Skip to main content

Algorithms in the wild: Experimental evidence from an online marketplace (2022)

September 06, 2022


  • Vito Stefano Bramante, University of Bologna
  • Emilio Calvano, University of Rome-Tor Vergata
  • Giacomo Calzolari, European University Institute
  • Maximilian Schaefer, Yale University, University of Bologna


This blog article is derived from the authors’ paper titled Algorithms in the Wild: Experimental Evidence from an Online Marketplace, a project of the Economics of Digital Services (EODS) research initiative led by Penn’s Center for Technology, Innovation & Competition (CTIC) and The Warren Center for Network & Data Services. CTIC and the Warren Center are grateful to the John S. and James L. Knight Foundation for its major support of the EODS initiative.


The use of repricing software has become ubiquitous in recent years. One example is the case of gasoline markets in which the widespread adoption of repricing software by large companies has attracted the attention of both academia and the media (Assad et al, 2020; The Economist, 2017). The use of repricing software is not limited to large firms, however. The emergence of peer-to-peer marketplaces such as Amazon, eBay, and Airbnb has given rise to a rich landscape of software companies that offer affordable off-the-shelf repricing solutions for small sellers.

In the context of online platforms, sheer speed is often advertised as the main benefit of using repricing software: Automatically monitoring and reacting to rival prices allows to relentlessly undercut slower sellers, who set prices manually. This, in turn, allows sellers to capture consumer demand because online platforms prominently feature the seller with the lowest price. While fierce price competition is likely to benefit customers through lower prices, the possibility that repricers might be powered by intelligent algorithms has also raised concerns that they might autonomously learn sophisticated strategies to charge higher prices, i.e., that they might autonomously learn to collude.

In fact, the possibility that algorithms might have the ability to collude has already drawn the attention of competition authorities around the globe. For example, the topic of algorithmic collusion was discussed at the 7th session of the FTC Hearings on Competition and Consumer Protection in the 21st Century in November 2018. White papers of the Organization for Economic Cooperation and Development (OECD) (2017) and the UK’s Competition and Markets Authority (CMA) (2021) also discussed the subject.

On the research side, a recent experimental study by Calvano et al. (2020) has provided the proof-of-concept that even simple artificial intelligence algorithms have the capability of sustaining collusion by autonomously learning to punish rivals who deviate from the collusive agreement. This lends credibility to the concerns of policymakers, especially since many repricing companies advertise the use of artificial intelligence algorithms.

But empirical evidence assessing the real-world effects of repricing software remains scarce. While the finance literature has started studying the subject of algorithmic trading already one decade ago, it does not directly speak to collusion (Henderschott et al., 2011; Chaboud et al., 2014). Additionally, it is questionable how findings from sophisticated financial markets, where sellers and buyers both use software, extend to consumer mass markets, where typically only the seller-side uses software. Assad et al. (2020) is among the first studies finding evidence consistent with a chilling effect on competition in a consumer market.

One shortcoming of existing real-world studies is the imperfect ability of researchers to identify the algorithms in use, as companies are naturally reluctant to reveal their pricing technology. This raises questions about the sophistication of the adopted repricing technologies and, ultimately, leaves open the possibility that the reported findings might not capture the effect of machine intelligence. For example, it has been pointed out that high prices might be the consequence of the algorithm’s failure to optimize, providing an interesting contrast to the prevailing narrative (Cooper et al., 2015).

Our study sets out to address this shortcoming by leveraging the environment created by peer-to-peer marketplaces, which allows researchers to implement real-world repricing software in a controlled yet realistic environment. To this end, we create two seller accounts on a major online marketplace, stock them with goods, and engage in sales activity. Additionally, we create our own repricing software, which allows us to create arbitrarily sophisticated opponents against which we let commercial repricing software compete.

One major benefit of this setting is that it enables the controlled implementation of commercial repricing software and the extensive monitoring of the market environment. This allows to better rule out possible alternative explanations for increasing prices such as changes in demand and supply conditions.

Additionally, the implementation of our own repricing software allows unprecedented insights into the typical abilities of commercial repricing software. For instance, we can assess how commercial repricing software fares against different types of opponents such as (i) simple ad-hoc mechanical repricing rules, (ii) sophisticated human opponents who instruct their software to play certain pre-determined strategies, (iii) short-sighted artificial intelligence algorithms, and (iv) far-sighted artificial intelligence algorithms.

The latter case of far-sighted algorithms is of particular interest because it offers the possibility to assess whether commercial repricing software can autonomously pick up “invitations” to collude when confronted with a far-sighted algorithm, which rewards forgoing short-term profits from low prices for an ultimately more profitable long-term strategy with high prices. This offers a compelling and unique method to test the ability of commercial software to learn collusive agreements.

The benefits of this research approach, however, are to be weighed against the complications researchers face when dealing and interacting with actual markets. Approaches based on simulations, lab experiments, and empirical analysis, developed so far, do not require to interact directly with markets. We had instead to address a series of novel issues that are uncommon to academic research.

Thus far, our research effort has predominantly focused on creating our own repricing software and on demonstrating that it can successfully compete against mechanical repricing rules and one selected commercial repricing software. Implementing our own repricing software in real marketplaces poses a series of engineering challenges that researchers typically do not have to confront such as significant time-delays between ordering the implementation of a new price and the actual implementation of the new price on the platform.

Additionally, real world environments require algorithms that can quickly adapt to a wide range of potential opponent strategies because sellers cannot afford extended periods of learning in which the algorithm implements suboptimal strategies, which would hurt their profits. Algorithms that quickly learn smart strategies pose a significant challenge that researchers in laboratory environments can typically afford to ignore. For example, Calvano et al. (2020) allow for several thousands of periods of learning.

Our current findings, which constitute the result of our first major pilot study, have been encouraging. We were successful in implementing our own short-sighted artificial intelligence repricing software. Using mechanical repricing rules and one selected commercial repricing software as opponents, we demonstrate that our artificial intelligence repricing software is successful in quickly learning strategies that perform successfully against both type of opponents. Our current findings also suggest that the commercial repricing software we selected for our pilot study is not particularly sophisticated and thus likely unable to implement collusive strategies.

The implementation of far-sighted algorithms, which learn sophisticated strategies quickly, remains a challenge. In this respect, we have recently achieved a proof-of-concept in the laboratory setting (ignoring engineering challenges faced in real-world environments) by demonstrating that such strategies can be learned quickly without sellers incurring prohibitive costs.

Looking ahead, far-sighted strategies capable of implementing collusive outcomes still need to be deployed in the real-world environment. Once this is achieved, we will have all the ingredients to fully implement our approach and let commercial repricing software compete against increasingly sophisticated versions of our own repricing software. Our goal is to test a variety of commercial repricing software to provide a comprehensive overview of the typical capabilities of this emerging technology, which is increasingly available to small sellers.


Assad, S., Clark, R., Ershov, D., & Xu, L. (2020). Algorithmic pricing and competition: Empirical evidence from the German retail gasoline market.

Calvano, E., Calzolari, G., Denicolo, V., & Pastorello, S. (2020). Artificial intelligence, algorithmic pricing, and collusion. American Economic Review, 110(10), 3267-97.

Chaboud, A. P., Chiquoine, B., Hjalmarsson, E., & Vega, C. (2014). Rise of the machines: Algorithmic trading in the foreign exchange market. The Journal of Finance, 69(5), 2045-2084.

CMA (2021). Algorithms: How they can reduce competition and harm consumers. (

Cooper, W. L., Homem-de-Mello, T., & Kleywegt, A. J. (2015). Learning and pricing with models that do not explicitly incorporate competition. Operations research, 63(1), 86-103.

Hendershott, T., Jones, C. M., & Menkveld, A. J. (2011). Does algorithmic trading improve liquidity?. The Journal of finance, 66(1), 1-33.

OECD (2017). Algorithms and Collusion: Competition Policy in the Digital Age (

The Economist (2017). Price Bots can collude against consumers.(