u e n t l y , the deviation o f the m o n e y
s t o c k from its long-run target path will b e
small, b o t h in magnitude a n d duration.

The issue of interest in this article is
whether the Federal Reserve's role as
lender of last resort has affected its role in
controlling the nation's money supply. To
examine this issue, the behavior of extended
credit is analyzed from November 1982
through July 1991- I chose this period
because extended credit was issued very
infrequently before November 1982, since
financial-sector difficulties were relatively
rare. The November 1982 to July 1991
period includes a time of consistent Federal
Reserve operating procedures—mainly the
borrowed-reserves operating regime.7
Empirical results suggest that the Federal
Reserve's role as lender of last resort did
not alter the course of monetary policy—
whether measured by movements in interest
rates or by movements in the money supply.
Before examining these results, I offer
some background on the Federal Reserve's
use of extended credit in the next section.
Then, I describe the statistical techniques
used to judge the potential conflict between
lender-of-last-resort and monetary policy

Chart 2
Nonborrowed Reserves and Extended Credit
Billions of dollars

•

Extended Credit

•

Nonborrowed Reserves

1983

1984

1985

1986

1987

1988

1989

1990

1991

SOURCE: CITIBASE, Citibank Economic Database.

objectives, and I present the results. The
final section offers conclusions from and
policy implications of these findings.

The Federal Reserve's Use of Extended Credit
Under the Federal Reserve's Regulation A,
extended credit is available to banks and
other depository institutions when similar
assistance is not readily available from other
sources and where exceptional circumstances or practices are adversely affecting
an individual depository institution. Regulation A states, "Exceptional circumstances
would include situations where an individual depository institution is experiencing financial strains arising from particular
circumstances or practices affecting that
institution—including sustained deposit
drains, impaired access to money market
funds, or sudden deterioration in loan
repayment performance." Applicants for
extended credit are required to make full
use of other reasonably available sources of
funds, including spccial industry lenders
before turning to the discount window.
Furthermore, to ensure effective coordination, requests for extended credit will be
reviewed jointly by the Federal Reserve
and other responsible supervisory agencies.

Finally, Federal Reserve assistance is provided only if a plan for eliminating the
institution's liquidity shortfall is in place or
is being worked out with the Federal
Reserve and other relevant authorities. An
interest rate above the basic discount rate
may be applied to loans made under other
extended credit, subject to review and
determination by the Board of Governors.8
Provisions of extended credit can affect
monetary policy because they can affect
the amount of total reserves outstanding.
Total reseives are the sum of borrowed
reserves, which include extended credit,
and nonborrowed reserves, which are
supplied by open market operations. On
average, extended credit is a very small
component of total reserves—less than 3
percent. However, extended credit tends to
be very volatile. In fact, extended credit
exhibits roughly forty times more volatility
than does the category of nonborrowed
reserves.9 Chart 1 shows movements in
extended credit over the period of this
analysis. While individual borrowers are
not publicly identified as recipients of
extended credit, the timing of sudden swings

Title 1 o f The Federal Deposit Insurance Corporation Improvement Act of 1991 places limits on Federal
Reserve discount window advances to undercapitalized insured depository institutions. T h e Federal
Reserve can n o w lend to an undercapitalized institution for a maximum of sixty days in any 120-day
period. Discount window loans to undercapitalized
institutions may exceed this limit provided that the
institution's regulator or the chairman of the Federal
Reserve certifies the institution's viability. Lending to
undercapitalized institutions could continue beyond
sixty days without this certification, but the Federal
Reserve would b e exposed to potential losses from
such loans. A similar set of criteria applies to Federal
Reserve loans to critically undercapitalized institutions
that extend for more than five days. These discount
window provisions are effective two years after date
of enactment of the legislation.
8

Volatility here is expressed in terms of standard
deviations. The standard deviation of changes in
extended credit is 90.2, while the standard deviation
of changes in nonborrowed reserves is 2.4.
9

Chart 3

conceivably make it more difficult for the
monetaiy authorities to achieve the desired
level of nonborrowed reserves plus extended
credit, thereby providing the potential for
conflict between the lender-of-last-resort and
monetaiy policy functions. Chart 2 (on
previous page) shows that the Federal
Reserve attempted to offset the provision
of extended credit with reductions in the
supply of nonborrowed reserves. During
those periods of rapid growth in extended
credit, nonborrowed reserves appeared to
shrink. More formal statistical tests are
available to determine if the Federal Reserve
was successful in these defensive open
market operations.

Response of Federal Funds Rate
to Movements in Extended Credit
Basis points
20

-10'

-i
1

r—
23

Months
SOURCE: CITIBASE, Citibank Economic Database.

in extended credit suggests three possibilities: that the big spike in extended credit
in 1984 corresponded to the difficulties
associated with Continental Illinois National
Bank, that the run-up in extended credit
in the late 1980s coincided with banking
difficulties in Texas, and that the greater
use of extended credit in 1990 coincided
with banking difficulties in New England.
Provisions of extended credit add to total
reserves but do not necessarily alter the
course of monetary policy. The Federal
Reserve can adjust its open market operations to offset the provision of extended
credit. Open market operations affect the
supply of nonborrowed reserves and are
undertaken to keep monetary policy on
course. If the amount of extended credit
increases, then the provision of nonborrowed reserves through open market
operations would be reduced correspondingly. However, the volatility of extended
credit over very short periods of time could

The Lender of Last Resort and Monetary
Policy: Is There A Conflict?
The policy variables I analyzed to judge
the effect on monetary policy that arises
from the provision of extended credit are
movements in the federal funds rate and
the narrow monetary aggregate, Ml. 10 This
combination allows me to judge the effect
of the volatility of extended credit on both
interest rates and a particular measure of
the money supply. I chose the federal

Chart 4
Response of M1 Growth
to Movements in Extended Credit

Months
10

M l consists mainly of currency plus c h e c k i n g

accounts.

SOURCE: CITIBASE, Citibank Economic Database.

Chart 5
Forecast Error in Predicting Federal Funds Rate

SOURCE: CITIBASE, Citibank Economic Database.

funds rate because a regime of targeting
borrowed reserves can be viewed as a
variation on a federal funds rate targeting
procedure.11 Although the Federal Reserve
now sets target ranges for the broad monetary aggregate, M2, I chose the narrow
monetary aggregate because this measure
of the money supply tends to be more
closely related to reserve growth than are
the broader monetary aggregates since the
major components of Ml are subject to
reserve requirments.12 In the statistical
analysis, the investigation centers on what
independent effect movements in extended
credit exerted on both the federal funds
rate and Ml. To accomplish this task, it is
necessary to control for other factors that
might have affected these variables. These
other factors include the supply of nonborrowed reserves, economic activity, and
prior movements in the funds rate and Ml
themselves.13
Charts 3 and 4 show what happens
over time to the federal hinds rate and Ml
growth from shocks or innovations in
extended credit, again after accounting for
reserves, economic activity, and previous
movements in the funds rate and in Ml.
Neither the federal funds rate nor the
narrow monetary aggregate appears to be
affected by movements in extended credit.
Chart 3 shows that changes in extended
credit caused, at most, about a four-basis-

point increase in the federal funds rate.
Chart 4 shows that changes in extended
credit led to, at most, less than a one-tenthof-one-percent increase in Ml. These
results indicate that the Federal Reserve
was successful in not allowing its role of
lender of last resort to conflict with its
monetary policy objectives.
Another gauge of the impact of extended
credit on monetary policy can be found in
Charts 5 and 6. These charts show how
much of the error in predicting both the
federal funds rate and Ml results from
shocks in extended credit and in the other
potential influences on these variables. As
evident in Chart 5, only 3 percent of the
prediction error in the federal hinds rate
can be explained by shocks or innovations
in extended credit, while from Chart 6 only
2 percent of the error in forecasting Ml is
accounted for by the volatility in extended
credit. Past innovations in the federal hinds
rate and Ml themselves account for most
of the errors in trying to forecast these
variables. This is often the case with economic variables that are highly correlated
over time, as both of these variables are.

Conclusions
Inherent in a fractional-reserve banking
system is the potential for liquidity problertis

For more on this procedure, see Thornton (1988).

See Gilbert (1992).

More formally, a four-variable vector autoregression
(VAR) is estimated. VARs are atheoretical dynamic
models that use only the observed time series properties of the data to test formally theories that imply
particular behavior of the variables in the model. To
investigate whether defensive open market operations
completely offset issues of extended credit, Butkiewicz
and Lewis ( 1 9 9 1 ) estimate a bivariate VAR model that
includes only measures of reserves. These authors
find that the Federal Reserve does offset its use of
extended credit with defensive open market operations. For more on the VAR model estimated in this
article, see the Appendix, "A Vector Autoregression
Model of Extended Credit."
13

Chart 6
Forecast Error in Predicting M1

SOURCE: CITIBASE, Citibank Economic Database.

to develop at banks and other depository
institutions, if financial difficulties emerge.
The Federal Reserve's role as lender of last
resort represents a vital central bank function needed to ensure the stability of the

banking system. A lender of last resort stands
ready to avert potential panics by providing
reserves to banks experiencing liquidity
pressures. The central bank's role of lender
of last resort, though, could conflict with
its monetary policy objectives. Evidence
presented here suggests that the Federal
Reserve's responsibility for providing emergency liquidity to financial institutions did
not conflict with its role in conducting monetary policy. The Federal Reserve was successful in mitigating die effect on reserves when
providing liquidity to troubled depository
institutions. Neither the federal funds rate
nor the money supply was significantly
affected by shocks to extended credit. These
results suggest that efforts to deal with the
banking difficulties of the past decade proceeded without significantly altering the
course of monetary policy.

Appendix
A Vector Autoregression Model of Extended Credit
To determine whetherthe Federal Reserve's lender-of-last-resort function conflicted
with its monetary policy goals, two vector autoregression (VAR) models were estimated. VAR models are atheoretical models that use only the observed time series
properties of the data to study relationships among different economic variables. The
results in this article are derived from two VAR models, one that captures the effect of
extended credit on the federal funds rate and the other that captures how extended
credit may have affected M1. Because it is necessary to control for other factors that
might affect the federal funds rate and the money supply, each VAR contains four
variables. In addition to including extended credit and the policy variable of interest
(either the federal funds rate, or M1), each model includes a measure of economic
activity, as well as the amount of reserves supplied by the Federal Reserve through its
open market operations. These variables are present because they can also affect the
path of interest rates and the money supply. The VAR models estimated are
(A1)

[ECON CREDIT FEDFUNDS

RESERVES]

and
(A2)

[ECON M1 CREDIT

RESERVES],

where ECOA/isthe Industrial Production Index, CREDITls extended credit, FEDFUNDS
is the federal funds rate, RESERVES is the amount of nonborrowed reserves, and M1
is the narrow monetary aggregate. Monthly data for all variables were obtained from
CITIBASE. 1 Unit root tests indicated that all variables were nonstationary in their levels,
indicating that each series contained one unit root. Tests for the presence of cointegration along the lines suggested in Engle and Yoo (1987) revealed that cointegration is not indicated in the models estimated. Finally, a test for appropriate lag
length, described in Sims (1980) suggested the use of relatively short lag lengths forthe
VARs. Therefore, the VAR models were estimated as
(A1.1)

ECON, = /}„ + tfiECON,,
: m

+i[3,sRESERVESl_l

±PMFEDFUNDSH
/1
=

+ f PM3CREDIT„

+ ev.

1 The use of monthly averages of daily figures for extended credit could overstate the volatility of extended
credit. Under a regime of targeting borrowed reserves, extended credit can be treated like nonborrowed
reserves in determining the needed volume of open market operations. That is, the projected demand for total
reserves minus a targeted level for borrowed reserves equals the sum of nonborrowed reserves plus
extended credit that must be supplied. A better measure of the volatility of extended credit would be the daily
volatility of such credit, particularly late in each reserve maintenance period. Such data are not publicly
available, however.

(Continued on the next page)

Appendix—Continued

A Vector Autoregression Model of Extended Credit
(A1.2)

CREDIT, = a 0 + £ a,CREDIT,:
M
+tcci+aFEDFUNDS,,
FEDFUNDS,

RESERVES,

, + £ 8MECON,
i=i

+£ 8m<RESERVESh
10
=

+ e*.

= q0 + t &RESERVES,+£
/=1

+£ i+9CREDITt_,
i=0

e2r

i=0

= <50 + £ S,FEDFUNDS,
/-1

+£ 8h9CREDIT,,
/0
(A1.4)

al:,ECONri

+£aM,RESERVES+

i=0

(A1.3)

+£
11
=

/=i

+£ ^FEDFUNDS,+
10
=

^ECON,_,

e4(.

and
(A2.1)

ECO/V, = a0 + ta.ECON,
+ t aMM1t_i
;=1

(A2.2)

±aM3CREDIT,
i=0

M1, = bQ + iblM1,-,

i-0

taMRESERVES,_i

, + e 1r

ibi„ECON,i
M

bi+9RESERVES,_i

)

ibM4CRED/T,_i+e2r

CREDIT, = dB + t d.CREDIT,+

i=i

+id,t9RESERVES,,

i-o

(A2.4)

+t
(A2.3)

RESERVES,

r=i

+ id„uM1t_,
;-o

= g0 + igiRESERVESll
/=i

+£ g, aCREDITt,
10
=

+ £ giUM1,,
i=0

dMECONM
+e3t.
+ £g,+4ECO/Vw
/=i

+ e 4( .

Qualitatively identical results were obtained w h e n a maximum of two lags of each
variable were included in the models. Also, since the ordering of the variables in the
VAR can affect the results, different specifications of Models A1 and A2 were
estimated. The results were not sensitive to the ordering of the variables. Charts 3 and
4 are the impulse functions calculated from the VAR models, while Charts 5 and 6 are
the two-year-ahead squared prediction errors, or the variance decompositions,
derived from models A1 and A2.

References
Bagehot, Walter (1873), Lombard Street: A
Description of the Money Market (London:
H.S. King).
Butkiewicz, James L., and Kenneth A. Lewis
(1991), "Bank Bailouts and the Conduct of
Monetary Policy," Southern,
Economic
Journal 58 (October): 501-09.
Engle, Robert F., and Byung Sam Yoo (1987),
"Forecasting and Testing in Co-Integrated
Systems," Journal of Econometrics 35
(May): 143-59.
The Federal Reserve System (1990), The
Federal Reserve Discount Window (Washington, D.C.: Board of Governors of the
Federal Reserve System).
, Federal Reserve Bulletin, various
issues (Washington, D.C.: Board of Governors of the Federal Reserve System).
Garcia, Gillian, and Elizabeth Plautz (1988),
The Federal Reserve: Lender of Last Resort

(Cambridge: Ballinger Publishing Company).
Gilbert, R. Alton (1992), "Loan Growth
Exerts a Strong Influence on M2 and M3 in
1991," Federal Reserve Bank of St. Louis,
Monetary Trends, January.
Humphrey, Thomas M. (1975), "The Classical Concept of the Lender of Last Resort,"
Federal Reserve Bank of Richmond Economic Review, January/February, 2-9.
Sims, Christopher A. (1980), "Macroeconomics and Reality," Econometrica 48 (January): 1-48.
Thornton, Daniel L. (1988), "The BorrowedReserves Operating Procedure: Theory and
Evidence," Federal Reserve Bank of St.
Louis Review 70 (January/February): 30-54.
Thornton, Henry (1807), An Enquiry into
the Nature and Effects of the Paper Credit
of Great Britain (Philadelphia: James Humphreys).

Full text of Financial Industry Studies : August 1992

FRASER