Adaptive Behaviour in an Oligopoly

Advances in game theory have provided an impetus for renewed investigation of the strategic behaviour of oligopolists as players in repeated games. Marketing databases provide a rich source of historical evidence of such behaviour. This paper uses such data to examine how players in iterated oligopolies respond to their rivals’ behaviour, and uses machine learning to derive improved contingent strategies for such markets, in order to provide insights into the evolution of such markets and the patterns of behaviour observed. The paper is an application of repeated games and machine learning to adaptive behaviour over time in the retail market for ground coffee. Using empirical data on the weekly prices and promotional instruments of the four largest and several smaller coffee sellers in a regional U.S. retail market, and using a market model to predict sellers’ market shares and profits in response to others’ actions in any week, we examine the adaptive strategic behaviour of the three largest sellers. We model the sellers’ strategic behaviours as finite automata with memory of previous weeks’ actions, and use the Axelrod/Forrest representation of the action function, mapping state to action. We use a genetic algorithm (GA) to derive automata which are fit, given their environment, as described by their rivals’ actions in the past and the implicit demand for coffee.