Corporate social responsibility (CSR) research has recently begun to focus on the CSR performance of business groups, with the scope shifting from group members to business groups in general. This paper focuses on whether business groups with centralized decision rights tend to disclose more CSR information and investigates the heterogeneous effect of the number of subsidiaries. Using a dataset for listed groups in China from 2010 to 2020, our empirical test discovered that centralized decision rights could promote group CSR disclosure. For groups with many subsidiaries, centralization makes a more significant contribution to promoting CSR disclosure. The mechanism test revealed that this positive relationship between centralization and disclosure relies on efficient internal capital market allocation, a reduction in rent-seeking behavior of subsidiaries, and reputational concerns. Furthermore, we observed that the centralized decision rights influence on disclosure varies across different aspects of CSR, with a negative impact on “Social Contribution” and a positive impact on “Shareholder Responsibility”, “Employee Responsibility”, “Supplier, Customer, and Consumer Responsibility” and “Environmental Responsibility”. Centralized decision rights promote more CSR disclosures with voluntary disclosures, while regulatory disclosures have no significant effect. We research the allocation of decision rights and group CSR disclosure.

China’s Shanghai-Hong Kong Stock Connect and Shenzhen-Hong Kong Stock Connect programs make it possible for investors to trade stocks within specified limits through the two stock exchanges. The A-H share exchange stock market is crucial to the opening of the Mainland market, but few studies have paid attention to the market risks of such stocks. Using deep learning and BP neural network algorithm, this study constructs a three-dimensional A-H share interconnection market risk prediction index system including stock price fundamental indicators, technical indicators, and macro indicators based on the CES300 Index. Taking the CES300 Index return as the output layer indicator, a BP neural network with a 21-10-1 structure is constructed, and the tan-sigmoid transfer function and the LM optimization algorithm training function are used for network training to predict the return of the A-H share interconnected stock market. The mean square error (MSE) converges to 10−6, and the goodness of fit R reaches 0.9928 and validates the prediction accuracy of the BP neural network model. It provides an efficient and accurate risk prediction model for the A-H share interconnected market, which facilitates the interactive development of the Mainland and Hong Kong markets.