TwinMarket Financial AI Simulation

A multi-agent financial market simulator for behavioral finance, emergent risk, and policy stress testing.

Financial and social simulation
TwinMarket Multi-agent simulation Financial AI Behavioral finance Policy sandbox
GitHub Project site Paper
TwinMarket BDI architecture

TwinMarket uses LLM-powered virtual investors to simulate financial markets from the bottom up. Instead of treating the market as an equation, it models agents with beliefs, desires, intentions, information, social signals, and trading behavior.

The project received Best Paper recognition at the ICLR 2025 Financial AI Workshop and is connected to the lab's broader agenda on agentic economies and economic world models.

Research Storyline

Agents
Start from investor cognition

Each virtual investor carries beliefs, desires, intentions, information access, and social influence rather than acting as a fixed equation.

Market
Let aggregate patterns emerge

Prices, volatility, herding, and bubbles become outcomes of many interacting agents rather than manually scripted market curves.

Policy
Turn the market into a sandbox

Researchers can perturb information flow, agent populations, or regulation and observe counterfactual market behavior.

EconWM
Connect to agentic economies

TwinMarket is a concrete financial prototype for the broader Economic World Models agenda: agents, environments, co-evolution, and real-world alignment.

What The Simulator Studies

Micro-to-macro emergence

Individual investor choices, overconfidence, herding, information delay, and social influence can aggregate into market-wide patterns and instability.

Stylized market facts

The simulator tests whether agent societies can reproduce empirical regularities that appear in real markets but are hard to capture with simple rules.

Risk and policy stress tests

Because the simulated world is controllable, researchers can perturb information flow, regulation, and agent composition to study counterfactual outcomes.

TwinMarket market simulation output

Display Figures

TwinMarket BDI design
The BDI architecture makes each trading decision traceable to agent beliefs, desires, intentions, and market signals.
TwinMarket bubble behavior
Market-level phenomena emerge from interacting virtual investors rather than fixed aggregate equations.
Economic World Model architecture
TwinMarket is one concrete sandbox inside the larger Economic World Models architecture.

Paper Trail

System
TwinMarket: A Scalable Behavioral and Social Simulation for Financial Markets

Builds the market simulator, investor agents, and experimental setting for financial social simulation.

Repository
Blueprint
From Economic Agents to Agentic Economies

Generalizes the TwinMarket-style sandbox into a wider architecture for economic digital twins and agentic economies.

Economic World Models

Why It Matters

  • Financial systems are social systems; agentic simulation makes individual behavior visible instead of hiding it inside aggregate curves.
  • It offers a practical bridge from LLM agents to social science experiments, financial risk analysis, and economic digital twins.
  • It gives the Economic World Models agenda a concrete, high-impact market sandbox.