Product Rankings, AI Pricing Algorithms, and Collusion

Reinforcement learning (RL) based pricing algorithms have been shown to tacitly collude to set supra-competitive prices in oligopoly models of repeated price competition. We investigate the impact of ranking systems, a common feature of online marketplaces, on algorithmic collusion in prices. We study experimentally the behavior of algorithms powered by Artificial Intelligence (deep Q-learning) in a workhorse duopoly model of repeated price competition in the presence of product rankings. Through extensive experiments, we find that the introduction of the ranking system significantly mitigates the tacit collusion that stems from RL based pricing. The ranking system increases the incentives for the RL agents to deviate from a collusive price which in turn requires more complicated punishment strategies to prevent deviation and sustain collusive prices. These punishment strategies are harder to learn for RL algorithms in non stationary environments and the high collusive prices are not sustained as a result. The ranking system's mitigation effect is moderated by the horizontal differentiation between the products offered by the firms and the stickiness of product ranks. In particular, when products are more horizontally differentiated from each other and when past sales have a larger influence on product ranks (sticky ranking), the prices charged by the two firms are higher and the ranking system's mitigation effect is weaker. However, in both cases, prices in the presence of ranking are lower than that in the absence of ranking. Our analysis sheds light on the impact of ranking systems on consumer welfare and on design of ranking systems to prevent algorithmic pricing collusion