Algorithmic Collusion under Observed Demand Shocks

TL;DR

This study uses simulations to show that Q-learning algorithms in markets with observable demand shocks develop demand-dependent pricing patterns, often leading to collusive, supracompetitive profits, influenced by the discount factor and information availability.

Contribution

It demonstrates how demand observability influences the learning and pricing strategies of Q-learning agents, revealing demand-contingent pricing and collusive behavior in simulated markets.

Findings

Demand observability induces demand-contingent pricing patterns.

High discount factors lead to procyclical pricing, low to countercyclical.

Algorithms sustain supracompetitive profits despite demand shocks.

Abstract

This paper examines how the observability of demand shocks influences pricing patterns and market outcomes when firms delegate pricing decisions to Q-learning algorithms. Simulations show that demand observability induces Q-learning agents to adapt prices to demand fluctuations, giving rise to distinctive demand-contingent pricing patterns across the discount factor $\delta$, consistent with Rotemberg and Saloner (1986). When $\delta$ is high, they learn procyclical pricing, charging higher prices in higher demand states. In contrast, at low $\delta$, they lower prices during booms and raise them during downturns, exhibiting countercyclical pricing. Q-learning agents also autonomously sustain supracompetitive profits, indicating that demand observability does not hinder algorithmic collusion. I further explore how the information available to algorithms shapes their learned pricing behavior. Overall, the results suggest that, through pure trial and error, Q-learning algorithms internalize both the stronger deviation incentives during booms and the trade-off between short-term gains and long-term continuation values governed by the discount factor, thereby reproducing the cyclicality of pricing patterns predicted by collusion theory.