Redes neurais recorrentes e expoente de Lyapunov aplicados a séries temporais financeiras

Abstract

The study of financial market asset pricing is considered one of the most relevant subjects in this segment. The possibility to obtain profit during the intraday oscillations turned the predictability problem relevant to this arear. Although economists, academics and industry professionals have different views about the possibility of predicting prices, some recent studies consider that there is some degree of the predictability in financial series. This evidence considers the prices series as a chaotic and non-linear system where, at least, short term predictability may takes place. In this way, Machine Learning and Deep Learning methods have been used together as decision support systems to forecast the movement direction of prices. In this work, the problem of predicting the next minute is addressed from the classification perspective. Five assets of the Brazilian Stock Exchange were chosen based on the ticks’ liquidity in the period. A three-class supervision method was applied and ten technical indicators were used as attributes for the assets chosen. Two data sets were constructed: one using the continuous value of the indicators, and other using discretized values. Using Support Vector Machine, Random Forest, MLP and LSTM three methods were applied to the classification process. The first method compares the results between the continuous and the discretized data sets. Using the classifier that presented the best result, the second experiment added four attributes using the Lyapunov exponent calculation. The objective of this experiment was to investigate if the attributes could contribute to improve the classification of all assets. Finally, a final experiment used the calculation of the maximum Lyapunov exponent as a training control for the classifier. The objective, in this case, was to exclude portions of the series where the value of the Lyapunov exponent indicated chaos. The result showed that although indicator discretization contributed positively for most classifiers, LSTM networks showed a significant improvement in continuous data sets with superior performance in most cases. In general, it was concluded that each asset benefited from a set of specific methods. The experiments demonstrated that there is no general method that presents an optimal performance for all assets in the defined period studied.