• No results found

Algorithmic Pricing under EU Competition Law

N/A
N/A
Protected

Academic year: 2021

Share "Algorithmic Pricing under EU Competition Law"

Copied!
54
0
0

Bezig met laden.... (Bekijk nu de volledige tekst)

Hele tekst

(1)European Competition Law and Regulation (LLM track). ALGORITHMIC PRICING UNDER EU COMPETITION LAW. Agnieszka Bartłomiejczyk. Master Thesis written under the supervision of Prof. dr Rein Wesseling. Amsterdam, 27 July 2018.

(2) ABSTRACT. This paper examines the issue of algorithmic pricing from the perspective of EU competition law. It attempts to analyse whether the existing EU competition regulation is sufficient to address the four scenarios which may result from the use of algorithmic pricing identified by A. Ezrachi and M. Stucke in “Virtual Competition”, namely: the scenario where algorithms are used to implement pre-existing explicit collusion, the algorithm-enhanced huband-spoke, the algorithm-fuelled tacit collusion and the algorithmic tacit collusion. In order to answer the above question, this paper describes the basic concepts relevant to the subject matter of this paper – algorithms and pricing algorithms. Then, throughout the analysis of the relevant academic literature, it describes the four collusion scenarios under consideration. Further, it determines what are the provisions of EU competition law relevant for the identified collusion scenarios, and how do these provisions address the identified scenarios. This involves analysis of the EU legislation – Article 101(1) and 102 TFEU, as well as the EU and UK case law. The findings of this analysis indicate that the existing EU competition law may be able to address the first two of the abovementioned challenges, although with some limitations, but does not provide a remedy for the remaining two scenarios, namely the algorithm-fuelled tacit collusion and the algorithmic tacit collusion. However, there are other alternative measures which may be used to address the indicated issues.. 2.

(3) TABLE OF CONTENTS:. INTRODUCTION ............................................................................................................................... 4 I.. ALGORITHMIC PRICING ........................................................................................................... 6 1. ALGORITHMS ................................................................................................................................. 6 2. POSSIBLE EFFECTS OF THE USE OF PRICING ALGORITHMS ........................................................................ 8 i. Efficiencies ............................................................................................................................. 8 ii. Challenges .............................................................................................................................. 9. II.. ANTICOMPETITIVE ALGORITHMIC SCENARIOS ....................................................................... 11 1. ALGORITHMS USED TO IMPLEMENT PRE-EXISTING EXPLICIT COLLUSION ................................................... 11 2. ALGORITHM-ENHANCED HUB-AND-SPOKE ......................................................................................... 14 3. ALGORITHM-FUELLED TACIT COLLUSION AND ALGORITHMIC TACIT COLLUSION ......................................... 17 i. “Classic” tacit collusion ........................................................................................................ 17 ii. Algorithm-fuelled tacit collusion .......................................................................................... 18 iii. Algorithmic tacit collusion .................................................................................................... 19. III. HOW ARE THE CHALLENGES RAISED BY PRICING ALGORITHMS ADDRESSED BY THE EXISTING EU COMPETITION LAW? ................................................................................................................ 22 1. ALGORITHMS USED TO IMPLEMENT PRE-EXISTING EXPLICIT COLLUSION ................................................... 22 i. Article 101(1) TFEU .............................................................................................................. 23 ii. Conclusion ............................................................................................................................ 27 2. HUB-AND-SPOKE .......................................................................................................................... 28 i. Article 101(1) TFEU .............................................................................................................. 28 ii. Conclusion ............................................................................................................................ 34 3. TACIT COLLUSION .......................................................................................................................... 36 i. Article 101(1) TFEU .............................................................................................................. 36 ii. Article 102 TFEU ................................................................................................................... 39 iii. Merger control ..................................................................................................................... 43 iv. Conclusion ............................................................................................................................ 45 IV.. CONCLUSION .................................................................................................................... 47. BIBLIOGRAPHY .............................................................................................................................. 50. 3.

(4) INTRODUCTION. This paper examines the issue of algorithmic pricing from the perspective of EU competition law. The subject matter of this paper derives from the ongoing discussion among the competition law scholars and authorities all around the globe on algorithmic pricing and the way in which it may affect the market dynamics and competition. The relation between algorithms and competition law was the main subject of numerous articles, books, research papers of prominent scholars such as Salil Mehra, Ariel Ezrachi, Maurice Stucke, Nicolas Petit, as well as conferences and international roundtables, such as the OECD Roundtable on Algorithms and Collusion held in June 2017. There were also some antitrust cases in which algorithmic pricing played a significant role, discussed in the further sections of the paper. Reactions of scholars and representatives of competition authorities stem from calm to rather cautious. According to the EU Commissioner for Competition – Margrethe Vestager: “we certainly shouldn't panic about the way algorithms are affecting markets. But we do need to keep a close eye on how algorithms are developing.”1 Director General for Competition - Johannes Leitenberger, further adds that: “we should not let ourselves get overwhelmed by risks real or perceived. We should be ready to seize the opportunities that are offered by digitisation.”2 Leitenberger also argues that EU Commission is not without means, that EU’s competition law rules are “robust and flexible” and “their rationale remains sound.” 3. 1. Margrethe Vestager, ‘Speech. Algorithms and competition. Bundeskartellamt 18th Conference on Competition,. Berlin’. (16. March. 2017),. <https://ec.europa.eu/commission/commissioners/2014-. 2019/vestager/announcements/bundeskartellamt-18th-conference-competition-berlin-16-march-2017_en>, accessed: 9 January 2018. 2. Johannes. Leitenberger,. ‘Competition. at. the. digital. frontier.’. (24. April. <http://ec.europa.eu/competition/speeches/text/sp2017_06_en.pdf>, accessed: 21 January 2018, p. 1. 3. Ibid, p. 3.. 4. 2017),.

(5) In a similar vein, other available sources argue that “algorithm-based competition is NOT a thing”4 and that “to a large extent, pricing algorithms can be analysed by reference to the traditional reasoning and categories used in EU competition law.”5 On the other hand, some scholars seem a bit more sceptical as to the capability of the EU competition law to address the challenges raised by algorithmic pricing, arguing that “the real issue is designing new tools to address the new.”6 Plenty of academic literature has been devoted to the development of new tools that may be used to address the challenges raised by algorithmic pricing. This paper attempts to determine whether the existing EU competition regulation is sufficient to address the challenges raised by algorithmic pricing. While this paper does not aim to disregard or underestimate other challenges associated with algorithmic pricing, it will focus on the four collusion scenarios identified by Ezrachi and Stucke in “Virtual Competition,” described in Section II of the paper. The structure of the paper is as follows: Section I discusses the basic concepts relevant to the subject matter of this paper. It briefly describes what algorithms are, as well as identifies the possible benefits and challenges which may result from the use of algorithmic pricing by the businesses. Section II throughout the analysis of the relevant academic literature describes the four collusion scenarios, as identified by Ezrachi and Stucke and further discussed by other scholars. Section III entails a descriptive research. On the basis of considerations in Section II, this section analyses what are the provisions of EU competition law relevant for the identified collusion scenarios, and how do these provisions address the identified scenarios. Section IV concludes all the previous remarks and briefly reflects on the alternative methods of addressing the challenges raised by the use of algorithmic pricing.. 4. Thibault Schrepel, ‘Here’s why algorithms are NOT (really) a thing’ (Concurrentialiste, May 2017). <https://leconcurrentialiste.com/2017/05/15/algorithms-based-practices-antitrust/> accessed: 21 January 2018. 5. OECD, Algorithms and Collusion – Summaries of Contributions, 16 June 2017, DAF/COMP/WD(2017)2, p. 4.. 6. Ariel Ezrachi and Maurice Stucke, Virtual Competition: The Promise and Perils of the Algorithm-Driven. Economy (2016), Harvard University Press, p. 218.. 5.

(6) I.. ALGORITHMIC PRICING. This section defines the basic concept behind algorithmic pricing – the concept of an “algorithm”, and lists various types of algorithms. The second chapter identifies the main reasons for the use of algorithmic pricing by the businesses, as well as lists the main challenges associated with algorithmic pricing.. 1. Algorithms. Definition. Simple Google search provides a myriad of definitions of an algorithm, most of. which come down to the fact that in essence algorithms can be described as mathematical recipes,7 sets of specific rules or instructions,8 or structured decision-making processes that use a set of rules or procedures to automatically supply outcomes on the basis of data input and decisional parameters9. Pursuant to the definition provided in the online version of Encyclopaedia Britannica an algorithm is a “systematic procedure that produces—in a finite number of steps—the answer to a question or the solution of a problem.”10 Types of algorithms. There are many different types of algorithms, depending on their. purpose – the question or the problem they are supposed to provide the answer to. This section will outline certain types of algorithms which may be used by businesses in the process of setting prices, using the classification provided for during the OECD Roundtable on Algorithms and Collusion.11 Those include the following:. Laura Martignon, ‘Algorithm’, International Encyclopedia of the Social & Behavioral Sciences, 2nd edition,. 7. Volume 1, <https://doi.org/10.1016/B978-0-08-097086-8.43002-3>, p. 529. 8. Simonetta Vezzoso, Competition by Design (2017). Faculty of Law, Stockholm University Research Paper No.. 39. <https://ssrn.com/abstract=3075199>, p. 2. 9. Michal S. Gal and Niva Elkin-Koren, Algorithmic Consumers, Harvard Journal of Law & Technology, Volume. 30, Number 2, Spring 2017, p. 5-6. 10. Encyclopaedia Britannica, ‘Algorithm’ (13 June 2006), <https://www.britannica.com/science/algorithm>,. accessed: 10 May 2018. 11. OECD,. Algorithms. and. Collusion:. Competition. Policy. in. the. Digital. Age. <www.oecd.org/competition/algorithms-collusion-competition-policy-in-the-digital-age.htm> p. 27-32.. 6. (2017),.

(7) i.. Monitoring algorithms used for collection, screening and analysis of data;. ii.. Parallel algorithms or dynamic pricing algorithms that automatically react to any changes in market conditions and adjust prices accordingly;12. iii.. Signalling algorithms that disclose and disseminate information,;13. iv.. Self-learning algorithms that use machine learning and deep learning technologies.14 Three different types of machine algorithms may be distinguished: (i) supervised learning algorithms that use a sample of labelled data to learn a general rule, (ii) unsupervised learning algorithms that use a sample of unlabelled data to identify hidden structures and patterns, and (iii) reinforcement learning algorithms that perform tasks in dynamic environments and learn through trial and error.15 Scholars identify two general classes of learning algorithms that may be used in algorithmic pricing: (i) an estimation-optimization algorithms that estimate the environment faced by a firm and then determine what conduct performs best for that estimated environment, and (ii) reinforcement learning algorithms that figure out what action or policy function is best based on how different actions or policy functions have performed in the past.16. Pricing algorithms. Pricing algorithm may be described as a code describing how prices are. assigned to market conditions.17 In this paper, the concept of a “pricing algorithm” will be understood as encompassing any algorithms used by businesses in the process of setting prices, whether it is one of the abovementioned categories, a combination of them or even a type of algorithm not mentioned above. Other terms, such as “software,” may be used interchangeably.. 12. Ibid, p. 27.. 13. Ibid, p. 32.. 14. Ibid, p. 31.. 15. Ibid, p. 9.. 16. Joseph E. Harrington Jr, Developing Competition Law for Collusion by Autonomous Price-Setting Agents (22. August 2017), <https://ssrn.com/abstract=3037818>, p. 59. 17. Ibid, p. 59.. 7.

(8) 2. Possible effects of the use of pricing algorithms. i.. Efficiencies. Efficiencies. There are many reasons why businesses might want to switch to algorithmic. pricing. Although at the current state of development computers cannot fully substitute human employees in all fields, there are many things that they can improve or even enable. Optimisation of business processes. As noted by OECD, algorithms allow for optimisation. of business processes.18 Businesses can use algorithms to substitute or assist human employees in many routine tasks, which computers can simply do faster and on a larger scale. This is even more significant now in the age of Big Data when companies are gathering more and more various data about their customers, market characteristics, and competitors. Algorithms allow to monitor, gather and analyse immense amounts of both historical and real-time data, recognise patterns and make predictions in speed and scope unattainable to humans. The speed with which computers operate, allows to implement dynamic pricing strategies, better target their customers through various degrees of price discrimination, or to immediately react to their competitors’ behaviour or their own changes in costs. Furthermore, as pointed out by some scholars, algorithms increase the consistency of decision-making – unlike humans, they always make the same decisions on the same dataset.19 As argued by Ezrachi and Stucke algorithms are unlikely to be affected by human biases.20 Artificial Intelligence. Self-learning. The self-learning algorithms are not only following the. instructions provided by humans, but they are learning how to solve problems of their own.21 Through constant learning on the basis of trials and errors, the self-learning algorithms can develop strategies and analyse scenarios which would take years for humans to come up with. 18. OECD, Algorithms and Collusion: Competition Policy in the Digital Age (n 11), p. 11.. 19. Robert Sloan, and Richard Warner, When is an Algorithm Transparent? Predictive Analytics, Privacy and Public. Policy, p. 4 20. OECD, Algorithmic Collusion: Problems and Counter-Measures - Note by A. Ezrachi & M. E. Stucke (21 June. 2017), DAF/COMP/WD(2017)25, <https://one.oecd.org/document/DAF/COMP/WD(2017)25/en/pdf>, accessed: 9 January 2018, para. 11. 21. Andrew McAfee and Erik Brynjolfsson, Machine, Platform, Crowd: Harnessing Our Digital Future (2017),. W.W. Norton&Company, p. 17.. 8.

(9) Research shows that they can go beyond the input – beyond what humans “taught” them, beyond the human capabilities, in areas where human expertise is lacking.22 This may allow the businesses to implement possible pricing strategies that humans simply cannot execute.23 ii.. Challenges Notwithstanding that the use of algorithmic pricing (and algorithms in general) seems to. create many benefits for businesses, it may also create some problems. There is no empirical evidence of how the algorithms will affect the competition in real markets,24 it may be that in fact they do more good than bad. Nevertheless, analysis of the literature available on this matter allows to identify three main challenges associated with algorithmic pricing. Facilitators of anticompetitive conduct. First, algorithms may facilitate anticompetitive. conduct.25 Thanks to their ability to analyse immense amounts of data almost in real-time, algorithms may facilitate and/or magnify the implementation of a pre-existing explicit collusion (e.g. through the use of monitoring algorithms allowing cartels to monitor the deviations from fixed price and retaliate quickly), or facilitate reaching a tacit collusion (e.g. through the use of dynamic pricing algorithms allowing competitors to adjust their prices automatically to rapidly changing market conditions). In other words, “if the goal is to do bad things, automated systems and algorithms could help you do bad things faster.” 26 Impact on market dynamics and competition. Second, algorithms may change the market. dynamics and competition “by creating incentives and mechanisms to collude that did not exist otherwise.”27 The risk is that the use of algorithms makes the market concentration – the number. 22. David Silver, Julian Schrittwieser, and others, Mastering the game of Go without human knowledge (19 October. 2017), Nature, Volume 550, <https://www.nature.com/articles/nature24270>, p. 354. 23. David. J.. Lynch,. Policing. the. digital. cartels. (8. January. 2017),. The. Financial. <https://www.ft.com/content/9de9fb80-cd23-11e6-864f-20dcb35cede2>, accessed: 21 January 2018. 24. OECD, Algorithms and Collusion: Competition Policy in the Digital Age (n 11), p. 33.. 25. Ibid.. 26. Lynch (n 23).. 27. OECD, Algorithms and Collusion: Competition Policy in the Digital Age (n 11), p. 33.. 9. Times,.

(10) of competitors in the market, less relevant,28 thus allowing for collusion in markets where collusion was not considered a significant risk before, e.g. in non-oligopolistic markets.29 Algorithmic collusion. Third, algorithms may enable the so-called “algorithmic collusion” –. new forms of coordination that were not observed or even possible before.30 Some may call it a “legal sci-fi”,31 but considering the recent developments in machine learning, that kind of approach seems ignorant. Just a year ago people thought it would take at least decades for the artificial intelligence (“AI”) to be so developed as to win with humans at such a complex and intuitive game as Go. However, last year AlphaGo, a programme developed by DeepMind Technologies, proved everyone wrong and won 4 out of 5 games with an 18-time world champion in Go, by using very creative and inventive moves.32 With that in mind, is it really too far-fetched to consider that algorithms will be able to create new forms of co-ordination not previously known to humans? Reinforcement learning algorithms through endless trial and errors may be able to develop new strategies which up to then were either overlooked by humans or not possible without the combination of algorithmic decision-making and the use of Big Data.. 28. Ibid, p. 21.. 29. Ibid, p. 33.. 30. Ibid, p. 18-19.. 31. Thibault Schrepel, Here’s why algorithms are NOT (really) a thing, Concurrentialiste, May 2017,. <https://leconcurrentialiste.com/2017/05/15/algorithms-based-practices-antitrust/>, accessed: 21 January 2018. 32. See: Gary Krieg, Kevin Proudfoot, Josh Rosen (producers), Greg Kohs (director), AlphaGo (motion picture),. USA, 2017.. 10.

(11) II.. ANTICOMPETITIVE ALGORITHMIC SCENARIOS. This section will describe and analyse the different collusion scenarios identified by Ezrachi and Stucke in their eye-opening book “Virtual Competition”, and then further discussed and developed by other scholars. The authors listed four different collusion scenarios that may result from the use of pricing algorithms, namely: algorithms used to implement pre-existing collusion (“The Messenger Scenario”), hub-and-spoke agreements, algorithm-fuelled tacit collusion (“The Predictable Agent”), and algorithmic tacit collusion (“Digital Eye”).33. 1. Algorithms used to implement pre-existing explicit collusion. The first anticompetitive scenario entails the use of algorithms to implement pre-existing explicit collusion. In principle, “humans agree to collude and machines execute the collusion, acting as mere intermediaries or messengers.”34 In this scenario, algorithms are used as a tool to execute an agreement between humans. This scenario may entail any type of anticompetitive agreement – both horizontal and vertical. Horizontal agreements. In case of horizontal agreements, the use of algorithmic pricing may. have the following effects. First, by increasing the market transparency and facilitating the exchange of information, the use of algorithms (e.g. monitoring and signalling algorithm) may facilitate the implementation of price fixing agreements (cartels).35 Second, by facilitating the monitoring of deviations and increasing the speed of detection of and retaliation to the deviation, it may limit the incentives for cartel members to deviate from the fixed price, and thus make cartels more stable.36 Third, following the conclusion that the algorithms make. 33. Ariel Ezrachi and Maurice Stucke, Virtual Competition: The Promise and Perils of the Algorithm-Driven. Economy (2016), Harvard University Press, p. 36-37. 34. Freshfields. Bruckhaus. Deringer,. ‘Pricing. algorithms:. the. digital. collusion. scenarios’,. <https://www.freshfields.com/globalassets/our-thinking/campaigns/digital/mediainternet/pdf/freshfields-digital--pricing-algorithms---the-digital-collusion-scenarios.pdf>, accessed: 9 January 2018, p. 1. 35. European Commission, Guidelines on the applicability of Article 101 of the Treaty on the Functioning of the. European Union to horizontal co-operation agreements, (2011/C 11/01), (“Horizontal Guidelines”) para. 59. 36. Commission Staff Working Document, Preliminary Report on the E-commerce Sector Enquiry, 15 September. 2016, SWD(2016)312 final, p. 175. 11.

(12) collusion less dependent on the number of competitors in the market,37 it could be argued that they may lead to the emergence of cartels in markets where otherwise they would not be as successful. Even though the discussions about pricing algorithms and competition law emerged only in recent years, EU Commission’s inquiry into the e-commerce sector shows that about half of the retailers track online prices of their competitors, out of which 67% use automatic software programmes for that purpose. 8% of the latter also use automatic price adjustment software, and 27% use both software and manual methods.38 Other reports suggest that algorithmic pricing is not only limited to online environments. Thanks to the use of electronic price tags it is also used by the brick-and-mortar retailers.39 Vertical agreements. The use of algorithms may help execute various types of vertical. restrictions, starting from retail price maintenance (“RPM”) to most-favoured nation-clauses (“MFNs”). As noted by the EU Commission in the submission to the Roundtable on Algorithms and Collusion, in case of RPM, the use of algorithms may have threefold effects. RPM. First, price monitoring algorithms may facilitate detection of retailers’ deviations from a. fixed or minimum resale price by the manufacturers,40 therefore making the price fixing more effective.41 According to the data acquired by the Commission during the E-commerce Sector Enquiry, approximately 18% of retailers reported that manufacturers monitor their retail prices.42 Two-thirds of the manufacturers admitting to monitoring retail prices use manual tracking, second most popular method is the use of the price-tracking software.43 The increased accessibility and popularity of algorithms and the ease and speed of monitoring with the use of algorithms might encourage more manufacturers to resort to a practice of RPM.. 37. OECD,. Algorithms. and. Collusion:. Competition. Policy. in. the. Digital. Age. (2017),. <www.oecd.org/competition/algorithms-collusion-competition-policy-in-the-digital-age.htm>, p. 21. 38. European Commission, ‘Preliminary Report on the E-commerce Sector Enquiry’ (n 36), p. 174-175.. 39. Oxera, When algorithms set prices: winners and losers. Discussion paper, 19 June 2017, p. 7. 40. OECD,. Algorithms. and. Collusion. –. Note. from. the. European. Union,. 21-23. DAF/COMP/WD(2017)12, paras 13-14. 41. European Commission, Guidelines on Vertical Restraints (2010/C 130/01), para. 48.. 42. European Commission, ‘Preliminary Report on the E-commerce Sector Enquiry’ (n 36), p. 170.. 43. Ibid, p. 171.. 12. June. 2017,.

(13) Second, the use of algorithms may lead to turning a recommended price into a fixed resale price.44 Under the EU competition law, recommended prices set by the suppliers are not a hardcore restriction of Article 101(1) TFEU45, provided that they do not amount to a fixed or minimum sale price as a result of pressure from, or incentives offered by, any of the parties, and provided that they are within the set market share thresholds.46 As noted by the EU Commission, the use of monitoring algorithms might disincentive the retailers to deviate from the recommended price, and essentially turn it to a fixed resale price.47 Third, widespread use of monitoring algorithms may lead to the adherence to the fixed or minimum resale prices of one manufacturer by seemingly unrelated market participants. Where one retailer complies with the price set by a manufacturer, other retailers who monitor him through algorithms may adjust their prices accordingly, thus spreading the fixed or minimum resale prices all over the market.48 In consequence, this might lead to increase of the market price. MFNs. As to the MFNs, similarly as in case of RPM, the use of algorithms may facilitate. monitoring of compliance with the clauses. As suggested by Ezrachi, this may lead to “industrywide price uniformity.”49 As noted by the EU Commission even without algorithmic pricing, the use of MFNs may make price fixing more effective, as well as make maximum or recommended price work as fixed resale price.50 Facilitation of monitoring of compliance with MFNs may amplify that effect.. 44. OECD, Algorithms and Collusion – Note from the European Union, (n 40), para. 15.. 45. Consolidated versions of the Treaty on European Union and the Treaty on the Functioning of the European. Union, OJ C326 , 26 October 2012 P.0001–0390 (‘TFEU’). 46. Commission Regulation (EU) No 330/2010 of 20 April 2010 on the application of Article 101(3) of the Treaty. on the Functioning of the European Union to categories of vertical agreements and concerted practices, art. 4(a). 47. OECD, Algorithms and Collusion – Note from the European Union, (n 40), para. 15.. 48. Ibid., para. 16.. 49. Ariel Ezrachi, The competitive effects of parity clauses on online commerce, European Competition Journal,. 2015, Volume 11, Nos. 2-3, p. 499. 50. European Commission, Guidelines on Vertical Restraints (2010/C 130/01), para. 48.. 13.

(14) 2. Algorithm-enhanced hub-and-spoke. This chapter will describe the second scenario identified by Ezrachi and Stucke – the algorithmenhanced hub-and-spoke. “Classic” hub-and-spoke. The hub-and-spoke scenario is unique neither to the online. environment or antitrust.51 Hub-and-spoke collusion entails two elements: a vertical one – a series of single vertical agreements between each of the competitors and a third party (e.g. a supplier, or retailer), and a horizontal one – an indirect exchange of information between the competitors. Although this is not a new concept, it has not yet been defined under EU law. However, UK case law provides a legal test for a finding of a hub-and-spoke collusion. It involves five conditions: (1) competitor A discloses to supplier its future pricing intentions; (2) company A may be taken to intend or foresee that the supplier will pass that information to other companies that may be A’s competitors; (3) the supplier passes that information to A’s competitor; (4) A’s competitor may be taken to know the circumstances in which the information was disclosed to the supplier by company A; (5) A’s competitor uses the information in determining its own future pricing intentions.52 The above structure is presented in Figure 1. The exchange of information may also be reciprocal (as in Figure 2.).53 Fig. 1.. Fig. 2. THIRD PARTY. COMPETITOR A. THIRD PARTY. COMPETITOR B. COMPETITOR A. COMPETITOR B. 51. Ezrachi and Stucke, Virtual Competition (n 33), p. 46-47.. 52. Tesco Stores Ltd, Tesco Holdings Ltd and Tesco Plc v OFT (2012), Competition Appeal Tribunal (“CAT”),. 1188/1/1/11, [2012] CAT 31, para. 57. 53. Argos Limited & Littlewoods Limited v OFT (2006), Court of Appeal, 2005/1071, 1074 and 1623, [2006] EWCA. Civ 1318, para. 141.. 14.

(15) Algorithm-enhanced hub-and-spoke. According to Ezrachi and Stucke, one of the possible. effects of the algorithmic pricing is that, as the creation and improvement of algorithms is costly, instead of investing in their own pricing algorithm, companies may resort to outsourcing their pricing to an algorithm of an upstream supplier.54 What distinguishes “classic” hub-andspoke from this scenario is that under the latter the horizontal collusion may be the effect, but not necessarily the aim of competitors’ action.55 This scenario can be observed at two levels: output – algorithm, and input – data.56 The first scenario (presented in Figure 3.), which Ezrachi and Stucke describe as a “de-facto hub-andspoke structure”,57 entails the use of the same algorithm or identical algorithms by the competitors, but no exchange of the data input.58 Consequently, as competitors pricing will be determined through the same set of rules, their market behaviour may become aligned, even without any direct interaction between them.59 Fig. 3. THIRD PARTY. COMPETITOR A. COMPETITOR B. In the second scenario (Figure 4.), which resembles the “classic” hub-and-spoke more, the alignment results from the use of the same algorithm and data to determine the price. It entails each of the competitors providing the algorithm with data, which then the algorithm uses to determine the prices for each of the competitors. Ezrachi and Stucke argue that by outsourcing. 54. Ezrachi and Stucke, Virtual Competition (n 33), p. 47-48.. 55. Ibid, p. 48.. 56. OECD, Algorithmic Collusion: Problems and Counter-Measures - Note by A. Ezrachi & M. E. Stucke (21 June. 2017), DAF/COMP/WD(2017)25, <https://one.oecd.org/document/DAF/COMP/WD(2017)25/en/pdf>, para. 32. 57. Ibid.. 58. Ezrachi and Stucke, Virtual Competition (n 33), p. 47-48.. 59. Ibid.. 15.

(16) their pricing to an algorithm supplier who promises to maximise their profits, the competitors surely know that the algorithm will make use of their own and their competitors’ information in assessing the prices.60 Fig. 4. THIRD PARTY’S ALGORITHM. COMPETITOR A. 60. COMPETITOR B. Ibid, p. 49.. 16.

(17) 3. Algorithm-fuelled tacit collusion and algorithmic tacit collusion. Scope. The next two collusions scenarios to be discussed are the scenarios identified by. Ezrachi and Stucke as “The Predictable Agent” and “Digital Eye”.61 They will be further referred to as algorithm-fuelled/enhanced tacit collusion and algorithmic tacit collusion respectively. Although they are in fact two separate scenarios, due to both of them including tacit collusion, and the second being the amplified version of the first one, they will be analysed under one chapter. i.. “Classic” tacit collusion. Tacit collusion. Similarly to the hub-and-spoke scenario, tacit collusion is a concept which is. neither new nor unique to algorithmic competition. It is interchangeable with concepts such as: “tacit coordination”, “coordinated effects” and “conscious parallelism”.62 Essentially, tacit collusion occurs where under certain market conditions the competitors are “able to behave in a parallel manner and derive benefits from their collective market power without, or without necessarily, entering into an agreement or concerted practice in the sense of Article 101 TFEU.”63 This concept should be distinguished with those of “oligopolistic interdependence” or “unconscious parallelism” which also describe parallel behaviour between the companies. In brief, the concept of oligopolistic interdependence covers not only the collusive oligopolies (e.g. those prone to tacit collusion), but also the non-collusive ones, where unconscious parallelism may occur.64 As explained by Mezzanotte, unconscious parallelism occurs where oligopolists “dampen the incentive to maximize short-term profits”, thus dampening or muting. 61. Ezrachi and Stucke, Virtual Competition (n 33), p. 36-37.. 62. Richard Whish and David Bailey, Competition law, Eighth Edition, Oxford University Press, 2015, p. 594.. 63. Ibid, p. 594.. 64. Nicolas Petit, The „Oligopoly Problem” in EU Competition Law, <http://ssrn.com/abstract=1999829>, p. 12,. 22-24.. 17.

(18) competition, but without colluding.65 Market conditions. Tacit collusion requires the following market conditions, typical for. oligopolistic markets, to occur and be sustainable: „First, the coordinating firms must be able to monitor to a sufficient degree whether the terms of coordination are being adhered to. Second, discipline requires that there is some form of credible deterrent mechanism that can be activated if deviation is detected. Third, the reactions of outsiders, such as current and future competitors not participating in the coordination, as well as customers, should not be able to jeopardise the results expected from the coordination.”66 ii.. Algorithm-fuelled tacit collusion. Algorithm-fuelled tacit collusion. According to Ezrachi and Stucke under certain market. conditions an industry-wide use of pricing algorithms unilaterally developed by each of the companies may lead to tacit collusion.67 This scenario entails each of the companies developing a profit maximising algorithm, which will i.a. monitor price changes and react to them accordingly.68 In this scenario there is no agreement or cooperation between the parties – it is the result of a series of unilateral decisions by each of the parties to use a pricing algorithm.69 Importantly, this scenario entails anticompetitive intent – each of the firms is aware that the industry-wide use of algorithmic pricing may facilitate tacit collusion.70 Impact of algorithms on market conditions. Ezrachi and Stucke argue that the industry-. wide use of pricing algorithms may change the market conditions and allow for tacit collusion in markets where it did not previously occur.71 There are several reasons for that, including the following. First, as algorithms are unlikely to be affected by human biases, they may increase. 65. Felix Mezzanotte, Using Abuse of Collective Dominance in Article 102 TFEU to Fight Tacit Collusion: The. Problem of Proof and Inferential Error, World Competition 33, no. 1 (2010), p. 79, 87. 66. European Commission, Guidelines on the assessment of horizontal mergers under the Council Regulation on. the control of concentrations between undertakings (2004/C 31/03), (‘Horizontal Merger Guidelines’) para. 4. 67. Ezrachi and Stucke, Virtual Competition (n 33), p. 56.. 68. Ibid, p. 61.. 69. Ibid, p. 59.. 70. Ibid, p. 56.. 71. Ibid, p. 60.. 18.

(19) the stability of the market necessary for tacit collusion.72 Second, the use of pricing algorithms increases market transparency, due to digitalisation and accessibility of market data.73 This will allow the algorithms to study the pricing patterns – “engage in ‘predictive analytics”,74 analyse market data and react to changes in market conditions. Furthermore, thanks to the availability of data and the algorithm’s ability to monitor immense amounts of it in a short period of time, as noted by OECD, algorithms may allow for monitoring of and adjustment to much larger number of competitors75 and enable collusion in markets where it was not considered a significant risk before (e.g. in non-oligopolistic markets).76 Third, as pricing algorithms will increase the speed of price monitoring and adjustment,77 they will also increase the speed of retaliation.78 As pointed out by Ezrachi and Stucke companies will have less incentive to discount, as they will not gain anything from price discounting, because their competitors may match their price (retaliate) even before the customers notice the discount.79 Therefore, pricing algorithms may constitute a deterrent mechanism for price discounting.80 iii.. Algorithmic tacit collusion. Algorithmic tacit collusion. According to Ezrachi and Stucke two technological. advancements will allow to amplify the scenario of algorithm-enhanced tacit collusion and expand it “beyond price, oligopolists markets, and easy detection.” First, it is the computers being able to process high volumes of data in real-time, thus giving the firms “a God-like view. 72. Ibid, para. 11.. 73. Ibid, p. 61.. 74. Ibid, p. 61.. 75. OECD, Algorithms and Collusion: Competition Policy in the Digital Age (n 37), p. 21.. 76. Ibid, p. 33.. 77. Ezrachi and Stucke, Virtual Competition (n 33), p. 62.. 78. OECD, Algorithmic Collusion: Problems and Counter-Measures - Note by A. Ezrachi & M. E. Stucke, (n 56),. para. 9. 79. Ezrachi and Stucke, Virtual Competition (n 33), p. 63, 65.. 80. OECD, Algorithmic Collusion: Problems and Counter-Measures - Note by A. Ezrachi & M. E. Stucke, (n 56),. para. 9.. 19.

(20) of the marketplace”, and second, the use of Artificial Intelligence (“AI”).81 There are several assumptions behind this scenario, differentiating it from that of algorithm-enhanced tacit collusion. Transparency. Ezrachi and Stucke assume that each firm will have the ‘God-like view’ of the. market resulting either from the fact that no company without it could remain on the market, or that companies will share this technology knowing that it will ultimately lead to tacit collusion.82 The increased transparency will discourage deviation, and may even lead to computers being able to anticipate and react to a price change before it occurs,83 or to assess whether its rival’s actions were mistakes.84 Moreover, it may be that people will not even be aware of the ongoing tacit collusion.85 No anticompetitive intent. In this scenario the algorithms are also set at profit maximisation. and avoiding any illegal activity. The difference is that unlike in the previous scenario, here the algorithm is independent, it is designed to reach the set goal by identification of the optimal strategy to reach it through of market observation.86 It is not commanded to collude (either tacitly or explicitly).87 There is no anticompetitive intent at the human level; humans may not even be aware of the ongoing tacit collusion. At best they may be aware of the possibility of such result, but not necessarily of its probability.88 In this context, it is worth mentioning that this scenario also does not entail any express collusion between computers.89 Artificial Intelligence. Unlike the previous scenario, this one involves widespread use of. Artificial Intelligence. As pointed out by Ezrachi and Stucke, it is not the human intent, but the use of self-learning algorithms combined with enhanced market transparency that is the source of collusion in this scenario.90. 81. Ezrachi and Stucke, Virtual Competition (n 33), p. 71.. 82. Ibid, p. 73.. 83. Ibid, p. 72-73.. 84. Ibid, p. 75.. 85. Ibid, p. 71.. 86. Ibid, p. 74.. 87. Ibid, p. 74, 78.. 88. Ibid, p. 78.. 89. Ibid.. 90. Ibid.. 20.

(21) In the context of the two analysed scenarios, it should be noted that currently there is no empirical evidence, no case law in which algorithms were the cause of tacit collusion, or where tacit collusion occurred without the human factor. It is also not quite clear how machine learning algorithms may reach a collusive outcome.91 It may as well be that both of the above scenarios are purely hypothetical, one may even call it a “legal sci-fi.”92. 91 92. OECD, Algorithms and Collusion: Competition Policy in the Digital Age (n 37), p. 31. Thibault Schrepel, Here’s why algorithms are NOT (really) a thing, Concurrentialiste, May 2017,. <https://leconcurrentialiste.com/2017/05/15/algorithms-based-practices-antitrust/>, accessed: 21 January 2018.. 21.

(22) III.. HOW. ARE. THE. CHALLENGES. RAISED. BY. PRICING. ALGORITHMS. ADDRESSED BY THE EXISTING EU COMPETITION LAW?. This section of the paper entails a descriptive research. On the basis of the considerations made in Section II, it analyses what are the provisions of EU competition law relevant for the identified collusion scenarios, and how do those provisions address those scenarios. Consequently, this section attempts to answer the question whether the currently existing EU competition law is a sufficient tool to address the problems listed in Section II above.. 1. Algorithms used to implement pre-existing explicit collusion. Scope. This chapter analyses whether and how does the existing EU competition law address. the scenario where algorithms are used to implement pre-existing explicit collusion. Analysis of this scenario conducted in Section II above leads to one simple conclusion. Under this scenario the algorithms do not create new anticompetitive agreements, they simply facilitate and amplify the execution of the anticompetitive agreements already known to and addressed by the EU competition law, by replacing or supplementing the previous ways to e.g. communicate with competitors or monitor them. It is without a doubt that both the horizontal and vertical anticompetitive agreements invoked in Section II fall within the scope of Article 101(1) TFEU. This chapter determines whether the involvement of algorithms to implement such agreements changes that. It also describes a “real-life” example of how the scenario under consideration was addressed by one of the European competition authorities - Competition and Markets Authority (“CMA”).. 22.

(23) i.. Article 101(1) TFEU. Liability. As most available sources confirm, with regard to liability, anticompetitive practices. under this scenario can be addressed by the already existing competition law.93 Article 101(1) TFEU prohibits „all agreements between undertakings, decisions by associations of undertakings and concerted practices which may affect trade between Member States and which have as their object or effect the prevention, restriction or distortion of competition within the internal market.” It does not refer in any way to the means that the undertakings use to implement the agreements or concerted practices. In fact, to fall within the scope of Article 101(1) TFEU, an agreement does not actually have to be implemented.94 Given that the competition authorities can establish the existence of an anticompetitive agreement, the issue of whether the companies actually implement it and what means do they use to implement it – algorithms, or personal acts of companies’ employees, does not affect that finding. As rightly pointed out by OECD, algorithms should be assessed together with the infringement they facilitate or execute.95 Where the agreement that algorithms implement is prohibited under Article 101(1) TFEU, the use of algorithms does not make the conduct any more or any less illegal.96. 93. i.a.. OECD,. Algorithms. and. Collusion:. Competition. Policy. in. the. Digital. <www.oecd.org/competition/algorithms-collusion-competition-policy-in-the-digital-age.htm>,. Age p.. 33;. (2017), Ariel. Ezrachi and Maurice Stucke, Virtual Competition: The Promise and Perils of the Algorithm-Driven Economy (2016), Harvard University Press, p. 39; Maureen Ohlhausen., Should We Fear The Things That Go Beep In the Night? Some Initial Thoughts on the Intersection of Antitrust Law and Algorithmic Pricing (23 May 2017), p. 2. 94. See i.a.: Horizontal Guidelines, para. 29, PVC (Case IV/31.865) Commission Decision 94/599/EC (1994), OJ. L239/14, para. 30. 95. OECD, Algorithms and Collusion: Competition Policy in the Digital Age (n 93), p. 33.. 96. As in: Ohlhausen (n 93), p. 11. 23.

(24) Investigation. On the other hand, many scholars argue that this scenario creates some new. challenges when it comes to detection and evidentiary techniques.97 However, as to detection of e.g. cartels, it could also be argued to the contrary. In the EU, cartels are discovered in various ways: through leniency applications by the cartel members, whistleblowing, individual complaints, as well as active detection by the authorities. Analysis of EU Commission’s decisions in which it imposed fines for cartel infringements, leads to a conclusion that the single most powerful discovery tool is, in fact, the leniency programme. The number of decisions in which immunity was granted under the European Commission's leniency programme to the first cooperating undertaking constituted over 75% of analysed decisions adopted within the 2001-2015 period, and over 91% in the 2011-2015 period.98 With that in mind, it is difficult to understand which of the potential effects as identified in Sections I and II above, could affect the “popularity” of leniency programme. It may be, as argued by Ezrachi and Stucke, that by increasing the distance between the human and the illegal conduct, computers may reduce the guilt.99 Although there is no empirical data showing whether guilt constitutes an important incentive for leniency applicants, it seems like the possibility to avoid or decrease fines might be more relevant in incentivising companies to file a leniency application. As to the detection techniques, the use of algorithms poses one important challenge: unlike humans, pricing algorithms will not leave evidence of communication. On the other hand, some authors note that, unlike in case of human collusion, the price-determining strategy is actually written down in the algorithm’s code, and thus, in principle, it can be accessed.100 Nevertheless, taking into account that the scenario under consideration involves prior explicit agreement on the human level, the competition authorities might confine to using the current evidentiary. 97. i.a.: OECD, Algorithms and Collusion: Competition Policy in the Digital Age (n 93), p. 33; Ohlhausen (n 93),. p. 2; Thibault Schrepel, Here’s why algorithms are NOT (really) a thing, Concurrentialiste, May 2017, <https://leconcurrentialiste.com/2017/05/15/algorithms-based-practices-antitrust/>, accessed: 21 January 2018. 98. Wouter P.J. Wills, The Use of Leniency in EU Cartel Enforcement: An Assessment after Twenty Years, World. Competition, Volume 39, Issue 3, September 2016, p. 9-10. 99. Ezrachi and Stucke, Virtual Competition (n 93), p. 42.. 100. Joseph E. Harrington Jr, Developing Competition Law for Collusion by Autonomous Price-Setting Agents (22. August 2017) < https://ssrn.com/abstract=3037818>, p. 2-3.. 24.

(25) techniques to prove the existence of an anticompetitive agreement (as was done in the case described below). Case law. Example of the use of pricing algorithms to implement pre-existing collusion can. be found in the Online sales of posters and frames case investigated and sanctioned by the CMA,101 where two Amazon UK posters and frames sellers (GBE and Trod) agreed not to undercut each other’s prices in certain circumstances.102 Initially, the parties implemented the agreement manually, but after a while they agreed that “logistically it is going to be difficult to follow the pricing effectively on a daily basis” and resorted to the use of automated repricing software configured to give effect to their agreement.103 GBE configured its software to undercut competing products, but match Trod’s price where there was no cheaper third party seller. Trod commanded its software to undercut competitors while ignoring GBE’s prices.104 CMA’s investigation commenced following a leniency application by GBE.105 CMA relied on the UK competition law act and the EU case law and concluded that the arrangement between the parties amounts to a prohibited agreement and/or concerted practice.106 Importantly, CMA’s assessment of the parties’ liability was no different than in cases not involving algorithms. CMA based its assessment on evidence of communication between the parties (mostly e-mails), as well as internal e-mails of the parties’ employees. As to the analysis of the software used by the parties, CMA confined to evidence in the form of e-mail communication regarding setting up the algorithm, as well as witness statements of the software provider.107 CMA did not actually analyse or inspect the software itself.. 101. Online sales of posters and frames, Case 50223, Decision of the CMA (12 August 2016).. 102. Ibid, para. 3.45.. 103. Ibid, paras 3.46, 3.62-3.63.. 104. Ibid, paras 3.77-78, 3.81.. 105. Ibid, p. 5.. 106. Ibid, p. 77.. 107. Ibid, p. 19-44 (paras 3.45-3.102).. 25.

(26) Inspecting the algorithm. Some changes in the investigation techniques might be necessary,. should a competition authority need to inspect the algorithm, e.g. to determine the duration of an infringement, or whether each of the participants took part in the agreement, or what was the impact of the infringement on the market. Article 20(1)(b) of Regulation 1/2003108 regarding the EU Commission’s powers of inspection gives the Commission the power „to examine the books and other records related to the business, irrespective of the medium on which they are stored.” This includes the power to examine electronic information, search the IT-environment (servers, computers, and other mobile devices) and all storage media of the inspected undertaking. The EU Commission may also use its own dedicated software and hardware for investigation.109 There is no doubt that the EU Commission is empowered to inspect companies’ software/algorithms. The issue is whether it is actually possible – whether the national competition authorities have the means, expertise, time and money to do it. The difficulties in this context arise from the nature of algorithms – their complexity and opacity.110 Even if the investigators have the access to algorithm’s source code, it does not necessarily mean they will be able to understand it and predict the effect of the use of such an algorithm.111 Another issue is the access to the dataset used by the algorithm. Ezrachi and Stucke propose that algorithms could be audited in a “sandbox” or a “collusion incubator”, where their effects on the market could be observed. The two scholars acknowledge that this would require a high level of expertise.112. 108. Council Regulation (EC) No 1/2003 of 16 December 2002 on the implementation of the rules on competition. laid down in Articles 81 and 82 of the Treaty, L001, 4 January 2003, P.0001-002. 109. European Commission, Explanatory note on Commission inspections pursuant to Article 20(4) of Council. Regulation No 1/2003, 11 September 2015, paras 9-10. 110. Simonetta Vezzoso, Competition by Design (2017). Faculty of Law, Stockholm University Research Paper No.. 39. <https://ssrn.com/abstract=3075199>, p. 12. 111. Robert Sloan, Richard Warner, When is an Algorithm Transparent? Predictive Analytics, Privacy and Public. Policy, p. 5. 112. OECD, Algorithmic Collusion: Problems and Counter-Measures - Note by A. Ezrachi & M. E. Stucke (21 June. 2017), DAF/COMP/WD(2017)25, <https://one.oecd.org/document/DAF/COMP/WD(2017)25/en/pdf>, para. 86.. 26.

(27) ii.. Conclusion The above considerations show that the involvement of algorithms in the analysed scenario. in no way changes that the explicitly colluding parties may be held liable for agreements infringing Article 101(1) TFEU. The pricing algorithms here are used as a tool to implement or execute the main infringement, and it is that infringement that triggers the application of Article 101(1). On the other hand, the use of algorithmic pricing may cause some problems in the context of investigation of the infringement. While the involvement of pricing algorithms does not seem to have any correlation with some discovery techniques such as the Commission’s Leniency Programme or whistleblowing, it may affect active detection by the authorities. Although not all cases may require an actual inspection of the algorithm itself, in cases where this will be necessary it may put quite a burden on the authorities. Auditing the algorithms requires not only a high level of expertise in computer science that competition authorities might be lacking, but also significant time and costs to carry out such an investigation.. 27.

(28) 2. Hub-and-spoke. This chapter considers whether and how does the current EU competition law address the algorithm-enhanced hub-and-spoke scenario. As discussed in Section II above, there are two variants of this scenario – one that results from the use of identical algorithms to determine prices, and one that results from the use of the same algorithm and dataset to determine prices. The following subsection discusses whether Article 101(1) TFEU may be used to address the “classic” hub-and-spoke structure, and whether the involvement of algorithms changes the conclusion to that. i.. Article 101(1) TFEU. Article 101(1) TFEU. Although the concept of hub-and-spoke has not been defined under EU. law, a “classic” hub-and-spoke scenario may be addressed under Article 101(1) TFEU, as information exchange between competitors. As indicated in the Horizontal Guidelines, information exchange between competitors may be direct or indirect.113 To fall within the scope of Article 101(1) TFEU information exchange must establish or be part of an agreement, a concerted practice or a decision by an association of undertakings.114 Concerted practice. Under the EU law, a concerted practice is defined as “a form of. coordination between undertakings by which, without it having reached the stage where an agreement properly so-called has been concluded, practical cooperation between them is knowingly substituted for the risks of competition.”115 Information exchange can amount to a concerted practice if the exchanged information is strategic, “if it reduces strategic uncertainty in the market thereby facilitating collusion.”116 Importantly, analysis of EU case law allows for a conclusion that under Article 101(1) TFEU both the spokes and the hub may be held liable for a concerted practice.. 113. Horizontal Guidelines, p. 55.. 114. Ibid, p. 60.. 115. Ibid, p. 60.. 116. Ibid, p. 61.. 28.

(29) The hub as a ‘collusion facilitator’. As to the liability of the hub, there is one CJEU’s. judgment that deserves attention. In light of CJEU’s judgment in AC-Treuhand v European Commission,117 the hub could be held liable as a collusion facilitator – a company acting as an intermediary in an infringement executed by other companies active in a different market than the intermediary. This was recently confirmed in Icap Plc v. Commission.118 An intermediary company may be held liable for infringement of Article 101(1) TFEU, under the following conditions. First, the intermediary must have been aware of the conduct planned or put into effects by the competitors or must have been able to reasonably foresee it and was prepared to take the risk. Second, the intermediary must have intended to contribute by its own conduct to the common objectives pursued by the competitors, to the implementation and continuation of anticompetitive conduct. Third, the intermediary must have actually contributed to or facilitated the anticompetitive conduct pursued by the competitors. The participation in the infringement can be either active or passive. The Court held that a mere “presence of an undertaking in meetings at which anticompetitive agreements were concluded, without that undertaking clearly opposing them” is in itself indicative of collusion, as it “encourages the continuation of the infringement and compromises its discovery.”119 It should also be noted that in AC-Treuhand CJEU referred to the “essential role” that the intermediary had in the infringements and to the fact that the “very purpose” of the services provided by the intermediary was “the attainment, in full knowledge of the facts, of the anticompetitive objectives in question.”120 Some scholars wondered whether this is just CJEU’s description of the facts of the case or whether those conditions are part of a legal test for determining the liability of the cartel facilitators.121 Then, in Icap Plc v. Commission, CJEU referred to Icap’s participation as “significant” as it made it possible to amplify the infringement to a much greater extent, and argued that Icap “should have expected, if necessary after taking appropriate legal advice, its conduct to be declared incompatible with the EU competition. 117. Case C-194/14 P, AC-Treuhand AG v European Commission (2015), EU:C:2015:717. . 118. Case T-180/15, Icap plc and Others v European Commission (2017), ECLI:EU:T:2017:795.. 119. AC-Treuhand AG (n 117), paras 30-31.. 120. Ibid, paras 37-38.. 121. Gianni De Stefano, AC-Treuhand Judgment: A Broader Scope for EU Competition Law Infringements?, Journal. of European Competition Law & Practice, 2015, Vol. 6, No. 10, p. 690.. 29.

(30) rules”.122 This seems to show that CJEU in AC-Treuhand CJEU did not establish a legal test; however, it is for the future case law to decide that. Liability of the spokes. As to the liability of the spokes, there are two CJEU’s judgments that. deserve attention – Eturas and VM Remonts. In Eturas,123 the administrator of an online travel booking system sent through an internal messaging system a message to several of the travel agencies offering travel bookings on their websites using the method determined by that system. The message concerned a reduction of discounts for online bookings, which was then implemented into the system through a technical restriction, limiting the discounts that could be applied to the range indicated in the message, requiring the agencies to take additional steps to offer greater discounts. Then, several of the travel agencies started offering discounts within that range.124 CJEU examined whether this amounts to a concerted practice. Advocate General - A. Szpunar argued that there was no hub-and-spoke collusion in this case, as it involved merely a “message which was conveyed simultaneously to all undertakings concerned by their common trading partner”. According to Szpunar, for a finding of a hub-and-spoke collusion, it would be necessary to establish a certain “state of mind of the parties involved”, it is not enough to establish that there was disclosure of sensitive market information between a distributor and its supplier.125 Instead, where the sender of the information is not a competitor, finding of a horizontal collusion is contingent upon the recipients’ awareness that the information comes from a competitor or is also communicated to a competitor.126 CJEU did not refer to hub-and-spoke collusion, however, in essence, it confirmed Szpunar’s arguments. CJEU held that circumstances such as in this case could justify a finding of a concerted practice between the competitors, if (i) the competitors were aware of the content of the message or it can be established on the basis of other objective indicia that they tacitly assented to the anticompetitive conduct; (ii) there was subsequent conduct on the market; (iii). 122 123. Icap plc and Others (n 118), paras 197-198. Case C-74/14, "Eturas" UAB and Others v Lietuvos Respublikos konkurencijos taryba (2016),. ECLI:EU:C:2016:42. 124. Ibid, paras 6-11, 26, 43.. 125. Case C-74/14 (n 123), Opinion of AG Szpunar (16 July 2015), ECLI:EU:C:2015:493, para. 65.. 126. Ibid, para. 50.. 30.

(31) there was a relationship of cause and effect between the concertation and the subsequent conduct.127 VM Remonts. In VM Remonts case CJEU held that an undertaking may be held liable for a. concerted practice on account of the acts of its independent services supplier, if even one of the following conditions is met: “the service provider was in fact acting under the direction or control of the undertaking concerned, or that undertaking was aware of the anti-competitive objectives pursued by its competitors and the service provider and intended to contribute to them by its own conduct, or that undertaking could reasonably have foreseen the anti-competitive acts of its competitors and the service provider and was prepared to accept the risk which they entailed.”128 As CJEU explained, the first condition would be met for example where the service supplier had little or no autonomy or flexibility in deciding how the service was carried out.129 Second one – where the undertaking intended to disclose information to its competitor through the service supplier, or where it expressly or tacitly consented to that.130 Algorithm-enhanced hub-and-spoke: input and output. The above considerations show. that Article 101(1) TFEU may be applied in a “classic” hub-and-spoke scenario, provided that the parties have a certain required state of mind. The analysis whether the involvement of algorithms changes that conclusion will begin with the analysis of the scenario involving the use of the same algorithm and dataset to determine prices. Liability of the competitors. In light of the above case law, to hold the spokes liable under. Article 101(1), the competition authorities would have to establish not only the conduct – the exchange of information, but more importantly the relevant state of mind of the parties. As to the latter, Ezrachi and Stucke suggested that if the supplier’s advertising materials promise profit maximisation, then “surely” each of the competitors must know that the algorithm will make use of the competitor’s information in price determination.131 This involves two 127. "Eturas" UAB and Others v Lietuvos Respublikos konkurencijos taryba (n 123), paras 42, 44-45.. 128. Case C-542/14, SIA ‘VM Remonts’ (formerly SIA ‘DIV un KO’) and Others v Konkurences padome (2016),. ECLI:EU:C:2016:578, para. 33. 129. Ibid, para. 27.. 130. Ibid, para. 33.. 131. Ezrachi and Stucke, Virtual Competition (n 93), p. 49.. 31.

(32) assumptions, first – that the competitors are aware that they all use the same algorithm supplier, second – that they are aware of the exchange of information. This seems to constitute a certain state of mind required under the Eturas judgment and to fulfill the second condition listed in VM Remonts case, as explained above. In light of the VM Remonts judgment, it would be enough to prove that an undertaking either intended to disclose information to its competitor through the algorithm or algorithm supplier, or that it expressly or tacitly consented to that. This would pose some difficulties, as absent any evidence stating otherwise, it will be difficult for the competition authorities to prove that a company outsourcing its algorithm to an independent supplier was aware of what does the algorithm do with the supplied information or of the information fed to its algorithms. Supplier’s advertising materials promising profit maximisation are certainly not enough to prove that. Competition authorities would need e.g. communication between the competitors and the supplier proving that the competitors were aware of the planned or ongoing exchange of information, a contract between the supplier and one of the competitors, or witness statements. On the other hand, it seems like it would be enough to prove either that the algorithm supplier was acting under the direction of the competitor or that the competitor could reasonably have foreseen the anticompetitive conduct and was prepared to accept that risk. It is unlikely that a company would decide to outsource its pricing to an algorithm supplier, not knowing how exactly it would work. Depending on the wording of the advertising materials or of the agreement between competitor and supplier, this could be enough to prove that the company could have foreseen the anticompetitive conduct. In this context, it is worth mentioning the words of M. Vestaeger, who said that companies had better know how the algorithms used by them work, because they will be held responsible for what the algorithms do.132 One other difficulty for the competition authorities might be proving the existence of information exchange. Absent any explicit evidence stating to the above, it might be necessary to audit the algorithm (the code) and the dataset it had access to. If the case involves a selflearning algorithm, the problem would be its “black box” nature, where it is only possible to 132. Margrethe Vestager, ‘Speech. Algorithms and competition. Bundeskartellamt 18th Conference on Competition,. Berlin’. (16. March. 2017),. <https://ec.europa.eu/commission/commissioners/2014-. 2019/vestager/announcements/bundeskartellamt-18th-conference-competition-berlin-16-march-2017_en>, accessed: 9 January 2018.. 32.

(33) access the dataset and the output of the algorithm without understanding which variables contributed to such outcome.133 It could be that the algorithm came up to the pricing decisions regardless of the information exchange, that it derived to that particular outcome due to market analysis. Algorithm supplier’s liability. As to the liability of the algorithm supplier, in light of the AC-. Treuhand and Icap judgments, there are three things a competition authority would have to prove to hold the hub liable for contributing to or facilitation of an anticompetitive objective among the competitors. First, and most importantly, the competition authorities must establish the existence of common anticompetitive objectives pursued by the competitors, which is fundamental to hub’s liability for contribution or facilitation of it. This was already analysed in the two preceding paragraphs. To put it simply, if the facilitated conduct is not illegal, the hub will not be held liable for its facilitation. Then, it must be proven that the hub was aware of the planned or already occurring collusion, or could have reasonably foreseen it and accepted that risk. In the outlined scenario, this should not be a problem, given that the algorithm supplier is the one supervising or controlling the algorithm, and could be presumed to be aware of how the algorithm actually determines pricing for the competitors, and thus of the information exchange. Lastly, it must be proven that the algorithm supplier intentionally and actively or passively participated in the conduct, for which the evidence that the algorithm supplier provided its services to the competitors should be enough. In this regard, an algorithm supplier could argue that he was not the one who participated in the infringement, that it was the algorithm. However, to quote Germany’s Federal Cartel Office President – “algorithms aren’t written by god in the heavens. Companies can’t hide behind algorithms.”134 As long as computers are not granted legal personality to be held liable for their own conduct, and as long as the algorithm is acting under the direction or control of the algorithm supplier,135 it will be the algorithm supplier who will be responsible for its actions. Even if the algorithm used is one that is self-learning – makes. 133. Sloan and Warner (n 111), p. 5.. 134. Tom Sims, Airlines can't blame computer models for higher fares, German cartel chief says (Reuters, 28. December 2017), <https://www.reuters.com/article/us-germany-airlines-pricing/airlines-cant-blame-computermodels-for-higher-fares-german-cartel-chief-says-idUSKBN1EM0KL>, accessed: 21 January 2018. 135. See: VM Remonts (n 128).. 33.

(34) autonomous pricing decisions, someone is still responsible for providing the algorithm with the information or for designing it in a certain way. Algorithm-enhanced hub-and-spoke: output The application of Article 101(1) TFEU to. the first version of the algorithm-enhanced hub-and-spoke, which involves neither direct nor indirect exchange of information between the competitors might prove difficult, if not impossible. Here the alignment is only the effect of competitors’ unilateral decisions to use the same algorithm supplier, there is no horizontal conduct or intent. It would be absurd to sanction the companies e.g. for outsourcing their pricing to the best algorithm supplier, just because its competitors might do the same. However, there are other possible versions of this scenario. First, where the competitors intentionally choose a pricing algorithm of an upstream supplier being aware of the fact that other competitors did the same. Should it be established that this was done intentionally, this could amount to a practice facilitating tacit collusion discussed in the following chapter. Second, where the industry-wide use of the single algorithm results from the fact that one of the competitors discloses its pricing algorithm and others start using it. Even if this was a genuinely public disclosure of information, this could amount to an exchange of strategic information between competitors, as this could be understood as a disclosure of the method of price calculation, and thus could amount to the disclosure of price related information.136 The disclosure of pricing algorithm could be understood as an invitation to collude, which would make the fact that the disclosure was genuinely public and unilateral irrelevant.137 This modification could also amount to a practice facilitating tacit collusion under Article 101(1) and 102 TFEU, discussed in the following chapter.. ii.. Conclusion. The above analysis shows that Article 101(1) TFEU may in fact be used to address the “classic” hub-and-spoke structure and hold both the hub and the spokes liable for a concerted practice. The above conclusion changes slightly after adding algorithms to the equation. EU competition. 136. See: Horizontal Guidelines, paras 86, 94.. 137. Ibid, para. 63.. 34.

(35) law may address the algorithm-enhanced hub-and-spoke, but only to some extent. In case of the scenario involving the competitors both using the same algorithm and dataset, although there may be some evidentiary difficulties, the application of Article 101(1) TFEU is pretty straightforward. Both the competitors and the algorithm supplier may be held liable for a concerted practice. In the words of M. Ohlhausen: Is it ok for a guy named Bob to collect confidential price strategy information from all the participants in a market, and then tell everybody how they should price? If it isn’t ok for a guy named Bob to do it, then it probably isn’t ok for an algorithm to do it either.138 On the other hand, the scenario involving the competitors using the same algorithm, but not the dataset (without any exchange of information), although it may result in a similar alignment of competitors’ pricing, it does not fall within the scope of Article 101(1) TFEU given the lack of horizontal element.. 138. Ohlhausen (n 93), p. 10. 35.

(36) 3. Tacit collusion. This chapter considers whether and how are the third and fourth scenario identified by Ezrachi and Stucke addressed by the current EU competition law. The following two subsections discuss whether the competition law can address tacit collusion and then algorithm-fuelled and algorithmic tacit collusion through an ex-post intervention under Article 101(1) TFEU or Article 102 TFEU. The third subsection determines whether the usefulness of the ex-ante method – merger review changes when considered in the context of algorithm-enhanced and algorithmic tacit collusion. i.. Article 101(1) TFEU. Scope. This subsection begins with a general analysis of the applicability of Article 101(1). TFEU to tacit collusion. Then, it considers whether the conclusion of that analysis will be in any way affected by the fact that the two scenarios in question involve pricing algorithms. Concerted practice. Since the scenarios of tacit collusion and algorithmic tacit collusion. involve neither any agreement, nor a decision by association of undertakings, Article 101(1) TFEU may only address them, where the tacit collusion amounts to a concerted practice, understood as “a form of coordination between undertakings by which, without it having reached the stage where an agreement properly so-called has been concluded, practical cooperation between them is knowingly substituted for the risks of competition”.139 However, the analysis of the available case law allows for a conclusion that tacit collusion in itself does not amount to a concerted practice. In ICI v. Commission, CJEU explicitly stated that “parallel behaviour may not by itself be identified with a concerted practice”.140 This was confirmed in Suiker Unie case, where CJEU held that Article 101(1) TFEU “does not deprive. 139. Horizontal Guidelines, para. 60.. 140. Case 48/69, Imperial Chemical Industries Limited v Commission (1972), ECLI:EU:C:1972:70, para. 66.. 36.

Referenties

GERELATEERDE DOCUMENTEN

Figure 9 Results with 95% confidence intervals for different number of directional sectors used by AMD in highway and urban scenarios for: delivery ratio (a), delay (b), and

Findings indicate that in terms of engagement, experiencing high levels of energy in their work, being dedicated to their work, and not having an avoidant coping strategy

When Enlightenment premises were declared dead in the Cold War era, many social scientists such as Marshall and Parsons, in search for a place for social science during the Cold War

The request for proposals made it clear that the main aim of the project was to investigate the relationship between the regulation of access and incentives on the part of operators

‘Perilous Deviations from FRAND Harmony – Operational Pitfalls of the 2015 IEEE Patent Policy ’, Proceedings of the IEEE SIIT 2015, 9th International Conference on Standardiz- ation

Commission Decision on the application of Article 106(2) of the Treaty on the Functioning of the European Union to State aid in the form of public service compensation granted

Commission Decision of 20 December 2011 on the application of Article 106(2) of the Treaty on the Functioning of the European Union to State aid in the form of public

Because in its case law on the concept of undertaking the ECJ attaches great value to the extent managing bodies can infl uence the level of benefi ts, it may be assumed that