Résumé
Accurate stock market prediction is of paramount importance for traders. Professional ones typically derive financial market decision-making from fundamental and technical indicators. However, stock markets are very often influenced by external human factors, like sentiment information that can be contained in online social networks. As a result, micro-blogs are more and more exploited to predict prices and traded volumes of stocks in financial markets. Nevertheless, it has been shown that a large volume of the content shared on micro-blogs is published by malicious entities, especially spambots. In this paper, we introduce a novel deep learning-based approach for financial time series forecasting based on social media. Through the Generative Adversarial Network (GAN) model, we gauge the impact of malicious tweets, posted by spambots, on financial markets, mainly the closing price. We compute the performance of the proposed approach using real-world data of stock prices and tweets related to the Facebook Inc company. Carried out experiments show that the proposed approach outperforms the two baselines, LSTM, and SVR, using different evaluation metrics. In addition, the obtained results prove that spambot tweets potentially grasp investors' attention and induce the decision to buy and sell.