Collusion by Algorithm: The Role of Unobserved Actions

We analyze the effects of better algorithmic demand forecasting on collusive profits. We show that the comparative statics crucially depend on the whether actions are observable. Thus, the optimal antitrust policy needs to take into account the institutional settings of the industry in question. Moreover, our analysis reveals a dual role of improving forecasting ability when actions are not observable. Deviations become more tempting, reducing profits, but also uncertainty concerning deviations is increasingly eliminated. This results in a u-shaped relationship between profits and prediction ability. When prediction ability is perfect, the 'observable actions' case emerges.