Algorithmic Choice and Collusion

The increased popularity of (self-learning) pricing algorithms has raised concerns regarding the possibility that those algorithms could learn to collude tacitly on non-competitive prices. This has been substantiated by initial research (e.g., Calvano et al., 2020; Klein, 2021), which suggests that such algorithmic collusion poses significant challenges in purely algorithmic markets. However, the dynamics of these algorithmic markets in comparison to those with human actors remain less explored. Existing studies either contrast algorithmic with human collusion within significantly different market environments (Calvano et al., 2020) or analyze market data to distinguish between the two, sacrificing the controlled environment necessary for understanding the underlying mechanisms (Assad et al., 2020). Moreover, the interaction between algorithmic and human sellers in the same marketplace, observed in sectors like online product sales on platforms like Amazon (L. Chen et al., 2016) and the gasoline market (Assad et al., 2020), and findings on algorithmic and human collusion in isolation might not be indicative for outcomes in markets with a mixed composition. While there is first evidence of collusion in these "mixed" markets, those studies usually focus on algorithms that allow for collusion by design and are not self-learning (Normann and Sternberg, 2021). Addressing this gap, Werner (2023) provides novel insights into the extent of the problem posed by self-learning algorithmic collusion in contrast to traditional human-driven markets through a controlled experimental setup. This study investigates the interactions between pricing algorithms and human sellers across various market compositions. Participants are assigned specific roles wherein they must either completely delegate their pricing decisions to an algorithm or set prices independently, without the ability to choose whether they want to adopt an algorithm. Overall, evidence suggests that self-learning algorithms may enable tacit collusion more effectively than humans (Calvano et al., 2020; Klein, 2021; Werner, 2023). Additionally, existing literature highlights a tendency for algorithm aversion among users, who show a preference for maintaining control over imperfect algorithms, even though these algorithms frequently outperform human decision-making (Dietvorst et al., 2018). Our research builds upon the existing body of literature to provide insights into the dynamics of adopting self-learning algorithms for pricing decisions and their impact on pricing outcomes in a Bertrand market. Specifically, we aim to obtain insights on whether and to what extent firms will delegate pricing, either partially or completely, to algorithms and how this affects market prices and potential collusion levels. To assess these questions, a between-subject design will be applied, in which each participant only plays one treatment. Participants are divided into three treatment groups: the first without the choice to use pricing algorithms, the second with the choice to fully delegate pricing decisions to a self-learning algorithm, and the third with the choice of access to algorithmic price recommendations while retaining decision-making autonomy. The self-learning pricing algorithm, consistent across all treatments and participants, follows the model presented in Werner (2023). In the treatments where participants have the option to either delegate their pricing decisions or receive algorithmic price recommendations, they are provided with comprehensive information about the algorithm's objectives and past performance. In all treatments, participants have the opportunity to familiarize themselves with the pricing process during a simulation round, where the other firm is played by the algorithm. Our lab experiment, programmed in oTree (D. L. Chen et al., 2016), involves about 300 participants in five super games to allow for learning of the participants, simulating a Bertrand market with two firms selling homogenous goods. Within this market, each firm has to set prices ranging from 0 to 5 to sixty computerized consumers, each willing to purchase one unit during each period, with a maximum willingness to pay of 4. Each round has a continuation probability of 95%. The group matching will be held constant within each super game. Across super games, participants will be rematched into two-firm markets within matching groups of 10. The super games conclude with participants incentivized to disclose their beliefs regarding their counterparts' use of algorithms. Additionally, participants will complete questionnaires to evaluate algorithm aversion, demographic factors, and measures of risk, trust, and reciprocity.