Sentiment-Enhanced Stock Price Prediction: A Novel Ensemble Model Approach
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Stock Price Prediction, Sentiment Analysis, Finance, NLP, BERT
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: This paper evaluates existing stock prediction methods, uses BERT for sentiment analysis, and introduces the FB-GAN model, showing improved stock price prediction compared to other models, except for GRU and headline-summary combined
Abstract: Stock price prediction remains a formidable challenge within the realm of financial markets, wherein a multitude of models and methodologies have been under exploration to prognosticate the dynamic behaviour of equities. This research endeavour encompasses an exhaustive examination of extant stock prediction systems, entailing a meticulous assessment of their merits and demerits, concurrently pinpointing discernible lacunae and avenues for enhancement. Subsequently, we harnessed the capabilities of BERT, an exemplar in the domain of natural language processing, to conduct sentiment analysis across a heterogeneous corpus of news articles pertinent to the subject stocks. Additionally, an ancillary sub-experiment was conducted to ascertain the relative impact of three distinct categories of news articles, namely headlines, summaries, and a composite amalgamation of the two, on the efficacy of stock price prediction.
The outcome of this investigative pursuit was the generation of sentiment scores for each trading date, which were subsequently integrated as input features in the training of a neural network. Through a comparative analysis of various neural network models, including but not limited to RNN, LSTM, GAN, and WGAN-GP, we discerned that the WGAN-GP model exhibited the most favourable predictive performance. Building upon these findings, we introduced the FB-GAN model, an ensemble architecture comprising WGAN-GP, which capitalizes on the fusion of historical stock price data and market sentiment scores for enhanced stock price prediction.
Subsequently, a comprehensive evaluation of our approach was undertaken vis-à-vis established models, gauging its performance against five prominent equities, namely Amazon, Apple, Microsoft, Nvidia, and Adobe. In summation, this research makes a compelling case for the integration of BERT-based sentiment analysis within the ambit of stock price prediction. Our initial hypothesis regarding the significant influence of market sentiment on stock price prediction was validated, and our proposed FB-GAN model outperformed all other models. Furthermore, incorporating both the headline and summary of the news article contributed to enhanced stock price prediction compared to utilizing either the headline or summary in isolation.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 8284
Loading