Proposta de um sistema de negociação para auxiliar na tomada de decisão no mercado de ações utilizando redes neurais dinâmicas auto-regressivas

Abstract

Since the equity of the companies have been divided into easy and quick trading, that is, with liquidity, there is a great interest of researchers and investors around the world to predict the future behavior of prices in order to obtain recommendations of purchase or sale of such assets with the ultimate objective of achieving profitability or preserving their assets. In this work, the Nonlinear AutoRegressive model (NAR) was used to predict a few steps ahead of the historical series of the Index of the São Paulo Stock Exchange (Ibovespa) in order to embed such predictions into a trading system. Precise prediction was not a primary requirement of the network, but rather the possibility of revealing a short-term (weeks) trend. It was proposed the IMMN index that is calculated on the predictions made by the networks and this parameter is used to identify extreme points (maximum or minimum local) that allowed to support the decisions of purchases and sales. Operating the system in a simulated way on the historical series for determination of annual performances and the total period, the trading system obtained a higher return than the growth of the index in the total period observed for the eight combinations of conceived networks, and on average the trading system presented a return of 26% while the index, only 3%. The work was carried out on the historical value of the index, however, such an effect could be transported roughly to the practical world through the purchase or sale of shares in such proportions as to reproduce the index or through the purchase and sale of a similar investment fund to the index.