Study finds AI pricing algorithms rarely collude unless identical versions used
New research shows multi-armed bandit algorithms converge to fair prices in most market scenarios.
Researchers Martin Bichler, Julius Durmann, and Matthias Oberlechner from TU Munich have released a paper investigating whether online optimization algorithms used in automated pricing can lead to tacit collusion, a growing concern for digital market regulators. The study focuses on multi-armed bandit (MAB) algorithms, which require minimal information and are realistic models for automated pricing agents. They prove analytically that mean-based online learning algorithms converge to correlated rationalizable actions, which in Bertrand competition settings translates to Nash equilibrium or prices very close to it.
In numerical experiments, most MAB algorithms—including those not mean-based—also converge to competitive outcomes. The only exception occurs when all sellers deploy identical symmetric versions of certain algorithms like UCB (Upper Confidence Bound), which can lead to temporary supra-competitive prices. However, this collusive effect diminishes as the number of competitors increases. The authors conclude that sustained algorithmic collusion is unlikely when independent firms use different learning algorithms, providing reassurance for regulators and managers adopting algorithmic pricing.
- Mean-based online learning algorithms provably converge to competitive Nash equilibrium in repeated price competition.
- Only symmetric UCB algorithms used by all sellers cause temporary supra-competitive prices; effect weakens with more competitors.
- Study uses minimal-information multi-armed bandit models, making results directly applicable to real-world automated pricing systems.
Why It Matters
Regulators can relax: diverse AI pricing agents naturally avoid collusion without explicit intervention.