A new research paper published on ArXiv introduces PAWS (Policy-driven Agentic World Simulation), a dataset designed to enhance multi-agent financial simulations. PAWS addresses a critical gap in current models: the lack of connection between policy interventions, public communication, and historically aligned evidence.
The Architecture of PAWS
The dataset encompasses 36 verified U.S. financial and economic policy episodes, supported by 12,727 policy-linked news records. It details 65,291 stakeholder actions, each grounded in source material and represented by a multi-layer event frame. These frames capture interaction modes, financial-action subtypes, and semantic attributes, while resolving entities to normalized organizations.
Validation and Historical Fidelity
To ensure data integrity, independent AI and human reviewers evaluated 2,522 stratified samples, achieving an initial agreement rate of 89.4%. The utility of PAWS is demonstrated through case studies of significant economic events, including the 2008 short-selling ban and the 2001 decimalization. These studies successfully recovered documented policy timelines and associated market patterns in both data-rich and data-sparse environments.
Identified Challenges
A replay study conducted using the dataset highlighted that high overall accuracy can sometimes mask a failure to detect rare but critical stakeholder actions. The researchers identify action timing and calibration as the primary hurdles in simulating policy-response cascades. PAWS aims to provide an auditable substrate for evaluating how agent influence and action-outcome alignment function within historically grounded financial contexts.
