Stock Price Prediction Using Sentiment Analysis

Abstract

We investigate the influence of financial news headline sentiment on the predictability of stock prices using Long Term Short Term Memory (LSTM) networks. The investigation is performed on intraday data with specific lag-times between published article headlines and realised stock prices. FinBERT, a natural language processing model which is fine-tuned specifically for financial news is used to perform sentiment analysis on the company related news headlines. Two base models, one with only historical stock price data as inputs and the other with both historical stock price data and sentiment data from the original BERT model is tested. An alternative model with have both historical stock price data and sentiment data from the fine tuned FinBERT model as additional features. A comparison is performed on both the base and alternative models using Root Mean Square Error (RMSE) and mean absolute error (MAE) as performance metrics. The results suggest that the use of news headline sentiment features from FinBERT significantly improve the predictive performance of LSTM networks in intraday stock price prediction. FinBERT features are also found to outperform features based BERT model trained on a general corpus, illustrating the positive effect of domain specific fine tuning for Large Language models.