Discussant's response to "Neural nets versus logistic regression: A comparison of each model's ability to predict commercial bank failures";

              Discussant's Response to "Neural Nets Versus Logistic Regression: A Comparison of Each Model's Ability to Predict Commercial Bank Failures"
Miklos A. Vasarhelyi
Rutgers University - Newark
Bell, Ribar and Verchio [1990] apply two alternative methodologies in the prediction of commercial bank failure. This discussion first examines the na-ture of the work; second, explores issues in neural network methodology; and third, concludes with the discussion of other relevant issues such as al-ternative approaches and paths for future work.
On the Nature of the Work
The bank failure problem has been extensively explored in the literature of accounting and finance. Consequently, there is a wide body of knowledge about the problem and substantial insight on analytical methods that help in the prediction of failure. This makes it an ideal arena for competitive method-ological testing allowing for comparison, not only among the methodologies in question but also with an external body of literature.
The study uses an extensive sample from the 1983-1988 period for draw-ing failed banks and chooses through a sample estimation procedure. Part of the sample is held out for model testing purposes. The method is quite standard and has been used in many similar studies. Some more recent stud-ies have used the jacknife/bootstrap method in order to avoid having to hold a large part of the data as a holdout sample. This approach could be adopted in this study leading to a different set of basic assertions. Nonfailed banks were chosen through a stratified sampling procedure for group pairing.
The authors used 28 prediction variables for failure prediction, very much in line with the literature. In these types of studies, you should always be con-cerned with two issues: over-fitting and missing variables. This study presents a relatively large sample, thereby decreasing some concerns with the first issue. The variables used are the standard financial variables that appear in most studies. These do not include any potential "soft" causes for failure (e.g. poor management, fraud) and/or macro variables.
In summary, the problem context and approach relate to a large set of studies in the literature and are an ideal setting for evaluating competing sta-tistical methodologies. It might have been desirable that the authors further discuss the literature and the main results. The logit approach has been used extensively in research during the recent years while neural networks are a new and forthcoming area. Consequently, they are to be discussed next
54