Data-Driven Economic Agent-Based Models ⋆
Marco Pangallo and R. Maria del Rio-Chanona · 2024
Read Data-Driven Economic Agent-Based Models ⋆ by Marco Pangallo and R. Maria del Rio-Chanona (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-16591.txt.
Introduction
Agent-based models of the economy, societies of simulated households and firms, have long been accused of floating free of real data. Marco Pangallo and Maria del Rio-Chanona argue in 2024 that the field has turned a corner, and they sketch how far each model has actually tied itself to evidence.
Their scale asks two questions: are the agents' starting values drawn from real micro-data, and do the model's motions track real time series? They review the craft of calibration, initialization, and live data feeding, then tour applications that produced new knowledge and shaped policy. The paper doubles as a manifesto and a field guide for economists weighing agent models against equilibrium rivals.
Catalog record
| CREATOR | Marco Pangallo and R. Maria del Rio-Chanona |
|---|---|
| YEAR | 2024 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 47 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2412-16591
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Data-Driven Economic Agent-Based Models ⋆ Marco Pangallo and R. Maria del Rio-Chanona Abstract Economic agent-based models (ABMs) are becoming more and more data-driven, establishing themselves as increasingly valuable tools for economic research and policymaking. We propose to classify the extent to which an ABM is data-driven based on whether agent-level quantities are initialized from real-world micro-data and whether the ABM’s dynamics track empirical time series. This paper discusses how making ABMs data-driven helps overcome limitations of traditional ABMs and makes ABMs a stronger alternative to equilibrium models. We review state-of-the-art methods in parameter calibration, initialization, and data assimilation, and then present successful applications that have generated new scientific knowledge and informed policy decisions. This paper serves as a manifesto for data-driven ABMs, introducing a definition and classification and outlining the state of the field, and as a guide for those new to the field. 1 1 footnotetext: We would like to thank the co-organizers and participants of the workshop “Data-driven economic agent-based models” ( https://sites.google.com/view/wddeabm ), where we discussed many of the ideas presented in this chapter. 1 Introduction Economic modeling is essential to understand how the economy works and to address pressing global challenges. While econometric studies provide an initial understanding of causal effects, theoretical economic models enable policymakers to simulate scenarios like tax reforms, industrial policy, or economic crises and their impacts on employment, inflation, GDP, and inequality. Theoretical models serve two additional purposes: first, they help us anticipate how agents might respond in new scenarios like climate change or the rapid adoption of artificial intelligence, and second, to project the effects of policies that could reshape agent behaviors in ways not observed historically ( Muth 1961 ) . Traditionally, economic models have relied on the concept of equilibrium, where the state of the economy is derived from solving a set of equations typically based on agents’ maximizing behavior. This approach can model how agents respond to novel situation that may alter their behavior, as agents re-maximize in the new setting. Since equilibrium models are based on equations, they can often be solved analytically, making them clear and transparent. Furthermore, they provide a consistent and unifying framework that reduces ad-hoc modeling assumptions. However, these benefits come at a cost: equilibrium models rely on simplifications that can misrepresent real-world dynamics, heterogeneity, and human behavior. For example, the use of a representative agent limits the ability to capture diverse behaviors, while assumptions of perfect rationality can overlook heuristics common in decision-making. As a result, even though policy counterfactuals in equilibrium models are internally consistent, they may miss critical real-world phenomena, such as sudden crises, inequality, and bounded rationality, which limits their reliability for policy analysis. In the last 30 years, Agent-Based Models (ABMs) have emerged as an alternative to equilibrium models ( Axtell and Farmer 2024 ) . ABMs are non-equilibrium in the sense that all model variables are obtained as a map from previously computed values, rather than by solving equilibrium equations ( Pangallo 2024 ) . For instance, cash available to an agent may update following sales and purchases, but neither of these is determined by setting demand equal to supply. Non-equilibrium dynamics makes it possible to include all sort of realistic aspects. For example, an ABM can simulate decision-making through heuristics ( Artinger et al. 2022 ) —such as households adjusting spending based on recent income changes rather than from market clearing conditions. As another example, ABMs can easily capture real-world dynamics and transient states, which is crucial for modeling sudden events like the COVID-19 pandemic. However, this flexibility comes with trade-offs: without equilibrium assumptions, ABMs may lose mathematical tractability and often lack closed-form solutions. The flexibility in modeling choices can also make ABM rules appear ad hoc or arbitrary, especially in earlier applications where ABMs served as thought experiments rather than calibrated models of actual economies, limiting their reliability for policy recommendations. In this chapter, we argue that the recent trend toward data-driven ABMs is helping overcome their traditional limitations, making them a compelling alternative to equilibrium models. First, agent-level micro-data enable us to observe behaviors directly, reducing the number of arbitrary assumptions. For example, in an economic-epidemic model simulating economic and mobility decisions ( Pangallo et al. 2023 ) , mobility data allow us to avoid detailed assumptions about daily activities like bringing kids to school and commuting, which are not crucial for an economic model. Second, ABMs can now faithfully represent a given economy by building synthetic populations that accurately reflect individual and household attributes, or by reconstructing the production networks linking firms. This detailed representation improves the quantitative reliability of policy counterfactuals, offering more grounded results than studies in abstract economies. Third, ABMs can now track empirical time series data, allowing for in-sample calibration and out-of-sample testing. In some cases, data-driven ABMs have been able to forecast macroeconomic variables, even after major shocks like the COVID-19 pandemic ( Poledna et al. 2023 ; Pichler et al. 2022 ) . This focus on time series fitting and forecasting increases the scientific validity of ABMs by allowing empirical testing of model assumptions. Assumptions that do not improve model performance can be streamlined or removed, simplifying models and making them more transparent. The tendency to more closely link ABMs and data has been greatly accelerating in the last ten years. In the following, we first discuss in what precise sense economic ABMs are becoming more and more data-driven, and then review new methods and applications, including success stories of recent data-driven ABMs, and highlighting challenges and opportunities for future research. 2 Definitions 2.1 A technical definition of ABMs Although there exist several conceptual definitions of ABM ( Jennings 2000 ; Bonabeau 2002 ; De Marchi and Page 2014 ; Delli Gatti et al. 2018 ) , to assess what makes an ABM data-driven we need a technical definition. Here is our proposal. An ABM is composed of N\gg 1 model units, indexed by i=1,\dots,N , that for simplicity we call agents . 1 1 1 Model units could be agents, i.e. autonomous units with “agency”, or features of the environment. It is important that they are distinct discrete units, not a “continuum” of identical agents. Each agent is characterized by a vector of time-dependent variables , denoted by \mathbf{x}_{i,t} at time t , and a vector of fixed attributes represented by \mathbf{a}_{i} . At the system level, we define a vector of model-wide variables \mathbf{y}_{t} , which evolve over time, and a vector of constant parameters \mathbf{\theta} that govern the behavior of the entire system. An ABM implicitly defines a probability distribution \mathbb{P} that determines the current values of both the set of agent-specific vectors \{{\mathbf{x}_{i,t}}\}_{i=1}^{N} and the model-wide vector \mathbf{y}_{t} , given the past values of these vectors, the agents’ attributes, and the parameters, as follows 2 2 2 While N\gg 1 requires that \{{\mathbf{x}_{i,t}}\}_{i=1}^{N} is non-empty, other components in Eq. ( 1 ) can be empty sets. For example, the model may not include model-wide variables, making \{\mathbf{y}_{\tau<t}\} an empty set. Similarly, some models may not have agent-specific fixed attributes or parameters. \{{\mathbf{x}_{i,t}}\}_{i=1}^{N},\mathbf{y}_{t}\sim\mathbb{P}\left(\{{\mathbf{x}_{i,\tau<t}}\}_{i=1}^{N},\{\mathbf{y}_{\tau<t}\},\{\mathbf{a}_{i}\}_{i=1}^{N},\boldsymbol{\theta}\right). (1) This definition captures in what sense the model is non-equilibrium, as model variables can be obtained as a map of