Previsão de preços de ações no período intradiário por meio de focused time lagged feedforward networks

Abstract

Stock price prediction is a subject of great interest for both market agents and scientific and academic community. At the same time, this problem is considered to be one of the most challenging in time series forecasting, due to its highly dynamic nature. A large amount of researches have proposed to address the issue. Some of them, with very promising results, adopt the Focused Time Lagged FeedForward Network (FTLFN), a type of Artificial Neural Network (ANN) that offers memory mechanisms capable of detecting temporal patterns. In many cases, however, the neural network’s capacities are not properly explored, being limited to tests with a minimum set of parameters. Besides, most of the studies on stock price prediction focus on low-frequency periods, such as days or months. On the other hand, due to the ease of access to information in the last decades and the automation of trades in stock market, these are getting more oftenly executed over short-term horizons, like hours, minutes or seconds. Therefore, there is a need to expand the knowledge related to forecasts in this scenario. With that in mind, this research has the objective of investigating the FTLFN’s potential on stock price forecasting over the intraday period. Its short-term memory and hidden layer size are widely and de eply explored, so the impact of different configurations on the accuracy results could be measured. Also, Technical Analysis indicators are built and utilized as input signals to the network, with their possible contributions to stock prediction being verified. From a general perspective, the work proposes the extention of the understanding regarding the FTLFN’s forecasting capabilities, as well as its use with intraday financial time series, which still require further exploration in literature. The obtained results show that, as human investors do, also FLTFNs are capable of taking enormous advantage from input signals’ history on providing better prediction quality within the proposed context. The same cannot be said for the supporting Technical Analysis indicators chosen, since they mostly increase forecasting errors. Evidences are presented based on the experimentation over several sets, bringing robustness to the conclusions and allowing the methodology and the results to serve as base for future researches related to predictions on high-frequency trading scenarios.