PAWS
A policy-driven multi-agent financial world simulation dataset covering U.S. financial and economic policy events. It provides historically grounded, auditable data for evaluating agent influence, policy-response cascades, and alignment between agent behaviors and outcomes.
Timeline (last 90 days)
PAWS (Policy-driven Agentic World Simulation) covers 36 verified U.S. financial and economic policy events.
PAWS is associated with 12,727 policy-related news records and 65,291 stakeholder behaviors with source support.
Every behavior in PAWS is linked to the news evidence supporting it, rather than being generated from nothing.
PAWS characterizes each behavior through a multi-layer event frame, capturing four dimensions: interaction patterns, financial behavior families and subtypes, semantic attributes, and conditional mappings.
PAWS is validated on 2,522 stratified-sampled behavior samples, with independent AI reviewers and human reviewers achieving an initial agreement rate of 89.4% on interaction pattern judgments.
PAWS validates its policy timeline reconstruction capability through two case studies: the 2008 short-selling ban and the 2001 stock price decimalization reform.
PAWS can reconstruct documented policy timelines and corresponding market patterns in both information-dense and information-sparse scenarios.
PAWS's replay study reveals that high accuracy rates may mask a model's failure to detect rare stakeholder behaviors.
The PAWS study points out that action timing and calibration are the core challenges in policy agent simulation.
PAWS performs normalized entity resolution, mapping actors to standardized organizational entities and aligning actions with the context of daily market returns.
All Facts (10)
PAWS (Policy-driven Agentic World Simulation) covers 36 verified U.S. financial and economic policy events.
50%UnverifiedPAWS is associated with 12,727 policy-related news records and 65,291 stakeholder behaviors with source support.
50%UnverifiedEvery behavior in PAWS is linked to the news evidence supporting it, rather than being generated from nothing.
50%UnverifiedPAWS characterizes each behavior through a multi-layer event frame, capturing four dimensions: interaction patterns, financial behavior families and subtypes, semantic attributes, and conditional mappings.
50%UnverifiedPAWS is validated on 2,522 stratified-sampled behavior samples, with independent AI reviewers and human reviewers achieving an initial agreement rate of 89.4% on interaction pattern judgments.
50%UnverifiedPAWS validates its policy timeline reconstruction capability through two case studies: the 2008 short-selling ban and the 2001 stock price decimalization reform.
50%UnverifiedPAWS can reconstruct documented policy timelines and corresponding market patterns in both information-dense and information-sparse scenarios.
50%UnverifiedPAWS's replay study reveals that high accuracy rates may mask a model's failure to detect rare stakeholder behaviors.
50%UnverifiedThe PAWS study points out that action timing and calibration are the core challenges in policy agent simulation.
50%UnverifiedPAWS performs normalized entity resolution, mapping actors to standardized organizational entities and aligning actions with the context of daily market returns.
50%