Trading strategies optimization using a Genetic Algorithm under the Directional Changes paradigm

This thesis is rooted in the Directional Changes (DC) paradigm, focusing on exploring its efficacy in developing profitable trading strategies. In traditional practice, market prices are typically sampled at fixed time intervals to construct physical time data. A method rooted in trading decisions based on this type of data, specifically Technical Analysis (TA), is constrained by the information generated through the selection of these fixed intervals, such as daily or hourly. However, the DC paradigm is an event-driven approach that distinguishes itself from the traditional physical time. It offers a complementary approach for extracting information from data. In the DC paradigm, price movements are recorded when specific events occur, instead of employing fixed intervals. The determination of these events relies on a threshold value, represented as θ, which determines which changes in value qualify as significant and which should be neglected, according to the trader. In this thesis, we begin by introducing our DC-based trading strategies. In the formulation of these strategies, we leveraged two main components of the DC paradigm, namely, scaling laws and indicators. We evaluated the performance of each strategy by employing a single θ. In order to enhance the performance of trading strategies, we proposed a method that utilizes a Genetic Algorithm (GA) to optimize these strategies. As part of our experimental validation, the results of the method were compared against each trading strategy to determine whether there was an increase in performance. Additionally, we added widely adopted TA strategies from the finance field for comparison, which rely on physical time. In these comparisons, highly used performance metrics were utilized. Results indicate that, when applied to a single θ, certain strategies demonstrated profitability, while others did not. Notably, the method we introduced exhibited superior performance in comparison to both individual strategies and conventional TA strategies. Following this, to investigate whether exposing each strategy to different DC-profiled data generated by various thresholds can enhance the performance, we conducted experiments for each strategy using multiple thresholds. Then, we assessed how various DC profiled data contributed to performance improvements by evaluating each strategy’s performance relative to the previous stage. The results at this stage show that using multiple θ improved the performance of certain strategies compared to testing with a single θ. At the final stage, we performed a more fine-grained optimization via GA, which simultaneously employed these strategies with distinct DC-profiled data, each characterized by varying θ. For the final experimental validation, we compared the performances of the previous two stages with this one. In doing so, we again included widely adopted TA strategies from the finance field, which rely on physical time. The results in this final stage demonstrated that the method, which combines multiple strategies and thresholds, not only improved the performance metrics from the previous two stages but also outperformed the TA strategies.