Datasets Comparison

We use Chinese A-share listed companies from 2007 to 2021 as the initial sample. Since the available sentiment-related data began in 2007, we select 2007 as the first year of our sample. We exclude observations from the financial industry, those with missing variables, and those labeled as special treatment (ST), and finally obtain 11729 firm-year observations. Our data is obtained from the China Stock Market Accounting Research (CSMAR) database and the Chinese Research Data Services (CNRDS) database. To mitigate the impact of outliers, all continuous variables are winsorized at the 1% and 99% levels.

Finance

Creative Commons Attribution 4.0 International

We use Chinese A-share listed companies from 2009 to 2025 as our initial sample. Because ESG-related data first became available in 2009, we take 2009 as the first year of the sample period. We exclude observations in the financial industry, those with missing variables, and those labelled as special treatment, yielding a final sample of 45,713 firm-year observations. The data are obtained from the China Stock Market and Accounting Research database, the Wind database, and the Chinese Research Data Services (CNRDS) database. To mitigate the influence of outliers, all continuous variables are winsorised at the 1% and 99% levels. We measure corporate ESG performance using the Huazheng ESG rating, published by Sino-Securities Index Information Service (Shanghai) Co., Ltd., an independent third-party rating agency. We apply a Chinese financial sentiment model to the textual content of the MD&A. The model is pre-trained on Chinese text (bert-base-chinese) and fine-tuned on a labelled Chinese financial-news sentiment corpus (the publicly available hw2942/bert-base-chinese-finetuning-financial-news-sentiment-v2 model). Because we restrict attention to the MD&A, standardised content such as greetings and the company profile, which appear in other sections of the annual report, is excluded by construction; within the MD&A, we remove non-textual elements such as tables, figures, and numerical exhibits and segment the narrative into sentences. The model then tokenises each sentence at the character level using the WordPiece vocabulary of bert-base-chinese and classifies it as positive, negative, or neutral. Because classification is performed at the sentence level by a Transformer encoder, the effects of negation, degree adverbs, and other context-dependent cues are captured implicitly through the model's self-attention over the entire sentence rather than through hand-coded rules. Moreover, because the model is trained on Chinese financial text, it accommodates finance- and industry-specific terminology that general-purpose or English-based lexicons do not. Based on these sentence-level classifications, we construct the managerial sentiment index (MS) as the proportion of positive sentences minus the proportion of negative sentences in the MD&A, with higher values indicating more positive managerial sentiment.

Finance, Corporate Governance, Corporate Social Responsibility

Creative Commons Attribution 4.0 International