Multimodal deep learning for international investment arbitration outcome prediction and bilateral investment...

International investment arbitration has expanded at a remarkable pace over the past two decades, generating pressing demand for robust outcome prediction tools that can guide strategic decisions. This study presents a multimodal deep learning framework that fuses textual, numerical, and visual data to predict arbitration outcomes in the investor-state dispute settlement context. Our attention-based fusion architecture channels legal documents, macroeconomic indicators, and visual evidence through dedicated encoders capable of capturing intricate cross-modal dependencies that shape tribunal reasoning. Evaluated on 1,247 arbitration cases drawn from major international institutions, the multimodal model attains an overall accuracy of 86.7%, surpassing single-modality counterparts by 7.8% points and conventional machine learning baselines by 14.6% points. Feature importance analysis reveals that the quality of legal argumentation, dispute monetary value, and arbitrator panel composition rank among the most decisive determinants of outcomes. Beyond their technical value, these findings equip investors, host states, and legal counsel with evidence-based tools for strategic planning, while simultaneously foregrounding normative questions about fairness, transparency, and equitable access to predictive technologies in dispute resolution.