Working Paper: CEPR ID: DP15504
Authors: Justin Johnson; Andrew Rhodes; Matthijs Wildenbeest
Abstract: Using both economic theory and Artificial Intelligence (AI) pricing algorithms, we investigate the ability of a platform to design its marketplace to promote competition, improve consumer surplus, and even raise its own profits. We allow sellers to use Q-learning algorithms (a common reinforcement-learning technique from the computer-science literature) to devise pricing strategies in a setting with repeated interactions, and consider the effect of platform rules that reward firms that cut prices with additional exposure to consumers. Overall, the evidence from our experiments suggests that platform design decisions can meaningfully benefit consumers even when algorithmic collusion might otherwise emerge but that achieving these gains may require more than the simplest steering policies when algorithms value the future highly. We also find that policies that raise consumer surplus can raise the profits of the platform, depending on the platform's revenue model. Finally, we document several learning challenges faced by the algorithms.
Keywords: algorithms; artificial intelligence; collusion; platform design
JEL Codes: K21; L00
Edges that are evidenced by causal inference methods are in orange, and the rest are in light blue.
| Cause | Effect |
|---|---|
| pricedirected prominence (PDP) (E30) | lower prices in competitive markets (D41) |
| pricedirected prominence (PDP) (E30) | increase consumer surplus (D11) |
| pricedirected prominence (PDP) (E30) | lower consumer surplus (D11) |
| dynamic pricedirected prominence (dynamic PDP) (C69) | destabilize collusion (D74) |
| dynamic pricedirected prominence (dynamic PDP) (C69) | substantial price drops (P22) |
| dynamic pricedirected prominence (dynamic PDP) (C69) | increase in consumer surplus (D11) |
| policies designed to raise consumer surplus (D18) | enhance platform profits (L21) |