Modelagem e aplicação de técnicas de aprendizado de máquina para negociação em alta frequência em bolsa de valores

• Ano de defesa: 2015
• Autor(a) principal: Everton Josué da Silva
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; UFMG
• Palavras-chave em Português: Negociação em alta frequência; Algoritmos de negociação; Aprendizado de máquina; Previsão do mercado de ações; Processo de tomada de decisão; Bolsa de valores; Computação; Modelagem de dados; Aprendizado do computador
• Link de acesso: http://hdl.handle.net/1843/ESBF-9XYGE7

Algorithmic trading has performed an important rule in the electronic stock market. However, these algorithms or strategies without any forecasting capability are not safe to perform trading. In this context, the stock market prediction was always an interesting research topic for the researchers, mainly due to its capacity of making profit on stock trading and/or in order to understand the information originated from the stock markets data. Many machine learning algorithms and statistic models have been proposed by researchers for forecast of stock price and stock price movement. In this works, it was developed and implemented a trading system, that include a trend signal generator based on machine learning techniques and a directional trading strategy, which perform their operations by identifying a short term trend. After developing many experiments with different machine learning techniques, it was opted for developing models using neural networks called Multilayer Perceptron (MLP) and an ensemble model, which combine two MLPs, to predict uptrends. These models act as a support for the trading algorithm proposed by this work. The algorithm uses the model output for taking decisions on performing trading. This works main objective was to model and use machine learning techniques to maximize the directional trading strategys return. Using a massive volume of tick data, it was conducted back-testing and simulation in a realistic simulator of the Sao Paulos stock market. From the empirics results obtained, it was demonstrated that the machine learning techniques were capable of increasing the effectiveness of the decision making process. It was demonstrated that the predictions precision and the results obtained from the realistic simulation are better with the ensemble approach. The achieved results opened new research opportunities: 1) Improving the forecasting models to reduce the false-positive numbers. This reduction directly impacts on the financial results obtained in simulation, because it is going to increase the trading strategy hit rate; 2) Using machine learning techniques in support of other types of high frequency trading strategies.