【报告摘要】The application of large language models has driven the development of sentiment analysis. Common natural language processing methods, such as LDA and BERT, can extract topics from large amounts of text data but cannot distinguish the importance of these extracted topics to the research question. This limitation can be addressed by using ChatGPT. The powerful text comprehension ability of ChatGPT enables it to extract meaningful topics from texts and determine whether these topics are important to the research question. Through topic modeling of the text data from the exchange rate markets validates the superior performance of ChatGPT compared to LDA and BERT. What's more, it is found that ChatGPT can identify topics associated with significant events and assign high importance scores to these topics. This paper further leverages the information on topic importance to construct a novel interval-valued sentiment index (TIS index). The TIS index not only incorporates the relative importance of various events that impact exchange rate fluctuations but also captures the continuous variation of market sentiment within an interval. Empirical results demonstrate that the TIS index can significantly improve the forecasting accuracy for exchange rates across various interval-based models.

【报告人简介】魏云捷，中国科学院数学与系统科学研究院副研究员，入选国家高层次人才特殊支持计划青年拔尖人才。研究方向：经济预测理论与方法、经济政策分析。获成思危优秀科研成果奖、第一届中国科技青年论坛二等奖等奖项。兼任中国系统工程学会理事、北京青年科技人才协会理事。