Analysis of Stock Price Movement Prediction with Financial News on Pre-trained Language Model

抄録

Financial news is undeniably crucial for investment decisions, it is considered an effective way to predict stock movement in the natural language processing (NLP) field as well. Since financial news corpus always comes with very few features and lots of noise, the models must be capable of handling ultra-long texts and being strong in feature extraction. The mainstream NLP patterns are generally based on pre-trained language models (PLMs), but PLMs are not good at processing ultra-long texts, so will the PLMs outperform the traditional statistical models in this task? We built several typical NLP patterns for experimental comparison and proposed a text extraction algorithm to improve the ultra-long text handling problem. According to the results, the PLMs have no significant advantage in analyzing ultra-long news corpus, and the algorithm we proposed can improve the accuracy of the PLMs by about 3%.