Financial ratios are routinely used as predictors in modelling tasks where accounting information is required. The purpose of this paper is to discuss such use, showing how to improve the effectiveness of ratio-based models.
First, the paper exposes the inadequacies of ratios when used as multivariate predictors. It then develops a theoretical foundation and methodology to build accounting-based models. Experiments then verify that this methodology outperforms the conventional methodology.
From plausible assumptions about the cross-sectional behaviour of accounting data, the paper shows that the effect of size, which ratios remove, can also be removed by modelling algorithms, which facilitates the discovery of meaningful predictors and leads to markedly more effective models.
The paper covers cross-sectional modelling only, accounting identities and other interactions between line items are ignored, the methodology is especially appropriate in tasks where the effectiveness of the model is seen as a valued quality.
The need to select ratios among many alternatives is avoided, models become more accurate and robust, less biased and less likely to generate missing values, model construction is less arbitrary.
The paper provides a solid foundation for accounting-based modelling, by developing a whole new methodology that can end the uncritical use of modelling remedies currently prevailing and release the full relevance of accounting information when utilised to support investments and other value-bearing decisions.
