PAWS Introduces a Policy-Driven Dataset for Financial Agent Simulation

Summary

PAWS is a dataset designed to support historically grounded simulation of how financial and economic policy interventions propagate through public communication, institutions, and stakeholders. It covers 36 verified U.S. policy episodes, 12,727 policy-linked news records, and 65,291 source-grounded stakeholder actions. Each action is connected to supporting news and encoded in an event frame describing its interaction mode, financial-action family and subtype, semantic attributes, and mappings to external taxonomies. The dataset also normalizes entities and aligns actions with daily market-return context, enabling policy-agent simulation replay. In a review of 2,522 stratified action samples, independent AI and human reviewers initially agreed on interaction mode labels 89.4% of the time before adjudication. Case studies of the 2008 short-selling ban and 2001 decimalization recovered documented policy timelines and related market patterns in both dense and sparse news conditions. However, replay experiments showed that high overall accuracy can conceal failures to detect rare stakeholder actions, making action timing and calibration important unresolved challenges. PAWS is intended as an auditable basis for evaluating agent influence, policy-response cascades, and the alignment between actions and outcomes in financial simulations.