Source-linked AI summary
EconAgent: Large Language Model-Empowered Agents for Simulating Macroeconomic Activities
Nian Li, Chen Gao, Mingyu Li, Yong Li, Qingmin Liao
TL;DR
Existing macroeconomic agent models struggle to customize heterogeneous decision mechanisms and to incorporate past experiences, market dynamics, and multifaceted factors. EconAgent introduces an LLM-empowered simulation framework with perception and memory modules, and experiments report more realistic decisions and more reasonable classic macroeconomic phenomena than rule-based or learning-based agents.
Problem
Existing agent models make customizing heterogeneous decision mechanisms difficult and often overlook past periods, market dynamics, and multifaceted macroeconomic factors.
Method
EconAgent combines a macroeconomic environment for work and consumption decisions with perception-based heterogeneous agents and a memory module for reflecting on prior experiences and market dynamics.
Results
EconAgent reproduces classic macroeconomic phenomena more reasonably than traditional rule-based or learning-based agents, while exhibiting human-like decision patterns and rapid environmental adaptation.
Takeaways & Limitations
The results support LLMs as a promising avenue for simulating more realistic macroeconomics and studying macroeconomic phenomena and regularities.
Takeaways & Limitations
The current system simulates household labor and consumption behavior but does not incorporate more complex agents such as firms or their pricing and hiring behaviors.
Abstract
from arXiv · showhide
The advent of artificial intelligence has led to a growing emphasis on data-driven modeling in macroeconomics, with agent-based modeling (ABM) emerging as a prominent bottom-up simulation paradigm. In ABM, agents (e.g., households, firms) interact within a macroeconomic environment, collectively generating market dynamics. Existing agent modeling typically employs predetermined rules or learning-based neural networks for decision-making. However, customizing each agent presents significant challenges, complicating the modeling of agent heterogeneity. Additionally, the influence of multi-period market dynamics and multifaceted macroeconomic factors are often overlooked in decision-making processes. In this work, we introduce EconAgent, a large language model-empowered agent with human-like characteristics for macroeconomic simulation. We first construct a simulation environment that incorporates various market dynamics driven by agents' decisions regarding work and consumption. Through the perception module, we create heterogeneous agents with distinct decision-making mechanisms. Furthermore, we model the impact of macroeconomic trends using a memory module, which allows agents to reflect on past individual experiences and market dynamics. Simulation experiments show that EconAgent can make realistic decisions, leading to more reasonable macroeconomic phenomena compared to existing rule-based or learning-based agents. Our codes are released at https://github.com/tsinghua-fib-lab/ACL24-EconAgent.
1 Introduction
EconAgent addresses limitations of rule-based and learning-based macroeconomic agents by using LLM-enabled perception and memory to model heterogeneous decisions and past market dynamics. Experiments report more realistic decisions and more reasonable classic macroeconomic phenomena.
- Motivation: Existing ABM approaches use predetermined rules or learned models, but customizing decision mechanisms for heterogeneous agents remains difficult.These models also typically emphasize current-period situations, limiting treatment of past periods and multifaceted macroeconomic factors.
- Approach: EconAgent uses an LLM-empowered agent with a perception module that automatically produces heterogeneous agents with distinct decision-making mechanisms.The simulation environment includes labor and consumption market dynamics driven by work and consumption decisions, alongside fiscal and monetary policies.
- Approach: A memory module lets agents reflect on past individual experiences and market dynamics when making decisions.This directly targets the multi-period dynamics that existing models often overlook.
- Results: EconAgent reproduces classic phenomena such as consumption-market inflation and labor-market unemployment more reasonably than traditional rule-based or learning-based agents.The experiments also report human-like decision patterns and rapid adaptation to internal and external environmental changes.
- Significance: The work integrates LLMs into macroeconomic simulation and reports the emergence of classic macroeconomic phenomena and regularities.The authors present this integration as a research avenue for studying macroeconomics.
2 Framework Overview
EconAgent provides a macroeconomic simulation environment in which heterogeneous agents make work and consumption decisions across interconnected labor, consumption, financial, and government-taxation components. The framework models market adjustments, redistribution, and macroeconomic feedback while acknowledging limits on tax coverage and firm simulation.
- Framework Overview: The simulation includes labor, consumption, financial markets, and government taxation, with agents deciding whether to work and how much to consume.These decisions influence fiscal revenues and broader market dynamics.
- Framework Overview: Conventional simulations use limited predetermined factors and representative rules, restricting flexibility in modeling heterogeneous decision-making mechanisms.The framework identifies heterogeneous decisions and multifaceted economic influences as central challenges.
- Government Taxation: Government taxation uses progressive 2018 U.S. federal brackets, and collected revenue is evenly redistributed among agents before savings are updated.The simulation considers income tax only and excludes other taxes such as value-added tax.
- Productivity and Consumption: Working agents contribute monthly productivity to essential-goods production, while consumption is constrained by available goods inventory and agents’ resources.Demand is met only when supply is sufficient, and consumption continues until every agent has consumed once.
- Market Dynamics: Labor and goods markets adjust to supply-demand imbalances: shortages raise wages to stimulate production, and higher labor costs increase goods prices.The wage and price adjustments are bounded by maximum rates of change.
- Financial Market: The bank adjusts interest rates annually using inflation and unemployment, with rates defined from annual goods prices and monthly observations.The framework adopts a Taylor rule and uses inflation and unemployment adaptation coefficients.
3 EconAgent
EconAgent combines heterogeneous agent profiles, prompt-based economic perception, and memory of recent market dynamics to support sequential macroeconomic decisions. Its modules incorporate economic factors and past experiences into adaptive work and consumption behavior.
- Perception Module: Real-world agent profiles support heterogeneous household behavior in the simulated economy.Profiles include names, ages, jobs, wages, and tax-related characteristics aligned with 2018 U.S. data.
- Economic Environment: Dynamic prompts use changing offers, wages, and shortages to expose agents to evolving economic conditions.An unemployment prompt can increase an agent’s inclination to work in the subsequent month.
- Memory Module: The memory module maintains 2L + 1 conversations covering the previous L months and quarterly reflections on economic phenomena.Reflections address labor, consumption, and financial markets and their dynamics.
- Memory Module: After reflection, agents adjust strategies in response to past market dynamics and future uncertainties, modeling macroeconomic trends through experience.The memory module connects prior experiences with subsequent adaptive decisions.
- Action Module: Prompts incorporate living costs, future aspirations, economic trends, income, and savings into work and consumption decisions.The LLM returns work and consumption propensities in the range [0, 1].
4 Experiments
Experiments compare EconAgent with rule-based and learning-based agents across macroeconomic indicators, regularities, decision-making, heterogeneity, and external shocks. EconAgent produces more plausible aggregate dynamics and human-like adaptive behavior, while perception and reflection materially shape simulation outcomes.
- Experimental Setup: EconAgent is evaluated against representative rule-based and learning-based agents across macroeconomic indicators and regularities.The experiments examine four research questions covering comparative behavior, module effects, interpretability, and external intervention.
- Macroeconomic Emergence: EconAgent produces more stable and numerically plausible indicators than the baselines.The baselines generate anomalous indicators and large fluctuations, while EconAgent produces more reasonable inflation, unemployment, nominal GDP, and growth-rate dynamics.
- Macroeconomic Regularities: EconAgent correctly manifests both the Phillips Curve and Okun’s Law, with correlations of -0.619 and -0.918, respectively.The rule-based baseline instead displays an incorrect positive Phillips-Curve relationship.
- Ablation Study: Removing perception makes unemployment and inflation fluctuations appear too stable, while removing reflection produces inflation anomalies near 15% initially.The results indicate reduced sensitivity without perception and insufficient long-term economic-environment awareness without reflection.
- Decision-Making and External Intervention: EconAgent exhibits heterogeneous, economically interpretable decisions and qualitatively reproduces COVID-19’s unemployment surge.Consumption propensity increases with age, high unemployment significantly lowers consumption propensity, and simulated unemployment remains elevated after 2021 without government intervention.
5 Related Work
Related work positions ABM as a bottom-up macroeconomic simulation paradigm and identifies limitations in existing agent models. EconAgent extends this line of work with reasoning and planning abilities for multi-step, multi-agent economic simulations.
- Agent-Based Modeling: ABM models macroeconomic activity through interactions among diverse agents without assuming a predetermined economic equilibrium.Compared with empirical statistical and DSGE models, ABM supports nonlinear behaviors and policy simulation.
- EconAgent: EconAgent is introduced with reasoning and planning abilities for simulating macroeconomic activities.This contribution applies LLM-empowered agents to macroeconomic simulation.
- LLM-Empowered Agents: LLM-empowered agents offer autonomous adaptive reactions, human-like planning, and interaction with other agents or humans.These capabilities motivate their use as a basis for simulation agents.
- Research Gap: Existing LLM-agent studies generally examine one-step or few-step behavior for a few agents rather than multi-step behavior in multi-agent environments.The paper identifies multi-step, multi-agent simulation as its focus.
6 Conclusion
The paper integrates LLMs with macroeconomic simulation through EconAgent, combining perception, reflection, and context-based decision-making. It reports more reasonable reproduction of classic macroeconomic phenomena than traditional agents.
- EconAgent integrates LLMs with macroeconomic simulation through perception, reflection, and decision-making based on real-world economic contexts.
- The framework aims to simulate more realistic macroeconomics using LLM capabilities.
- Classic macroeconomic phenomena are reproduced more reasonably than by traditional rule-based or learning-based agents.
7 Limitations
The simulation currently models only household labor and consumption behavior, excluding firms and their pricing or hiring decisions. Its results remain confined to stylized facts rather than policy optimization or accurate forecasting.
- The environment currently simulates only households’ labor and consumption behavior, excluding firms and their pricing or hiring decisions.
- Adding firms could support more realistic systems and richer regularities such as the Beveridge curve and procyclicality or countercyclicality.
- Current results are confined to replicating stylized facts, while policy optimization and accurate forecasting remain unresolved goals.
- More realistic responses to subtle policy changes are still required for policy optimization.
- Urban-scale implementation requires faster inference and lower computational demands from LLMs.
A EconAgent
Agent profiles are initialized with age, monthly wage, tax, and job information. Jobs are aligned with wage ranges, and prompts provide additional economic context for the agents.
- Agent profiles include age and monthly wage distributions, with 2018 U.S. federal tax brackets and rates represented alongside them.
- A full economic prompt supplies additional factors and details beyond those described in the main text.
- The prompt instantiates an individual profile with age, location, taxation, income redistribution, date, and prior employment context.
Baselines
The baselines encode consumption and work decisions using memory, income, savings, prices, interest rates, and utility functions. Their parameters and simulation costs are also documented.
- Consumption: LEN makes consumption memory-based, incorporating current income, accumulated savings, and goods prices.
- Consumption: CATS makes consumption non-memory-based, relating it to current income while targeting a desired savings-to-income ratio.
- Consumption: Interest rates are included in consumption rules to provide agents with fiscal-policy perception.
- Work: Work propensity increases with higher expected income, lower savings, or lower interest rates.
- AI-Economist: AI-Economist uses an MLP policy network whose inputs include wages, interest rates, goods prices, and tax rates, with utility positively related to savings and consumption but negatively related to labor.
- Simulation parameters: The reported main-text baseline results use β = 0.1, γ = 0.1, and h = 1 after grid search.
C Additional Results
Additional experiments examine quarterly and long-horizon macroeconomic behavior, sensitivity to agent count, and simulation robustness. EconAgent produces stable, plausible results across repeated simulations and remains insensitive to the number of agents.
- Quarterly Indicators: Quarterly macroeconomic indicators show conclusions similar to annual indicators.
- Fluctuated Unemployment Rate: After 20 years, the unemployment rate returns to a lower level during five additional simulation years.
- Sensitivity and Robustness: 300 agents produce stable and plausible inflation rates similar to simulations with 100 agents.Other indicators show the same pattern, making results insensitive to the number of agents.
- Sensitivity and Robustness: Across five simulations, EconAgent consistently yields more stable and plausible results than the other agent models.The simulations show no significant differences in stability or numerical scale.