Building a Shapley FinBERTopic System to Interpret Topics and Articles Affecting Stock Prices

Financial sentiment analysis can help investors make more efficient decisions. Finan-cial sentiment analysis involves analyzing text data to extract sentiments and assists market participants in making informed investment decisions. In recent years, there have been many efforts to improve analysis accuracy using machine learning and deep learning. Particularly FinBERT, which is pre-trained on financial datasets and excels in analyzing financial documents. However, the output may require more information to make decisions. It has the disadvantage of being difficult to use for critical decision-making because of the need for help explaining the output results. This study applies the BERTopic and SHAP to financial sentiment analysis by FinBERT. It proposes a Shapley FinBERTopic system, which contributes to better investor decision-making by explicitly showing the impact of news article topics and individual words on stock prices. The results of the experiments with this system showed an increased interpretability of the results compared to using FinBERT alone. A more detailed analysis will be conducted in the future.