Impact of Price Inflation on Algorithmic Collusion Through Reinforcement Learning Agents

TL;DR

This work investigates how inflation shocks shape emergent coordination and profitability in reinforcement-learning–driven pricing within a differentiated Bertrand framework. By embedding inflation into the Markov decision process and using a Deep Q-Network with an inflation-aware payoff, the authors assess how macroeconomic volatility affects pricing strategies and market competitiveness. They find that inflation tends to raise profits and foster non-competitive dynamics, but robust punishment mechanisms to deter deviations are not consistently learned. The study highlights regulatory considerations for AI-driven pricing under inflation and points to future work with more complex agent populations and data-driven policy design.

Abstract

Algorithmic pricing is increasingly shaping market competition, raising concerns about its potential to compromise competitive dynamics. While prior work has shown that reinforcement learning (RL)-based pricing algorithms can lead to tacit collusion, less attention has been given to the role of macroeconomic factors in shaping these dynamics. This study examines the role of inflation in influencing algorithmic collusion within competitive markets. By incorporating inflation shocks into a RL-based pricing model, we analyze whether agents adapt their strategies to sustain supra-competitive profits. Our findings indicate that inflation reduces market competitiveness by fostering implicit coordination among agents, even without direct collusion. However, despite achieving sustained higher profitability, agents fail to develop robust punishment mechanisms to deter deviations from equilibrium strategies. The results suggest that inflation amplifies non-competitive dynamics in algorithmic pricing, emphasizing the need for regulatory oversight in markets where AI-driven pricing is prevalent.

Figures (7)

• Figure 1: Annual Inflation Rate in the United States, 2010--2024.
• Figure 2: Proposed Markov Decision Process representing Bertrand Competition with increasing costs of production.
• Figure 3: Neural Network Architecture for Agents.
• Figure 4: Comparison of profit evolution across experimental settings with and without inflation. The blue and orange lines depict the profitability levels $\nabla$ under inflationary and non-inflationary conditions, respectively. Given the variance in prices set by individual agents, the figure shows a 1000-timestep moving average along with its corresponding confidence interval. For reference, the Nash and Monopoly equilibrium outcomes are also included, represented by the green and red lines, respectively.
• Figure 5: Prices over costs (%) for experiments with and without inflation using both in-sample and out-of-sample data. For each configuration, the empirical joint distribution of action pairs is visualized as a percentage heatmap using a logarithmic color scale. The x-axis corresponds to the pricing decisions made by Agent 0, while the y-axis represents those of Agent 1. In Figure \ref{['fig:heatmap_no_inflation_train']}, for instance, 80% of the observed timesteps show both agents selecting prices that are 60% above their respective costs.