Classifying trend movements in the MSCI U.S.A. capital market index—A comparison of regression, arima and neural network methods

This paper describes our initial results in applying neural networks to forecast the MSCI U.S.A. Capital Market Index. The objective is to test the ability of an non-parametric learning network to provide valuable information to a global portfolio manager, who needs to assess investment opportunities in equity markets in order to shape a one month ahead asset allocation. Primarily, the objective is to test the directional classification properties of the method with secondary objectives of higher magnitude prediction and lower RMS error. The system achieved fairly good results on the directional classification criteria as well as the other criteria mentioned both in absolute terms and in comparison with multiple linear regression and two ARIMA models.