The stock market is an important segment of the economy that circulates a large volume of assets. Several factors may affect the stock market transactions, leading to fluctuations in the stock values that may pose a problem for those who seek to forecast future stock values and maximize their profits. This issue is more serious when the stock values present the concept drift phenomenon, which means that the stock value’s patterns change over time. In this work, we aimed to evaluate whether machine learning-based predictors that incorporate mechanisms to deal with concept drift are suitable for stock market forecasting. To do so, a historic database of stock prices of 10 companies, negotiated in the Brazilian stock exchange and collected over 20 years, was used. We compared the performance of predictors based on different paradigms, with and without mechanisms to deal with concept drift, and the results showed that, although the strategies that handle concept drift demand longer computational times, they also tend to present smaller prediction errors. The highlight was the EOS-D approach, which had the best performance in 6 of the 10 stocks analyzed considering one-to-one comparisons.

股票市场是经济的重要组成部分，流通着大量资产。有几个因素可能会影响股票市场交易，导致股票价值波动，这可能会给那些寻求预测未来股票价值并最大化利润的人带来问题。当股票价值出现概念漂移现象时，这个问题就更加严重，这意味着股票价值的模式随着时间的推移而发生变化。在这项工作中，我们旨在评估基于机器学习的预测器是否适合股票市场预测，该预测器包含了处理概念漂移的机制。为此，我们使用了 20 多年来在巴西证券交易所协商收集的 10 家公司股票价格的历史数据库。我们比较了基于不同范式的预测器的性能，无论是否有处理概念漂移的机制，结果表明，尽管处理概念漂移的策略需要更长的计算时间，但它们也往往会出现更小的预测误差。亮点是 EOS-D 方法，考虑到一对一比较，该方法在所分析的 10 只股票中的 6 只中表现最佳。