Agent-Based Model of Competitive Strategy
- Introduced by
- Jason P. Davis, Kathleen M. Eisenhardt & Christopher B. Bingham (simulation-methods roadmap)
- Year
- 2007
- Type
- Agent-based computational simulation of competitive strategy
- Family
- Strategic management / computational modeling
What does it do?
An agent-based model of competitive strategy represents firms as autonomous, heterogeneous, adaptive agents whose decision rules and local interactions generate emergent industry-level dynamics that no single firm designs.
Davis, Eisenhardt, and Bingham's 2007 roadmap for developing theory through simulation places this kind of computational modeling in the sweet spot between inductive case research and formal mathematics, well suited to longitudinal, nonlinear, and interactive strategy phenomena. Instead of solving for an equilibrium, the analyst builds firms with strategies, lets them compete over many simulated periods, and studies the market structures, survival patterns, and performance dispersions that emerge. The method gives strategy researchers a controlled laboratory for theory building about competitive dynamics that are too complex and path-dependent for closed-form analysis.
The intuition behind the method
Real competition is a process: firms watch rivals, adjust prices and products, learn, imitate, and occasionally fail, and the resulting industry structure emerges from countless local moves rather than from any grand plan. Trying to capture this with a single equation usually forces you to assume away the very heterogeneity, adaptation, and feedback that make competition interesting. An agent-based model takes the opposite route. You build many simulated firms, each with its own state and behavioral rules, place them in an environment where their fortunes depend on what rivals do, and press play. By running the model thousands of times under different assumptions, you can see which firm behaviors lead to concentration or fragmentation, which strategies survive turbulence, and how small differences in rules cascade into very different industry outcomes — insights that emerge from the interaction itself rather than being built into a solved equilibrium.
History
Computational simulation entered strategy research as scholars sought tools for dynamics that defied both case studies and closed-form models. Early agent-based and behavioral simulations of organizational learning and adaptation in the management literature showed that firm-level rules could generate non-obvious aggregate patterns. Davis, Eisenhardt, and Bingham's 2007 Academy of Management Review article consolidated this tradition into an explicit methodological roadmap, positioning simulation between theory-creating inductive methods and theory-testing empirical methods and detailing how to build, run, and validate simulations for theory development. Their framework helped legitimize agent-based and computational modeling as a disciplined way to theorize about competitive dynamics, market evolution, and other complex, interactive strategy phenomena.
When to use it
Use an agent-based model of competitive strategy when the phenomenon involves many heterogeneous, interacting firms whose adaptation and feedback produce emergent, path-dependent industry outcomes that resist closed-form analysis. It is well suited to building theory about competitive dynamics, imitation and differentiation, market evolution, and how firm-level rules aggregate into market structure, especially when longitudinal and nonlinear processes are central. It is most valuable when empirical data are too thin or the system too complex for econometric estimation but the underlying mechanisms can be specified as rules. It is less appropriate when a tractable analytical or equilibrium model already captures the logic, when rich data permit direct estimation, or when the question is descriptive rather than about generative mechanisms, since agent-based results are conditional on the modeling assumptions.
Limitations and considerations
- Results are only as credible as the assumed agent rules and interaction structure, which are hard to validate against data.
- Many free parameters and design choices create a large space that can be tuned to produce almost any outcome.
- Emergent findings can be sensitive to implementation details, random seeds, and the number of replications.
- Generalization is uncertain because conclusions are conditional on the specific model rather than on observed populations.
Source conditions do not replace independent scientific review or validation by an analysis engine.
Strengths
- Captures heterogeneity, adaptation, and feedback among firms that equilibrium and regression models typically assume away.
- Generates emergent industry-level structure from specified micro-level strategy rules, supporting mechanism-based theory.
- Provides a controlled virtual laboratory for experiments that are impossible to run on real markets.
- Handles longitudinal, nonlinear, and path-dependent competitive dynamics that closed-form analysis cannot.
Applications
- Studying how firm imitation versus differentiation rules shape long-run market concentration.
- Exploring which competitive strategies survive in turbulent versus stable environments.
- Modeling the co-evolution of firm capabilities and industry structure over many periods.
- Generating and refining theoretical propositions about competitive dynamics before field testing.
- Examining how heterogeneity in firm decision rules produces dispersion in performance and survival.
How it works
How it works
The model begins by specifying the agents and their behavioral rules. Each firm is defined by an internal state — resources, position, beliefs — and a set of decision rules that map what it observes to its next action, such as pricing, investing, imitating, or exiting. Davis, Eisenhardt, and Bingham stress that good simulation theory starts from a clear, theoretically grounded construct and a small set of rules rather than arbitrary complexity. Heterogeneity across agents is deliberate, because differences in strategy and capability are what drive the competitive variation the model is meant to explain.
Formula
a_i = (\,s_i,\, \theta_i,\, f_i\,),\quad x_i^{t+1} = f_i(s_i^t,\, \theta_i,\, \mathbf{x}_{-i}^t)Next the analyst defines how agents interact and how the environment responds. Firm payoffs depend not only on their own choices but on rivals' choices and on a shared environment — demand, technology, a market landscape — that the firms collectively alter as they act. Davis and colleagues emphasize representing the interaction structure and feedback explicitly, because the theoretical payoff of simulation comes precisely from interdependence and nonlinearity. Specifying the payoff function and the environment's law of motion fixes the competitive logic within which strategies will succeed or fail.
Formula
\pi_i^t = \Pi\bigl(x_i^t,\, \mathbf{x}_{-i}^t;\, E^t\bigr),\quad E^{t+1} = g(E^t,\, \{x_i^t\})The model is then run forward over many periods and, crucially, many times. Because agent rules and the environment can include randomness, a single run is just one realization; the analyst executes many replications across random seeds and across the space of parameter values. Davis, Eisenhardt, and Bingham describe this disciplined experimentation as the engine of simulation-based theory, where the researcher systematically varies assumptions and observes the consequences rather than reading a single trajectory as if it were the truth.
Formula
\{\mathbf{x}^t\}_{t=0}^{T} \sim \text{sim}(\theta,\, E^0,\, \omega),\quad \omega \in \Omega \ \text{(random seeds / replications)}From the runs the analyst measures emergent, industry-level outcomes that were not programmed in directly, such as concentration, survival rates, and the dispersion of firm performance. These aggregate patterns, averaged over replications to separate signal from stochastic noise, are the dependent variables of the study. The point Davis and colleagues make is that simulation generates novel theory by revealing how micro-level firm behavior aggregates into macro-level competitive structure, so the measurement step is where the emergent regularities the theory will explain are captured.
Formula
Y(\theta) = \mathbb{E}_\omega\bigl[\phi(\{\mathbf{x}^t\})\bigr],\quad \phi = (\text{HHI},\, \text{survival rate},\, \text{dispersion of }\pi)Finally the model is validated and used for controlled experiments. The analyst compares emergent outcomes to known empirical patterns or stylized facts to establish credibility, then conducts virtual experiments by perturbing parameters and rules to map how outcomes respond. Davis, Eisenhardt, and Bingham frame this as the heart of theory development through simulation: once the model reproduces plausible dynamics, systematically varying its assumptions yields propositions about competitive strategy that can later be tested with field data.
Formula
\frac{\partial Y}{\partial \theta_k} \approx \frac{Y(\theta + \Delta e_k) - Y(\theta)}{\Delta},\quad \text{compare } Y(\theta) \ \text{to empirical patterns}Common pitfalls
- Adding complexity for realism until the model is as opaque as the reality it was meant to illuminate.
- Reading a single simulation run as a result instead of averaging over many replications to separate signal from noise.
- Skipping validation against empirical patterns, so the model's dynamics are merely plausible-looking artifacts.
- Treating simulation output as empirical evidence about real firms rather than as theory-building conditional on assumptions.
Frequently asked questions
How does an agent-based model differ from game-theoretic analysis of competition?
Game theory typically solves for an equilibrium under assumptions of optimization and common knowledge, yielding clean closed-form predictions but struggling with many heterogeneous, boundedly rational, adapting firms. An agent-based model instead specifies behavioral rules and lets outcomes emerge from repeated interaction, with no equilibrium imposed. Davis, Eisenhardt, and Bingham position simulation between formal modeling and inductive research precisely because it can represent process, heterogeneity, and nonlinearity that equilibrium analysis abstracts away, at the cost of analytical tractability and the assurance that a stable solution exists.
If results depend on the assumptions, what makes simulation more than storytelling?
Discipline. Davis and colleagues argue that credible simulation rests on a clear theoretical construct, transparent rules, systematic experimentation across the parameter space, validation against empirical patterns, and replication to separate signal from stochastic noise. The contribution is not a claim about a particular real market but a set of internally consistent propositions showing how specified micro-mechanisms generate macro-outcomes. Those propositions are then meant to be tested with field data, so simulation builds theory rigorously rather than merely illustrating a hunch.
How many runs are needed and how is the model validated?
Because agent rules and the environment include randomness, outcomes must be averaged over many replications per parameter setting, with enough runs that the estimated emergent measures stabilize. Validation proceeds on two fronts: checking that the model's emergent dynamics reproduce known stylized facts or empirical regularities, and conducting sensitivity analysis to confirm conclusions are robust to reasonable changes in assumptions. Davis, Eisenhardt, and Bingham treat this combination of replication, validation, and sensitivity testing as essential to credible theory development through simulation.
Related methods
Sources
Davis, J. P., Eisenhardt, K. M., & Bingham, C. B. (2007). Developing Theory Through Simulation Methods. Academy of Management Review, 32(2), 480-499.
Open DOI referenceCopied.