Core Concepts
Sophisticated pricing algorithms used by rideshare platforms can exhibit emergent collusive behavior, which is influenced by the responsiveness of the underlying market.
Abstract
The paper investigates the potential for algorithmic collusion in a complex rideshare market model, using the state-of-the-art Proximal Policy Optimization (PPO) reinforcement learning algorithm. The authors extend a previous mathematical program network (MPN) based rideshare model to a temporal multi origin-destination setting and study the collusive tendencies of the pricing algorithms under different market responsiveness conditions.
The key findings are:
In a responsive market, where drivers can quickly adjust their allocation across platforms, the pricing algorithms converge to a competitive equilibrium, with commissions being driven towards the rates, reducing platform profits.
In a lagging market, where drivers have higher inertia and are slower to adjust, the pricing algorithms learn to collude by keeping commissions low, allowing the platforms to extract higher profits.
This suggests that market characteristics, such as the responsiveness of the supply side, can significantly influence the emergence of algorithmic collusion, even when the learning algorithms themselves are held constant. The authors argue that this has important implications for regulatory bodies investigating potential algorithmic collusion, as markets with high inertial properties should be prioritized.
Stats
Increasing number of industries are gravitating towards dynamic pricing, including platform economies like rideshares.
Evidence suggests that algorithmic collusion does exist, with a study claiming an increase of almost 28% margins in the German retail gasoline market after adoption of dynamic pricing by duopolies.
The authors extend a previous MPN-based rideshare model to a temporal multi origin-destination setting and use PPO to solve for a repeated duopoly game.
Quotes
"With dynamic pricing gaining traction, much has been invested in devising complex models that can accurately predict an optimum price given informative attributes of any spatio-temporal frame. However, most of the sophisticated machine learning models are effectively a black-box [4]."
"Perhaps of most importance is the fact that it's not entirely clear what causes a collusive behavior to emerge in disconnected algorithmic agents. Some works argue that certain hyperparameters give rise to collusion [5], while others argue that superior algorithms cause this behavior [6]."