| Resumo: |
The digitalization of the nancial market has led to the emergence of automated trading strategies: the trading robots. While it is possible to use just one algorithm to perform nancial operations, it has become common to create investment portfolios with several trading strategies. When performing market operations, it is possible to distribute the assets among the automated strategies in an arbitrary manner. Di erent con gurations of these distributions, however, may lead to completely distinct levels of returns. Given a set of assets, the optimal distribution problem may be studied from a multi-objective perspective, as several market indices can be used to evaluate the strategy portfolio. A set of automated trading strategies in capital markets may be combined into a portfolio, aiming to maximize returns and minimize losses. The best combination for the portfolio requires assigning optimal weights to each strategy, considering di erent indicators from nancial market. In this work, the application of evolutionary algorithms based on a lexicographical approach and based on NSGA-II are proposed to optimize a portfolio of automated strategies applied to the Brazilian futures market. The experiments study several nancial indicators, with di erent rankings, as well as optimization and time period conditions, using historical data from mini futures contracts of Ibovespa and U.S. Dollar index. Experiments were performed in order to adjust the hyperparameters (e.g. initial population, crossover rate), evaluating the impact of the chosen objective functions and the time-frame window size, as well as the accumulated capital over the period. After the objective function experiments, the group of functions that optimized the Sortino ratio had superior accumulated capital in both evolutionary algorithms. In the experiments with varying time-frame window sizes, the \Highest Return" and the \Nearest Extreme Objective Values" NSGA solutions produced the highest average returns and also the highest average accumulated capital in all scenarios. Moreover, all evaluated solutions outperformed both the IPCA and the Selic benchmarks. Short In-Sample periods managed to reduce risk and also raised the return-to-risk ratio in Out-of-Sample time-frame windows. Longer Out-of-Sample periods, however, were able to raise pro tability levels and the accumulated capital across the entire time series. |
|---|
