Evaluating Predictors of Abnormal Stock Returns Multivariate Models

A.J. Hoffman (S. Africa)


Neural modelling, stock return prediction, fundamental indicators, technical indicators


In recent years much research effort has been spent on the development of statistical and neural network techniques to predict abnormal stock returns. Much previous work focused on time series prediction of stock returns, comparing conventional techniques (ARMA or ARMAX) with the capabilities of neural or other artificial intelligence techniques. This paper models cross-sectional difference between all stocks, using a combination of fundamental and technical parameters found to be indicative of expected future returns. The most common basis for comparing different models is a combination of an error measure (e.g. MSE between the actual and predicted future returns), and the correlation coefficient between actual and predicted returns. This paper uses a different measure to assess the usefulness of the cross-sectional model, based on the returns of categories of stocks sorted according to the value of the predicted return. This measure is shown to be superior to the more conventional measures, as the models producing best returns over a period of comparison are not always the models displaying the largest correlations with actual returns. It is furthermore shown that the best models incorporating a number of fundamental and technical parameters perform better than any of the individual predictors, both in terms of average returns as well as in terms of consistency over different market conditions.

Important Links:

Go Back