Discovering the Optimal Set of Ratios to Use in Accounting-Based Models

Ratios are the prime tool of financial analysis. In predictive modelling tasks, however, the use of ratios raises difficulties, the most obvious being that, in a multivariate setting, there is no guarantee that the collection of ratios eventually selected as predictors will be optimal in any sense. Using, as starting-point, a formal characterization of cross-sectional accounting numbers, the paper shows how the Multilayer Perceptron can be trained to create internal representations which are an optimal set of ratios for a given modelling task. Experiments suggest that, when such ratios are utilized as predictors in well-known modelling tasks, performance improves on that reported by the extant literature