Sistema de gestão de risco em operações automatizadas na bolsa de valores baseado em aprendizado profundo

• Ano de defesa: 2022
• Autor(a) principal: Johnathas Araújo de Carvalho
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; Brasil; ICX - DEPARTAMENTO DE CIÊNCIA DA COMPUTAÇÃO; Programa de Pós-Graduação em Ciência da Computação; UFMG
• Palavras-chave em Português: Gestão de risco; Algotrading; Aprendizado profundo; Computação – Teses; Aprendizado do computador – Teses; Gestão de riscos – Teses; Algoritmos de computador – Teses
• Link de acesso: http://hdl.handle.net/1843/54641

Machine learning and deep learning algorithms are being used by several studies to predict future values in financial series or the direction of assets price movements, and proved to outperform the traditional econometric models. Nevertheless, there are not yet working papers that conduct systematical analysis of historical predictions in order to establish a risk management system (RMS) for automated trades (algotrading). This work employ deep learning algorithms, specifically the LSTM and Bi-LSTM networks, to predict daily trading ranges of assets (high and low prices). A statistical and systematical analysis of the predictions allow to estimate the hit probability, risk/gain ratio, as well to measure more adequate exposure sizes. The range predictions were applied to filter suggested operations of seven automated trading strategies based on moving average and oscillatores (RSI and MACD). Backtestings were performed upon 5 most negotiated assets in B3 between 2008 and 2019 (ABEV3, B3SA3, ITUB4, PETR4 and USIM5). The results pointed that, despite of the predictions uncertainty of the models, the appliance of RMS outperformed the strategies in terms of net return in 64% of the cases and demonstrated average gain of 0,31% in the net return of the strategies. Those strategies that made use of risk management system and had the best performance for each tested asset also had a net return over allocated capital greater than the baseline defined as buy-and-hold strategy.