Leveraging State-of-the-Art Topic Modeling for News Impact Analysis on Financial Markets: A Comparative Study

News impact analysis has become a common task conducted by finance researchers, which involves reading and selecting news articles based on themes and sentiments, pairing news events and relevant stocks, and measuring the impact of selected news on stock prices. To facilitate more efficient news selection, topic modeling can be applied to generate topics out of a large number of news documents. However, there is very limited existing literature comparing topic models in the context of finance-related news impact analysis. In this paper, we compare three state-of-the-art topic models, namely Latent Dirichlet allocation (LDA), Top2Vec, and BERTopic, in a defined scenario of news impact analysis on financial markets, where 38,240 news articles with an average length of 590 words are analyzed. A service-oriented framework for news impact analysis called "News Impact Analysis" (NIA) is advocated to leverage multiple topic models and provide an automated and seamless news impact analysis process for finance researchers. Experimental results have shown that BERTopic performed best in this scenario, with minimal data preprocessing, the highest coherence score, the best interpretability, and reasonable computing time. In addition, a finance researcher was able to conduct the entire news impact analysis process, which validated the feasibility and usability of the NIA framework.