sis of the Type I and Type II errors by individual case may have been useful. The au-
thor compared the discriminant model using the underlying ratios (described in the
foregoing) with a discriminant model using the factor scores and found that the per-
centage accuracy of classification was the same in both cases. This is an interesting
result for future researchers.
(b) Briones, Marín, and Cueto (1988). This study presents the results of empirical
research undertaken to build a multivariate model to forecast the possible failure of
financial institutions in Spain and their takeover by the monetary authorities or reg-
ulatory agencies.
During the period 1978–1983, Spain underwent a serious crisis in its financial in-
stitutions. Roughly 47% of all Spanish banks failed during this period; 21.4% of the
equity and 18.7% of the deposits were affected by the problem banks. Banco de Es-
pana (the Spanish equivalent of the Federal Reserve) working through Fondo de
Garantía de Depósitos (the Spanish equivalent of the Federal Deposit Insurance Cor-
poration) carried out the resolution of the banks through “administrative solutions.”
Legal solutions such as bankruptcy procedures were not used for fear of causing a
panic. A bank may thus be technically insolvent when it has a liquidity crisis or it
may be definitively insolvent when there is negative net worth. Since a “failed” in-
stitution can operate indefinitely with assistance from the regulators, the authors have
defined a bank to have failed if there was an intervention by Fondo de Garantía De-
pósitos.
The sample consisted of 25 failed banks and an equal number of nonfailed banks
paired up based on the five-year average size of deposits during the period prior to
intervention. The data sources were Anuario Estadístico la Banca Privada published
by the Consejo Superior Bancario and the memorandum of the Fondo de Garantía de
Depósitos. Both a univariate and multivariate approach were used in classifying the
failed and nonfailed groups.
In the univariate approach, the authors found that the mean values for the ratios
maintain a logical correspondence (the actual mean values obtained are not men-
tioned in the study, however). They also found that standard deviations of the failed
bank ratios tended to be generally higher. Profitability and liquidity measures were
found to be the most significant variables for forecasting failures in a univariate
analysis. The cutoff point for the individual ratio was fixed in a heuristic way, by a
process of trial and error. The costs of Type I and Type II errors were assumed to be
equal.
In the multivariate approach, discriminant analysis was used to develop models
using data of j year prior as the development sample (j = 1, 2, 3, 4, 5) and testing the
model on the data for all the years j. Since the ratios for a bank tend to be correlated
from one year to the next, the classification test on the other years does not constitute
a true out-of-sample (hold out) test. Some of the classification results presented are
nonsensical because if you used data for j = 2 to develop the model, you can not test
it on data of j = 1 because in real time that information would be nonexistent; only
j = 3, 4, and 5 would be!
The multiple discriminant analysis produced three and four variable models for
each year prior, resulting in a total of 10 alternative models to choose from. The com-
parison of the prediction accuracy using univariate analysis and the discriminant
analysis showed that univariate analysis actually did better than the discriminant
function in the first and the fifth year (Exhibit 10.5)—a surprising result. Most re-
10 • 22 BUSINESS FAILURE CLASSIFICATION MODELS