AI based stock market trending analysis

Abstract

There is an abundance of factors that can affect the value of the stock, making the price movement dynamic, non-static, and usually non-stationary. Applications of machine learning and deep-learning algorithms towards stock price forecasting have been explored extensively. In general, the trend of a stock's price is determined by the perception (i.e. sentiments) of the public towards it. Due to the wealth of data in our digital era, there have also been attempts to use sentiment analysis on news data to improve the performance of stock price forecasting. However, the techniques employed for these attempts are usually not state-of-the-art. This project aims to implement and use Bidirectional Encoder Representation from Transformer (BERT) model- which achieved state-of-the-art results for sentiment analysis in 2018- alongside a suitable stock price forecasting model to analyse if the inclusion of news sentiments will improve the stock price forecasting performance. BERT was implemented and evaluated using accuracy, precision, recall, and f1 score against 5 other baseline models for multi-class sentiment analysis (i.e. positive, negative, neutral). BERT achieved the best evaluation results of 0.957, 0.931, 0.964, and 0.947 for accuracy, precision, recall, and f1 score respectively. For stock price forecasting, a long-short-term-memory model was chosen based on its stock price forecasting performance among 4 other baseline models. The evaluation of the final model-comprising BERT and a multivariate LSTM- shows a small improvement in evaluation results for stock price forecasting when incorporating news sentiment acquired through BERT as compared to without sentiments.