MSFFN: Multi-scale feature fusion network for cross-border M&A success prediction based on global acquisition networks

Cross-border mergers and acquisitions (M&A) are crucial strategic instruments for enterprise internationalization, and their success prediction holds significant value for investors, managers, and policymakers. However, existing research predominantly focuses on single-scale features, neglecting the complexity that M&A success is jointly influenced by multi-level factors including national institutional environment, industry competitive landscape, and enterprise microeconomic capabilities. This study proposes a Multi-Scale Feature Fusion Network (MSFFN) that constructs a three-layer heterogeneous network architecture of country-industry-company, utilizing Graph Convolutional Networks, Graph Attention Networks, and GraphSAGE to extract features at respective scales with hierarchical attention mechanisms for adaptive fusion. Based on 92,788 M&A transactions of 119,263 enterprises from the Crunchbase database spanning 1990-2023 across 198 countries and 43 industries, experimental results demonstrate that MSFFN significantly outperforms 11 baseline methods, achieving 89.34% accuracy and 0.9241 AUC-ROC. Ablation experiments confirm that country-level, industry-level, and company-level features contribute 32.4%, 28.7%, and 25.3% of predictive capability respectively, with cross-scale interaction adding 13.6%. Centrality analysis identifies New York, London, and San Francisco as global M&A hubs. This research extends graph neural network applications in financial prediction and provides new perspectives for understanding global M&A network structures.