Stochastic Sentiment Analysis for Trading Signals Based on Financial News Headlines

This research project focuses on leveraging natural language processing (NLP) techniques to predict stock price movements based on Reuters News. The methodology involves data collection and preprocessing, feature engineering through tokenization, and the utilization of Bayesian Convolutional Neural Networks trained using Stochastic Gradient Langevin Dynamics. Historical stock prices are crawled and normalized returns are calculated to enhance the predictive model. The project aims to build a robust connection between Bayesian DNN and stock price prediction by incorporating the sentiment analysis of news headlines. Thinning models are employed for predicting future news sentiments. The implementation is carried out using Python 3, PyTorch, numpy, NLTK, and various crawler tools. The research demonstrates the application of the proposed methodology with backtesting on past news, showcasing promising results. The evaluation of the model reveals a noteworthy classification accuracy of 58.07%, surpassing the critical 50% threshold, positioning it as a valuable tool for market participants. Future work is outlined for further refinement, including the exploration of industry-specific sentiments and the enhancement of stopword lists to improve prediction accuracy.