This paper examines the different responses of local and foreign investors to host government risk perceptions in the context of extreme events. We develop COVID-19 attention indices that capture attention related to COVID-19 according to China Central Television (CCTV) news program and further construct the government’s risk perception (GRPC) measure about COVID-19. Given the cross-listed AH-shares in China, we find that GRPC caused the extreme movement of stock markets by applying the multi-quantile VaR Granger causality approach. The results show that the reaction of cross-listed stocks in the A-share market is more inflexible than that in the H-share market during the outbreak period of the pandemic, foreign investors follow GRPC as a weather vane than local investors, and both types of investors are more concerned about the pessimism of GRPC. In the period of epidemic normalization, local and foreign investors prefer the optimistic attitude conveyed by the Chinese government.

We use supervised machine learning to develop a Chinese language financial sentiment dictionary from 3.1 million financial news articles. Our dictionary maps semantically similar words to a subset of human-expert generated financial sentiment words. In article-level validation tests, our dictionary scores the sentiment of articles consistently with a human reading of full articles. In return validation tests, our dictionary outperforms and subsumes previous Chinese financial sentiment dictionaries such as direct translations of Loughran and McDonald’s (2011) financial words. We also generate a list of politically-related positive words that is unique to China; this list has a weaker association with returns than does the list of otherwise positive words. We demonstrate that state media exhibits a sentiment bias by using more politically-related positive and fewer negative words, and this bias renders state media’s sentiment less return-informative. Our findings demonstrate that dictionary-based sentiment analysis exhibits strong language and domain specificity.