Tacit Algorithmic Collusion: A Replication Study

Abstract

The adoption of pricing algorithms in markets has raised concerns about their potential to collude, even without explicit programming or communication. This paper reproduces and validates a study by Calvano et al. (2020), demonstrating that reinforcement learning-based pricing algorithms can learn to collude tacitly. After translating their original Fortran implementation into Python, I confirm the result that algorithms are able to collude and achieve supra-competitive prices and profit gains of 70-80% in equilibrium. I also confirm that they learn sophisticated punishment strategies by going into a price war before returning to equilibrium. I provide the Python replication code as open source, making it easier for subsequent research to reuse it jointly with the large existing open source codebase for reinforcement learning and pricing algorithms. The findings underscore the concerns and highlight the need for future regulation of algorithmic collusion.