Integrating Spatio-Temporal Information for Forecasting Financial Indicators of Listed Companies

The forecasting of financial indicators stands as a significant task in computational finance. It plays a crucial role in understanding a company's future operational conditions, providing investors with references for investment decisions. Existing methods predominantly focus on the independent historical financial data of listed companies without capturing the complex interactions among companies. This paper proposes an Adjustable Spatio-Temporal Forecasting Network (ASTFN), which integrates the historical financial features with the topological features of the cross-shareholding network. Focusing on predicting multivariate earnings and risk indicators in listed companies' financial data, ASTFN achieves optimal performance compared to all baseline models. Our proposed model reveals the crucial role of leveraging network relationships in financial indicator forecasting. By adjusting a parameter that balances temporal and spatial features, ASTFN can enhance the overall forecasting performance. Additionally, we analyze the impacts of network directionality and find that changing the direction of information passing from unidirectional to bidirectional improves the predictive accuracy of earnings indicators, but probably degrades the performance of risk indicators. It reveals the distinct propagation patterns for earnings and risk within the network.