Robust neural modeling for the cross-sectional analysis of accounting information

The performance of robust artificial neural network models in learning bivariate relationships between accounting magnitudes is assessed in this paper. Predictive performances of a number of modeling paradigms (namely, linear models, log–linear structures, classical ratios and artificial neural networks) are compared with regard to the problem of modeling a number of the most outstanding accounting ratio relations. We conduct a large scale analysis, carried out on a representative Spanish data base. Several model fitting criteria are used for each model class (namely, least squares, weighted least squares, least absolute deviations (LAD), and weighted LAD regressions). Hence, besides the standard (least squares-based) version of each model we test a robust (LAD) counterpart, in principle more adequate for distributions strongly affected by outliers, as typically appear in accounting ratio modeling. Our results strongly suggest that classical ratio models, although much used in practical applications, appear to be largely inadequate for predictive purposes, with linear models (both in their least squares and LAD variants) providing much more adequate specifications. In a number of cases, the linear specification is improved by considering flexible non-linear structures. Neural networks, because of their model-free regression capabilities, let us capture generic non-linearities of unknown form in the modeled relations, as well as providing – when properly trained – robust tools for modeling and prediction of general relationships.