The application of artificial neural networks in the prediction of earnings

Abstract

In the past three decades, earnings has been the most researched variable in accounting. Empirical research provided substantial evidence on the usefulness of earnings in stock market valuation and capital market research. Unfortunately, there has been little or weak empirical evidence in predicting earnings. From the empirical evidence provided thus far, it is debatable whether a number of underlying econometric and statistical issues were fully resolved, mainly because of the nature of accounting and financial data. Linear and logistic regression approaches were applied in the past for providing prediction models, but their predicting ability was not robust. In an attempt to improve the prediction ability and robustness of the prediction models and thus avoid the inherent econometric and statistical problems in prior research, a different approach was employed. This paper examines the application of artificial neural networks (ANNs) in the prediction of future earnings. The variable to be predicted was a dichotomous realization of the change in earnings per share, adjusted for the drift in the prior earnings changes, and the predictor variables which were used had been identified, in this study, to have the highest information content for the prediction. The multi-layer perceptron (MLP) feedforward neural network architecture was used, due to its suitability as a classifier and its implementation simplicity on a sequential computer. In contrast with prior applications of ANNs in accounting, and business in general, this study also focused in the selection of an efficient and robust training algorithm. The complexity and the size of the problem, in combination with the scattered nature of pooled accounting data, demanded that a training algorithm should guarantee convergence without oscillations and be relatively fast in order to be used in such a problem. Therefore, the conjugate gradient training algorithm, originally adapted for the efficient training of ANNs by Charalambous (1992), was employed. The algorithm considers the training of the network as a multi-variable function minimization problem, employing a line search in each iteration. The logistic regression method was also examined in parallel with the ANN approach, to compare the performance of the two methods in a problem which cannot be fully solved, since it is proved that earnings follow a random walk model. Comparing the performance of the two models examined, it was found that the ANN approach performed slightly better than the logit approach. Moreover, different accounting variables were identified as earnings predictors. Future research should investigate the usefulness of other accounting information to further improve the predictability of earnings. >