search using multivariate methods appears to come to the opposite conclusion be-
cause it is believed that the interaction or the substitution effects of one variable with
others provide better information than if the variables are considered sequentially.
The authors conclude that there is a close balance between the univariate ratio ap-
proach and the function approach and that both types of analysis can be viewed as
complementary.
More rigorous testing using a holdout sample will be needed to confirm that uni-
variate approach has predictive power comparable to the multivariate approach.
Coming to this conclusion based solely on original sample test results is premature
because of sampling bias in the results.
10.10 ITALY
(a) Altman, Marco, and Varetto (1994). This study presents the results of two inter-
esting innovations in the diagnosis of corporate financial distress. The first is the use
of a two-stage decision process employing two discriminant analysis models to fine-
tune the process used to grade companies into groups of healthy, vulnerable, and un-
sound companies. The second innovation is the application of neural networks (NN)
to solve the same problem. The study is also of interest because of the access of the
authors to a large and well-developed database of financial information on over 37,000
companies in Italy, as much as to the pooling of this data by a consortium of banks
that have thereupon been able to use the diagnostic system developed for medium- and
small-sized businesses in Italy. After trying various alternative approaches in neural
network modeling, the authors conclude that the linear discriminant model compares
well relative to neural networks. The main advantages of the discriminant model are
its consistency of performance and the modest cost in fine-tuning the model. Having
said that, the authors state that neural networks continue to hold promise especially in
situations where the complexity of the problem can be handled well by the flexibility
of NN systems and the capacity to structure them into simple, integrated families.
The study was carried out in the Centrale dei Bilanci (CB) in Turin, Italy. CB is
an organization established by the Banca d’Italia, the Associazione Bancaria Italiana
and over forty leading banks and special credit institutions in Italy. CB develops and
distributes tools for the member banks to use. One product was a linear discriminant
analysis-based model that is used in practice to improve credit analyst productivity
by pre-selecting the credits and for monitoring the uniformity of the judgments made
about businesses by the various branches of the bank.
The first part of the study describes the results of the new release of the system
10.10 ITALY 10 • 23Years Ratios Functions
1 90/95% 80/85%
2 75/80 80/85
3 75/80 75/80
4 75/80 75/80
5 80/85 75/80Exhibit 10.5. Overall Accurate Predictions—Comparison of Single Ratios with Dis-
criminant Functions.