Evaluation of neural network models to predict the sign of the variation of the IPSA

Authors

  • Antonino Parisi F. Universidad de Chile

Abstract

This paper analyzes the capacity of the neural networks to forecast the sign of weekly variations of IPSA. Several architectures of neural networks were used over the time period between January 11th of 1999 to October 22th of 2001, being the Recursive Ward Network the one with the best performance, reaching an outsample predictive capacity of 72% and an outsample accumulate yield for the IPSA portfolio of a 24.42%. The Recurrent Recursive Jordan & Elman Network achieved a forecast ability of 64% and a return of 21.33%; while the AR(1,1) model obtained a return of 18.31%, higher than the Ward Standard Network and Recursive MLP returns. Even though the first one had not statistical evidence of predictive capacity it would allow to conclude that a higher predictive capacity do not always implies higher yields. The Pesaman & Timmermann test (1992) gives evidence that the Recursive Ward Network and the Recurrent Recursive Jordan & Elman Network, in his recursive and standard version can forecast the index directional change for the Chilean case. Also, this models could produce higher returns than AR (1,1) model. This result supports the conclusions of Leung, Daouk & Chen (2000) about the prediction of movement directions can give greater capital gains that the forecasting of close values.

Keywords:

Test Directional Accuracy, Neural Networks, Multilayer perceptron, Predictive Capacity