Pierre Honoré and Guillaume Fabre, Bredin Prat
This is an extract from the second edition of the E-Commerce Competition Enforcement Guide - published by Global Competition Review. The whole publication is available here.
What is algorithmic pricing?
In early ninth century Persia, Muhammad ibn Musa al-Khwarizmi worked on what would later be called, after the Latinised version of his name, ‘algorithms’.
In the 11 centuries since then, algorithms have been slightly refined: today, they are virtually omnipresent. They help us find information from billions of pages on the internet (Google); keep up with our email (Gmail, etc.); track the news (Google and Facebook news feeds); find professional contacts (LinkedIn); friends (Facebook); and dates (Tinder). They match supply and demand on the online advertising market (Google), delivering ads based on our experience; or match supply and demand on the car-sharing market (Uber). They help us find films or series we might like (Netflix), or tell us which music to listen to (Deezer, Spotify, etc.), and they help us plan our travels (Google Maps, etc.).
Yet, as pointed out in an OECD report: ‘a universal definition [of the concept of algorithm] that is consensually accepted is still missing. At its core, an algorithm is, according to the Oxford English Dictionary: ‘a process or set of rules to be followed in calculations or other problem-solving operations, especially by a computer’. To put it differently, it is ‘a sequence of rules that should be performed in an exact order to carry out a certain task’, such as a food recipe or a music sheet.
Applied to pricing – the focus of this paper – an algorithm follows a set of rules to determine automatically at what price a product or service can be sold. Increasingly, this may involve data gathering, the analysis of data through artificial intelligence (AI), and the determination of a pricing algorithm, be it by engineers or by AI itself.
Algorithmic pricing can take a great variety of forms, and may, for instance, give rise to:
- Prices that vary automatically according to the cost of the products or services (to take a basic – unrealistic – example, retailers could price bread according to the price of wheat future prices).
- Prices that vary automatically according to supply and demand – as in the well-known case of Uber (Uber goes one step further as it tries to influence supply and demand by providing for ‘surge’ pricing, inciting drivers to go where demand exceeds supply).
- Prices that vary according to each consumer’s preferences, in particular when coupled with a sufficient amount of data. A consumer who will return twice or thrice to the same website to rent a car or book a hotel room can theoretically be charged at a different price than the consumer who visited the website just once. Or, as some companies try to sell to advertisers, a consumer whose email content shows a keen interest in renting cars could be charged more than a consumer whose email does not show such an interest. Arguably, with Amazon Echo or Google Home, the amount of information available to tailor prices to each individual consumer will get bigger and bigger.
Not all industries can afford using algorithmic pricing in the same way. For instance, brick-and-mortar retailers would have a very difficult time personalising prices to consumers. This is because, even in stores where prices are displayed digitally, it would be impossible to display different prices for two consumers at the same time or at close intervals. However, thanks to digital display, brick and mortar retailers can now automatically update their prices much more frequently and, for instance, adapt them to any evolution of their variable costs or other factors.
Conversely, some industries are more prone to using algorithmic pricing. This is in particular the case for industries with significant capacity constraints, such as airlines, which have been using yield management techniques (in other words, algorithmic pricing) to ensure they can maximise both the occupancy of flights and their profitability. Online retailers, for their part, do not face the hurdles that brick-and-mortar retailers do and may therefore use algorithms, for instance, to adapt prices both to consumers and to the evolution of their costs.
In the past, using a pricing algorithm was only available to large firms with sophisticated IT departments. Now, it is possible for any SME to buy a ready-made algorithmic service online. It appears that 53 per cent of respondents in the European Commission’s e-commerce sector inquiry track online prices of competitors. Moreover, 67 per cent of those respondents who track prices do so with software and 78 per cent of them subsequently adjust prices.
Most industries could use algorithmic pricing to adapt prices automatically to market price or, in other words, to the prices of their competitors. From this perspective, a company can follow three potential types of strategies:
- Price below the market. This could be used if a company’s strategy is to undercut competitors or maximise sales volumes. Potentially, the algorithm can adapt this strategy based on costs and inventory. For instance, it can price below market so long as the price covers both fixed and variable costs, or so long as only the variables costs are covered, or could even potentially follow this strategy regardless of the company’s costs. Or it can price below the market only so long as inventory is above a given threshold, thus providing sales velocity to overstocked products.
- Price at the market average. A company could adopt this strategy to maximise its turnover without reducing its profitability.
- Price above the market. A company could adopt this strategy if it wants to present its product or services as higher-end and preserve its aura of prestige, or if it wants to create a brand. Apple, for instance, may be pursuing such a strategy by pricing its iPhone X above prices for previous generations of iPhones.
From the point of view of consumers, algorithmic pricing may result in a number of outcomes: lower prices on the market in which a couple of competitors decide to use algorithms to lower prices (while avoiding a price war that would lead them to make losses); continuously stable prices (which some consumers may find reassuring); or higher prices, even potentially, continuously higher prices without any justification.
An infamous example of the last possibility lies in the story of The Making of a Fly – a rare 1992 book that is a reference work in evolutionary biology and is out of print. Only two sellers on Amazon’s marketplace had new copies. They both used algorithmic pricing. One of them priced it at 1.270589 times the price of its competitor. The other one priced the book at 0.9983 times the price of its competitor. Neither seemed to have put in place a restriction such as a maximum price on their algorithmic pricing. Accordingly, the first one kept pushing prices up (each 27 per cent price hike at a time) until the price reached over US$23 million.
In this context, it was only natural that antitrust regulators should take a keen interest in algorithmic pricing. More than previous technological innovation, it is at the heart of the topic on which antitrust lawyers focus (Commissioner Margrethe Vestager has specifically referred to The Making of a Fly story in speeches).
The intersection of pricing algorithms and competition law can be analysed from various points of view. This paper will focus on the application of Article 101 of the Treaty on the Functioning of the European Union (TFEU); not on the application of Article 102 TFEU or on the merger control aspects.
Algorithmic pricing and Article 101 TFEU
As is well known, Article 101 TFEU prohibits anticompetitive agreements, concerted practices and decisions by associations of undertakings. To stick to the tenets of classical competition law, we touch below on the intersection between algorithms and Article 101 TFEU in a horizontal context and in a vertical context.
The horizontal context
Recent case law shows that competitors may reach agreements to fix, set or otherwise align prices using algorithms, in cases that do not appear to challenge the usual application of antitrust law (in other words, companies enter into a ‘good old’ price cartel to increase profitability using an algorithm).
Other cases suggest that using algorithms, companies may enter into pricing agreements to offer a new service at lower prices – which may call into question the current quasi-impossibility of exempting price-fixing agreements between competitors.
Even more fundamentally, there are debates as to whether the notion of ‘concerted practice’ – as we know it – can withstand the rise of algorithms.
Price-fixing cartels and algorithms: the Posters cases
According to the US Department of Justice (DOJ), various sellers of posters on the Amazon marketplace ‘participated in conversations and communications with representatives of other poster-selling firms’ to agree on poster prices. To implement this agreement, the various undertakings ‘agreed to adopt specific pricing algorithms . . . with the goal of coordinating changes to their respective prices’ and a representative of a firm ‘wrote computer code that instructed Company A’s algorithm-based software to set prices of the agreed-upon posters in conformity with this agreement’. Implementation of the pricing agreement was monitored.
In that case, competitors reaching an agreement on using a pricing algorithm to align their pricing strategy is no different from a cartel sharing Excel spreadsheets or having a dedicated phone line to agree on prices.
The DOJ explicitly said so in 2015 when it adopted its sanction decision: ‘We will not tolerate anticompetitive conduct, whether it occurs in a smoke-filled room or over the internet using complex pricing algorithms.
Following the DOJ case, the UK Competition and Markets Authority issued a similar decision on 12 August 2016, also concerning the online sales of posters and frames, on facts similar to the US decision.
In both cases, antitrust authorities were able to rely on correspondence between the companies involved in the cartel (in both cases there was a leniency application). Various emails described the efforts to use a pricing algorithm to implement the pricing agreement, for instance, by ensuring that each company directed its pricing algorithm not to undercut the prices of the other members of the cartel.
Arguably, an agreement to use a pricing algorithm could be much more subtle than a cartel in a smoke-filled room. As the machine can do all the work, companies might not need to communicate so often, to exchange vast volumes of information to assess the implementation of the cartel, etc. In other words, algorithms could significantly reduce the paper trail. Exchanges of information could be reduced in some instances to a suggestion to use a given provider of pricing algorithms (which might be public information). Once all market operators use the same algorithm, they could easily set their respective algorithm at similar parameters to implement their agreement.
This could make it much more difficult for antitrust authorities to prove the existence of a hardcore agreement on prices.
However, the use of algorithms could also allow competition authorities to monitor more easily prices on various markets. Whenever they have a doubt or a hint that explicit collusion might be at work, competition authorities could automatically collect price data and its evolution. If they detect a significant correspondence in prices, they could use more traditional investigative tools to identify if collusion has taken place on the market.
Margrethe Vestager specifically referred to that possibility: ‘It is a hypothesis that not all algorithms will have been to law school. So maybe there is a few out there who may get the idea that they should collude with another algorithm who haven’t been to law school either [sic]. So of course, we would like to have our own algorithms to be out there, looking into the market, figuring out if there has been collusion taking place. Other regulators have already implemented such mechanisms, for instance, in Korea, where an algorithm measures the risk of bid-rigging by assessing data from Korean public agencies.
Exempted price-fixing agreements and algorithms: the Webtaxi case
At the other end of the spectrum, pricing algorithms may provide significant efficiencies and thus be exempted.
The Competition Authority in Luxembourg, by a decision dated 7 June 2018, gave an individual exemption to Webtaxi for an agreement to fix prices among competitors. Webtaxi had in fact implemented a system to assign taxis to requests for a ride and to determine the price with an algorithm.
The Competition Authority considered that the joint use of the pricing algorithm by all competing taxi companies did amount to a price agreement between competitors. However, the agreement could be exempted given that it was indispensable to achieve various efficiencies (less waiting time, more rides for drivers and lower prices for consumers).
This case shows that pricing algorithms may serve the overall welfare, first and foremost by allowing companies to lower prices, to the immediate benefit of consumers. In contrast to the Poster case, it shows that pricing algorithms may at times call into question well-established rules of antitrust law. This is because pricing agreements between competitors can ‘generally’ not be exempted: ‘Price fixing can generally not be justified, unless it is indispensable for the integration of other marketing functions, and this integration will generate substantial efficiencies. This did not preclude the Luxembourg Competition Authority from exempting what it considered to be a price-fixing agreement between competitors.
Perhaps – as was argued shortly after the decision was adopted – the specific reasoning of the Luxembourg decision could have been better drafted. However, the outcome of the Luxembourg case should be approved. Had the Luxembourg authority considered that Webtaxi was a price-fixing cartel and could not be exempted, this might have called into question the operation of other ride-sharing firms, which would have deprived consumers of a very innovative service at low prices.
The decision also opens the door to other decisions that could fully recognise the benefits of using pricing algorithms in other circumstances where they contribute to consumer welfare.
Could algorithms call into question the notion of concerted practices?
According to settled case law (the well-known Wood pulp case):
[A] concerted practice refers to a form of coordination between undertakings which, without having been taken to the stage where an agreement properly so-called has been concluded, knowingly substitutes for the risks of competition practical cooperation between them . . . the criteria of coordination and cooperation must be understood in the light of the concept inherent in the provisions of the Treaty relating to competition that each economic operator must determine independently the policy which he intends to adopt on the Common Market.
Further, the court made clear in the same case that parallel conduct that could be explained by market characteristics (in particular, price transparency on oligopolistic markets) did not necessarily amount to a concerted practice. The court concluded that the European Commission could not consider that, given the market characteristics, concertation was ‘the only plausible explanation for the parallel conduct’. On that basis, the court explained that the Commission had not brought forward a ‘firm, precise and consistent body of evidence’ that concertation had taken place.
The law as it stands is therefore clear: parallel conduct does not necessarily entail concertation or coordination. As long as each operator determines its commercial strategy independently, parallel conduct does not fall in the scope of Article 101 TFEU. By contrast, any type of concertation or coordination that affects the independence of operators in determining their commercial strategy falls within the scope of Article 101 TFEU.
In this context, the immediate effect of a pricing algorithm may be to invert the paradigm for which the Wood pulp case was the exception. At that time, the level of price transparency in markets was such that it would be an exceptional case in which parallel conduct could occur without concertation or coordination. Today, parallel conduct could take place in a greater number of markets. Data accumulation and the ability of AIs to process that data results in far greater transparency than before. That transparency can then be readily exploited by pricing algorithms. Even in the absence of concertation, many operators may unilaterally decide to adopt the same strategy on how to configure their pricing algorithms, thus achieving lawful parallel conduct.
For instance, various operators may decide to set an algorithm so that their pricing is always on par with the market. Alternatively, these operators may decide that the algorithm should follow the prices of the market leader, which may realise that and understand that it has no interest in decreasing prices. Or, the operators may programme the algorithm so that their prices fluctuate with indexes relevant to their costs (say, for instance, price of Brent Crude oil for an operator for whom oil is a significant cost), or with the predictable reaction of competitors based on past behaviour. In such circumstances, the market operators do not exchange information. They each independently programme an algorithm, without any concertation whatsoever. This is, according to the expression of the leading scholars on the topic, ‘tacit-collusion on steroids’.
These scholars also envisage a scenario even more remote from the standard of explicit collusion (or concerted practices). In this scenario, intelligent algorithms with access to a ‘God-like view of the market’ due to the accumulation of data have the ability to learn and adapt by doing. Those algorithms would be able, without human interaction, ‘to coalesce around a dominant strategy’ to maximise profits and thus engage in durable tacit collusion.
In those circumstances, there would be no concertation (nor even any human intervention), just machines doing the work for which they have been independently programmed.
Competition authorities have taken notice of this potential risk. On 10 May 2016, the French Competition Authority and the German Federal Cartel Office published a joint paper on data and its implications for competition law. This paper focused on the issue of data more than on pricing algorithms, which it touched on briefly. It described these various possibilities and concluded that ‘all in all, prosecuting such conducts [sic] could prove difficult: first, market transparency is generally said to benefit consumers when they have – at least in theory – the same information as the companies and second, no coordination may be necessary to achieve supra competitive results’. On 19 June 2018, these authorities announced a new joint initiative, focused on algorithms and noting that ‘in certain areas, the use of algorithms may also reduce the need for human interaction. In particular, high degrees of automatization and related machine-to-machine communication may pose new questions for competition authorities’.
In this context, is it time to change the law? In this regard, we would like to emphasise three points.
First, at this stage and to the extent of our knowledge, there is no hard evidence that such algorithms could actually frequently lead to parallel conduct or even less behave in a ‘God-like’ fashion. The likelihood that all operators on the market would, each independently, come up with the same algorithm or use an algorithm in the same way sounds a distant prospect given the number of potential commercial strategies (but perhaps the joint Franco–German initiative launched in June 2018 will provide further evidence). Even then, there is no evidence at this stage that algorithm-organised parallel conduct would lead to supra-competitive results (in other words to market outcomes in which consumers would be worse off than in a situation without algorithms determining prices, taking into account the fact that, as demonstrated in the Webtaxi case, algorithms may provide innovative services and lower prices).
Quite the opposite, in a 2012 sector inquiry on e-commerce, the French Competition Authority concluded that ‘E-commerce often offers consumers lower prices and more choice than the traditional retail sectors’. Although a bit less conclusive, the European Commission’s report on e-commerce also makes clear that e-commerce has increased price transparency and allowed consumers to benefit from increased price competition.
Second, most economic operators cherish legal certainty. Changing the application of Article 101 TFEU to encompass unilateral conduct leading to parallel behaviour as some have advocated would not be a mere technical adaptation of the rules – it would profoundly alter the paradigm for the application of the rules on competition. In particular, it would call into question the divide between unilateral conduct captured by Article 102 TFEU and coordination in all its forms captured by Article 101 TFEU. In this context, having clear and constant legal rules might be preferable to changing the rules to catch a few potentially harmful cases in which algorithms could theoretically allow unilateral conduct to create harmful parallel conduct.
Third, various authors have called for the creation of a regulatory framework to monitor algorithms in markets in which there is stable parallel conduct. Assuming that competition authorities gathered strong evidence of an enforcement gap (which, to our knowledge, they have yet to do), with a number of markets with parallel conduct detrimental to consumers due to algorithms (which is, again to our knowledge, far from established), could an ad hoc regulatory framework close any enforcement gap?
An advantage would be that it would not be necessary to change the current interpretation of key notions of Article 101 TFEU. In that regard, a parallel can be drawn with the creation of the merger control framework: when competition authorities noticed that there was a growing number of concentrated markets and realised that neither Article 101 nor Article 102 TFEU could be relied on to tackle the issue adequately, they opted for the creation of a merger control framework. They did not, for instance, change the interpretation of Article 102 to discard the notion of ‘abuse’ or to extend that notion to breaking point by considering that an acquisition of a competitor could amount to an abuse.
However, this solution also seems to have significant downsides. What would be the scope of such a framework? How would it be possible to limit its application only to cases where harm to consumer welfare is a very likely possibility? To follow the same analogy, the percentage of Phase 1 cases without commitments in the total number of mergers notified every year throughout the EU certainly calls for caution. Even then, such a regulatory framework would in essence create a barrier to entry. This barrier to entry would certainly reduce innovation and potential competition by maverick firms. Would a new entrant in a market be prevented from using algorithmic pricing to adapt its pricing policy to that of the leader, which for a time may be necessary to get established in the market? This strategy could then allow it to try to undercut its competitors. Given the risk that such a regulatory framework could entail for competition and innovation, it prima facie does not appear to be an adequate solution.
In the face of these many constraints and in the absence of certainty on the risk to consumer welfare created by pricing algorithms at this stage, we would suggest trying to find a solution only once it has been firmly established that there is indeed a problem. Only once there is significant evidence that the problem has materialised, would competition authorities be able to study it sufficiently to determine what the best solution might be.
The use of algorithms in a vertical context
Regulation (EU) No. 330/2010 (the Vertical Block Exemption Regulation) and the Commission’s guidelines on vertical restraints make it clear that suppliers may not set a minimum resale price to their distributors: this practice is presumed both to fall within Article 101(1) TFEU and to be unlikely to fulfil the conditions of Article 101(3) TFEU.
By contrast, suppliers may recommend prices, so long as that recommendation does not amount to imposing prices. A hardcore restriction is therefore found when the supplier issues a recommended price list that is applied by retailers and when competition authorities find evidence that the supplier intervened to ensure that retailers stick to the recommended prices (through monitoring retail prices and punishing deviance).
Since the adoption of Regulation 1/2003, the European Commission had taken little interest in pursuing vertical cases. It focused on adopting the new vertical block exemption regulation and the corresponding guidelines. This left it to national competition authorities to enforce such practices at their level. Some of these cases have given rise to complex cases with significant litigation. In France, for instance, the French Competition Authority adopted on 13 March 2006 a decision fining 13 perfumes suppliers and three retailers for resale price maintenance. The press release from the FCA made clear that ‘the investigators collected in-store prices. This shed light on the strict implementation of the agreement: to a significant extent, the agreed resale prices were implemented.
In a vertical context, the use of algorithms may intervene at both the supply level and the retail level.
At the supply level, algorithms may be used to set the suppliers’ price; to monitor retailers’ prices; and even potentially to adapt the suppliers’ prices should a retailer deviate from the recommended prices. For instance, an algorithm could automatically take out any discounts for retailers that do not implement the recommended prices.
At the retail level, the use of algorithms leading to harmonised prices among retailers could mean that only a few retailers applying the recommended prices could ultimately lead all retailers’ prices to converge towards the recommended prices.
From this point of view, the use of algorithms can make a retail price maintenance practice both easier to implement (the supplier can easily monitor and punish deviance) and more effective (since even retailers that resisted implementing the policy can in practice end up implementing the recommended retail price).
The European Commission has already tackled such a case. On 24 July 2018, it adopted its first decision fining retail price maintenance practices since 2003. In that decision, which follows from the e-commerce sector inquiry, the Commission fined four consumer electronics manufacturers (Asus, Denon & Marantz, Philips and Pioneer) for imposing online resale prices. Interestingly, the Commission noted, in its press release:
Many [retailers], including the biggest online retailers, use pricing algorithms which automatically adapt retail prices to those of competitors. In this way, the pricing restrictions imposed on low pricing online retailers typically had a broader impact on overall online prices for the respective consumer electronics products.
Moreover, the use of sophisticated monitoring tools allowed the manufacturers to effectively track resale price setting in the distribution network and to intervene swiftly in case of price decreases.
However, as with algorithm-enhanced concertation, the use of algorithms may facilitate both illegal practices (resale price maintenance) and their detection and prosecution. For instance, in the Perfumes case, the French Competition Authority had to collect retail prices in many stores over a long period of time. Now, competition authorities may merely programme an algorithm to monitor prices to determine whether recommended prices are being applied by retailers in a possibly illegal fashion.
Throughout this chapter, we have tried to show that the overall impact of algorithms on competition in the marketplace could easily be largely positive: not only do they allow firms to provide new services to consumers, they can be used in a number of ways to ensure that prices are closer to an optimum level. From this perspective, they could maximise consumer welfare.
They do, however, raise some risks from a competition point of view, as well as some questions.
First, and as far as the risks are concerned, in a horizontal context, they might help with the implementation of an anticompetitive agreement or concertation that has been agreed between the parties using more traditional ‘human’ means. In a vertical context, they may facilitate resale price maintenance by helping suppliers monitor the implementation of imposed prices. In these cases, the anticompetitive conduct may also be harder to prove, given that this conduct could leave only a limited paper trail (which was, paradoxically, not the case in the Poster case). However, at the same time, algorithms provide a useful tool for competition authorities to detect parallel conduct or resale price maintenance, before carrying out more traditional types of investigation to prove the illegal conduct.
Second, and as far as the questions are concerned, some have expressed a fear that algorithms could increase price transparency on the market to the extent that tacit collusion would become much more frequent and that actors on the market could more easily detect deviation. Competition authorities, in the current state of the law, can do little to sanction tacit collusion on the basis of Article 101 TFEU. This could therefore, from a theoretical point of view, amount to an ‘enforcement gap’. However, given the imperatives of legal certainty and various risks identified with taking action to close that potential enforcement gap, as well as the benefits flowing from algorithmic pricing in the short term (such as giving the possibility to mavericks to undercut the prices of their competitors), we would favour waiting for a sufficiently certain and convincing body of evidence to show that this risk is not merely a possibility but a reality that needs addressing.
Subscribe here for related content, breaking news and market analysis from Global Competition Review.
Global Competition Review covers the most crucial developments in competition law and enforcement worldwide with breaking news commentary and analysis five days a week, as well as original in-depth reports, interviews and features on antitrust enforcement worldwide.