This study proposes a prediction model based on a dynamic spatiotemporal causal graph neural network to address the challenges of complex dynamic dependencies, strong structural correlations, and ambiguous causal relationships in corporate revenue forecasting. The model constructs a time-varying enterprise association graph, where enterprises are represented as nodes and industry or supply chain relationships as edges. A graph convolutional network is used to extract structural dependency features, while a gated recurrent unit captures temporal evolution patterns, achieving joint modeling of structural and temporal features. On this basis, a causal reasoning mechanism is introduced to model and adjust potential influence paths among enterprises. A learnable causal weight matrix is used to describe the strength of economic transmission, suppress spurious correlations, and strengthen key causal paths. The model also employs multi-scale temporal aggregation and attention fusion mechanisms to dynamically integrate multidimensional information, enhancing adaptability to both long-term trends and short-term fluctuations. Experimental results show that the proposed model outperforms mainstream methods in multiple metrics, including MSE, MAE, MAPE, and RMAE, verifying its effectiveness in capturing corporate revenue dynamics, modeling economic causal dependencies, and improving prediction accuracy. This study establishes a unified framework that integrates spatiotemporal dependency modeling with causal structure reasoning, providing new insights and methodological foundations for intelligent forecasting in complex economic systems.

In corporate finance, deploying AI agents for financial modeling requires integrating diverse data streams into a reliable and auditable decision-making framework. Existing multimodal and large language models provide strong representations but often miss clear temporal ordering, cross-modal coordination, and adaptive rule enforcement in real-world financial practice. To address these gaps, this paper introduces HAFIN (Hierarchical Adaptive Financial Intelligence Network), a rule-aware framework built upon the Typhoon2-70B foundation. HAFIN combines multi-scale temporal modeling, cross-modal fusion, sparse expert routing, soft rule propagation, Bayesian uncertainty modeling, and selective attention in a single architecture. The model captures fine and coarse financial patterns through time-aware embedding and aligns data types through simple gated fusion. Sparse expert routing improves focus and efficiency. Soft rule propagation incorporates regulatory limits into learning. Bayesian inference refines uncertainty handling, and selective attention highlights salient signals. The result is a robust, interpretable, and compliance-driven system designed for high-stakes multimodal financial analysis, directly supporting the needs of financial modeling and finance transformation leaders.

This study proposes a dynamic graph enhancement framework designed for uncertainty conditions to address structural uncertainty, temporal nonlinearity, and multi-scale dependency in corporate financial health risk modeling, where accurately capturing the dynamic evolution of financial states under uncertainty has become a core scientific problem in the digital transformation of financial intelligence and corporate governance. The method starts from multi-source financial data and first performs multi-scale decomposition and normalization to achieve a hierarchical representation of financial time-series signals. Based on this, a dynamically updated enterprise correlation graph is constructed to capture potential risk transmission relationships among firms and to monitor industrial systematic risk. An uncertainty correction mechanism is then introduced to adaptively enhance the adjacency matrix, effectively mitigating modeling bias caused by data noise and structural perturbations. The main model adopts a time-aware graph convolutional network (T-GCN) to jointly learn temporal dependencies and graph topological features, combined with a contrastive learning constraint to strengthen feature consistency and the robustness of risk representation. Experimental results show that the proposed model significantly outperforms mainstream methods such as Transformer, GNN, GAT, and BiLSTM on multiple corporate financial datasets, achieving an average accuracy of 0.921 and improving the F1-Score by about 3.1%. These results verify its robustness and high-precision risk identification capability under dynamic uncertainty conditions. Overall, this study achieves dynamic modeling and structure-aware optimization of corporate financial health, providing an effective technical approach for financial risk monitoring, health assessment, and scenario analysis for financial forecasting and planning in complex economic environments.