Algorithmic Collusion is Algorithm Orchestration

We propose a fresh meta-game' perspective on the problem of algorithmic collusion in pricing games a la Bertrand. Economists have interpreted the fact that algorithms can learn to price collusively as tacit collusion. We argue instead that the co-parametrization of algorithms, in ways as are necessary to obtain algorithmic collusion, typically requires algorithm designers to engage in some form of explicit collusion or algorithm orchestration.’ In our model, the algorithm designers play a meta-game of parametrizing their algorithms, which then play repeated Bertrand competition. The strategic analysis at the meta-level reveals new equilibrium and collusion phenomena. (JEL: C62, C63, D43, L13)

💡 Research Summary

The paper introduces a novel “meta‑game” framework to rethink algorithmic collusion in repeated Bertrand pricing contests. Traditional antitrust literature treats the emergence of supra‑competitive pricing by learning algorithms as tacit collusion, assuming that algorithms independently discover cooperative strategies. The authors argue that such outcomes are only possible when the designers of the algorithms coordinate their hyper‑parameters—a process they term “algorithm orchestration.” In the meta‑game, each firm’s designer chooses a set of hyper‑parameters (e.g., exploration rate ε, discount factor γ) for a Q‑learning pricing agent. These agents then play a repeated Bertrand duopoly for T periods; demand is linear and the low‑price firm captures the whole market, with equal split on ties. The designers’ payoffs are the total profits earned by their agents over the T rounds, making the designers’ choices interdependent.

The authors formalize the meta‑game, define a best‑response mapping br(θ), and explore the payoff landscape over a two‑dimensional hyper‑parameter space. Computational experiments reveal two distinct regions. First, a competitive region where low ε and high γ lead to a pure Nash equilibrium in the meta‑game; agents quickly converge to near‑competitive pricing, yielding only modest supra‑competitive margins. Second, a cooperative region (the Pareto frontier) where higher ε and lower γ—especially when both firms adopt similar settings or asymmetric but complementary settings—produce sustained high prices and jointly higher profits. These cooperative outcomes require the designers to deliberately align their parameters, i.e., to orchestrate the algorithms, rather than stumbling upon collusion through independent learning.

The paper distinguishes between “collusion in the meta‑game” (parameter orchestration) and “price collusion in the underlying repeated game.” It shows that the former is a prerequisite for the latter under the studied Q‑learning Bertrand setting. The authors also discuss policy implications: regulating the algorithms themselves may be less effective than monitoring and possibly restricting the sharing or joint development of hyper‑parameters among firms. They propose a simple statistical test based on emergent price patterns to differentiate orchestrated from competitive algorithms.

Future research directions include extending the meta‑game to multi‑firm markets, non‑linear demand, other reinforcement‑learning methods (deep RL, policy gradients), and empirical validation using real platform data on algorithm parameter sharing. Overall, the meta‑game perspective provides a richer analytical tool for understanding when algorithmic pricing will be competitive versus when it will be coordinated, offering actionable insights for antitrust enforcement.