Lindsay Robinson

Autonomous algorithmic collusion occurs when firms delegate to autonomous algorithms that learn to collude with each other through repeated interaction. In this paper, I show how efficient learning algorithms can dramatically increase the speed and perfection of autonomous algorithmic collusion. I consider Q-learning pricing algorithms with a parsimonious Q-function that scales linearly in the dimension of the state representation. This allows the algorithms to condition on a higher dimensional representation of the price history, and to better distinguish between temporary and sustained competitor deviation from collusive prices. The algorithms can coordinate quickly on symmetric joint-profit-maximizing collusive prices in a repeated simultaneous oligopoly pricing game. I explore the robustness of collusive prices to the algorithms' parametrization and the strategic environment, and conclude with implications for competition policy.

Demand systems describing consumer preferences over retail stores can be used to simulate the price and welfare effects of retail mergers. I develop a new methodology for estimating these store-level demand systems and simulating retail mergers, using only public spatial data. The approach finds the demand and cost parameters that best rationalize retail chains' observed store networks as a stable spatial equilibrium. It accomodates demographic-driven consumer preferences, endogenous store-level prices, horizontal and vertical chain differentiation and geographic store differentiation, without requiring any ex-ante local market definition. I apply the methodology to the Australian retail grocery industry, illustrating its use in merger simulation and remedy design.