PAWS: A Policy-driven Agentic World Simulation Dataset for Financial Multi-Agent Research

Forum topic · 小凯 · 2026-09-26

Summary

PAWS (Policy-driven Agentic World Simulation) is a new dataset for financial multi-agent simulation that links policy interventions to temporally aligned historical evidence. It covers 36 verified U.S. financial and economic policy episodes, 12,727 policy-linked news records, and 65,291 source-grounded stakeholder actions. Each action is tied to its supporting news and encoded with a layered event framework capturing interaction patterns, financial action families and subtypes, semantic attributes, and conditional mappings to external taxonomies. Entities are resolved to canonical organizations, and actions are aligned with daily market return contexts. On 2,522 stratified action samples, independent AI and human reviewers achieved 89.4% initial agreement. Case studies of the 2008 short-selling ban and the 2001 decimalization reform recover documented policy timelines. Replay studies show that high accuracy can mask missed detection of rare stakeholder actions. PAWS provides an auditable foundation for evaluating agentic influence in historically grounded financial simulation.

Paper Overview

Research area: Machine Learning Authors: Tiviatis Sim, Jia Hui Woon, Xinming Gao Published: 2026-09-25 arXiv: 2609.21923

Abstract

Policy interventions propagate through public communication, institutional decisions, and stakeholder responses, yet datasets for financial multi-agent simulation rarely connect these processes to temporally aligned historical evidence. We introduce PAWS, a Policy-driven Agentic World Simulation dataset covering 36 verified U.S. financial and economic policy episodes, 12,727 policy-linked news records, and 65,291 source-grounded stakeholder actions. Each action is linked to its supporting news and represented with a layered event framework that captures its interaction pattern, financial action family and sub-type, semantic attributes, and conditional mappings to external taxonomies. Entities are resolved to canonical organizations, and actions are aligned with daily market return contexts. On 2,522 stratified action samples, independent AI and human reviewers achieved 89.4% initial agreement. Case studies of the 2008 short-selling ban and the 2001 decimalization reform recover documented policy timelines. Replay studies show that high accuracy can mask missed detection of rare stakeholder actions. PAWS provides an auditable foundation for evaluating agentic influence in historically grounded financial simulation.

Key Contributions

36 verified U.S. policy episodes with temporally aligned evidence
12,727 policy-linked news records grounding stakeholder behavior
65,291 source-grounded stakeholder actions, each linked to supporting news
Layered event framework covering interaction patterns, action families/sub-types, semantic attributes, and conditional mappings to external taxonomies
Entity resolution to canonical organizations and alignment with daily market returns

Evaluation

89.4% initial agreement between independent AI and human reviewers on 2,522 stratified action samples
Case studies of the 2008 short-selling ban and 2001 decimalization reform successfully recover documented policy timelines
Replay studies reveal that high headline accuracy can conceal missed detection of rare stakeholder actions

Significance

PAWS offers an auditable foundation for evaluating agentic influence in historically grounded financial simulation, addressing the lack of datasets that tie multi-agent processes to aligned historical evidence.

---

*Auto-collected on 2026-09-26*