Frameworks for trading stocks in the financial market using data mining techniques and multi-swarm optimization algorithms for dynamic and multimodal enviroments

Abstract

Financial time series behave similarly to a data stream, that is, a set of input elements that arrive continuously and sequentially over time. So, a time series may present concept drifts, which is the change in the data generating process. This phenomenon negatively affects forecasting methods that rely on observing the past behavior of the series to predict future values. Many papers report the use of data mining techniques and computational intelligence to predict the future direction of stock prices, uncovering patterns in time series data to support decision making for financial market operations. The traditional optimization algorithms proposed in the literature generally assume that the environment is static, assuming that the time series data generation distribution is the same over the period of interest. Another problem is that, sometimes these methods do not take into account the possibility that the function to be optimized has multiple peaks and, in this case, is represented by multimodal functions. However, multimodality is one of the known features of real-time financial time series optimization problems. Furthermore, several methods involving optimization algorithms have been proposed in the literature, however most of them do not consider real world problems. The main contribution of this work is a decision support system capable of dealing with concept changes and multimodality in the financial time series environment. To achieve this goal, we propose two modelos that aim to find patterns in financial time series, using multi-swarms to improve particle initialization, thus avoiding local optimum in the final optimization phase. In addition, the models use a validation step with the early stopping criteria to avoid overfitting. In contrast to the first proposed model, the second one considers two consecutive generations of populations to detect changes in time series, then a statistical test is used to check for changes in the environment to avoid false positives. Once a change is detected, the second model performs a series of actions to find new patterns, replacing obsolete ones. The patterns discovered by the models are used in conjunction with proposed investment rules to support decisions and help investors maximize profit on their stock market operations. Experiments using 82 stocks from the S&P100 index were tested with a confidence level of 95%, showing that the proposed method is able to improve results when concept drifts are considered.