Modelling Stock Returns With Neural Networks

Neural networks have attracted much interest in financial engineering but many multivariate data series remain diflcult to model. In this paper we use a non trivial problem in expsure analysis of share prices to multiple factors to explore the interrelationships among the numerous network and data engineering parameters and we highlight the importance of a careful choice of the indicators used as network inputs. We show how data pre-processing can improve generalisation performance by up to 30.5% and present a "time-sensitive" cost function, designed to take into account gradually changing input-output relationships. We give empirical evidence that when it is combined with the right leaMags in the indicators generalisation can be further improved by up to IO. 1 %.