Analysis of Small-Business Financial Statements Using Neural Nets

Abstract

This paper provides a tutorial in the use of artificial neural networks (ANNs) in accounting and finance in the form of a well-documented exploratory application of an ANN with automatic learning to the analysis and interpretation of the noisy data contained in the financial statements of a sample of small businesses. The application uses the learning algorithm of the Boltzmann machine (BM), which employs the technique of stochastic optimization called simulated annealing. Various experiments are run using balanced training sets and (un)balanced test sets. Since most assessment errors are boundary errors and the ratios of successful classifications for the test sets are clearly above chance for the balanced training/test sets, the BM with two- and three-class output separation appears to have practical value. As expected, the BM demonstrates better accuracy as the size of the training sample increases in a balanced fashion, and less accuracy as the test set is augmented in an unbalanced manner.