Enhancing stock trend prediction through financial news analysis

Abstract

Nowadays, precise stock trend prediction is of paramount importance in financial markets, as it aids investors and traders in making informed decisions and maximizing their profits. Financial news has emerged as a valuable and influential source of information for stock trend prediction due to its wide coverage of economic events, corporate announcements, and market developments, providing crucial insights into the factors impacting stock prices. In this context, various representation techniques such as TF-IDF, bag-of-words, word embeddings have been explored to effectively process financial news data. On the other hand, deep learning predictive models have shown promising results in capturing intricate patterns and relationships within the data.

Despite the potential benefits of using financial news in stock trend prediction, several challenges exist. Current challenges include the uncertainty and noise present in news data, the difficulty of interpreting the predictions made by deep learning models, and the need to improve prediction accuracy. Throughout this thesis, we address these challenges and propose novel techniques to enhance the accuracy, interpretability, and robustness of stock trend prediction through financial news analysis.

We first introduce a multi-instance learning framework to address the uncertainty of financial news and adaptively forecast stock trends. Our proposed MIL framework involves constructing news vectors and inferring the probability of each instance for each trend. The probability inferences of each instance aid in mitigating the unknown mixture of daily news. Additionally, to tackle the random news occurrences within each trading day, we aggregate all inferred instance probabilities to arrive at the overall prediction.

Next, we propose a method that integrates causal inference with deep learning to enhance the interpretability of stock trend prediction models. Causal inference enables a better understanding of the cause-and-effect relationships between financial news and stock trends. By identifying the causal factors that drive specific trends, our proposed model provides causality-based explanations for predictions and leads to more explainable and robust predictions.

As recent advancements of large language models (LLMs) show growing potential in many applications, we further explore the application of LLMs, particularly ChatGPT, in stock prediction using financial news data. To effectively harness the power of ChatGPT, we propose a novel method to filter out noise from the massive financial news streams. Our approach not only leverages the strengths of LLMs in understanding textual data but also addresses the challenges posed by noisy and irrelevant news contents.

Our experimental results on real-world financial datasets demonstrate the effectiveness of our proposed approaches in improving stock trend prediction accuracy and efficiency. We highlight the potential of financial news as a valuable resource for enhancing stock trend prediction models and underscore its significance in improving decision-making processes for investors and financial analysts.