Temporal Movement Prediction Based on Tweet Texts and Historical Price via Adversarial Training

Abstract

Temporal movement prediction is essential to finance and investing. The early works for stock movement prediction are mainly based on traditional machine learning algorithms. They have achieved good results, while these conventional machine learning methods require costly hand-craft engineering for massive features. Recently, the prevalence of deep learning has effectively promoted the ability to solve stock movement prediction, which has been widely studied in academia and industrial fields. Although deep-learning-based methods have achieved certain progress, there are two problems limiting performance improvement for the temporal movement prediction task. Firstly, current procedures mainly apply temporal factors like trading volume, market value, etc. They may ignore the influence of news and temporal sentiment information on temporal movement. Secondly, even if there exist researches of temporal movement prediction based on financial news and temporal sentiments, those methods would be probably overfitting. Thirdly, they are vulnerable to inherent stochasticity (e.g. black swan events) in the temporal scenario. To address these problems, this paper puts forward a temporal movement prediction model based on adversarial training. Specifically, this research employs Fast Gradient Method (FGM) algorithm during the training process, which can effectively avoid overfitting, deal with stochasticity, and thus improve the model performance in generalization and accuracy. The finally reported experimental results on a tweet text dataset and a historical price dataset demonstrate that our method outperforms baseline models and achieves competitive performance compared with some current advances.