Avaliação de redes neurais aplicadas à previsão de índices de mercados de ações

Abstract

Artificial neural networks have been utilized in modeling solutions for time series forecasting problems arisen in different financial area segments such as financial statements analysis, macro-economic indicators, currency market, stock quotations and market indexes. When dealing with such problems, the forecasting model quality is usually appraised in terms of the difference between the actual value and the network forecasted value. But financial area applications also require that related financial goals such as profitability and low risk exposure be considered. The commonly used mean square error generally does not grant those needs are met. It seems then to be necessary to establish additional criteria considering specific financial goals. The current paper shows the results obtained in experiments carried on to compare different neural network architectures to forecast the São Paulo Stock Exchange Ibovespa index using different training criteria and performance evaluation strategies based on business and financial goals