Applications of machine learning and sentiment analysis in financial forecasting

Abstract

Machine learning has gained popularity due to its ability to apply statistical techniques to large datasets, enabling complex analyses and progressively improving results. In the financial sector, it is used to enhance business performance in areas such as market studies, marketing, sales, and process automation, although there are still areas in finance that remain unexplored by these techniques. The main objective of this thesis is to investigate the capabilities of machine learning techniques and sentiment analysis in the field of finance and apply them to financial forecasting. Events such as dividend announcements create market situations that can be exploited to obtain abnormal returns. Specifically, the aim is to analyze to what extent daily data obtained from various news sources helps predict market reactions to different corporate events. Additionally, this doctoral thesis seeks to fill the existing gap in the field by creating an expert decision system that recommends the most suitable currency hedging strategy for the user. Chapter 1 raises the question of whether a dividend announcement has the capacity to generate a market signal that leads to changes in stock returns the following day, thereby generating abnormal returns. A sentiment analysis performed with ChatGPT is used to identify the tone of the news and link it to abnormal returns. The sample includes 394 companies listed in the S&P 500 index, from which 1,574 dividend announcements and 7,222 news articles were collected during the years 2022–2023. The study concludes that sentiment plays a key role in identifying market sentiment immediately after dividend announcements. Chapter 2 delves deeper into the impact of sentiment on the market following dividend announcements and investigates abnormal returns on an intraday scale. For this study, 4,682 news articles were collected from 1,258 dividend announcements made by 394 companies listed in the S&P 500 index, covering the period from January 2023 to January 2024. A logistic regression model is used to predict the trend of abnormal returns. This chapter concludes that the sentiment of financial news has a significant and immediate impact on abnormal returns, with a diminishing effect over time. Additionally, the profitability of the investment strategy using financial news sentiment outperforms the benchmark strategy of investing in all stocks. Finally, Chapter 3 focuses on minimizing losses caused by exchange rate fluctuations. This study uses 20 years of daily euro to US dollar exchange rate data, from 2002 to 2022, comprising 5,044 historical records. The study employs the random forest machine learning algorithm and a set of technical indicators to predict future exchange rate fluctuations. The predictions are then used to create a recommendation system that advises users on hedging possibilities. First, the random forest model achieves 79% accuracy in predicting the next day's exchange rate trend. Second, it develops an expert decision system that helps companies reduce costs related to managing foreign exchange exposure. This thesis contributes to the existing literature on the applications of machine learning in finance. The results demonstrate that machine learning methods can be successfully implemented in the financial domain.