Several studies have demonstrated a superiority of neural networks as models for financial diagnosis. It has been proposed that the ability of these models to represent intermediate abstractions is a main reason for their superior performance. It has also been proposed that the represented intermediate abstractions resemble the diagnostic concepts used by skilled diagnosticians, and thus, that they have cognitive relevance. In this paper, we investigate these propositions applying an experimental methodology to obtain valid data sets of financial diagnoses. A multilayered perceptron model is developed and validated using cross validated performance measures and analyses of error term distributions and internal representations. The hidden units of the connectionist model represent intermediate abstractions explaining the model's superior performance, but the cognitive relevance of these intermediate abstractions is not obvious.
展开▼