Algorithms and-colllusion-competition-policy-in-the-digital-age

ALGORITHMS AND COLLUSION
Competition policy in the digital age

Algorithms and Collusion:
Competition Policy in the Digital Age

2 │
Algorithms and Collusion: Competition Policy in the Digital Age
Please cite this publication as:
OECD (2017), Algorithms and Collusion: Competition Policy in the Digital Age
www.oecd.org/competition/algorithms-collusion-competition-policy-in-the-digital-age.htm
This work is published under the responsibility of the Secretary-General of the OECD. The opinions
expressed and arguments employed herein do not necessarily reflect the official views of the OECD or
of the governments of its member countries or those of the European Union.
This document and any map included herein are without prejudice to the status or sovereignty over
any territory, to the delimitation of international frontiers and boundaries and to the name of any
territory, city, or area.
© OECD 2017

│ 3
Foreword
The combination of big data with technologically advanced tools, such as pricing
algorithms, is increasingly diffused in everyone’s life today, and this is changing the
competitive landscape in which many companies operate and the way in which they make
commercial and strategic decisions. While the size of this phenomenon is to a large extent
unknown, a growing number of firms are using computer algorithms to improve their
pricing models, customise services and predict market trends. This phenomenon is
undoubtedly associated to significant efficiencies, which benefit firms as well as
consumers in terms of new, better and more tailored products and services.
However, a widespread use of algorithms has also raised concerns of possible anti-
competitive behaviour as they can make it easier for firms to achieve and sustain
collusion without any formal agreement or human interaction. This paper focuses on the
question of whether algorithms can make tacit collusion easier not only in oligopolistic
markets, but also in markets which do not manifest the structural features that are usually
associated with the risk of collusion.
This paper discusses some of the challenges algorithms present for both competition law
enforcement and market regulation. In particular, the paper addresses the question of
whether antitrust agencies should revise the traditional concepts of agreement and tacit
collusion for antitrust purposes, and discusses how traditional antitrust tools might be
used to tackle some forms of algorithmic collusion. Recognising the multiple risks of
algorithms and machine learning for society, the paper also raises the question of whether
there is need to regulate algorithms and the possible consequences that such a policy
choice may have on competition and innovation.
This paper was prepared by Antonio Capobianco, Pedro Gonzaga and Anita Nyeső of the
OECD Competition Division as a background note at the OECD Competition Committee
Roundtable on "Algorithms and Collusion" that took place in June 2017
www.oecd.org/daf/competition/algorithms-and-collusion.htm. This report contributes to
the OECD Going Digital project which provides policy makers with tools to help
economies and societies prosper in an increasingly digital and data-driven world. For
more information, visit www.oecd.org/going-digital.

│ 5
Table of contents
1. Introduction ....................................................................................................................................... 7
2. Algorithms: How they work and what they are used for............................................................... 8
2.1 Concepts and definitions................................................................................................................ 8
2.2 Applications of algorithms by businesses.................................................................................... 11
2.3 Applications of algorithms by governments ................................................................................ 13
3. Pro-competitive effects of algorithms ............................................................................................ 14
3.1 Algorithms and supply-side efficiencies...................................................................................... 15
3.2 Algorithms and demand-side efficiencies.................................................................................... 17
4. Algorithms and the risk of collusion.............................................................................................. 18
4.1 Collusion - concepts and definitions............................................................................................ 19
4.2 Impact of algorithms on the relevant factors for collusion .......................................................... 20
4.2.1 Structural characteristics of the industry............................................................................... 20
4.2.2 Demand and supply factors................................................................................................... 22
4.2.3 The impact of algorithms on the likelihood of collusion ...................................................... 23
4.3 Role of algorithms as facilitators of collusion ............................................................................. 24
4.3.1 Monitoring algorithms........................................................................................................... 26
4.3.2 Parallel algorithms................................................................................................................. 27
4.3.3 Signalling algorithms ............................................................................................................ 29
4.3.4 Self-learning algorithms........................................................................................................ 31
5. Algorithms and challenges for competition law enforcement ..................................................... 33
5.1 Algorithms and tacit collusion..................................................................................................... 34
5.2 The notion of agreement: does it need revisiting? ....................................................................... 36
5.3 The scope for antitrust liability.................................................................................................... 39
5.4 Possible alternative approaches to algorithmic collusion ............................................................ 40
5.4.1 Market studies and market investigations ............................................................................. 40
5.4.2 Ex ante merger control .......................................................................................................... 41
5.4.3 Commitments and possible remedies.................................................................................... 42
6. Algorithms and market regulation ................................................................................................ 42
6.1 Arguments in favour of regulating algorithms............................................................................. 43
6.1.1 Risks of “algorithmic selection” beyond collusion ............................................................... 43
6.1.2 Market failures ...................................................................................................................... 45
6.2 Possible regulatory interventions................................................................................................. 46
6.2.1 Institutional options to govern algorithms............................................................................. 46
6.2.2 Measures on algorithmic transparency and accountability ................................................... 47
6.2.3 Regulations to prevent algorithmic collusion........................................................................ 49
7. Conclusions ...................................................................................................................................... 51

6 │
Notes ..................................................................................................................................................... 53
References ............................................................................................................................................ 56
Annex 1. Mathematical derivation of collusion ................................................................................ 65
Tables
Table 1. Collusive patterns and corresponding feature engineering technologies ................................ 14
Table 2. Impact of algorithms on collusion........................................................................................... 23
Table 3. Summary of the roles of algorithms in implementing collusion ............................................. 32
Table 4. Risk categories of algorithmic selection.................................................................................. 44
Figures
Figure 1. Machine vs deep learning algorithms .................................................................................... 10
Figure 2. Relationship between artificial intelligence, machine learning and deep learning ................ 11
Figure 3. Illustration of monitoring algorithm ...................................................................................... 27
Figure 4. Illustration of parallel algorithms........................................................................................... 29
Figure 5. Illustration of signalling algorithms....................................................................................... 31
Figure 6. Illustration of collusion as a result of deep learning algorithms ........................................... 32
Boxes
1. Libratus, the poker playing robot.................................................................................................................. 10
2. Modern applications of deep learning........................................................................................................... 12
3. The Korean bid-rigging indicator analysis system (BRIAS) ........................................................................ 13
4. Artificial intelligence in the insurance industry............................................................................................ 15
5. Pricing algorithms and dynamic pricing ....................................................................................................... 16
6. Price comparison websites............................................................................................................................ 18
7. The 2010 “Flash Crash” in financial markets ............................................................................................... 25
8. Monitoring fuel prices with computer vision algorithms.............................................................................. 26
9. Algorithmic collusion in the Amazon marketplace....................................................................................... 28
10. The airline tariff publishing company case................................................................................................... 30
11. Individualised pricing and standard antitrust tools ....................................................................................... 34
12. The “oligopoly problem” .............................................................................................................................. 35
13. The notion of “agreement” for antitrust purposes......................................................................................... 37
14. Unfair competition standards and Section 5 US FTC Act ............................................................................ 38
15. Findings on pricing software of the EC sector inquiry on e-commerce........................................................ 40
16. USACM’s principles for algorithmic transparency and accountability ........................................................ 48
17. Accountability and the “right to explanation” in the European GDPR......................................................... 49

│ 7
1. Introduction
The importance of algorithms in today’s life cannot be understated. According to some
scientists, algorithms are so pervasive in modern society that they track, predict and
influence how individuals behave in nearly all aspects of life (Hickman, 2013 and
O’Neal, 2016). Although few would dispute the great benefits offered by algorithms,
especially in terms of improved automation, efficiency and quality, for both firms and
their customers, there are questions about the extent to which human decision-making
will be supported (or even replaced in certain cases) by machines and the implications of
the automation of decision-making processes for competition.
Scientists have identified this dilemma clearly. In 2015, over 70 scientists and artificial
intelligence (AI) experts signed an open letter calling for more research on the societal
impacts of these new technologies.1
These scientists believe that AI has great potential
benefits for society, including eradicating diseases and poverty, but they have identified a
need for more concrete research on how to prevent potential “pitfalls”: in other words,
researchers must not create something which cannot be controlled. They have called for
an interdisciplinary research effort bringing together disciplines ranging from economics,
law and philosophy to computer security and, of course, various branches of AI itself.
The impact of data-driven innovation on competition and social well-being has been well-
documented by the OECD.2
Examples of recent policy discussions include the benefits of
data-related disruptive innovations in financial markets (OECD, 2016g), land transport
(OECD, 2016e) and the legal sector (OECD, 2016f), recognising the risks of consumer
harm in the absence of an adequate competition framework that disciplines these new
market realities. In 2016, the OECD Competition Committee Hearing on Big Data
attempted to identify the more broad implications of data-driven business models for
competition law enforcement and market regulation (OECD, 2016a). One of the concerns
highlighted was that the use of data and computer algorithms may enable new forms of
collusion, a topic that is analysed in more detail here.3
This paper describes how algorithms are changing the competitive landscape by offering
opportunities to firms to achieve collusive outcomes in novel ways that do not necessarily
require the reaching of an agreement in the traditional antitrust sense, or may not even
require any human interaction. Economic theory suggests that there is a considerable risk
that algorithms, by improving market transparency and enabling high-frequency trading,
increase the likelihood of collusion in market structures that would traditionally be
characterised by fierce competition. The analysis in this paper shows that algorithms
might facilitate tacit co-ordination, a market outcome that is not covered by competition
law, by providing companies with automated mechanisms to signal, implement common
policies, as well as monitor and punish deviations. The paper also emphasises how
algorithms can make tacit collusion more likely, both in oligopolistic markets with high
barriers to entry and a high degree of transparency, and in markets where traditionally
tacit collusive outcomes would be difficult to achieve and sustain over time, widening the
scope of the so-called “oligopoly problem”.

8 │
In light of these findings, this paper reviews some of the possible responses that
competition law enforcers could give to this new phenomenon, including: (1) conducting
market studies and sector inquiries to evaluate whether algorithms actually pose a
significant competition problem; (2) reviewing merger control to account for the impact
of algorithms on the likelihood of coordinated effects; and (3) designing remedies to
prevent companies from programming specific codes that could be recognised as
facilitating practices for tacit collusion. Acknowledging the limitations of current antitrust
tools, the paper discusses whether it is necessary to revisit some traditional antitrust
concepts, by developing a more clear notion of agreement for antitrust purposes and even
by reconsidering the position of competition policy towards tacit collusion (with its pros
and cons). Lastly, while the increased risk of tacit collusion and other market failures
might support some regulatory interventions, this paper warns about the risks that
regulating algorithms could pose on competition and innovation.
The paper is structured as follows. Section 2 provides a brief overview on the main
concepts behind algorithms and their programming principles, such as artificial
intelligence, machine learning and deep learning. It also provides an overview of how
algorithms are used by businesses and governments. Section 3 is dedicated to the benefits
and possible efficiency gains from algorithms. Section 4 sets out some of the anti-
competitive risks of algorithms for collusion, by changing market conditions and
providing companies with new tools to coordinate strategies. Section 5 addresses the
possible competition law enforcement challenges in the current legal and economic
framework, and discusses possible solutions emphasising the difficulties in identifying
the adequate level of intervention. Section 6 identifies the potential benefits and risks of
alternative regulatory interventions. Finally, Section 7 offers some concluding remarks.
2. Algorithms: How they work and what they are used for
This section provides a brief insight into the basic concepts and programming principles
behind algorithms, discussing as well some of their applications by businesses and
governments. Applications by consumers and the benefits they create on the demand side
will be explored in section 3.
2.1 Concepts and definitions
Although the concept of “algorithm” has existed for a long time, far before the creation of
the first computers, a universal definition that is consensually accepted is still missing
(Moschovakis, 2001). Intuitively, an algorithm is a sequence of rules that should be
performed in an exact order to carry out a certain task. Thus, an algorithm is an instance
of logic that generates an output from a given input, whether it is a method to solve a
mathematical problem, a food recipe or a music sheet. Given the lack of precision of
these intuitive notions, this paper uses a more formal definition proposed in the literature:
An algorithm is an unambiguous, precise, list of simple operations applied
mechanically and systematically to a set of tokens or objects (e.g.,
configurations of chess pieces, numbers, cake ingredients, etc.). The initial state
of the tokens is the input; the final state is the output. Wilson and Keil (1999)

│ 9
Algorithms can be represented in multiple ways, such as plain language, diagrams, codes
or even programmes that can be read and executed by a machine. With the evolution of
computer science, algorithms have been developed to automatically perform repetitive
tasks involving complex calculations and data processing that could be costly to execute
for human beings. Recent developments in artificial intelligence and machine learning
have brought algorithms to a new level, allowing computers to solve complex problems,
make predictions and take decisions more efficiently than humans, frequently achieving
desirable policy goals for society.
Artificial intelligence refers to the broad branch of computer science that studies and
designs intelligent agents, who should be able to carry out tasks of significant difficulty in
a way that is perceived as “intelligent” (Swarup, 2012). This concept was coined by John
McCarthy in 1956, who defined it as “the science and engineering of making intelligent
machines”. At the initial stages of AI, machines were programmed with extensive lists of
detailed rules in order to attempt to replicate human thoughts, which could easily become
a burdensome process. AI became a more effective tool after the development of
algorithms that teach machines to learn, an idea that evolved from the study of pattern
recognition and learning theory, and which would establish the new branch of machine
learning.
Machine learning (ML) is a subfield of AI which designs intelligent machines through
the use of algorithms that iteratively learn from data and experience. According to
Samuel (1959), machine learning gives “computers the ability to learn without being
explicitly programmed”. Machine learning algorithms can be classified into three broad
categories, depending on their learning pattern (Anitha et al., 2014):
 Supervised learning, where the algorithm uses a sample of labelled data to learn
a general rule that maps inputs to outputs.
 Unsupervised learning, where the algorithm attempts to identify hidden
structures and patterns from unlabelled data.
 Reinforcement learning, where the algorithm performs a task in a dynamic
environment, such as driving a vehicle or playing a game (Box 1), and learns
through trial and error.
Regardless of the learning method, conventional machine learning systems have some
limitations in their ability to process raw data (LeCun et al., 2015). Indeed, raw databases
can have such a large dimension that, before running a machine learning algorithm, it is
often necessary to extract from the raw data the features that are relevant for the
underlying problem – a process known as “feature engineering”. Features can be
numerical variables or strings that are either a subset of the original dataset or a
construction made from combinations of the original variables. Identifying and
constructing the relevant features can be a time-consuming and expensive process that
must be manually executed by humans. Alternatively, automatic feature extraction can be
performed through deep learning models (Figure 1).

10 │
Figure 1. Machine vs deep learning algorithms
Type the subtitle here. If you do not need a subtitle, please delete this line.
Source: Moujahid (2016).
Box 1. Libratus, the poker playing robot
Libratus is a machine learning program designed by Tuomas Sandholm and Naom Brown
from Carnegie Mellon University (CMU) to play a complex version of poker, the no-limit
Texas hold’em. In order to master poker strategy, the algorithm behind Libratus uses a
reinforcement learning method that can be decomposed into a three stage process:
In a first stage, after entering into the program the rules of the game, Libratus plays against
itself trillions of poker hands. While the first hands are played randomly, Libratus starts
improving the algorithms at fast speed based on a method of trial and error. This means that
Libratus does not actually rely on observed data to learn.
In a second stage Libratus plays against other humans and, taking into consideration their
particular actions and behaviours, it selects the strategies learned during the first stage that
are best suited to beat the opponents.
Finally, in a third stage, as human players start finding patterns in the behaviour of the
machine and adapting their strategies, Libratus removes those patterns by randomising some
actions. In other words, Libratus learns how to bluff.
Libratus was put to the test in January 2017, when it played a total of 120 000 hands in a
tournament against top world players. In the course of the tournament, the machine was
competing against the human players by day and, based on the new collected data, it
improved its strategy overnight, by correcting any imperfections that the human players were
able to figure out. The success of the machine was unprecedented, earning a total of
USD 1.776.250 in chips, as compared to all the other players who ended the game with a
negative balance.1 Given that the big blind of the game was set to USD 100, Libratus had a
winning rate of 14.7 big blinds per 100 hands, which is very high and statistically significant.
Games have long served as useful tools to improve and measure new developments in
artificial intelligence. Yet, the particular application of AI to poker might enable also
alternative applications. As a dynamic game of imperfect information with million hand
combinations, poker partially reflects the complexity and uncertainty of real world problems.
In particular, the ability of Libratus to undertake complex reasoning in an uncertain scenario,
to interpret past information that might be intentionally misleading and to anticipate the
impact of its own actions of the other players’ decisions suggests that AI can be also used to
solve more complex “human-like” interactions and decision making.
1. Spice and Allen (2017)
Source: Metz (2017).

│ 11
Deep learning is a subfield of machine learning (Figure 2) that enables computer systems
to learn using complex software that attempts to replicate the activity of human neurons
by creating an artificial neural network. Goodfellow et al. (2016) point out that deep
learning is able to model powerful abstractions in data. While traditional machine
learning algorithms are linear, deep learning algorithms are structured in a hierarchy of
increasing complexity and abstraction. As a result, deep learning enables computers to
learn faster and more accurately than conventional machine learning.
Despite recent advances and the enormous potential of deep learning to solve the most
complex problems, the lack of feature extraction implies that there is no way to know
which features or information were used by the algorithm to convert inputs into outputs.
In other words, regardless of the quality of the results produced, deep learning algorithms
do not provide programmers with information about the decision-making process leading
to such results.4
Figure 2. Relationship between artificial intelligence, machine learning and deep learning
2.2 Applications of algorithms by businesses
The growing importance of algorithms for society is a direct consequence of their
enhanced adoption by companies, not only in online markets, but also in many other
high-tech industries. In this context, Stucke and Ezrachi (2016) have introduced the
concept of “algorithmic business” to refer to the use of complex algorithms to improve
business decisions and automatise processes for competitive differentiation. Although
some businesses are still at the very early stages of adopting algorithms, an increasing
number of companies nowadays rely on them for predictive analytics and optimisation of
business processes.
 Predictive analytics involves the development of algorithms to measure the
likelihood of future outcomes based on the analysis of historical data. Predictive
models can be used to estimate demand, forecast price changes, predict customer
behaviour and preferences, assess risks and forecast endogenous or exogenous
shocks that might affect the market environment, such as the entry of new firms,
variations in exchange rates or even natural disasters. All this information can be
extremely valuable to improve decision-making, enabling companies to plan
more efficiently their business strategies and to develop innovative and
customised services that would not be possible otherwise.
 Algorithms can also be implemented to optimise business processes, allowing
businesses to gain a competitive advantage by reducing production and
transaction costs, segmenting consumers or by setting optimal prices that
effectively respond to market circumstances. The particular ability of algorithms
to optimise processes is a result of their automated nature and great computation
Artificial Intelligence
Machine Learning
Deep Learning

12 │
power, which allows them to process large datasets, react fast and incur in a
lower cost than what would be observed if the same tasks were performed by
humans.
The employment of algorithms for predictive analysis and optimisation of business
processes has multiple practical applications that are observed across many industries,
such as fraud prevention, supply-chain optimisation, targeted advertising, product
recommendation, corporate security and dynamic pricing. In addition, there are peculiar
industry-specific applications of algorithms that, particularly when based on machine and
deep learning principles, are bringing breakthrough data-driven innovations and are
revolutionising existing markets (Box 2).
Box 2. Modern applications of deep learning
While the initial attempts to develop the first deep learning algorithms can be traced back to
the mid-twentieth century,
1
for many decades the practical applications of deep learning were
very limited, possibly due to the insufficient processing speed of computers and the lack of
funding to support research. However, as digitalisation made computers faster and more data
became available in recent years, computer scientists have come up with algorithms that
outperform humans in tasks that were once thought impossible to be carried out by a
machine. Some of the areas where state-of-the-art applications of deep learning have been
recently developed include:
Health: In the health sector, image recognition algorithms were created to detect the
melanoma cancer at an early stage from the analysis of moles and lesions, as well as to
assist in brain cancer surgery by detecting cancerous cells to be removed.
2
Additional
applications in health include the analysis of foetal ultrasound images to assist in pre-natal
care
3
and the examination of radiographs to quantify the severity of knee osteoarthritis.
4
Civil and mechanical engineering: Artificial neural networks have been applied for multiple
purposes, such as to predict the dispersion of dust in drilling operations,
5
to forecast the
structural response of buildings under earthquakes,
6
as well as to predict traffic speed
conditions.
7
Finances: In financial markets, neural networks and genetic algorithms have been designed
to generate buy and sell signals for portfolios of stocks
8
and to predict corporate bankruptcy,
9
where they were shown to be more accurate and sometimes even easier to implement than
traditional techniques.
Biology: Some scientists have been deploying and testing complex algorithms in order, for
instance, to automate the detection and classification of deep-sea animals
10
or to estimate
the concentration of chlorophyll in the leaves of medicinal and aromatic plants.
11
Sound and image: Deep learning can be used to automatically colour black and white images
by recognising the objects and their context in the photograph and, more impressively, to add
sounds to silent movies by analysing what is happening in the scenes.
12
1. Foote (2017); 2. Hemsoth (2016); 3. O’Shea et al. (2016); 4. Anthony et al. (2016); 5. Nagesha et al.
(2016); 6. Suryanita et al. (2016); 7. Lee and Teknomo (2016); 8. Er and Hushmat (2016); 9. Jones et al.
(2016); 10. Hollis et al. (2016); 11. Odabas et al. (2016); 12. Brownlee (2016).
The widespread adoption of algorithms by businesses is not only transforming the way in
which companies operate and interact with each other, but it is also significantly affecting
the evolution of markets towards global digitalisation triggering a domino effect that
promotes a wider use of algorithms in an industry. As argued by Stucke and Ezrachi
(2016), as companies use algorithms to become more efficient, other companies feel
pressure to digitalise their operations and to develop similar algorithms themselves. In

│ 13
turn, as more companies rely on automated systems, computer scientists are encouraged
to develop new technologically-advanced programming principles and companies have
strong incentives to come up with novel business applications for algorithms.
2.3 Applications of algorithms by governments
In addition to the rapid spread of algorithms in the business world, there is a growing
interest towards the application of algorithms by government agencies, especially for
detecting crimes. In some countries, like the United States, there have been developments
towards a more data-driven approach to detect patterns of criminal behaviour (Rudin,
2016). In a recent initiative, PhD students from MIT have collaborated with crime
analysts to design a machine learning algorithm called “Series Finder”, which uses
historical criminal data to detect housebreak patterns and build a modus operandi
(Umamaheswari et al., 2016). This algorithm was considered a powerful tool to assist the
police detecting series of crimes and identifying likely suspects.
Box 3. The Korean bid-rigging indicator analysis system (BRIAS)
The BRIAS is an automatic quantitative analysis system that predicts the probability of bid
rigging, by analysing large amounts of bidding data from public agencies in Korea. Since
2013, the system collected bidding information from 51 central government agencies, 246
local governments and 26 public enterprises. The system is designed to quantify the
possibility of bid rigging by weighting different types of information, such as the rate of
successful bids, the number of companies that participated in the auction and the bid prices.
The BRIAS operates in three phases, from the gathering of the data and input, to the
generation of the results:
In a first phase, BRIAS collects all bid-related data and information concerning large scale
bidding contracts awarded by central and local administrations. All data and information are
collected within 30 days of the tender award.
In a second phase, the system analyses the data received and it automatically generates
scores on the likelihood of bid rigging, by assessing each relevant factor for the analysis. To
each of these factors a weighted value is assigned and the scores of each evaluation item
are then added up.
In a last step, the bidding opportunities are screened by BRIAS according to their score.
Tenders which score above a certain threshold are flagged by the system for in-depth review
by the Korean Fair Trade Commission (KFTC).
In order to design the system and to identify the criteria to detect possible bid rigging
conspiracies, the KFTC relied on its past enforcement experience and used pre-determined
red flags for collusion as a benchmark. On the basis of these markers, the system was
designed to give a higher score when: (1) the successful winning rate of a company is high;
(2) there are few bidders in the tender process; (3) there are several bidders whose bid is
higher than the estimated price; (4) non-competitive bidding processes are used; and (5)
there is a large gap between the winning and the losing bids. The system, however, is
effective only if the weighting system is correctly balanced.
Source: Korea’s submission to the Roundtable on Ex officio Cartel Investigations,
www.oecd.org/daf/competition/exofficio-cartel-investigation-2013.pdf, Report to the OECD Council on the
implementation of the Recommendation of the Council on Fighting Bid Rigging in Public Procurement.
Similarly to catching criminals, data-driven approaches have been proposed and
sometimes applied to detect collusion, particularly through the use of screening methods.
Although competition authorities generally rely on reactive detection tools to detect and

14 │
investigate cartels, such as leniency programmes, a combination of both reactive and
proactive detection measures – including screens – is considered to be more effective.5
A
number of competition agencies have already reported using screens to detect bid rigging
cases, which has only been possible due to the availability of extensive and reliable
bidding data on public tenders, as well as algorithms able to detect bidding anomalies and
suspicious bidding patterns across large data sets.6
That is the case, for instance, of the
Korea Fair Trade Commission (KFTC), which has in several occasions succeeded in
detecting bid rigging conspiracies by screening procurement bidding data (Box 3).
Despite some difficulties observed, it is likely that there will be significant developments
in this area in the near future, as screening technology improves and data are more readily
available. For example, Akhgar et al. (2016) describe the potential of applying machine
learning algorithms to identify hidden relationships as a possible indicator of collusion in
public tenders. They argue that through feature engineering it is possible to provide a
sufficiently rich set of attributes for machine learning methods to detect even
sophisticated collusive tendering schemes. This involves labelling collusive bidding
patterns and associating to each of them a corresponding set of features (Table 1) that can
be computed and used by the machine learning algorithm. The use of algorithms to detect
collusive bidding, and more generally possible cartel conduct, will open unprecedented
possibilities to put technology at the service of public antitrust enforcement.
Table 1. Collusive patterns and corresponding feature engineering technologies
Collusive Pattern Numeric or Nominal Feature
Same company wins most of the
time
Number and fraction of wins, by region, by sector, etc.
Bid rotation Time series descriptors of winners
Few or no new participants Histogram of participation distribution
Bidding does not erode target
price
Bid to reserve prize statics
Participant withdraw Withdraw count by bid and participant
Source: Akhgar et al. (2016).
3. Pro-competitive effects of algorithms
Data-driven marketplaces are generally associated with significant efficiencies both on
the supply and demand side. Algorithms are no exception.
On the supply side, algorithms help increasing transparency, improving existing products
or developing new ones. The OECD’s work on disruptive innovation has shown how
market entry has been promoted by the ability of firms to develop new offerings based on
algorithms (OECD, 2016e; OECD, 2016f and OECD, 2016g). This triggers a virtuous
mechanism whereby companies are under constant pressure to innovate, thus promoting
dynamic efficiencies (OECD, 2015a). The rise of algorithms on the supply side can also
promotes static efficiencies by reducing the cost of production, improving quality and
resource utilisation, as well as streamlining business processes.

│ 15
On the demand side, algorithms may significantly affect the dynamics of the market by
supporting consumer decisions: they better organise information so that it can be accessed
more quickly and more effectively and they provide new information on dimensions of
competition other than prices, such as quality and consumers’ preferences. Thus,
algorithms have the potential to create positive effects on consumers and social welfare.
3.1 Algorithms and supply-side efficiencies
Supply-side efficiencies allow firms to lower their production costs by improving the
allocation of resources. This is reflected in lower prices to consumers. In the past, it could
take time to find patterns and create data trends in order to determine the optimal
decisions. Now algorithms can perform this task in a matter of seconds. Deep learning
techniques enable companies to optimise their commercial strategies instantaneously
following trials and feedback. There is a rapid progress in self-learning algorithms to
assist in almost every field of business operations, especially planning, trade and logistics
(Box 4).
Box 4. Artificial intelligence in the insurance industry
For many years, insurance markets have been unable to keep up with technological
developments,
1
being frequently perceived as a sector where traditional business practices
prevail. Indeed, according to a survey by Willis Towers Watson (2017), “[a]lmost three-
quarters of global insurers (74%) believe their sector has failed to show leadership in digital
innovation”. This lack of innovation, along with the moral hazard traditionally observed in
insurance business models, may result in high premiums, as well as in slow and bureaucratic
claiming procedures in case of an event. Nonetheless, digitalisation appears to have recently
started reaching the insurance industry, as new start-ups attempt to disrupt the market and
incumbents are pressured to make substantial investments in research and development.
1
One of these new online start-ups, Lemonade, is attempting to transform the industry by
providing insurance services exclusively through a mobile app and a website, while
successfully introducing AI in several phases of its business model. This allows the company
to save resources and run its operation more efficiently. In order to acquire an insurance
policy, a new customer must interact and provide personal information to a “chatbot”, while
algorithms are being run in the background to cross that information with other databases,
assess the risk of the customer and make an automatic offer. If any eventuality occurs, the
claim is either rapidly addressed by another AI bot or, in more complex cases, it is addressed
by a human. The speed at which claims are handled is also a result of Lemonade charging a
flat fee to consumers and donating any unclaimed premiums to charity causes, therefore
eliminating corporate incentives to delay or deny claims.
2
Late last year a customer called Brandon claimed for a stolen coat. He answered a few
questions on the app and recorded a report on his iPhone. Three seconds later his claim was
paid – a world record, says Lemonade. In those three seconds ‘A.I. Jim’, the firm’s claims
bot, reviewed the claim, cross-checked it with the policy, ran 18 anti-fraud algorithms,
approved it, sent payment instructions to the bank and informed Brandon. The real-life Jim
(Hageman), Lemonade’s chief claims officer, was driving home for Christmas at the time.
3
1. Ralph (2017).
2. More information about Lemonade’s business model can be found at their corporative website
www.lemonade.com/faq#service.
3. The Economist (2017).
Certain algorithms can offer a wide range of quality effects. They can assist in improving,
refining or developing products and services in a number of ways. Search engines, for

16 │
example, use data to deliver more relevant and high-quality search results. By learning
from user search queries and clicks, search engines can identify the most relevant results
for a particular query and can also use the data to provide additional “value-added”
services to users. Some e-commerce sites use past purchase information and browsing
history to make personalised shopping recommendations for users. Online media outlets
use browsing history and personal information to recommend other articles that may
interest a reader.
The perceived benefits from algorithms are also due to the fast growing use of dynamic
pricing, which enable consumers and providers alike to see and act on fast changing
prices in many business areas, such taxi fares, sports tickets or hotel rooms, to name just a
few. Pricing algorithms allow for constant adjustment and optimisation of individual
prices based on many factors, including available stock and anticipated demand (Box 5).
Pricing algorithms learn through trial and error and through finding patterns from a great
volume and variety of data, leading to optimal pricing. As companies collect more user
data and algorithms have more opportunities to experiment (such as presenting items and
suggesting other purchases), pricing becomes more dynamic, differentiated and
personalised (Schumpeter, 2016).
Box 5. Pricing algorithms and dynamic pricing
Pricing algorithms are commonly understood as the computational codes run by sellers to
automatically set prices to maximise profits,
1
being particularly common in the airline, hotel
booking, road transport, electricity and retail industries. As compared to standard pricing
strategies, pricing algorithms have the advantage of processing large amounts of data that
are incorporated into the optimisation process, allowing them to react fast to any change in
market conditions. Given their automated nature, pricing algorithms are particularly useful to
implement continuous price changes over time – dynamic pricing – or to charge different
price to consumers based on their personal characteristics – price discrimination.
Dynamic pricing algorithms have been recognised to improve market efficiency,
2
by allowing
companies to react instantaneously to changes in supply conditions – such as stock
availability, capacity constraints or competitors’ prices – as well as to fluctuations in market
demand. This way, by guaranteeing that the market is constantly in equilibrium, dynamic
pricing does not only prevent unsatisfied demand and excess of supply, but under certain
competitive conditions it also guarantees that all mutually beneficial transactions between
consumers and suppliers are exhausted. Still, dynamic pricing strategies can be so
successful for algorithmic sellers, that they have been criticised to make it challenging for
non-algorithmic sellers to compete and for consumers to make decisions under constant
price fluctuations, unless they also use algorithms to facilitate decision-making.
3
The implementation of algorithms in digital markets has also been argued to facilitate first
degree or ‘perfect’ price discrimination, by allowing companies to price consumers based on
their location, browsing history, previous purchases and other private information.
4
First
degree price discrimination is recognised to improve efficiency, by allowing companies to
supply services at lower prices to underserved consumers with low willingness to pay. Still,
there are also concerns that implementing algorithmic price discrimination to such an extent
could result in a lower consumer surplus overall, as well as in undesirable forms of
discrimination based on gender or race.
5
1. Chawla et al. (2007).
2. Weiss and Mehrotra (2001).
3. Chen et al. (2016).
4. OECD (2016a) and OECD (2016b).
5. OECD (2016b).

│ 17
3.2 Algorithms and demand-side efficiencies
Algorithms do not only help companies improving their business processes, but also
assist consumers in their purchasing decisions, resulting thereby in significant demand-
side efficiencies. Gal and Elkin-Koren (2017) introduced the concept of “algorithmic
consumers” to describe the shift in the decision-making process allowing consumers in
data-driven ecosystems to outsource purchasing decisions to algorithms. Due to the fact
that algorithms can be used to compare prices and quality, to predict market trends and to
make fast decisions, they can significantly reduce search and transaction costs, help
consumers to overcome biases and make more rational choices, strengthening buyer
power. While a wide variety of algorithms are available to assist in purchasing decisions,
a new generation of algorithms can be used to make and execute decisions for the
consumer by directly communicating with other systems through the internet. In this case,
the algorithm automatically identifies a need, searches for an optimal offer and executes
the transaction. These algorithms are known as “digital half” or “digital butlers”.
The speed, level of information and analytical sophistication that can be provided by
digital butlers can also relate to parameters other than the price of the offer; their ability
to rely on additional competitive variables can be very important in the decision-making
process that consumers go through. Such algorithms allow consumers to access not only
comparative price information (Box 6), but also a wider set of other quality dimensions of
the offers that can be compared for competitive purposes. Gal and Elkin-Koren (2017)
also point out that, by avoiding consumer biases that could otherwise result in non-
optimal decisions, algorithms help consumers overcome manipulative marketing
techniques. Algorithms may also increase the equality among consumers, due to their
autonomous operation: consumers who do not know how to take advantage of online
purchasing tools can easily rely on digital butlers to optimise purchasing strategies on
their behalf.
Besides pure demand-side efficiencies, algorithmic consumers also have an effect on
suppliers’ incentives to innovate and compete. By allowing consumers to compare a
larger number of offers, algorithms can potentially lead to consumers switching and
therefore increase the competitive pressure on suppliers. Furthermore, algorithms can
expand the sets of variable on which competition takes place, since they can check and
compare many more variables. For example, algorithms are able to cover considerations
such as effects on market structures and collective action problems; they can recognise
forms of co-ordination between suppliers (i.e. potentially identifying instances of
collusive pricing) and diversify purchasing proportions to strengthen incentives for entry
(i.e. help sponsoring new entrants). Another possible use could be the aggregation of
demand by pooling consumers and creating buying platforms, which could increase buyer
power and solve some collective action problems on the demand side.

18 │
Box 6. Price comparison websites
The development of algorithms has improved the ability to offer price comparison services
either via search engines or comparison platforms. Price comparison websites (PCW) make
it easier for consumers to compare the available offers and find the best alternative.
Comparison platforms can also contribute to level the playing field and intensify competitive
pressure. In order to produce such benefits, consumers must be aware of the availability of
alternative options, understand their needs, have low switching costs, and have the
incentives to search and switch. When this is the case, PCWs can facilitate a transparent
market environment in which customers obtain comparable information on the available
offers. By collecting and aggregating information on products and services, PCWs can
reduce the asymmetry of information and improve information flows. This can make it harder
for suppliers to take advantage of ill-informed customers. PCWs can also weaken the power
of sellers to segment and price discriminate.
PCWs can significantly reduce search cost when shopping for a product or service. Likewise,
in the case of switching between providers, using interactive tools to compare offers can
result in notable savings. Research conducted by the Financial Conduct Authority revealed
that, in recent years, there has been a significant growth in the use of PCWs in real estate,
travel and motor insurance industries. Apart from easy and fast access to information and
cost savings, consumers identified the awareness-raising of new brands or providers among
the significant benefits from using PCWs.
1
By reducing the search cost for finding the best possible product or service, PCWs have the
potential to bring a non-negligible increase in consumer surplus. Moreover, being informed
about available products or services and their characteristics – especially price and quality –
allows customers to make well-informed purchasing decisions, which can result in increased
competition among suppliers.
2
1. Financial Conduct Authority, Price Comparison Website: Consumer Market Research (June 2014), 7,
prepared by Atticus Market Research Consultancy www.fca.org.uk/publication/research/price-
comparison-website-consumer-research.pdf.
2. While there is a lot of information available online this still involves search costs and consumers may
suffer from information overload. On the limited ability of consumers to process too much information and
on why information intermediaries may not always be the perfect solution to information asymmetries
(Armstrong, 2008).
4. Algorithms and the risk of collusion
There is no doubt that automated computer algorithms can be a powerful tool to extract
value from the increasing amount of data collected in the digital economy, potentially
fostering market efficiency, innovation and even promoting competition. Nonetheless,
whenever new technological tools revolutionise deeply the way companies operate and
interact with each other, there is the risk that some market players use their enhanced
power to achieve private interests that are not aligned with social goals. The discussion
that follows will focus on the risks that algorithms can raise for competition without
disregarding the great benefits that they have brought to society. The question is not
whether algorithms should be banned, but rather to understand the risk they may pose to
the competitive process and to identify solutions which are compatible with incentives to
innovate.
While algorithms might be used to implement virtually any anti-competitive conduct that
is typically observed in traditional markets, a particular concern highlighted in the

│ 19
literature is the risk that algorithms may work as a facilitating factor for collusion and
may enable new forms of co-ordination that were not observed or even possible before.
This is referred to as “algorithmic collusion”.
This section introduces traditional definitions of collusion that are fundamental for the
discussion and evaluates the impact of algorithms on market conditions identified in the
economic literature that affect the likelihood of collusion. The section then looks at
several ways in which algorithms might be used to actually facilitate the implementation
of an anti-competitive agreement.
4.1 Collusion - concepts and definitions
In the literature, the term “collusion” commonly refers to any form of co-ordination or
agreement among competing firms with the objective of raising profits to a higher level
than the non-cooperative equilibrium,7
resulting in a deadweight loss. In other words,
collusion is a joint profit maximisation strategy put in place by competing firms that
might harm consumers. In order to reach and sustain a collusive equilibrium over time,
competitors must put in place a structure to govern their interaction, enabling them (1) to
agree on a “common policy”; (2) to monitor the adherence to this common policy; and (3)
to enforce the common policy by punishing any deviations.8
Economists usually distinguish between two forms of collusion, explicit and tacit.9
 Explicit collusion refers to anti-competitive conducts that are maintained with
explicit agreements, whether they are written or oral. The most direct way for
firms to achieve an explicit collusive outcome is to interact directly and agree on
the optimal level of price or output.
 Tacit collusion, on the contrary, refers to forms of anti-competitive co-ordination
which can be achieved without any need for an explicit agreement, but which
competitors are able to maintain by recognising their mutual interdependence. In
a tacitly collusive context, the non-competitive outcome is achieved by each
participant deciding its own profit-maximising strategy independently of its
competitors. This typically occurs in transparent markets with few market
players, where firms can benefit from their collective market power without
entering in any explicit communication.10
Contrary to the economic approach, which considers collusion a market outcome, the
legal approach focuses on the means used by competitors to achieve such a collusive
outcome. For this reason, competition laws generally do not prohibit collusion as such,
but prohibit anti-competitive agreements. If collusion is the result of such as agreement
then an infringement of the law can be successfully established. Although there is great
variance in how jurisdictions interpret the notion of agreement, they traditionally require
some sort of proof of direct or indirect contact showing that firms have not acted
independently from each other (the so-called “meeting of the minds”).
The distinction between explicit and tacit collusion shows how, under certain market
conditions (i.e. transparent markets with few sellers and homogenous products), supra-
competitive price strategies may be the normal outcome of rational economic behaviour
of each firm on the market. It is for this reason that tacit collusion or conscious
parallelism falls outside the reach of competition laws on agreements between
competitors. However, from a policy perspective, such a tacitly collusive outcome may

20 │
not be desirable, as it confers firms the ability to significantly suppress output or raise
prices to the detriment of consumers as much as an explicit agreement would.11
Rules on anti-competitive agreements generally are not designed to tackle individual and
rational market strategies of a single firm, even if the overall result of similar individual
strategies has an equivalent outcome to a cartel. However, between explicit collusion
(which should always be regarded as illegal under competition rules) and mere conscious
parallelism (which should fall outside the reach of competition law as it does not entail
any form of co-ordination between competitors), there is a grey area of business
behaviour which goes beyond conscious parallelism but at the same time does not involve
an express agreement between competitors. This is a situation which may arise
particularly in oligopolistic markets where competitors are able to coordinate on prices
and increase the likelihood of a tacitly collusive outcome by engaging in facilitating
practices which make co-ordination easier (e.g. by facilitating communications) and more
effective (e.g. by facilitating detection of cheating and the administration of punishment
of deviations) (OECD, 2007).
In order to address these intermediate forms of co-ordination, some jurisdictions have
stretched the concept of “agreement” for antitrust purposes and looked at whether an
agreement can be inferred from evidence suggesting that competitors have not acted
independently. Even in the absence of an explicit agreement, for example, courts have
established an infringement of the competition law if evidence of parallel conduct was
accompanied by other factors (so called “plus factors”), which indicated that the parallel
conduct was indeed inconsistent with unilateral behaviour and rather the result of co-
ordination among the parties. Examples of such plus factors include communications
revealing an intention to collude or engagements in facilitating practices, such as
information exchanges. Some jurisdictions, in particular those in the European Union,
also rely on the concept of “concerted practice”, which allows them to deal with practices
that do not amount to an agreement but nevertheless have the effect of replacing effective
competition with practical co-operation between competitors.
4.2 Impact of algorithms on the relevant factors for collusion
Regardless of the means used by companies, there is a concern that algorithms make
collusive outcomes easier to sustain and more likely to be observed in digital markets.
This section discusses how the use of algorithms and the increasing availability of
business data online are changing market conditions and how that may potentially
enhance the likelihood of collusive outcomes. Section 5 will then discuss in further detail
what this means for competition enforcement and the challenges algorithms pose under
the legal framework that most jurisdictions use to assess competitors’ co-operation.
Economists have identified the most relevant factors that may increase the likelihood of
collusion in a given market.12
These factors can be divided into structural characteristics,
demand-side characteristics and supply-side characteristics. This section discusses how
computer algorithms are changing in particular the structural and supply characteristics,
so as to make digital markets and some traditional markets more prone to collusion.
4.2.1 Structural characteristics of the industry
The number of firms and barriers to entry are traditionally identified as two of the
most important structural characteristics that affect the risk of collusion. A large number
of firms not only makes it harder to identify a “focal point” for co-ordination, but it also

│ 21
reduces the incentives for collusion as each player would receive a smaller share of the
supra-competitive gains that an explicit or tacit collusive arrangement would be able to
extract. Similarly, in the absence of entry barriers, collusion can hardly be sustained over
time, as any increase in profits will increase the incentives to deviate from the collusive
equilibrium and will attract new entrants. Supra-competitive profits will therefore be
rapidly eroded.
It is unclear how algorithms may affect these two structural characteristics.
 Some of the most typical industries, where algorithms are used to set dynamic
prices, segment consumers or improve product quality, have a small number of
large players, as it is the case of search engines, online marketplaces, discount
stores, booking agencies, airlines, road transport and social networks. However,
many of these industries are also characterised by natural barriers to entry, such
as economies of scale, economies of scope and network effects, which allow
companies to grow, collect large amounts of data and develop more accurate
algorithms. Therefore, it is difficult to assert whether algorithms are the cause or
the effect of barriers to entry.
 Also the impact of algorithms on the likelihood of entry is not univocal. On the
one hand, as discussed in OECD (2016a), algorithms can be used to identify any
market threats very fast, for instance through a phenomenon known as now-
casting, allowing incumbents to pre-emptively acquire any potential competitors
or to react aggressively to market entry. On the other hand, the increasing
availability of online data resulting from the use of algorithms may provide
useful market information to potential entrants and improve certainty, which
could reduce entry costs.
 Finally, one peculiar aspect of algorithms is that it makes the number of
competitors in the market a less relevant factor for collusion. In traditional
markets, collusion is more easily sustainable if there are few competitors, as it is
easier to find terms of co-ordination, to monitor deviations and implement
effective punishment mechanisms among fewer firms. Algorithms can allow co-
ordination, monitoring and punishment to take place also in less concentrated
markets as their ability and speed in collecting and analysing data makes the
number of firms to monitor and agree with less relevant. In other words, the small
number of firms is an important but not a necessary condition for algorithmic
collusion to take place.
Other two important structural characteristics are market transparency and frequency
of interaction, both of which make industries more prone to collusion. While transparent
markets allow companies to monitor each other’s actions and detect deviations from an
agreement, frequent interactions enable them to retaliate fast and punish aggressively any
deviators. Unlike with the number of firms and entry barriers, algorithms are very likely
to enhance these two factors for collusion, posing therefore a threat to competition.
Focusing first on market transparency, the effective implementation of algorithms as a
central component of a business model requires the collection of detailed real-time data
that can be automatically analysed and converted into action. Therefore, in order for firms
to benefit from the strong predictive capacity and efficient decision-making rules of
algorithms, they have an enormous incentive not only to gather market information, but
also to actually develop automated methods that allows them to collect and store the data
into computer systems, ready to be processed, without the need for human action. This

22 │
can be done using online cookies, smart cards, bar codes, voice recognition, radio
frequency identification and other technologies.
Then, as soon as a few market players make an investment in technology to benefit from
an “algorithmic competitive advantage”, the remaining firms of the industry have a strong
incentive to do the same, risking otherwise being driven out of the market. The result is
an industry where all market participants constantly collect and observe in real-time
rivals’ actions, consumers’ choices and changes in the market environment, creating thus
a transparent environment that is prone to collusion. Also, as Ezrachi and Stucke (2016)
suggest, the realisation of this fact may encourage firms to share their data stream online,
allowing all players to have access to the same information set.
The increase of market transparency is not only a result of more data being available, but
also of the ability of algorithms to make predictions and to reduce strategic uncertainty.
Indeed, complex algorithms with powerful data mining capacity are in a better place to
distinguish between intentional deviations from collusion and natural reactions to changes
in market conditions or even mistakes, which may prevent unnecessary retaliations.
With respect to the frequency of interaction, the advent of the digital economy has
revolutionised the speed at which firms can make business decisions. Unlike in a brick
and mortar business environment where price adjustments are costly and take time to
implement, in online markets prices can in principle be changed as frequently as the
manager wishes. If automation through pricing algorithms is added to digitalisation,
prices may be updated in real-time, allowing for an immediate retaliation to deviations
from collusion. In fact, the combination of machine learning with market data may allow
algorithms to accurately predict rivals’ actions and to anticipate any deviations before
they actually take place.
The joint report of the French and German competition authorities clearly identifies the
risks for competition caused by an artificial increase in market transparency and in the
frequency of interaction that is possible thanks to algorithms:
Even though market transparency as a facilitating factor for collusion has been
debated for several decades now, it gains new relevance due to technical
developments such as sophisticated computer algorithms. For example, by
processing all available information and thus monitoring and analysing or
anticipating their competitors’ responses to current and future prices,
competitors may easier be able to find a sustainable supra-competitive price
equilibrium which they can agree on. Autorité de la Concurrence and
Bundeskartellamt (2016)
4.2.2 Demand and supply factors
The likelihood of collusion in a given industry can be also affected by demand factors. In
particular, market stagnation characterised by declining demand and the existence of
business cycles may hinder collusion. This is because firms have strong incentives to
profitably deviate when demand is high and reducing the costs of retaliation in future
periods when demand is low.
The use of algorithms by consumers can enable them to improve their decision-making
process and buy products at periods of low demand (when they are usually cheaper).
However, this section is specifically focused on the risks associated with the use of
algorithms by firms which should not directly affect market demand. Therefore, it is

│ 23
considered here that the use of algorithms as part of business practice does not affect
significantly collusion through demand factors.
Supply factors, on the other hand, can play a major role in the sustainability of collusive
arrangements. One of the most relevant supply-side characteristics is innovation. The
innovative nature of the market reduces the present value of collusive agreements, as well
as the ability of less innovative firms to retaliate. In this matter, algorithms are naturally
an important source of innovation, allowing companies to develop non-traditional
business models and extract more information from data, in order to improve product
quality and customisation. In industries where the algorithm is a source of competitive
advantage, as it may be the case for search engines, navigation apps and matching
platforms, companies may actually face a greater competitive pressure to develop the
best-performing algorithm.
Similarly, if algorithms allow companies to differentiate their services or the production
process in such a way that leads to cost asymmetry, collusion might be again harder to
sustain, due to the inherent difficulties of finding a focal point to coordinate and as a
result of the low incentives for the low-cost firms to collude. This suggests that some
supply characteristics of digital markets may somehow counterbalance the enhanced risk
of collusion resulting from more transparent markets where companies react fast.
4.2.3 The impact of algorithms on the likelihood of collusion
The previous analysis suggests that it may be quite hard to evaluate whether algorithms
increase or reduce the likelihood of collusion, as they have changed structural market
conditions and supply side factors that, all together, could have a positive, negative or
ambiguous impact on the sustainability of collusion. Table 2 summarises the main
relevant factors for collusion and identifies, for each, the expected impact of algorithms.
Table 2. Impact of algorithms on collusion
Relevant factors for collusion Impact of algorithms on the likelihood of collusion
Structural characteristics Number of firms ±
Barriers to entry ±
Market transparency +
Frequency of interaction +
Demand variables Demand growth 0
Demand fluctuations 0
Supply variables Innovation -
Cost asymmetry -
Note: + positive impact; - negative impact; 0 neutral impact; ± ambiguous impact.
There are other considerations that make the impact of algorithms on the likelihood of
collusion uncertain. For example, in highly dynamic markets where companies have
distinct sizes, sell differentiated products and play heterogeneous business strategies, tacit
collusion might be very hard to achieve due to the lack of a natural focal point.
Algorithms have no impact at all on some of the other factors that the economic literature
has identified as making markets more prone to collusion and which are related to the
nature of the product, the characteristics of the firms and the existence of alternative
competitive constraints. For instance, collusion is likely to arise not only in concentrated
markets, but also in markets which involve homogenous products. If products are
differentiated, co-ordinating prices and other keys terms of sale must be quite difficult

24 │
despite the use of algorithms by the participating firms. Likewise, if competitors have
different cost levels, different capacity utilisation ratios, market shares and customer
loyalty, collusion is harder to achieve and sustain over time regardless of the fact that
competitors may be using algorithms. Finally, the existence of fringe competitors and
buyers with strong buying power, all of whom may also be using their own pricing
algorithms to disrupt any attempt to reach terms of coordination may make algorithmic
collusion an unlikely outcome.
Despite the apparently ambiguous effects, algorithms appear to have changed more
substantially the structural characteristics that raise competition concerns, namely market
transparency and frequency of interaction, as compared to other structural characteristics
or demand and supply factors. Annex 1 proposes a simple exercise using the standard
modelling of collusion in the economic literature, which mathematically shows that, in a
perfectly transparent market where firms interact repeatedly, when the retaliation lag
tends to zero, collusion can always be sustained as an equilibrium strategy.
The consequence from this model is the following: if markets are sufficiently transparent
and firms can adjust their decisions very fast, for instance by changing prices in real time,
collusion is always sustainable, regardless of the potential counter-balancing effect of
other factors, such as the number of firms in the industry or the risk that innovations will
disrupt the market in the future. The intuition for this result is straightforward, as the
combination of perfect market transparency with very frequent interaction entirely
eliminates the profitability of deviations, which can be easily identified and immediately
retaliated.
Naturally, the assumptions underlying the model are strong and not necessarily verified in
reality, as it may be quite difficult to observe perfect market transparency and
instantaneous retaliation. In addition, the fact that collusion can be sustained as an
equilibrium strategy does not necessarily imply that it will. As Selten (1973) has shown,
in industries with a high number of firms each player has an incentive to stay out of the
cartel and benefit from the so-called ‘cartel umbrella’. This may give companies an
incentive to wait for others to initiate an agreement and ultimately result in failure to
coordinate.
Still, there is a clear risk that current changes in market conditions may facilitate anti-
competitive strategies, such as collusion and other market manipulations. The stock
exchange is a particularly good example of market where stock prices are transparent,
securities are transacted at high speed and, accordingly, market manipulations have been
observed despite the existence of strong regulatory institutions (Box 7).
4.3 Role of algorithms as facilitators of collusion
Taking into consideration the recent evolution of the digital economy and previous
experience in some markets, competition law enforcers should be at least alerted to the
risk that collusion might become easier to sustain and more likely to be observed when
algorithms are involved. Once it is asserted that collusion might be easier to sustain in
digital markets characterised by high transparency and frequent interaction, the natural
question that follows is how companies can actually establish collusion and put in place
the necessary structures to coordinate strategies, allocate gains and enforce the agreement.

│ 25
Box 7. The 2010 “Flash Crash” in financial markets
On 6 May 2010 financial markets suffered one of the fastest and strongest shocks in history
that would become known as the “Flash Crash”, when several stock indices such as the S&P
500, Nasdaq 100 and Dow Jones deeply collapsed and rebounded in the short time frame of
36 minutes.1 During that period, more than 8000 securities temporarily dropped between 5%
and 15%, and 300 other securities were transacted at prices more than 60% away from their
original value, with the market closing at losses around 3% from the previous day.2 Although
the exact cause of the Flash Crash is still subject of enormous debate, most theories suggest
that algorithmic trading activity and automated execution programs had a fundamental role in
setting up the conditions that would eventually lead to the crash.
According to a joint report by the US Commodity Futures Trading Commission and the US
Securities & Exchange Commission,2 the Flash Crash was triggered by a mutual fund
company that used an automated execution algorithm to sell 75.000 E-Mini S&P 500
contracts (E-Minis) valuated in nearly USD 4.1 billion, in a market context of high price
volatility and reduced liquidity during the European debt crisis. The execution algorithm was
programmed to sell the E-Minis at an execution rate proportional to the trading volume in the
market, regardless of the price or timing of the transactions. Most sell orders were absorbed
by high frequency traders who, using automated algorithms themselves, rapidly accumulated
E-Mini contracts that they started aggressively selling and rebuying among each other. As
the trading volume increased in the market, the algorithm of the mutual fund company
submitted more sale orders and fostered a liquidity crisis, until an automatic mechanism of
the stock market interrupted transactions for five seconds, apparently allowing algorithms to
reboot and prices to recover.
Five years after the event, the US Department of Justice charged an independent British
trader, Navinder Singh Sarao, who may have also contributed to the Flash Crash by
designing “spoofing” algorithms to manipulate the market.3 According to a press release of
the Department of Justice,4 Sarao used the automated trading program to bid large sale
orders of E-Minis and to cancel them just before execution, sending a pessimistic signal to
the market and making prices temporarily drop. By purchasing the E-Minis at low prices and
selling them after the prices recovered, Sarao gained USD 900 thousand during the Flash
Crash and, using similar illegal strategies over four years, he made around USD 40 million in
profit.5
Despite the lack of consensus about what actually triggered the Flash Crash, an empirical
study by Kirilenko et al. (2010) suggests that the incident was at least facilitated by high
frequency trading, which “contributes to flash-crash-type events by exploiting short-lived
imbalances in market conditions” and is a root cause of fragility of automated markets. This
case certainly illustrates how certain market conditions, such as high frequency trading, may
enable automated execution algorithms to distort and manipulate markets.
1. Kirilenko et al. (2010).
2. CFTC and SEC (2010).
3. Brush et al. (2015).
4. DOJ (2015b).
5. Stafford and Croft (2016).
One of the main risks of algorithms is that they expand the grey area between unlawful
explicit collusion and lawful tacit collusion, allowing firms to sustain profits above the
competitive level more easily without necessarily having to enter into an agreement. For
instance, in situations where collusion could only be implemented using explicit
communication, algorithms may create new automatic mechanisms that facilitate the
implementing of a common policy and the monitoring of the behaviour of other firms
without the need for any human interaction. In other words, algorithms may enable firms
to replace explicit collusion with tacit co-ordination.

26 │
This section describes in more detail how the use of algorithms might increase the risk of
tacit collusion and discusses a non-exhaustive list of the possible roles of algorithms in
governing such structures and achieving the collusive outcome.13
4.3.1 Monitoring algorithms
The most obvious and simple role of algorithms as facilitators of collusion is in
monitoring competitors’ actions in order to enforce a collusive agreement. This role may
include the collection of information concerning competitors’ business decisions, data
screening to look for any potential deviations and eventually the programming of
immediate retaliations.
The collection of data might be the most difficult step out of this process. Even if pricing
data is publicly available it does not necessarily mean that a market is transparent.
Companies that take a part in a conspiracy still need to aggregate that data from all
competitors in an easy-to-use format that can be regularly updated. This is already done
by some price comparison websites, also known as aggregators, which either receive data
directly from online companies or, instead, use web scraping, an automated process to
extract data from websites using software applications such as internet bots. As new
automatic data collection processes become available, these technologies will likely be
extended from electronic commerce to traditional markets (Box 8). As a result, colluding
companies will be able to increasingly monitor each other’s actions using sophisticated
algorithms.
Box 8. Monitoring fuel prices with computer vision algorithms
The petrol industry is a notable example of a highly studied market where, despite firms
transacting a relatively homogeneous product, price dispersion tends to be observed across
competing petrol stations.
1
While fuel price comparison websites already exist in many
jurisdictions, these aggregators often incur in significant costs related to collecting manually
prices from different petrol stations. However, Dong et al. (2008) have proposed and tested a
new application for “wireless sensor network” (WSN) technologies, which could be used to
implement an automatic data collection system that would allow consumers and firms to
monitor fuel prices in close-to-real time.
WSNs consist of networks of a large number of distributed nodes or devices, which are
equipped with sensor technology to detect physical phenomena, such as light or heat.
2
The
system proposed by Dong et al. (2008) involves using a network of mobile phones equipped
with GPS and video cameras owned by individuals who would voluntarily share information
through a mobile application. In order to achieve this purpose, Dong et al. (2008) developed
a prototype computer vision algorithm that is automatically triggered when a mobile phone is
close to a petrol station and which is able to detect and read fuel prices from the images of
price boards collected by the mobile camera. After testing the computer vision algorithm
using 52 images, the system was shown to achieve “a hit rate of 92.3% for correctly detecting
the fuel price board from the image background” and to read “the prices correctly in 87.7% of
them.”
Although the system in Dong et al. (2008) ultimately relies on the willingness of a sufficient
number of users to provide access to their mobile cameras and GPS signals, similar
algorithms can be developed in the future to take advantage of existing networks of devices,
such as cameras in public places. As wireless sensor networks become more common, it will
be increasingly easier to use algorithms to monitor prices even in brick and mortar industries
as it was never possible before.
1. Yilmazkuday and Yilmazkuday (2016) and Haucap et al. (2015).
2. IEC (2014).

│ 27
The data collected by automatic
collection methods can then be
monitored and combined with a
pricing algorithm that automatically
retaliates to deviations from an
agreed price. For instance,
companies may program pricing
algorithms to effectively execute
trigger strategies, which consist in
setting the agreed price as long as
all the rivals do the same, but
reverting to a price war as soon as
any firm deviates (see Figure 3).
Naturally, because algorithms are
very fast at detecting and punishing
deviations, firms do not have any
incentive to actually deviate.
Hence, unlike traditional cartels, it
is very unlikely that price wars
between algorithms will be actually
observed, except if triggered by
algorithmic errors.
Recently, one of these algorithms malfunctioned. The matching led to a price
war, fuelled by algorithms and proceeding at breakneck speed. Before it was
detected, some vendors had lost thousands of dollars. Fortunately, the vendors
noticed a huge spike in traffic as consumers attempted to buy large screen TVs
worth thousands – for mere pennies. Love, 2015
In conclusion, monitoring algorithms may facilitate illegal agreements and make
collusion more efficient, by avoiding unnecessary price wars. Still they do not eliminate
the need for explicit communication during the establishment and implementation of the
cartel. Therefore, while competition law enforcers should certainly be vigilant and alerted
to the role of monitoring algorithms, as long as prices and other trading conditions are
concerted by human beings, this behaviour could be prevented using traditional antitrust
tools.
4.3.2 Parallel algorithms
One of the difficulties of implementing a cartel in highly dynamic markets is the fact that
continuous changes in supply and demand require frequent adjustments of prices, output
and other trading conditions. As a result, companies have to frequently renegotiate the
collusive agreement through meetings, phone calls, emails or using third-parties, all of
which pose a risk of detection. An alternative solution for colluding companies is to
automatise their decision process so that prices react simultaneously to any changes in
market conditions, replicating thereby a scenario of conscious parallelism.
As already discussed, dynamic pricing algorithms have been implemented, for instance,
by airlines, hotel booking services and transportation network companies to efficiently
adjust supply to periods of lower or higher demand, resulting in pro-competitive effects.
However, concerns for competition might arise if companies start sharing the same
dynamic pricing algorithm, which may be programmed not to compete against other
Figure 3. Illustration of monitoring algorithm

28 │
firms, but to set anti-competitive prices (Box 9). Such algorithms would allow companies
not only to collude, but also to have their prices automatically reacting to market changes
without the need to engage in further communications.
Box 9. Algorithmic collusion in the Amazon marketplace
In 2015, the US Department of Justice (DOJ) charged a seller in the Amazon marketplace,
David Topkins, for coordinating with other sellers the prices of posters sold online between
September 2013 and January 2014. According to the details of the investigation released by
the DOJ, David Topkins and the other conspirators designed and shared among each other
dynamic pricing algorithms, which were programmed to act in conformity with their
agreement. As stated by the Assistant Attorney General Bill Baer:
Today’s announcement represents the Division’s first criminal prosecution against
a conspiracy specifically targeting e-commerce (…) We will not tolerate
anticompetitive conduct, whether it occurs in a smoke-filled room or over the
Internet using complex pricing algorithms. American consumers have the right to
a free and fair marketplace online, as well as in brick and mortar businesses.
1
Although algorithmic pricing is not illegal and is a common practice in many industries, it was
the existence of an agreement to jointly implement the algorithm that made the case for the
DOJ. Since then, researchers have been able to detect many other algorithmic sellers in the
Amazon marketplace by using target price times series,
2
but no new cases of collusion have
been identified. Indeed, to the best of our knowledge, this is so far the only competition case
of algorithmic collusion being detected by any competition authority and resulting in criminal
prosecution, having thus a fundamental role in the growing interest by academic and
practitioners in the risks of algorithms for competition.
1. DOJ (2015a).
2. Chen et al. (2016).
While sharing pricing algorithms with rivals is a more clear-cut violation of competition
rules, there may be more subtle ways to coordinate parallel behaviour without actually
engaging in explicit communication. For instance, concerns of co-ordination could arise if
firms outsourced the creation of algorithms to the same IT companies and programmers.
This might create a sort of “hub and spoke” scenario where co-ordination is, willingly or
not, caused by competitors using the same “hub” for developing their pricing algorithms
and end up relying on the same algorithms, or very close versions of them, to develop
their pricing strategies (Ezrachi and Stucke, 2015). Similarly, a collusive outcome could
be achieved if most companies were using pricing algorithms to follow in real-time a
market leader (tit-for-tat strategy), who in turn would be responsible for programming the
dynamic pricing algorithm that fixes prices above the competitive level (see Figure 4).
Despite the simplicity of tit-for-tat strategies, they have proved to be very effective in
solving the iterated Prisoner's Dilemma game, where players have to repeatedly select a
cooperative or non-cooperative choice on each move. In a computer tournament
organised by Axelrod (1984), professional game theorists were invited to write an
algorithm that would play the iterated Prisoner’s Dilemma game against all the other
algorithms submitted and against its own twin. The winner of the tournament, achieving
co-operation most often, corresponded precisely to the simple version of the tit-for-tat
strategy. More importantly, this algorithm was successful despite the fact that participants
in the tournament did not communicate with each other or coordinated in any way the
programmed decision rules.

│ 29
Figure 4. Illustration of parallel algorithms
Note. Firm 1 is the leader and firm 2 is the follower.
4.3.3 Signalling algorithms
In highly dynamic markets where companies have distinct sizes, sell differentiated
products and play heterogeneous business strategies, tacit collusion might be very hard to
achieve due to the lack of a natural focal point. In order to avoid explicit communication,
companies may attempt to reveal an intention to collude and coordinate more complex
cooperative strategies through signalling and unilateral price announcements.14
As
recognised by Judge Posner during the course of an antitrust litigation:
If a firm raises price in the expectation that its competitors will do likewise, and
they do, the firm’s behaviour can be conceptualized as the offer of a unilateral
contract that the offerees accept by raising their prices.15
It is difficult to establish bright lines on how to treat signalling under competition laws,
given the potential pro and anti-competitive effects of such conduct which obliges
competition agencies to assess if, on balance, the anticompetitive effects of unilateral
disclosure of information outweigh any efficiency enhancing effects (OECD, 2012).
Greater transparency in the market is generally efficiency enhancing and, as such,
welcome by competition agencies. But it can also produce anticompetitive effects by
facilitating collusion or providing firms with focal points around which to align their
behaviour if transparency only goes to the benefit of suppliers.

30 │
Box 10. The airline tariff publishing company case
During the early 1990’s, the US DOJ investigated tariff fixing activities in the airline industry,
where the cartel members were able to implicitly coordinate tariffs using a third party centre
and sophisticated signalling mechanisms. The case is described in detail in Borenstein
(1999).
In the US airline industry, airline companies send fare information on a daily basis to the
Airline Tariff Publishing Company (ATPCO), a central clearinghouse that compiles all the
data received and shares it in real time with travel agents, computer reservations systems,
consumers and even the airline companies themselves. The database published by ATPCO
includes, among other things, information about prices, travel dates, origin and destination
airports, ticket restrictions, as well as first and last ticket dates, which indicate the time range
when the tickets at a particular fare are for sale.
According to the case presented by the DOJ, airline companies were using first ticket dates
to announce tariff raises many weeks in advance. If the announcements were matched by
the rivals, when the first ticket date arrived all companies would simultaneously raise the
tariff. Some of the co-ordination strategies were more complex, involving the use of fare code
numbers and ticket date footnotes to send signals or negotiate multimarket co-ordination.
According to the DOJ’s case it was the existence of a fast data exchange mechanism to
monitor tariffs and react rapidly to price changes that enabled companies to collude without
explicitly communicating. As tacit collusion is not forbidden by competition law and any
explicit co-ordination was very hard to prove in a criminal case, eventually the DOJ reached a
settlement agreement with the airline companies, under which the latter agreed to stop
announcing most price increases in advance, with the exception of a few circumstances
where early announcements could enhance consumer welfare.
1
All of the airline defendants’
fares had to be concurrently available for sale to consumers.
1. United States v. Airline Tariff Publishing Co., 1994-2 Trade Cas. (CCH) ¶70,687 (D.D.C. Aug. 10,
1994).
Source: Extract from OECD (2016a).
Although signalling may be observed virtually in any market, it usually does not come
without a cost. Whenever a firm increases the price to indicate an intention to collude, if
most competitors do not receive the signal or intentionally decide not to react, the
signalling firm loses sales and profits. This risk might encourage firms to wait for other
competitors to signal, eventually leading to delay or even failure to coordinate
(Harrington and Zhao, 2012). Algorithms might reduce or even entirely eliminate the cost
of signalling, by enabling companies to automatically set very fast iterative actions that
cannot be exploited by consumers, but which can still be read by rivals possessing good
analytical algorithms. There may be several ways to achieve this. For instance, firms may
program snapshot price changes during the middle of the night, which won’t have any
impact on sales but may be identified as a signal by rivals’ algorithms. Or, alternatively,
companies may use algorithms to publicly disclose a lot of detailed data that is used as a
code to propose and negotiate price increases, as it was observed in the US case on
Airline Tariff Publishing Company (Box 10).
The evidence from the ATPCO case shows that signalling can be very effective not only
in establishing an informal cartel, but particularly in supporting negotiation between
companies whose interests are not necessarily aligned. When empowered with
technologically advanced algorithms, this informal negotiation process may become even
faster and more efficient. Similarly to the study conducted by Axelrod (1984), Henderson
et al. (2003) used a tournament-based approach to evaluate the performance of alternative

│ 31
negotiation algorithms, some of which interacted autonomously to obtain the best deal
while at the same time accommodating the interests of the other part.
Figure 5 illustrates how a general
signalling algorithm could be set in
order to establish and negotiate the
terms of collusion before actually
engaging in price co-ordination. As
portrayed in the flowchart, each
firm continuously sends new
signals (for instance, offers to raise
prices) and monitors the signals
sent by the other competitors.
When all players finally reach an
agreement by sending the same
signal, they fix the agreed price
until a new successful negotiation
takes place.
4.3.4 Self-learning algorithms
Finally, the most complex and
subtle way in which algorithms can
achieve collusive outcomes is through the use of machine learning and deep learning
technologies, which may potentially enable a monopoly outcome even without
competitors explicitly programming algorithms to do so. In other words, there is the risk
that some algorithms with powerful predictive capacity, by constantly learning and
readapting to the actions of other market players (who may be human beings or artificial
agents themselves), will be able to collude without the need for any human intervention.
It is still not clear how machine learning algorithms may actually reach a collusive
outcome. But once it has been asserted that market conditions are prone to collusion, it is
likely that algorithms learning faster than humans are also able through high-speed trial
and error to eventually reach a cooperative equilibrium. Even when an infinite number of
anti-competitive prices can be sustained, it is also likely that self-learning algorithms can
more easily determine the price that maximises joint profits and which harms consumers
the most.
Some studies in game theory have analysed the ability of machine learning to achieve
cooperative outcomes. In particular, Hingston and Kendall (2004) simulated an
evolutionary game scenario where a population of model-based adaptive agents and non-
adaptive agents play the Iterated Prisoner’s Dilemma. In their simulation, the adaptive
players that learn fast perform better than non-adaptive agents. More recently, Agrawal
and Jaiswal (2012) proposed also a machine learning algorithm for the Iterated Prisoner’s
Dilemma that is shown to perform strictly better than the tit-for-tat strategy discussed
before.
It is hard to know whether self-learning algorithms are already leading to collusive
outcomes in digital markets or to detect when that happens, as machine learning may
result in collusion being observed in effect and substance, but not in format – designated
here as virtual collusion. Indeed, by relying on machine learning to move business
decisions from humans to computers, managers do not only avoid any explicit
communication during the initiation and implementation stages of collusion, but are also
Figure 5. Illustration of signalling algorithms

32 │
released from the burden of creating any structures, such as signalling mechanisms, that
could be seen by authorities as facilitating practices of collusion (OECD, 2007).
If companies go one step further and implement
deep learning algorithms to automatically set
prices and other decision variables, collusion
becomes even harder to prevent using traditional
antitrust tools. The way a deep learning
algorithm works can be simplistically illustrated
with the concept of a “black box” (see Figure 6),
as it processes raw data in a complex, fast and
accurate way, resembling the human brain, and
delivers an optimal output without revealing the
relevant features that were behind the decision
process. Therefore, by relying on deep learning,
firms may be actually able to reach a collusive
outcome without being aware of it, raising
complex questions on whether any liability
could ever be imposed on them should any
infringement of the law be put in place by the
deep learning algorithm.
In conclusion, this section describes a non-exhaustive list of algorithms that pose multiple
different risks for market competition. Table 3 summarises the main characteristics of
each type of algorithm and the role they may play in implementing a collusive
arrangement. The next section addresses these challenges in more detail and discusses
how they could affect competition policy in the future.
Table 3. Summary of the roles of algorithms in implementing collusion
Role in implementing collusion
Monitoring algorithms Collect and process information from competitors and eventually punish deviations.
Parallel algorithms
Coordinate parallel behaviour, for instance by programming prices to follow a leader;
sharing pricing algorithms; or using the same third party algorithm.
Signalling algorithms
Disclose and disseminate information in order to announce an intention to collude and
negotiate the common policy.
Self-learning algorithms
Maximise profits while recognising mutual interdependency and readapting behaviour to
the actions of other market players.
Figure 6. Illustration of collusion as
a result of deep learning algorithms

│ 33
5. Algorithms and challenges for competition law enforcement
The current growth in the use of algorithms combined with developments in machine
learning have induced many changes in digital markets and fuelled a wide debate on what
such developments may mean for competition agencies and their enforcement activities.16
Without disregarding the many pro-competitive efficiencies which are commonly
associated with automation, such as cost reduction, quality improvement and better
allocation of resources, the analysis of the previous section alerts to the risk that
algorithms might distort digital markets by creating incentives and mechanisms to collude
that did not exist otherwise. This is not only the case for oligopolistic markets where
algorithms can concur to make collusive outcomes more likely and stable over time, but
also for non-oligopolistic markets where collusion was not considered a significant risk
that competition law should be too concerned with.
Although the use of algorithms by companies is widespread in certain industries, the use
of complex algorithms based on deep learning principles may still be relatively rare
across traditional sectors of the economy. At the moment, there is still no empirical
evidence of the effects that algorithms have on the actual level of prices and on the degree
of competition in real markets. This is an area where future research will be certainly
welcomed to inform policy choices that governments will be facing. However, as more
processes become automated and more transactions digitalised, one could expect the use
of algorithms to be increasingly more common in the future. And with that, one could
also expect challenges will arise for agencies when enforcing standard antirust tools
(Box 11) as well as an increase in the risks associated with collusion.
Before addressing some of the open questions for enforcement in an algorithmic business
environment, it is important to make a distinction between instances where algorithms
amplify conduct which is already covered under the current legal framework and
algorithms which to some extent create new risks related to behaviours not covered by the
current antitrust rules. Under the first scenario, the discussion is rather straightforward, as
algorithms ought to be assessed together with the main infringement that they help
enforcing. While detecting the existence of an infringement17
and proving such an
infringement might still be complex because of the presence of an algorithm, agencies
can nevertheless rely on existing rules on anti-competitive agreements, concerted
practices and facilitating practices, which offer agencies a framework to assess algorithms
either on their own or as practices ancillary to a main infringement. As such, the
challenges for agencies are left to understanding how the technology works and how the
algorithm can facilitate or support the main antitrust infringement.
On the other hand, the discussion on what competition agencies can do in order to address
algorithms’ risks for competition is more complex under the second scenario, when such
conduct is not covered by standard antitrust rules on co-operation between competitors.
This is the issue of algorithms achieving a tacitly collusive equilibrium without any need
for contact between competitors or without putting in place any facilitating practice.

34 │
Algorithms can amplify the so called “oligopoly problem” and make tacit collusion a
more frequent market outcome.18
Accordingly, this section focuses on the more complex challenges brought by algorithms
under the second scenario. After discussing the relationship between algorithms and tacit
collusion, this section questions whether the notion of agreement (which does not cover
pure unilateral business conduct) should be revisited and debates the scope for antitrust
liability. Finally, this section debates the antitrust tools that could be used to address, at
least partially, the competition concerns brought by algorithms. While it is still uncertain
how competition policy can be adapted to react to algorithmic collusion, it is the purpose
of this paper to frame the ongoing debate by identifying the main challenges raised by
algorithms for antitrust enforcement and to anticipate possible solutions for future
discussion.
Box 11. Individualised pricing and standard antitrust tools
Algorithms can raise specific challenges for enforcers which go beyond proving an anti-
competitive agreement. The ability of pricing algorithms to discriminate between buyers
based on a number of pre-determined variables may have important consequences on the
general application of economic competition tools in markets where pricing algorithms are
used by suppliers to establish individualised prices.
Current antitrust analysis largely depends on the ability of the enforcer to observe prices and
to infer conclusions from the way in which companies price goods and services. An analysis
that is often crucial in an antitrust case consists in looking at whether the supplier has the
incentive and ability to raise prices above a competitive level; this is a key question, for
example, when trying to define markets based on a standard SSNIP (Small but Significant
and Non-Transitory Increase in Price) test, measure market power of a firm or assess the
impact of a merger based on a unilateral effects analysis. Typically this requires a
comparison between the market price given the conduct under review and a hypothetical (or
pre-conduct) competitive market price.
However, it is not clear how these standard economic antitrust tools will be applicable when
individualised prices are made possible by algorithmic pricing. These pricing mechanisms are
capable of changing prices based on supply and demand at such a speed and with so many
interactions between them that a standard price comparison will be quite a difficult exercise.
The challenges that these new forms of pricing mechanisms present to agencies are still
largely unexplored, as enforcement has not yet been confronted with such questions, but
they can arise more frequently in the future as cases involve more often automatic
individualised pricing models by the firms under investigation.
5.1 Algorithms and tacit collusion
If assessing algorithms ancillary to a main infringement is a rather straightforward case
from a legal perspective, the analysis of algorithms which escape the traditional legal
framework is more complex, because they allow collusion to take place without the need
for any communication or contact between competitors. Algorithms are changing the
characteristics of digital markets, increasing transparency, the velocity of business
decisions and the ability of companies to immediately respond to rivals’ actions. Under
these circumstances, algorithms may make firms’ actions interdependent without the need
for explicit communication or interaction, increasing thereby the risk of tacit collusion
and subsequently leading to higher price levels.

│ 35
Box 12. The “oligopoly problem”
The “oligopoly problem”, an expression sometimes attributed to Posner (1969), refers to the
concern that high interdependence and mutual self-awareness in oligopolistic markets might
result in tacit collusion, an outcome which is socially undesirable but that falls out of the reach
of competition law. Competition authorities in some jurisdictions have attempted to extend
antitrust tools to address the oligopoly problem, using in particular two distinct solutions:
 Ex ante merger control rules to prevent structural changes which favour
coordinate effects.
 Ex post rules to prevent unilateral conducts that promote oligopolistic
interdependence, such as facilitating practices, under the notion of joint
dominance.
On the other hand, some critics of the oligopoly problem argue that it is rather unlikely that
oligopoly structures unavoidably lead to tacit collusion, questioning therefore the need for
antitrust intervention in this area. Whish and Bailey (2012) identify some of the counter-
arguments to the oligopoly problem:
 The interdependence of oligopolists is frequently overstated. Even in
symmetrical oligopolies a firm might be able to raise profits by cutting its
price, as rivals may take time to react or might be unable to expand their
capacity in order to meet the increased demand.
 Oligopolists frequently have different cost levels, produce differentiated
goods, have different market shares and can build customer loyalty, which
makes tacit collusion harder.
 Competitive edges of small market players may act as a competitive
constraint on the oligopolists.
 Even if oligopolists are able to sustain tacit collusion and earn supra-
competitive profits over a short period that would attract new entrants to the
market and increase competition in the long run, unless there are significant
barriers to entry.
Note. For a general discussion on competition policy in oligopoly markets, see OECD (2015b).
As a result, one can identify similarities between the concerns raised by algorithms and
the ones posed by the classical oligopoly problem (Box 12). It has been for long
recognised that in highly concentrated, stable and transparent markets the actions of firms
have a significant impact upon those of rivals. Therefore, after a period of repeated
interactions, firms become conscious that their respective strategic choices are
interdependent and that, by matching each other’s conduct, they can set prices at a supra-
competitive level, without actually communicating. In other words, the structure of some
markets is such that, through interdependence and mutual self-awareness, prices may rise
towards the monopoly level (Whish and Bailey, 2012).
So far, competition policy has provided limited solutions to the oligopoly problem, as it is
commonly understood that the particular conditions required to sustain tacit collusion are
rarely observed, with the exception of a few markets with a very small number of
competitors (possibly even only in a duopoly context), high degree of transparency and
high barriers to entry (Potters and Suetens, 2013). However, algorithms might affect some
characteristics of digital markets to such an extent that tacit collusion could become
sustainable in a wider range of circumstances possibly expanding the oligopoly problem
to non-oligopolistic market structures.19

36 │
In addition to the concerns typically evoked in the oligopoly problem, algorithms might
directly facilitate a non-competitive equilibrium by working as instruments that eliminate
the need for explicit communication or interaction between competitors. Indeed,
algorithms function as an “intermediary” between firms, by collecting and processing
market data, and rapidly responding to rivals’ actions. As noted by Mehra (2015),
algorithms might actually succeed in governing collusive structures better than humans,
given their greater accuracy in detecting price changes, the elimination of the element of
irrationality and the reduction of the chance that the collusive scheme is undermined by
mistake.
In light of all these considerations, one should question whether, with the recent
developments of algorithms, tacit collusion is likely to become a more common
phenomenon in digital markets. If that is the case this could ultimately result in
substantial consumer harm given the difficulties to address tacit collusion under the
current competition law frameworks. If future research supports this hypothesis,
competition law enforcers might want to consider if their current approach to the legal
treatment of tacit collusion needs adjusting.
5.2 The notion of agreement: does it need revisiting?
Because of the way in which competition laws are designed, identifying an “agreement”
between competitors is a prerequisite to enforce the law against collusive outcomes. In
most cases, the term agreement is broadly defined to ensure the widest possible reach of
the competition rules (Box 13).
In practice, the definition of agreement may still provide little guidance on whether more
subtle forms of communication fall in the scope of application of the competition rules.
For instance, signalling mechanisms such as unilateral public price announcements could
be seen as an invitation to reach a common policy, but it is questionable under the law of
many jurisdictions if they can amount (and under which circumstances) to an agreement.
In the absence of communication and explicit co-ordination, the applicability of
provisions on agreements is not straightforward. Mere parallel conduct, such as
simultaneous price increases by competitors, is insufficient to indicate co-ordination since
it can result from independent and rational behaviour.
This raises the concern of whether the need to address algorithmic collusion should
require a new definition of what is an agreement for antitrust purposes. This is not a new
question for antitrust scholars20
but the question has come up again in recent times in a
“renewed debate about whether classic oligopoly behaviour can be prosecuted as an
unlawful agreement” (Hay, 2013).

│ 37
Box 13. The notion of “agreement” for antitrust purposes
In the European Union, the Article 101 of the TFEU applies to all “agreements” and
“concerted practices”, although the TFEU does not provide clear-cut definitions of what is an
agreement or a concerted practice. According to the courts, an agreement reflects “a
concurrence of wills between economic operators on the implementation of a policy, the
pursuit of an objective, or the adoption of a given line of conduct on the market, irrespective
of the manner in which the parties’ intention to behave on the market in accordance with the
terms of that agreement is expressed.”
1
In other words, the concept of agreement
simultaneously involves the existence of a common will and some form of manifestation,
whether it is implicit or explicit. In the absence of a formal agreement, the category of
concerted practices can be applied. This involves, among other factors, direct or indirect
contacts intended to deliberately influence the conduct of other firms.
In the United States, the section 1 of the Sherman Act uses multiple terms to refer to an
agreement, including “contract”, “combination in the form of trust” and “conspiracy”.
According to the Supreme Court of the US, an agreement does not necessarily need to be
explicit or formal, as long as it involves “a unity of purpose or a common design and
understanding, or a meeting of minds”,
2
as well as “a conscious commitment to a common
scheme”.
3
This definition is, in principle, very broad and could potentially cover parallel
conduct. In practice, courts have required evidence that observed parallel conduct is indeed
the result of co-ordination among the parties and not mere oligopolistic interdependence (so-
called “plus factors”). An example of a “plus factor” required by courts is that the parties have
communicated their intentions to act in a certain way.
1. Case T-41/96, Bayer AG v Commission, [2000] ECR II-3383, para 173.
2. Interstate Circuit Inc v United States, 306 US 208, 810 (1939); Am. Tobacco Co. v. United States, 328
U.S. 781, 809–10 (1946), at 810.
3. Monsanto Co. v. Spray- Rite Serv. Corp., 465 U.S. 752, 768 (1984); In re Flat Glass, 385 F.3d at 357.
Kaplow (2011) argues that the current approach to horizontal agreements may be too
formalistic and incapable of addressing harmful interdependence among firms. He also
argues that a narrow view of the meaning of agreement is in contrast with a more
economically-based approach to competition law, stating that “successful interdependent
coordination that produces supra-competitive pricing leads to essentially the same
economic consequences regardless of the particular manner of interactions that generate
this outcome.” He has highlighted limitations of this approach, such as the reliance of US
doctrine on communication as an important condition to define the concept of agreement,
advocating thus for a broader interpretation of what constitutes an agreement. However,
the suggestion that competition law should capture standard oligopoly behaviour is not
without critics, as it is the case of Judge Posner who has recently warned against the
danger of the laws “treating tacit collusion as if it were express collusion.”21
Similar challenges can arise with algorithms. As the development of algorithms
increasingly allows for rapid and sophisticated interactions between competitors, who
might use complex codes as intermediates to reach a common objective, the notion of
agreement and its applicability to the digital economy becomes more unclear:
While an agreement among competitors to fix prices is per se illegal, computer
technology that permits rapid announcements and responses has blurred the
meaning of 'agreement' and has made it difficult for antitrust authorities to
distinguish public agreements from conversations among competitors.
Borenstein (1997)

38 │
In light of the role of algorithms in reaching and enforcing a common policy, some have
raised the question of whether the concept of agreement should be revisited in order to
incorporate other “meetings of minds” that are reached with the assistance of algorithms.
For instance, the signalling algorithms discussed above might result in very fast iterative
change of prices that eventually converge to a common value, resembling an actual
negotiation process between businessmen to implement a collusive agreement. Could
therefore be concluded that the fast adjustment of prices in reaction to competitors until
convergence is reached is tantamount to an agreement?
Likewise, the parallel algorithms mentioned above could potentially be seen as an alternative
automatic mechanism to implement an agreement. For instance, in a “follow-the-leader”
strategy, a firm could make an offer to collude by implementing an algorithm that imitates in
real-time the price of the market leader, while the leader could accept the offer by increasing
the price in reaction to the competitor’s algorithm. Alternatively, a firm may make an
invitation to collude by publicly releasing a pricing algorithm, while competitors would
accept the offer by using the same algorithm as part of their business strategy.
Box 14. Unfair competition standards and Section 5 US FTC Act
In the United States, under Section 5 of the FTC Act, the US Federal Trade Commission
(FTC) has the power to prohibit “unfair methods of competition”. The Supreme Court has held
that Section 5 extends beyond the Sherman Act and other US antitrust laws,
1
and over time
the FTC has relied on this power to tackle various conducts that have anti-competitive
effects, but would be challenging to pursue under cartel or monopolisation provisions. An
example is unilateral communications of information to competitors with anti-competitive
effects, or the so-called “invitations to collude”.
2
Section 5 is broadly constructed and offers flexibility to the FTC on which practices to tackle,
as it is a provision which is “principle-based” rather than “rule-based”.
3
Under the current
legal standard, the FTC would need to show that a practice (e.g. the use of an algorithm) is
unfair because (1) it causes or is likely to cause substantial injury to consumers; (2) it cannot
be reasonably avoided by consumers; and (3) is not outweighed by countervailing benefits to
consumers or to competition.
It has been suggested that statutes like Section 5 could be used to tackle algorithmic
collusion if the agency could show that, when developing the algorithms, defendants were
either motivated to achieve an anticompetitive outcome or were aware of their actions’
natural and probable anticompetitive consequences.
3
1. FTC v. Sperry & Hutchinson Co., 405 US 233 (1972) (Finding that the statute, its legislative history,
and prior cases empower the Commission under Section 5 to define and proscribe an unfair competitive
practice, even though the practice does not infringe either the letter or the spirit of the antitrust laws); and
FTC v. Indiana Federation of Dentists, 476 US 447 (1986) (“The standard of ‘unfairness’ under the FTC
Act is, by necessity, an elusive one, encompassing not only practices that violate the Sherman Act and
the other antitrust laws […] but also practices that the Commission determines are against public policy
for other reasons”) (dictum).
2. OECD (2015b) also notes that, because Section 5 can only be enforced by the FTC, an expert
administrative body, it limits the potentially burdensome combination of class actions and treble damages
that could result from private antitrust enforcement in this area. In stand-alone cases violating Section 5
(i.e. cases that have not been alleged to violate the other antitrust laws) the FTC has tended to restrict its
remedies to prohibiting future recurrence of the conduct (e.g. prohibiting future unilateral communications
of the type that gave rise to concerns) rather than seeking damages or punitive sanctions.
3. For example, see Ethyl Corp v FTC, 729 F.2d 128, 136 (2d Cir. 1984) (“Congress, in the process of
drafting Sec. 5, gave up efforts to define specifically which methods of competition and practices are
competitively harmful and abandoned a proposed laundry list of prohibited practices for the reason that
there were too many practices to define and many more unforeseeable ones were yet to be created by
ingenious business minds.”).
4. Ezrachi and Stucke (2017).

│ 39
At this point it is still very hard to draw firm conclusions on whether algorithmic
interactions (or a “meeting of algorithms”) should be treated similarly to a “meeting of
minds” under the definition of agreement covered by competition rules. However, a more
clear definition of agreement could not only reduce uncertainty by helping businesses
understanding which practices are illegal and which ones are acceptable, but also to
potentially address some of the concerns related to algorithmic collusion.
Because proving that a pure unilateral conduct such as conscious parallelism constitutes
an agreement restricting competition can be difficult under the legal standards, in
particular in jurisdictions where the concept of agreement has been interpreted narrowly,
some agencies may have the possibility to rely on legal standards such as “unfair
competition” which provide the enforcement agency with more flexibility (Box 14).
5.3 The scope for antitrust liability
A question discussed in the literature on algorithms and competition is whether antitrust
liability can be established when pricing decisions are made by a machine using an
algorithm rather than by human beings.
As Mehra (2016) puts it, “[i]n dealing with a robo-seller that takes anticompetitive
actions there are three choices in attributing responsibility: to the robo-seller itself, to the
humans who deploy it, or to no one.” While the third option (no liability) cannot be
considered a realistic one, as it would de facto provide impunity for anti-competitive
conduct put in place through the intermediary of an algorithm, the debate has highlighted
the challenges of attributing antitrust liability to individuals when commercial strategies
are delegated to an algorithm and humans have no ability to influence the way in which
such decisions are taken. Of course, most algorithms today still operate based on
instructions designed by human beings and there is no doubt that humans will be in most
cases responsible for the decisions made by algorithms. Based on the current stand of the
law, computer programs and algorithms are to be considered simply as tools, implying
that their decision can be directly attributed to their human operators.
As the European Commissioner Vestager stated in a recent speech:
The challenges that automated systems create are very real. If they help
companies to fix prices, they really could make our economy work less well for
everyone else. (…) So as competition enforcers, I think we need to make it very
clear that companies can’t escape responsibility for collusion by hiding behind a
computer program. Vestager (2017)
However, as AI develops further, the links between the agent (the algorithm) and its
principal (the human being) become weaker and the ability of algorithms to act and price
autonomously puts in question the liability of the individuals or firms who benefit from
the algorithm’s autonomous decisions. In such cases, determining liability mainly
depends on the facts at hand. As noted by Ezrachi and Stucke (2016), defining a
benchmark for illegality can be challenging since it requires assessing whether any illegal
action could have been anticipated or predetermined by the individuals who benefit from
the algorithm. Such assessment includes, for example, a careful consideration of the
programed instructions of the algorithm, available safeguards, reward structure and the
scope of its activities.
Furthermore, agencies should consider the extent to which humans can control the
activities of algorithms. It can be obvious that algorithms are designed by humans, but do

40 │
they intentionally create algorithms to harm consumers? Could liability be automatically
charged jointly and severally on the person who designed the algorithm, on the individual
who used it and on the person (or entity) who benefitted from the decision made by the
algorithm?22
These questions do not have clear-cut answers at the moment, but will likely
arise in courts of law as more antitrust cases involving independent algorithmic activities
will be litigated.
5.4 Possible alternative approaches to algorithmic collusion
Aside from more radical interventions (possibly of a legislative nature) that involve
reconsidering the legal approach to tacit collusion, revisiting the notion of agreement that
would fit algorithmic collusion or determining the scope of liability for AI systems, there
are some traditional measures that antitrust agencies can put in place to address at least
some of the competition concerns. Possible alternative approaches could include ex ante
measures such as the use of market studies, merger control enforcement, the use of
remedial actions or even a regulatory approach. This section briefly exposes the first three
approaches, while the last section of this paper focuses on the scope for market regulation
in more detail.
5.4.1 Market studies and market investigations
In general, a prerequisite for antitrust intervention against an anti-competitive agreement
is the existence of evidence of some form of competitors’ co-ordination which has a
negative impact on competition and poses a risk of consumer harm. Yet, when there are
signs that the market is not functioning well, but there are no indications of any co-
ordination among the market players, competition agencies may decide to engage in
market studies or sector inquires in order to understand why the market is failing and to
identify possible solutions.23
Hence, the use of market studies typically precedes other
enforcement actions.
Box 15. Findings on pricing software of the EC sector inquiry on e-commerce
An example of how market studies can assist agencies and governments identifying possible
concerns from algorithmic pricing for competition, as well as possible policy solutions, can be
found in the final Report of the European Commission to the Council and European
Parliament on the E-Commerce Sector Inquiry. Among the main findings, the Report
concludes that
(…) increased price transparency allows companies to monitor more easily
their prices. A majority of retailers track the online prices of competitors. Two
thirds of them use automatic software programmes that adjust their own
prices based on the observed prices of competitors. With pricing software,
detecting deviations from ‘recommended’ retail prices takes a matter of
seconds and manufacturers are increasingly able to monitor and influence
retailers’ price setting. The availability of real-time pricing information may also
trigger automatised price co-ordination. The wide-scale use of such software
may in some situations, depending on the market conditions, raise
competition concerns.
1
1. European Commission (2017).
Source. Ezrachi and Stucke (2017).

│ 41
With respect to algorithmic collusion, agencies could seek to examine whether algorithms
commonly result in coordinated effects and, if so, to attempt to identify the circumstances
and sectors under which algorithmic collusion is more likely to be observed. In this sense,
market studies (or sector inquiries) may support agencies’ efforts to understand the
market characteristics that can lead to collusive outcomes, whether they consist in high
transparency, predictability and frequent interaction or in any other structural
characteristics that have not been identified yet. Ezrachi and Stucke (2017) suggest that
“[s]uch approach may prove useful in helping agencies understand the new dynamics in
algorithm-driven markets and the magnitude of any competitive problems.”
The use of market studies can lead to recommendations for the government to engage in
regulatory interventions to address legal or structural restrictions to competition, as well
as to the opening of investigations when the cause of the concern is behavioural. Market
studies could also lead to advocacy efforts and recommendations to the business
community itself with the objective of fostering stronger compliance with competition
principles. This could result, for instance, in the adoption of self-regulation in the form of
codes of conduct, which companies would agree to comply with when designing and
using pricing algorithms.
Finally, in some jurisdictions, such as the United Kingdom, Iceland and Mexico,
competition authorities can go beyond market studies and engage in so-called market
investigations. These tools allow them to go a step beyond by issuing non-binding
recommendations (which is the typical outcome of a market study) and eventually to
impose structural or behavioural remedies. The advantage of market investigations is that
they offer the agency a degree of flexibility in restoring competition in the market that
would not be possible through other means.
5.4.2 Ex ante merger control
Another possible ex-ante approach consists in establishing a system capable of preventing
tacit collusion, through the enforcement of merger control rules in markets with
algorithmic activities. This may require agencies to consider lowering their threshold of
intervention and investigate the risk of coordinated effects not only in cases of 3 to 2
mergers, but potentially also in 4 to 3 or even in 5 to 4. Such an approach would allow
agencies to assess the risk of future co-ordination, going beyond the traditional duopolies
where tacit collusion is more easily sustainable, to include also cases where the use of
algorithms may facilitate collusion even in less concentrated industries:
One factor is if tacit collusion, because of algorithms, spreads beyond duopolies
to markets with as many as five to six significant players. The agencies can be
more sensitive to whether the elimination of a particular player would increase
significantly the risk of algorithmic tacit collusion. It may be preserving a market
of diverse sellers with different horizons for profits and different capacity
constraints. Ezrachi and Stucke (2017)
In order to effectively prevent algorithmic collusion, competition agencies should focus
their analysis particularly on the impact of the transactions on market characteristics such
as transparency and velocity of interaction, which are the factors that are mostly affected
by the use of algorithms. Ezrachi and Stucke (2017) also suggest that agencies may need
to reconsider the approach to conglomerate mergers when tacit collusion can be
facilitated by multimarket contacts. In particular they note that “one aspect of machine
learning is to discover correlations in large data sets. Thus, the algorithms can ascertain

42 │
and respond to punishment mechanisms in distinct product markets, which to the human
may appear unrelated.”
5.4.3 Commitments and possible remedies
Lastly, competition law enforcers could make tacit collusion harder to sustain through a
behavioural approach, by preventing oligopolists from putting in place harmful
mechanisms for the competitive process in so far as they facilitate collusion. Although
this has not been tested yet, one could argue that relying on certain types of algorithms
may amount to a facilitating practice that could trigger competition law enforcement. This
could be the case, for instance, if competitors facilitated a collusive equilibrium by
relying on monitoring, parallel or signalling algorithms.
In many jurisdictions, enforcement actions can also lead to the adoption of remedies in
the form of commitments (OECD, 2011a and OECD, 2016i). Remedies can resolve and
prevent the harm to the competitive process that may result from a unilateral conduct,
such as a facilitating practice. Although it could be a delicate task to find an adequate
remedy in case of algorithm-driven conducts, it is worth considering the use of remedies
to introduce special compliance or monitoring programs. A similar procedure should be
introduced for the so-called “notice-and-take-down” process, under which the online
hosts should post a notice and remove the content in response to court orders. This way, if
as a result of the use of algorithms any anti-competitive behaviour is detected, an
appropriate action can immediately be taken.
Another possible remedy could be the introduction of auditing mechanisms for
algorithms, which could guarantee that algorithms are programmed in a way to steer clear
of any competition concerns. However, as noted by Ezrachi and Stucke (2016), this can
fail in leading to a meaningful tool, since (1) algorithms do not necessarily include
instructions to collude, but rather to maximise profit; (2) auditing is not likely to keep
pace with the development of the industry, especially given the self-learning nature of
algorithms and (3) it may be hard to prevent algorithms from ignoring information that is
publicly available (“cheap talk” problem).
6. Algorithms and market regulation
From the previous discussion there seems to be scope for competition law enforcers to
tackle some of the anti-competitive effects of algorithms. Yet, as computation methods
become more complex and a phenomenon of “governance by algorithms” is increasingly
observed, this section debates whether competition law is enough to address most of the
existing concerns or whether some form of regulatory intervention is necessary.
In what follows, the general risks of algorithms for society are identified, as well as the
market failures that might prevent such risks from being addressed by self-correcting
mechanisms, therefore serving as arguments in favour of market regulation. Then, some
alternative regulatory approaches are presented, alerting to the risks that over-
enforcement could pose on market competition and innovation. Finally it is discussed the
challenge of designing market regulations with the specific purpose of preventing
algorithms from reaching collusive outcomes.

│ 43
Many of the considerations touched upon in the next sections go well beyond the issue of
collusion and even competition law enforcement in general. Discussions on the viability
and advisability of any regulatory options would also require a comprehensive and
careful consideration of a whole host of other policy areas. While such an exercise goes
far beyond the scope of this paper, these sections frame the competition concerns raised
by the spread use of algorithms in such a wider policy context.
6.1 Arguments in favour of regulating algorithms
The ongoing debate about regulating the digital economy seems a result, at least in part,
of the exponential growth of a few players in digital markets, which include some of the
biggest companies in the world by market capitalisation. The internet giants, as they have
become known, are responsible for the provision of multiple information goods, such as
operating systems, browsers, search engines, email and messaging, navigations maps,
electronic books, music distribution and social networks. In turn, the value added of these
services partially relies on complex proprietary algorithms that are implemented for
multiple ends: dynamic pricing, data mining, result ranking, user matching, product
recommendations and ads targeting, among others.
Despite the immeasurable value that the internet giants have brought to society in the
form of innovative technology and high-quality online services, their increasing
dimension and presence across multiple fundamental markets have attracted not only the
attention of the general public, but also of regulators and policy makers.24
The increasing
reliance of big online companies on secret algorithms poses a concern that the
organisation of the world’s information is, to some extent, controlled by automated
systems in the hands of a few market players.
6.1.1 Risks of “algorithmic selection” beyond collusion
Algorithms can determine the news that online users read, the multimedia content they
access, the products that they purchase and even the individuals that they meet or interact
with socially. As Domingos (2016) puts it, “these days, a third of all marriages start on
the Internet, so there are actually children alive today that wouldn’t have been born if not
for machine learning.”
The use of automated computer systems to organise and select relevant information –
“algorithmic selection” – is not necessarily undesirable, particularly if the decisions and
predictions determined by machines are considerably more efficient, accurate and
objective than those that any human being could make. Nevertheless, due to their vast
potential in supporting decision-making, algorithms are starting to be implemented in
fundamental areas that govern the structure of society, where computer errors or bias
could have major implications. Indeed, in the same way that algorithms perform actions
that cannot be feasibly executed by humans, computers are not infallible and may also
result in errors that are typical of automated systems, leading to a whole new range of
policy concerns. Saurwein et al. (2015) identifies multiple risk categories of algorithmic
selection, some of which are summarised in Table 4.
The risk categories in the table do not only pose social concerns, but can actually have
severe implications for the good functioning of the digital economy. Some of the risks,
such as the potential of algorithms as a tool for abuse of market power, have a direct
impact on competition. Other risks, including information bias, market manipulations and

44 │
violations of property rights, may also indirectly affect competition by creating barriers to
new entrants and reducing incentives to innovate:
The way that algorithms are used to make decisions automatically could even
undermine our democracy. These days, social media is a vital source of news.
(…) So the scope for social media algorithms to create an alternative reality, by
showing people one story after another that just isn't true, is a concern for us all.
Vestager (2017)
Table 4. Risk categories of algorithmic selection
Risk Description Examples
Abuse of
market power
Algorithms programmed to
facilitate anti-competitive
practices, such as collusion, as
well as exclusionary and
exploitative abuses.
 Allegations that search engines manipulate search
results in order to disadvantage competitors1
 Algorithmic co-ordination to fix prices in internet
marketplaces
Bias Information filters that reduce
variety and bias information
according to the preferences of
online users, leading to “echo
chambers”2 and “filter bubbles”.3
 Search engines that provide online readers with
news that match their own believes and
preferences
 Product recommendations for books and movies
with similar content to the ones previously acquired
 Social networks’ updates about the closest
contacts
Censorship Restrictions programmed to
control or block the content that
certain users are able to access.
 Content-control software used by companies to
block sites with particular religious, political and
sexual orientations
 Content-control software implemented by
governments in certain jurisdictions
Manipulation Manipulation of algorithms to
select information according to
business or political interests,
instead of its relevance or
quality.
 Creation of multiple accounts or repetitive
transactions in internet marketplaces in order to
manipulate feedback scores and influence rating
 Design of internet links to bias search engine
results in order to rank certain websites higher4
Privacy rights Automated systems that collect
personal data from users
(sometimes shared with third
parties), posing concerns for
data protection and privacy.
 “Instant Personalization” model adopted by
Facebook in 2010, which allowed service providers
to access users’ profiles5
 Collection of user’s location data in order to better
target advertisement
Property
rights
Use of algorithms to collect,
aggregate, display and share
information goods protected by
Intellectual Property rights.
 News aggregator services that redistribute
fragments of copyrighted articles5
 Unlicensed streaming sites for music and video
Social
discrimination
Automated data-decision
processes that, by considering
personal information in their
formulas, can result in
discriminatory outcomes.
 Pricing algorithms that discriminate based on social
and demographic characteristics, such as location
 Recidivism algorithms that could result in racial
discrimination
1. Patterson (2013).
2. Sunstein (2009).
3. Pariser (2011).
4. Bar-Ilan (2007).
5. Helft and Wortham (2010).
6. Quinn (2014).

│ 45
6.1.2 Market failures
In light of the risks brought by algorithms, a central question for policy makers is whether
market competition combined with existing laws on privacy, intellectual property and
fundamental human rights will suffice to dissipate such concerns, or whether some form
of regulatory intervention in necessary. In other words, will competition be sufficient to
drive out of the market fallible and erratic algorithms, while preserving those that are
efficient and enhance social welfare? While the answer to this question may depend on
the particular risk addressed, it is possible to identify at least three market failures that
could eventually compromise the ability of digital markets to self-correct:
1. Imperfect information resulting from lack of algorithmic transparency: The
lack of transparency in the way algorithms are programmed and run may limit
consumers’ ability to make valid and conscious choices among competing
products. Likewise, law enforcers may lack the necessary information or even
expertise to make sure that automated systems comply with existing regulations.
In part, this lack of transparency results from the fact that most algorithms are
trade secrets. Yet, even if companies publicly release or share their secrets with
some regulators, their long and complex program codes would be still extremely
hard to interpret. The effects of algorithms can be also particularly difficult to
evaluate when outcomes are highly variable and depend on individual
characteristics (Sandvig et al., 2014).
2. Data-driven barriers to entry: The development of good predictive algorithms
requires expensive complementary assets such as advanced data-mining and
machine learning software, as well as physical infrastructures such as data
centres, whose investment is subject to economies of scale. The ability of
algorithms to find new relations and patterns of behaviour also requires access to
a variety of data collected from multiple sources, resulting in economies of scope.
Thereby, small firms that do not have the necessary complementary assets or that
are not simultaneously present in multiple markets might face barriers to entry,
preventing them from developing algorithms that can effectively exert
competitive pressure (OECD, 2016a).
3. Spill-overs associated with information and knowledge: By being
programmed to select only the most relevant and useful information, algorithms
might reduce the variety of ideas and views that individuals are confronted with.
Sunstein (2009) and Pariser (2011) alert to the danger that algorithms tend to bias
information in a way that reinforces our own beliefs, leading to a phenomenon
known as ‘echo chambers’. This could arguably lead to an environment where
static efficiency is optimised but which is less prone to human creativity and
innovation. In the process of research and development, many great discoveries
were made by accident when scientists were confronted with data that were not
relevant for the objectives that they initially had in mind; or when they broke
with established knowledge and tested hypothesis that could appear risky or ill-
advisable for a machine. In economic terms, algorithms might fail to internalise
the spill-overs that a variety of knowledge and multi-disciplinary approaches
could have on the process of innovation.

46 │
6.2 Possible regulatory interventions
Recognising the prominent role of algorithms in the organisation and processing of the
world’s information, with consequences that go far beyond the limits of the digital
economy, academics and policy makers have engaged in a growing debate concerning the
need for new forms of regulatory intervention. Some of the central topics debated include
the institutional options to govern algorithms, alternative regulatory measures and the
risks associated with excessive regulation, all of which are addressed in turn.
6.2.1 Institutional options to govern algorithms
Saurwein et al. (2015) identify several options that have been either proposed or
implemented to govern algorithms, ranging from a continuum between market solutions
and state regulations. Each option has different limitations and might be more or less
appropriate to address each of the risk categories of algorithms identified in the last
section.
In one extreme of the continuum of governance options there are supply and demand-side
market solutions. The first consist in suppliers competing to provide better algorithms
across several dimensions, for instance by using machine learning to reduce bias or to
prevent algorithms from being manipulated. The second correspond to consumers taking
active actions, such as refusing to use certain services or relying on advanced technology
to protect themselves from algorithmic risks (for example, with tools for anonymisation
or to prevent censorship). Whenever possible, market solutions should be applied as they
are less likely to hamper innovation or to deter new entry.
Nonetheless, the market failures previously identified could serve as an argument for
alternative solutions, which include self-organisation (companies committing to good
principles and standards to improve their reputation), self-regulation, co-regulation and
state intervention. Along this range of governance options, many regulatory measures
have been proposed, including information measures, principles of “search neutrality”,
cybercrime regulations, data protection certification schemes, etc.
In the scope of state intervention options, some academics are currently advocating for
the establishment of new regulatory institutions to govern the digital economy. For
instance, Gawer (2016) has suggested the creation of a global digital regulator, a central
and independent agency that would be responsible for coordinating and supervising the
different regulatory aspects of internet and data. On the other hand, in order to address the
risks associated with algorithms and artificial intelligence, others have proposed the
establishment of a new AI regulatory regime:
(…) the scheme outlined (…) proposes legislation, the Artificial Intelligence
Development Act (‘AIDA’), that would create an agency tasked with certifying
the safety of AI systems. Instead of giving the new agency FDA-like powers to
ban products it believes to be unsafe, AIDA would create a liability system
under which the designers, manufacturers, and sellers of agency-certified AI
programs would be subject to limited tort liability, while uncertified programs
that are offered for commercial sale or use would be subject to strict joint and
several liability. Scherer (2016)
Whether the establishment of such regulatory agencies is feasible and socially desirable is
still an open question and one that spans a number of policy areas. So far governments
have adopted a market-oriented approach towards the digital economy, which has greatly

│ 47
contributed to the fast growth of online commerce and encouraged the development of
innovative services, as well as efficient and fast transactions. This market approach has
prevailed since the early stages of the internet, when the White House recommended a set
of principles to expand internet markets through private-sector leadership and minimum
regulatory restrictions on competition:
Unnecessary regulation of commercial activities will distort development of the
electronic marketplace by decreasing the supply and raising the cost of products
and services for consumers the world over. (…) Accordingly, governments
should refrain from imposing new and unnecessary regulations, bureaucratic
procedures, or taxes and tariffs on commercial activities that take place via the
internet. White House (1997)
Before taking any actions, policy makers should cautiously evaluate the risks of over-
enforcement, as excessive regulatory interventions could result in new barriers to entry
and reduce the incentives of companies to invest in proprietary algorithms, which so far
have brought a great value for society. In this matter, the OECD (2009) recommends
governments to evaluate the competitive impact of market regulations, emphasising that
“competition assessment of proposed public policies should be integrated in the policy
making process at an early stage.”
Keeping these important principles in mind, this section questions whether there is scope
to tackle algorithmic risks through market regulation, identifying a few alternative forms
of intervention and the respective risks for competition.
6.2.2 Measures on algorithmic transparency and accountability
Some of the regulatory interventions discussed in most recent debates appear to focus on
making algorithms more transparent and accountable for their effects. In the United
States, the FTC’s Bureau of Consumer Protection established the brand-new Office of
Technology Research and Investigation, which is responsible for conducting independent
studies and providing guidance in several topics, including algorithmic transparency. In
addition, the US Public Policy Council of the Association for Computing Machinery
(USACM) published a statement proposing a set of principles for algorithmic
transparency and accountability, which are intended to minimise harm while at the same
time realising the benefits of algorithmic decision-making (Box 16).
Recent developments in Europe seem to indicate a similar movement to make algorithms
more transparent and accountable for law infringements. In a recent speech at the
Bundeskartellamt, the EU Commissioner Vestager (2017) stated that businesses have the
obligation of programming algorithms to deliberately comply with data protection and
antitrust laws, which can be denominated as "compliance by design". The German
Chancellor Angela Merkel made also a public statement calling for companies like
Facebook and Google to publicly disclose their proprietary algorithms:
The algorithms must be made public, so that one can inform oneself as an
interested citizen on questions like: what influences my behaviour on the
internet and that of others? (…) These algorithms, when they are not
transparent, can lead to a distortion of our perception, they narrow our breadth
of information. Agerholm (2016)
One way to empower the public with a watchdog function is to have regulators reverse-
engineering algorithms in order to understand how their decision-making process
functions. However, enforcing algorithmic transparency and accountability might turn out

48 │
to be a challenging task in practice, especially when facing black box algorithms that
make inherently autonomous decisions and might contain implicit or explicit biases. The
sensible reaction to demand for more transparency about how these algorithms work may
not achieve the intended purpose, as making these complex algorithms fully transparent
can be extremely challenging. Merely publishing (or disclosing to a regulator) the source
code of the algorithm may not be a sufficient transparency measure. Complete
transparency would require that someone could explain why any particular outcome was
produced, but that might be an impossible task when machine learning systems have
made autonomous decisions that have not been instructed by anyone.
Box 16. USACM’s principles for algorithmic transparency and accountability
“1. Awareness: Owners, designers, builders, users, and other stakeholders of analytic
systems should be aware of the possible biases involved in their design, implementation,
and use and the potential harm that biases can cause to individuals and society.
2. Access and redress: Regulators should encourage the adoption of mechanisms that
enable questioning and redress for individuals and groups that are adversely affected by
algorithmically informed decisions.
3. Accountability: Institutions should be held responsible for decisions made by the
algorithms that they use, even if it is not feasible to explain in detail how the algorithms
produce their results.
4. Explanation: Systems and institutions that use algorithmic decision-making are
encouraged to produce explanations regarding both the procedures followed by the
algorithm and the specific decisions that are made. This is particularly important in public
policy contexts.
5. Data Provenance: A description of the way in which the training data was collected
should be maintained by the builders of the algorithms, accompanied by an exploration
of the potential biases induced by the human or algorithmic data-gathering process.
Public scrutiny of the data provides maximum opportunity for corrections. However,
concerns over privacy, protecting trade secrets, or revelation of analytics that might allow
malicious actors to game the system can justify restricting access to qualified and
authorized individuals.
6. Auditability: Models, algorithms, data, and decisions should be recorded so that they
can be audited in cases where harm is suspected.
7. Validation and Testing: Institutions should use rigorous methods to validate their
models and document those methods and results. In particular, they should routinely
perform tests to assess and determine whether the model generates discriminatory
harm. Institutions are encouraged to make the results of such tests public.”
Source. Quoted from USACM (2017).
On top of this, it is still not clear which authority or regulator would be best placed to
review and supervise algorithms in order to ensure transparency and accountability. And
if multiple agencies should be involved in this role, questions remain on how co-
ordination and co-operation between regulators would work to reconcile possibly
conflicting objectives of different policies. Indeed, one of the main difficulties of
regulating the digital economy is precisely the fact that online companies operate at the
interface of existing laws, such as privacy law, transparency law, data protection,
intellectual property rights, consumer protection and competition law. This may require
regulatory interventions under different policy areas (Box 17), which are enforced by

│ 49
multiple different agencies (Strowel and Vergote, 2016). Moreover, many online
companies operate beyond national borders, posing thereby a territorial challenge in the
design of regulations.
6.2.3 Regulations to prevent algorithmic collusion
It is still unclear at this point whether any regulations can be created to prevent machine
learning algorithms from autonomously reaching tacit co-ordination, at least not without
harming the competitive process in other ways. To the best of the Secretariat’s
knowledge, no solutions have been proposed so far in the antitrust literature to tackle this
conduct. Moreover, there are no competition cases or investigations providing supportive
evidence of this “virtual” form of collusion, making it hard to justify the creation of
regulations to prevent the negative impact of conducts that have not been observed yet.
Box 17. Accountability and the “right to explanation” in the European GDPR
In April 2016, the European Parliament adopted a comprehensive set of rules for the
collection, storage and use of personal information, the General Data Protection Regulation
(GDPR). Among the different regulatory measures adopted, which deal in particular with data
protection, the GDPR also introduces the right of citizens to seek and receive an explanation
for decisions made by algorithms, especially if they are using profiling techniques (Art. 22). In
particular Article 13, 14 and 15 GDPR (which regulate the right of access in specific
circumstances) specify that individuals have a right to ask for “the existence of automated
decision-making, including profiling, […] and, at least in those cases, meaningful information
about the logic involved, as well as the significance and the envisaged consequences of such
processing for the data subject.”
This regulatory framework highlights the importance that the EU legislator places on the
ability of humans to interpret decisions made by algorithms and provides a strong indication
to algorithm designers that they should be in a position to provide meaningful information
about the logical process involved in the algorithm itself. This form of transparency and
accountability is associated with the right of individuals “to obtain human intervention” and to
express their points of view if they wish to contest the decision made by an algorithm. The
GDPR is an attempt to use regulation to ensure that humans can intervene in algorithmic
decision-making and to promote forms of algorithmic design that can ensure compliance with
the regulatory framework.
1
1. Goodman and Flaxman (2016).
Source. Regulation 2016/679 on the protection of natural persons with regard to the processing of
personal data and on the free movement of such data, and repealing Directive 95/46/EC (General Data
Protection Regulation), OJ L119/1 of 4.5.2016.
Nonetheless, recognising that collusion among machine learning algorithms might be
very hard to detect, and given the fast speed at which digital markets have evolved in the
last years, it is important to anticipate what kind of regulations could be considered in the
future if this peculiar form of collusion ever becomes part of the market reality. Three
potential types of regulatory intervention are briefly mentioned here, as well as the risks
they could pose for competition.
1. Price regulation: Given that algorithms may lead to anti-competitive prices even
in the absence of traditional conducts that could amount to “plus factors” – such
as communication or signalling – policy makers could be tempted to introduce
maximum price regulations. Maximum prices are considered to pose significant
barriers to competition and, whenever possible, should be replaced with more
efficient alternative policies.25
Indeed, price regulations not only reduce

50 │
incentives to innovate or to provide high-quality products, but could actually
result in higher prices by creating a focal point for collusion in digital markets
that would otherwise be competitive.
2. Policies to make tacit collusion unstable: Policy makers might come up with
policies to change the structural characteristics of digital markets that most
facilitate collusion.26
For instance, in order to make markets less transparent,
policy makers could enforce systems of secret discounts or impose restrictions on
the information that can be published online; likewise, in order to reduce the high
frequency of interaction in digital markets, they could enforce lags on price
adjustments or require companies to compromise to any new offer for a minimum
time period (Ezrachi and Stucke, 2017). Unfortunately, such policies would also
be likely to result in severe restrictions to competition, by reducing the amount of
information available to consumers27
and by preventing fast price adjustments
from efficiently matching demand and supply.
3. Rules on algorithm design: Thirdly, policy makers could eventually consider
the creation of rules that restrict the way algorithms are designed, i.e. an
algorithmic version of Asimov’s Three Laws of Robotics. If the purpose is to
prevent companies from independently coordinating anti-competitive prices,
regulations could inhibit algorithms from reacting on particular features or
market variables that are necessary to sustain tacit co-ordination.28
As an
example, algorithms could be programed not to react to most recent changes in
prices; or, instead, to ignore price variations of individual companies, while still
accounting for average prices in the industry. This solution might also constrain
the ability of firms to develop innovative algorithms, though it is probably less
restrictive for competition than the two previous forms of intervention. On the
other hand, regulating algorithm design could also pose on some agencies the
additional burden of supervising whether companies are effectively complying
with the rules.
The list of potential regulatory interventions discussed here is not intended to direct
policy in any particular direction, but merely to set a framework for discussion and
promote future debate. In fact, if any regulations must be designed to make markets less
prone to collusion, policy makers should adopt a conservative approach, since such rules
could have many other unpredictable implications that could ultimately compromise the
good functioning of digital markets.

│ 51
7. Conclusions
Recent developments of the digital economy are challenging the established approaches
on which law enforcers and regulators rely to protect competition, secure market trust and
promote social wellbeing. At the frontline of most discussions by current practitioners is
the concern that the generalised integration of computer algorithms in modern business
models might create risks for competition that should not be underestimated. Without
disregarding the significant benefits that automated systems have brought to society, this
paper attempted to frame the debate around the potential impact of algorithms on
collusion, identifying some preliminary possible responses that competition law enforcers
could pursue and possible areas for further research.
This paper recognises, in particular, two major mechanisms through which algorithms
could challenge antitrust investigators. Firstly, algorithms are fundamentally affecting
market conditions, resulting in high price transparency and high-frequency trading that
allows companies to react fast and aggressively. These changes in digital markets, if
taken to a certain extent, could make collusive strategies stable in virtually any market
structure. Secondly, by providing companies with powerful automated mechanisms to
monitor prices, implement common policies, send market signals or optimise joint profits
with deep learning techniques, algorithms might enable firms to achieve the same
outcomes of traditional hard core cartels through tacit collusion.
From an enforcement perspective, one has to distinguish between cases where algorithms
are used by competitors as an ancillary tool to a wider collusive arrangement falling
within the traditional reach of competition rules on anti-competitive agreements, and
cases where algorithms allow firms to align business conduct in what looks very much
like conscious parallelism, a conduct with is not illegal under competition rules. While in
the first case the challenges for agencies are related to detecting possible anti-competitive
cases, understanding the technology and collecting evidence to meet the required legal
standard, the second scenario raises more difficulties because the current legal standard
does not allow intervention with the traditional rules on anti-competitive agreements
between competitors.
In light of these difficulties, competition law enforcers can take alternative courses of
action. Using traditional antitrust tools, they could firstly conduct market studies in order
to evaluate whether algorithmic collusion is a phenomenon commonly observed and, if
so, under which conditions and in what industries it should be expected. Then, if a
competition problem is actually identified, some solutions might involve the adaptation of
merger review to account for the impact of algorithms on coordinated effects; or the
design of behavioural remedies to prevent companies from using algorithms for collusion.
Finally, as research progresses and delivers consistent evidence that algorithms are indeed
significantly contributing to price increases through tacitly collusive interdependence,
policy makers might need to consider if more substantial revisions are needed to address
algorithmic collusion, including revisiting the concept of “agreement” as well as the
current legal treatment of tacit collusion.
Since algorithms can result in multiple other market failures and affect substantially the
selection and organisation of the world’s information, some commentators have been

52 │
giving an increasing attention to the potential need for a regulatory reform in the digital
economy, whose governance has so far been left to the market. This paper mentions a few
regulatory approaches that might be considered in the future to tackle algorithmic
collusion, such as price regulation, policies to make tacit collusion unstable and rules on
algorithm design. However, at this stage, there are still concerns that any regulatory
interventions might have severe negative impacts on competition that could outweigh
their potential benefits. If regulatory solutions are to be considered, competition concerns
would only be an element of such discussion and considerations going beyond the risk of
collusion would have to be factored in such discussions.
Given the multi-dimensional nature of algorithms, policy approaches should be developed
in co-operation with competition law enforcers, consumer protection authorities, data
protection agencies, relevant sectorial regulators and organisations of computer science
with expertise in deep learning. In conclusion, despite the clear risks that algorithms may
pose on competition, this is still an area of high complexity and uncertainty, where lack of
intervention and over regulation could both pose serious costs on society, especially given
the potential benefits from algorithms. Whatever actions are taken in the future, they
should be subject to deep assessment and a cautious approach.

│ 53
Notes
1. The full text of the letter is available at https://futureoflife.org/ai-open-letter/.
2. For an analysis of the impact of data-driven innovation on growth and well-being, see
OECD (2015a).
3. Although it cannot be excluded that algorithms may have an impact on how competition
laws are enforced outside the area of collusion, this paper only deals with the potential
risk that algorithms and AI can have on co-ordination of business behaviour between
competitors.
4. It is not unusual for companies to run simultaneously deep learning and traditional
machine learning algorithms, so that they can identify the best course of actions and,
simultaneously, to be aware of the features that were relevant for the final decision.
5. See OECD (2013) and International Competition Network (2010) for more information
about reactive methods (generated by an event, such as a complaint or leniency
application) and proactive methods (initiated by the competition authority) to detect bid
rigging.
6. Screening algorithms can be structural, involving the analysis of markets characteristics
and identifying those where collusion is more likely; or behavioural, involving the
analysis of market conduct to assess whether companies’ actions are consistent with
collusion. While structural and behavioural approaches are both useful to detect collusion
and can actually complement each other, they face some common challenges, such as the
lack of sufficiently detailed and validated data to identify suspicious signs of collusion.
Relevant and accurate data is necessary for all stages of the screening exercise, from its
design, to its implementation, up to the interpretation of its results. Accessing this
information is a key issue in any empirical methodology and lack of it exposes screens to
a serious risk of failure. For more information about behavioural and structural screens,
see OECD (2013).
7. Common definitions of collusion can be found in OECD (1993), O’Sullivan and Sheffrin
(2003) and Green et al. (2013).
8. Firms can collude on a variety of competitive variables. In most cases, co-ordination
involves keeping prices above the competitive level, but in other occasions collusion may
aim at limiting production or capacity investments. Firms may also co-ordinate by
dividing the market (for instance, by geographic area or customer characteristics) or by
allocating contracts in bidding markets. Usually, traditional concepts of collusion excludes
agreements that are pro-competitive in nature, such as some vertical agreements or joint
ventures in R&D. See OECD (2012b) and OECD (2015b).
9. See Green et al. (2013), Harrington (2012), Ivaldi et al. (2003), Tirole (2002), Posner
(2001) and Turner (1962).
10. The role of transparency as a relevant factor for collusion is more extensively discussed in
Ezrachi and Stucke (2016), Whish and Bailey (2012), OECD (2012b), OECD (2010),
Ivaldi et al. (2003) and Stigler (1964).

54 │
11. This is the so-called “oligopoly problem”, which has generated a large debate on whether
competition policy should be concerned with tacit collusion or not (see also Section 5 in
this paper and OECD, 2015b).
12. Relevant factors for collusion are discussed, for instance, in Ivaldi et al. (2003),
Levenstein and Suslow (2006), OFT (2005), Canoy et al. (2004), Jacquemin and Slade
(1989) and Stigler (1964).
13. See Ezrachi and Stucke (2016) for a detailed description of alternative collusive scenarios,
where the roles of algorithms identified here might apply.
14. Such strategies are not only used in more complex scenarios of differentiated products and
heterogeneous businesses but also when firms are trying to reach co-ordination in more
homogeneous and straight forward market contexts
15. In Re High Fructose Corn Syrup Antitrust Litigation Appeal of A & W Bottling Inc. et al.,
United States Court of Appeals, Seventh Circuit, 295 F3d 651, 2002, p. 2.
16. For a more detailed discussion see Mehra (2015), Ezrachi and Stucke (2015 and 2017) and
Graef (2016).
17. Detection may be in practice very challenging as enforcers would not be in a position to
determine if a particular pricing would be the ‘natural’ outcome of market dynamics or if
it was ‘artificially’ enhanced or created by algorithms. Ezrachi and Stucke (2017) suggest
that detection could be enhanced through auditing of algorithms by agencies or regulators
to “assess whether an algorithm was designed to foster a change in the market dynamics.
In essence, such approach resembles ex-ante merger appraisal - focusing on whether a
proposed action would lead to a harmful change in market structure. Accordingly,
algorithms could be activated in a ‘sand box’ where their effects will be observed and
assessed.” The authors themselves recognise the challenges linked to auditing of
algorithm, first and foremost because of the sheer number of algorithms that would
require auditing and because the degree of technical expertise required to understand their
functioning and effects. Today agencies are not equipped for such an analysis, especially
if the algorithm to audit falls into the deep learning type of algorithm.
18. Mehra (2015) notes that the competition concerns related to the use of algorithm are not
just similar to the oligopoly problem, but they can aggravate it by giving individual firms
the incentive to raise the price above the competitive level in the absence of co-ordination,
which can also contribute to the stability of this action in the long run. He emphasises a
number of factors to support his thesis: algorithm ensure greater accuracy in detection of
price changes; the minimal level of human collaboration can remove the element of
irrationality and reduce the chance that the collusive scheme is undermined by mistake;
and due to mass data collection and processing, price cuts are become more detectable and
also lower the possibility of price wars.
19. Apart from increasing the internal stability of markets, which relates to the strategic
interaction between existing incumbents, algorithms may also foster external stability,
which refers to the entrance of potential entrants who are not present in the market yet.
For that purpose, colluding companies can use algorithms to instantaneously target new
competitors with exclusionary practices, eliminating thereby the threat of entry, which is
the main poison of collusion.
20. The debate on the most appropriate notion of “agreement” for antitrust purposes is not
recent. Indeed in the United States this debate that goes back to arguments in the sixties
by Posner and Turner. Posner (1968) favoured a broad interpretation of the notion of
agreement that would reach interdependent pricing by oligopolists, even where they did
not enter into an agreement in the common sense of the term. Turner (1962), on the other

│ 55
hand, agreed that the term “agreement” could not be limited only to an classic explicit
agreement. However, he disagreed with Posner’s approach and believed that punishing
businesspeople for interdependent oligopolistic pricing would be problematic, since, as in
a competitive industry, they were simply rationally optimising their prices given market
realities.
21. See Aircraft Cheque Services et al. v. Verizon Wireless et al., No. 14-2301 (7th Cir. 2015).
22. A strict liability standard (i.e. business will always be responsible for what their
algorithms do) may induce counterproductive (chilling) effects on incentives to develop
new and more performant algorithms which have the potential to play an important role in
promoting innovation and a greater role in modern society. On the other hand, a strict
liability standard would promote incentives of developers and users of algorithms to
understand how algorithms (including deep learning algorithms) work and build in
safeguards to prevent algorithms from reaching anticompetitive outcomes. Such an
approach seems to be implicitly suggested by Vestager (2017): “[…] businesses also need
to know that when they decide to use an automated system, they will be held responsible
for what it does. So they had better know how that system works.”
23. According to OECD (2016d), market studies and sector inquiries are useful tools to
understand the dynamic of the market and to promote competition. Market studies are
used primarily for the assessment of markets and their competitive conditions. They are
mainly considered as an advocacy tool to issue recommendations to change laws and
regulations or as a pre-enforcement tool in case they reveal constraints to competition of a
behavioural nature.
24. See for instance Miller (2016), The Economist (2016) and Teffer (2016).
25. For a more detailed discussion about the barriers to competition resulting from price
regulations, see paragraph B1 of the checklist of the Competition Assessment Toolkit
(OECD, 2016c).
26. As discussed in detail in Section 4.2, collusion in digital markets might be facilitated by
market transparency and high frequency of interaction.
27. For further details about restrictions that limit the choice and information available to
customers, see paragraph D of the checklist of the Competition Assessment Toolkit
(OECD, 2016c).
28. According to Vestager (2017): “What businesses can – and must – do is to ensure antitrust
compliance by design. That means pricing algorithms need to be built in a way that
doesn't allow them to collude. Like a more honourable version of the computer HAL in
the film 2001, they need to respond to an offer of collusion by saying “I’m sorry, I’m
afraid I can’t do that.”

56 │
References
Agerholm, H. (2016), “Angela Markel says internet search engines are ‘distorting perception’ and
algorithms should be revealed”, Independent, www.independent.co.uk/news/angela-merkel-says-
internet-search-engines-endangering-debate-algorithms-should-be-revealed-a7383811.html.
Agrawal, A. and D. Jaiswal (2012), "When Machine Learning Meets AI and Game Theory", CS 229
Machine Learning Final Project, http://cs229.stanford.edu/proj2012/AgrawalJaiswal-
WhenMachineLearningMeetsAIandGameTheory.pdf.
Akhgar, B., P. S. Bayerl and F. Sampson (2016), Open Source Intelligence Investigation: From Strategy
to Implementation, Springer International Publishing.
Anitha, P., G. Krithka and M. D. Choudhry (2014), International Journal of Advanced Research in
Computer Engineering & Technology, Vol. 3, No. 12, pp. 4324-4331, http://ijarcet.org/wp-
content/uploads/IJARCET-VOL-3-ISSUE-12-4324-4331.pdf.
Anthony, J., K. McGuiness, N. E O Conner, K. Moran (2016), “Quantifying Radiographic Knee
Osteoarthritis Severity using Deep Convolutional Neural Networks”, CS229 Final Report,
http://cs229.stanford.edu/proj2016/report/SureshaDalalMahajan-
Automatically%20Quantifying%20Radiographic%20Knee%20Osteoarthritis%20Severity-
report.pdf.
Autorité de la Concurrence and Bundeskartellamt (2016), Competition Law and Data,
www.bundeskartellamt.de/SharedDocs/Publikation/DE/Berichte/Big%20Data%20Papier.pdf?__bl
ob=publ.
Axelrod, R. (1984), The Evolution of Cooperation, Basic Books, Inc., Publishers, New York.
Bar-Ilan, J. (2007), “Google Bombing from a Time Perspective”, Journal of Computer-Mediated
Communication, Vol. 12, No. 3, pp. 910-938, http://onlinelibrary.wiley.com/doi/10.1111/j.1083-
6101.2007.00356.x/full.
Borenstein, S. (1999), “Rapid Price Communication and Coordination: The Airline Tariff Publishing
Case”, in J. E. Kwoka Jr. and L. J. White (Eds.), The Antitrust Revolution: Economics,
Competition and Policy, Oxford University Press, New York, pp. 310-328,
http://global.oup.com/us/companion.websites/fdscontent/uscompanion/us/pdf/kwoka/9780195322
972_09.pdf.
Brown, N and T. Sandholm (2017), “Safe and Nested Endgame Solving for Imperfect-Information
Games”, Proceedings of the AAAI Workshop on Computer Poker and Imperfect Information
Games, www.cs.cmu.edu/~sandholm/safeAndNested.aaa17WS.pdf.

│ 57
Brownlee, J. (2016), “8 Inspirational Applications of Deep Learning”, Machine Learning Mastery,
http://machinelearningmastery.com/inspirational-applications-deep-learning/.
Brush, S., T. Schoenberg and S. Ring (2015), “How a Mystery Trader with an Algorithm May Have
Caused the Flash Crash”, Bloomberg, www.bloomberg.com/news/articles/2015-04-22/mystery-
trader-armed-with-algorithms-rewrites-flash-crash-story.
Canoy, M., P. Rey and E. van Damme (2004), “Dominance and Monopolization”, in M. Neumann and J.
Weigand (Eds.), The International Handbook of Competition, Edward Elgar Publishing, UK.
CFTC and SEC (2010), “Findings Regarding the Market Events of May 6, 2010”, Report of the Staffs of
the CFTC and SEC to the Joint Advisory Committee on Emerging Regulatory Issues,
www.sec.gov/news/studies/2010/marketevents-report.pdf.
Chawla, S., J. D. Hartline and R. Kleinberg (2007), “Algorithmic Pricing via Virtual Valuations”,
http://users.eecs.northwestern.edu/~hartline/papers/bayesian-pricing-EC-07.pdf.
Chen, L., A. Mislove and C. Wilson (2016), “An Empirical Analysis of Algorithmic Pricing on Amazon
Marketplace”, in Proceedings of the 25th International Conference on World Wide Web, pp. 1339-
1349, www.ccs.neu.edu/home/amislove/publications/Amazon-WWW.pdf.
DoJ (2015a), “Former E-Commerce Executive Charged with Price Fixing in the Antitrust Division's First
Online Marketplace Prosecution”, Justice News of the US Department of Justice, Office of Public
Affairs, www.justice.gov/opa/pr/former-e-commerce-executive-charged-price-fixing-antitrust-
divisions-first-online-marketplace.
DoJ (2015b), “Futures Trader Charged with Illegally Manipulating Stock Market, Contributing to the
May 2010 Market ‘Flash Crash’”, Justice News, www.justice.gov/opa/pr/futures-trader-charged-
illegally-manipulating-stock-market-contributing-may-2010-market-flash
Domingos, P. (2016), “Pedro Domingos on Machine Learning and the Master Algorithm”, EconTalk
Episode hosted by Russ Roberts, www.econtalk.org/archives/2016/05/pedro_domingos.html.
Dong, Y.F., S. Kanhere, C. T. Chou and N. Bulusu (2008), “Automatic Collection of Fuel Prices from a
Network of Mobile Cameras”, In S. E. Nikoletseas, B. S. Chlebus, D. B. Johnson and B.
Krishnamachari (Eds.), Distributed Computing in Sensor Systems, Lecture Notes in Computer
Science, Vol. 5067, Springer, Berlin, Heidelberg, www.cse.unsw.edu.au/~ctchou/papers/dong-
dcoss08.pdf.
Er, H. and A. Hushmat (2016), “The Application of Technical Trading Rules Developed from Spot
Market Prices on Futures Market Prices Using CAPM”, Eurasian Business Review, pp. 1-41,
http://link.springer.com/article/10.1007/s40821-016-0056-2.
Ezrachi, A. (2015), “The Competitive Effects of Parity Clauses on Online Commerce”, Oxford Legal
Studies Research Paper No. 55/2015, https://ssrn.com/abstract=2672541.
Ezrachi, A. and M. E. Stucke (2016), "Virtual Competition: The Promise and Perils of the Algorithm-
Driven Economy", Harvard University Press, United States.

58 │
Ezrachi, A. and M. E. Stucke (2017), “Two Artificial Neural Networks Meet in an Online Hub and
Change the Future (of Competition, Market Dynamics and Society)”, SSRN paper,
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2949434
European Commission (2017), “Final Report on E-Commerce Sector Inquiry”, 10 May 2017,
COM(2017) 229 final, http://ec.europa.eu/competition/antitrust/sector_inquiry_final_report_en.pdf
Foote, K. D. (2017), “A Brief History of Deep Learning”, Dataversity Education,
www.dataversity.net/brief-history-deep-learning/.
Gal, Michal S. and N. Elkin-Koren (2017), “Algorithmic Consumers”, Harvard Journal of Law and
Technology, Vol. 30, https://ssrn.com/abstract=2876201.
Gawer, A. (2016), “Competition Policy and Regulatory Reforms for Big Data: Propositions to Harness
the Power of Big Data while Curbing Platforms’ Abuse of Dominance”, note submitted to the
Hearing on Big Data of the 126th meeting of the OECD Competition Committee,
https://one.oecd.org/document/DAF/COMP/WD(2016)74/en/pdf.
Goodfellow, I., Y. Bengio and A. Courville (2016), “Deep Learning”, MIT Press,
http://www.deeplearningbook.org/.
Goodman, B. and S. Flaxman (2016), “European Union regulations on algorithmic decision-making and
a ‘right to explanation’”, 2016 ICML Workshop on Human Interpretability in Machine Learning,
New York, https://www.arxiv.org/pdf/1606.08813v2.pdf.
Graef, I. (2016), “Algorithmic Price Fixing Under EU Competition Law: How to Crack Robot Cartels?”,
Centre for IT & IP Law, www.law.kuleuven.be/citip/blog/algorithmic-price-fixing-under-eu-
competition-law-how-to-crack-robot-cartels/.
Green, E. J., R. C. Marshall and L. M. Marx (2013), “Tacit Collusion in Oligopoly”, in R. D. Blair and
D. D. Sokol (Eds.), The Oxford Handbook of International Antitrust Economics, Volume 2,
Chapter 19, pp. 464-497, Oxford University Press,
https://faculty.fuqua.duke.edu/~marx/bio/papers/tacitcollusion.pdf.
Harrington, J. E., Jr. (2012), “A Theory of Tacit Collusion”, Economic Working Paper, The John
Hopkins University, http://www.tse-
fr.eu/sites/default/files/medias/stories/SEMIN_11_12/ECONOMIC_THEORY/harrington.pdf.
Harrington, J.E., Jr. and W. Zhao (2012), “Signaling and Tacit Collusion in an Infinitely Repeated
Prisoners’ Dilemma”, Mathematical Social Sciences, Vol. 64, pp. 277-289,
http://assets.wharton.upenn.edu/~harrij/pdf/mss12.pdf.
Haucap, J., U. Heimeshoff and M. Siekmann (2015), “Price Dispersion and Station Heterogeneity on
German Retail Gasoline Markets”, DICE Discussion Paper 171,
www.econstor.eu/handle/10419/106709.
Hay, G. (2013), “Anti-competitive Agreements: The Meaning of ‘Agreement’”, Cornell Law Faculty
Working Paper No. 105, http://scholarship.law.cornell.edu/clsops_papers/105.
Helft, M. and J. Wortham (2010), “Facebook Bows to Pressure Over Privacy”, New York Times, New
York, www.nytimes.com/2010/05/27/technology/27facebook.html.

│ 59
Hemsoth, N. (2016), “The Next Wave of Deep Learning Applications”, The Next Platform,
www.nextplatform.com/2016/09/14/next-wave-deep-learning-applications/.
Henderson, P., S. Crouch, Walters R.J. and Q. Ni (2003), “Comparison of Some Negotiation Algorithms
Using a Tournament-Based Approach”, in J. G. Carbonell, J. Siekmann, R. Kowalczyk, J. P.
Müller, H. Tianfield and R. Unland (Eds.), Agent Technologies, Infrastructures, Tools, and
Applications for E-Services, Lecture Notes in Computer Science, Vol. 2592, Springer, Berlin,
Heidelberg.
Hickman, L. (2013), “How Algorithms Rule the World”, The Guardian,
https://www.theguardian.com/science/2013/jul/01/how-algorithms-rule-world-nsa.
Hingston, P. and G. Kendall (2004), “Learning Versus Evolution in Iterated Prisoner's Dilemma”, in
Proceedings of the Congress on Evolutionary Computation (CEC'04),
http://www.cs.nott.ac.uk/~pszgxk/papers/cec2004ph.pdf.
Hollis, D. J., D. Edgington and D. Cline (2016), “Automated Detection of Deep-Sea Animals”, Digital
Commons, http://digitalcommons.calpoly.edu/star/370/.
IEC (2014), “Internet of Things: Wireless Sensor Networks”, White Paper of the International
Electrotechnical Commission, http://www.iec.ch/whitepaper/pdf/iecWP-internetofthings-LR-
en.pdf.
International Competition Network (2010), Anti-cartel Enforcement Manual, Chapter 4 - Cartel Case
Initiation, http://www.internationalcompetitionnetwork.org/uploads/library/doc628.pdf.
Ivaldi, M., B. Jullien, P. Rey, P. Seabright and J. Tirole (2003), “The Economics of Tacit Collusion”,
Final Report for DG Competition, European Commission,
http://ec.europa.eu/competition/mergers/studies_reports/the_economics_of_tacit_collusion_en.pdf.
Jacquemin, A. and M. E. Slade (1989), “Cartels, Collusion, and Horizontal Merger”, in R. Schmalensee
and R. Willig (Eds.), Handbook of Industrial Organization, Vol. 1, Chapter 7, pages 415-473,
Elsevier.
Jones, S., D. Johnstone and R. Wilson (2016), “Predicting Corporate Bankruptcy: An Evaluation of
Alternative Statistical Frameworks”, Journal of Business Finance & Accounting, Vol. 44, No. 1-2,
pp. 3-34, http://onlinelibrary.wiley.com/doi/10.1111/jbfa.12218/full.
Kaplow, L. (2011), “On the Meaning of Horizontal Agreements in Competition Law”, California Law
Review, Vol. 99, No. 3, pp. 683-818, https://www.jstor.org/stable/23014697?seq=1 -
page_scan_tab_contents.
Kirilenko, A., A. S. Kyle, M. Samadi and T. Tuzun (2017), “The Flash Crash: The Impact of High
Frequency Trading on the Electronic Market”, Journal of Finance, Forthcoming,
http://onlinelibrary.wiley.com/doi/10.1111/jofi.12498/abstract.
LeCun, Y., Y. Bengio and G. Hinton (2015), “Deep Learning”, Nature, Vol. 521, pp. 436-444,
https://www.cs.toronto.edu/~hinton/absps/NatureDeepReview.pdf.
Lee, J. B. and K. Taknomo (2016), “Comparison of several short-term traffic speed forecasting models”,
arXiv:1609.02409v1, https://arxiv.org/abs/1609.02409v1.

60 │
Levenstein, M. C. and V. Y. Suslow (2006), “What Determines Cartel Success?”, Journal of Economic
Literature, Vol. 44, No. 1, pp. 43-95,
https://www.jstor.org/stable/30032296?seq=1#page_scan_tab_contents.
Love, J. (2015), “Technology is Out of Your Control and Why You Should Celebrate", IT World
Canada, accessed in 15-07-2017, www.itworldcanada.com/article/technology-is-out-of-your-
control-and-why-you-should-celebrate/377563.
Mehra, S. K. (2015), “Antitrust and the Robo-Seller: Competition in the Time of Algorithms”,
Minnesota. Law Review, Vol. 100, www.minnesotalawreview.org/wp-
content/uploads/2016/04/Mehra_ONLINEPDF1.pdf.
Metz, C. (2017) “Artificial Intelligence Is About to Conquer Poker—But Not Without Human Help”,
WIRED, accessed in 24-01-2017, https://www.wired.com/2017/01/ai-conquer-poker-not-without-
human-help/.
Miller, M. J. (2016), “Techonomy: Should Internet Giants Be Regulated?”, PC Magazine,
www.pcmag.com/article/349736/techonomy-should-the-internet-giants-be-regulated.
Moschovakis, Y. N. (2001), “What is an Algorithm?”, in B. Engquist and W. Schmid (Eds.),
Mathematics Unlimited — 2001 and Beyond, Springer, pp. 919–936 (Part II),
www.cs.cmu.edu/~cdm/pdf/Moschovakis01.pdf.
Moujahid, A. (2016), “A Practical Introduction to Deep Learning with Caffe and Python”,
http://adilmoujahid.com/posts/2016/06/introduction-deep-learning-python-caffe/.
Nagesha, K. V., K. R. Chandar and V. R. Sastry (2016), “Prediction of Dust Dispersion by Drilling
Operation Using Artificial Neural Networks”, International Journal of Prevention and Control of
Industrial Pollution, Vol. 1, No. 2, pp.1-13,
http://chemical.journalspub.info/index.php/JPCIP/article/view/78.
O’Neal, C. (2016), “How Algorithms Rule our Working Live”, The Guardian,
https://www.theguardian.com/science/2016/sep/01/how-algorithms-rule-our-working-lives.
O’Shea, K., S. Reid, G. Condous and C. Lu (2016), “Deep neural networks for predicting pouch of
Douglas obliteration based on transvaginal ultrasound sliding sign videos”, Ultrasound in
Obstetrics & Gynecology, Vol. 48, No. S1, pp. 381,
http://onlinelibrary.wiley.com/doi/10.1002/uog.17163/full.
O’Sullivan, A. and S. M. Sheffrin (2003), Economics: Principles in Action, Upper Saddle River, New
Jersey.
Odabas, M. S., N. Senyer, G. Kayhan and E. Ergun (2016), “Estimation of Chlorophyll Concentration
Index at Leaves using Artificial Neural Networks”, Journal of Circuits, Systems and Computers,
Vol. 26, No. 2, https://doi.org/10.1142/S0218126617500268
OECD (1993), “Glossary of Industrial Organisation Economics and Competition”, compiled by R. S.
Khemani and D. M. Shapiro, commissioned by the OECD,
www.oecd.org/regreform/sectors/2376087.pdf.
OECD (2001), “Roundtable on Price Transparency”, www.oecd.org/competition/abuse/2535975.pdf.

│ 61
OECD (2007), “Roundtable on Facilitating Practices in Oligopolies”,
www.oecd.org/daf/competition/41472165.pdf .
OECD (2009), “Recommendation of the OECD Council on Competition Assessment”,
www.oecd.org/daf/competition/OECD-Recommendation-CompetitionAssessment.pdf .
OECD (2010), “Information Exchanges Between Competitors Under Competition Law”,
www.oecd.org/competition/cartels/48379006.pdf.
OECD (2011a), “Roundtable on Remedies in Merger Cases”,
www.oecd.org/daf/competition/RemediesinMergerCases2011.pdf.
OECD (2011b), “Roundtable on Information Exchanges Between Competitors under Competition Law”,
www.oecd.org/daf/competition/cartels/48379006.pdf.
OECD (2012a), “Roundtable on the Role of Efficiency Claims in Antitrust Proceedings”,
www.oecd.org/competition/cartels/48379006.pdf.
OECD (2012b), “Roundtable on Unilateral Disclosure of Information with Anticompetitive Effects (e.g.
through press announcements)”, www.oecd.org/competition/cartels/49548583.pdf.
OECD (2013). “Roundtable on Ex Officio Cartel Investigations and the Use of Screens to Detect
Cartels”, https://one.oecd.org/document/DAF/COMP(2013)14/en/pdf.
OECD (2015a), Data-Driven Innovation: Big Data for Growth and Well-Being, OECD Publishing, Paris,
http://dx.doi.org/10.1787/9789264229358-en.
OECD (2015b), “Roundtable on Competition Enforcement in Oligopolistic Markets”,
https://one.oecd.org/document/DAF/COMP(2015)2/en/pdf.
OECD (2016a), “Big Data: Bringing Competition Policy to the Digital Era”,
OECD (2016b), “Roundtable on Price Discrimination”,
OECD (2016c), Competition Assessment Toolkit, Volume I: Principles, OECD, Paris,
www.oecd.org/daf/competition/46193173.pdf.
OECD (2016d), “Roundtable on The Role of Market Studies as a Tool to Promote Competition”,
https://one.oecd.org/document/DAF/COMP/GF(2016)4/en/pdf.
OECD (2016e), “Competition and Innovation in Land Transport”,
https://one.oecd.org/document/DAF/COMP/WP2(2016)6/en/pdf.
OECD (2016f), “Protecting and Promoting Competition in Response to ‘Disruptive’ Innovations in Legal
Services”, https://one.oecd.org/document/DAF/COMP/WP2(2016)1/en/pdf.
OECD (2016g), “Refining Regulation to Enable Major Innovations in Financial Markets”,
https://one.oecd.org/document/DAF/COMP/WP2(2015)9/en/pdf.

62 │
OECD (2016h), “Promoting Effective Competition in Public Procurement”,
https://one.oecd.org/document/DAF/COMP/LACF(2016)3/en/pdf.
OECD (2016i), “Commitment Decisions in Antitrust Cases”,
OFT (2005), “Predicting Cartels”, Economic Discussion Paper, Office for Fair Trading, report written by
P. Grout and S. Sonderegger, Crown Copyright,
http://webarchive.nationalarchives.gov.uk/20140402142426/http:/www.oft.gov.uk/shared_oft/repo
rts/comp_policy/oft773.pdf.
Patterson, M.R. (2013), “Google and search-engine market power”, Harvard Journal of Law &
Technology Occasional Paper Series, http://jolt.law.harvard.edu/antitrust/articles/.
Posner, R. (1968), “Oligopoly and the Antitrust Laws: A Suggested Approach”, Stanford Law Review,
Vol. 21, pp. 1562-1606,
http://chicagounbound.uchicago.edu/cgi/viewcontent.cgi?article=2862&context=journal_articles.
Posner, R. A. (2001), Antitrust Law, Chicago: University of Chicago Press.
Potters, J. and S. Suetens (2013), “Oligopoly experiments in the current millennium”, Journal of
Economic Surveys, Vol. 27, No. 3, 439–460,
http://onlinelibrary.wiley.com/doi/10.1111/joes.12025/abstract.
Quinn, D.J. (2014), “Associated Press v. Meltwater: Are Courts Being Fair to News Aggregators?”,
Minnesota Journal of Law, Science and Technology, Vol. 15, No. 2, pp. 1189-1219,
http://scholarship.law.umn.edu/cgi/viewcontent.cgi?article=1039&context=mjlst.
Ralph, O. (2017), “Insurance and the Big Data Technology Revolution”, Financial Times,
www.ft.com/content/bb9f1ce8-f84b-11e6-bd4e-68d53499ed71.
Samuel, A. L. (1959), "Some Studies in Machine Learning Using the Game of Checkers", IBM Journal
of Research and Development,
http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.368.2254&rep=rep1&type=pdf
Sandvig, C., K. Hamilton, K. Karahalios and C. Langbort (2014), “An Algorithm Audit”, in S. P.
Gangadharan, V. Eubanks and S. Barocas (Eds.), Data and Discrimination: Collected Essays,
Open Technology Institute, New America, https://na-
production.s3.amazonaws.com/documents/data-and-discrimination.pdf.
Saurwein, F., N. Just and M. Latzer (2015), “Governance of Algorithms: Options and Limitations”, info,
Vol. 17, No. 6, pp. 35-49, https://ssrn.com/abstract=2710400.
Scherer, M. U. (2016), “Regulating Artificial Intelligence Systems: Risks, Challenges, Competencies,
and Strategies”, Harvard Journal of Law & Technology, Vol. 29, No. 2, pp. 353-400,
http://jolt.law.harvard.edu/articles/pdf/v29/29HarvJLTech353.pdf.
Schumpeter (2016), “Flexible Figures, A Growing Number of Companies are Using 'Dynamic' Pricing”,
The Economist, www.economist.com/news/business/21689541-growing-number-companies-are-
using-dynamic-pricing-flexible-figures.

│ 63
Selten, R. (1973), “A Simple Model of Imperfect Competition, Where 4 Are Few and 6 Are Many”.
International Journal of Game Theory, Vol. 2, No.1, pp. 141-201,
http://dx.doi.org/10.1007/BF01737566.
Spice, B. and G. Allen (2017), “CMU Artificial Intelligence Is Tough Poker Player”, CMU News,
https://www.cmu.edu/news/stories/archives/2017/january/AI-tough-poker-player.html.
Stafford, P. and J. Croft (2016), “Flash Crash Trader Navinder Singh Sarao Loses US Extradition Case”,
Financial Times, www.ft.com/content/ace643de-920c-11e6-8df8-d3778b55a923.
Stigler, G. J. (1964), “A Theory of Oligopoly”, Journal of Political Economy, Vol. 72, No. 1, pp. 44-61,
http://home.uchicago.edu/~vlima/courses/econ201/Stigler.pdf.
Strowel, A. and W. Vergote (2016), “Digital Platforms: To Regulate or Not To Regulate? Message to
Regulators: Fix the Economics First, Then Focus on the Right Regulation”,
http://ec.europa.eu/information_society/newsroom/image/document/2016-
7/uclouvain_et_universit_saint_louis_14044.pdf.
Sunstein, C. (2009), Republic.com 2.0, PUP, Princeton, NJ.
Suryanita, R., H. Jingga and E. Yuniarto (2016), “The Application of Artificial Neural Networks in
Predicting Structural Response of Multistory Building in The Region of Sumatra Island”, ICoSE
Conference on Instrumentation, Environment and Renewable Energy, Vol. 2016,
http://knepublishing.com/index.php/KnE-Engineering/article/view/526/1619.
Swarup, P. (2012), “Artificial Intelligence”, International Journal of Computing and Corporate Research,
Vol. 2, No. 4, http://www.ijccr.com/july2012/4.pdf.
Teffer, P. (2016), “EU Not Planning One-Size Rules for Internet Giants”, EUObserver,
https://euobserver.com/economic/133082.
The Economist (2016), “Regulating Technology Companies: Taming the Beasts”, The Economist
Newspaper Limited, www.economist.com/news/business/21699465-european-governments-are-
not-alone-wondering-how-deal-digital-giants-taming.
The Economist (2017), “When Life Throws You Lemons: A New York Startup Shakes Up the Insurance
Business”, www.economist.com/news/finance-and-economics/21718502-future-insurance-named-
after-soft-drink-new-york-startup-shakes-up.
Tirole, J. (2002), The Theory of Industrial Organization, Cambridge, MA: MIT Press.
Turner, D. F. (1962), “The Definition of Agreement under the Sherman Act: Conscious Parallelism and
Refusals to Deal”, Harvard Law Review, Vol. 75, No. 4, pp. 655-706,
www.jstor.org/stable/1338567?seq=1#page_scan_tab_contents.
Umamaheswari, B., P. Nithya and N. S. Chandran (2016), “Survey on Web Crime Detection Using Data
Mining Technique”, International Journal of Advanced Research in Computer Engineering &
Technology, Vol. 5, No. 1, pp. 177-184, http://ijarcet.org/wp-content/uploads/IJARCET-VOL-5-
ISSUE-1-177-184.pdf.

64 │
USACM (2017), “Statement on Algorithmic Transparency and Accountability”, Association for
Computing Machinery, US Public Policy Council, Washington, DC,
www.acm.org/binaries/content/assets/public-policy/2017_usacm_statement_algorithms.pdf.
Vestager, M. (2017), “Algorithms and Competition”, Speech at the Bundeskartellamt 18th Conference on
Competition, Berlin, https://ec.europa.eu/commission/commissioners/2014-2019/vestager/
announcements/bundeskartellamt-18th-conference-competition-berlin-16-march-2017_en.
Weiss, R. M. and A. K. Mehrotra (2001), “Online Dynamic Pricing: Efficiency, Equity and the Future of
E-commerce”, Virginia Journal of Law and Technology, Vol. 6, No. 11,
www.citi.columbia.edu/B8210/read10/Online%20Daynamic%20Pricing.pdf.
Whish, R., and D. Bailey (2012), Competition Law, Oxford University Press,
https://books.google.co.in/books?id=9QDhoLiRX-
MC&printsec=frontcover&source=gbs_ge_summary_r&cad=0#v=onepage&q&f=false.
White House (1997), “A Framework for Global Electronic Commerce”, White House, United States.
Willis Towers Watson (2017), “Insurers under Pressure to Go Digital – Willis Towers Watson Survey”,
Press Release, www.willistowerswatson.com/en/press/2017/02/insurers-under-pressure-to-go-
digital.
Wilson, R. A. and F. C. Keil (1999), The MIT Encyclopedia of the Cognitive Sciences, MIT Press.
Yilmazkuday, D. and H. A. Yilmazkuday (2016), “Understanding Gasoline Price Dispersion”, The
Annals of Regional Science, Vol. 57, No. 1, pp. 223-252,
http://link.springer.com/article/10.1007/s00168-016-0775-4.

│ 65
Annex 1. Mathematical derivation of collusion
Proposition
In a perfectly transparent market where firms interact repeatedly, if collusion does not
decrease the payoff of any player and the retaliation lag tends to zero, collusion can
always be sustained as an equilibrium strategy.
Proof
This formal demonstration applies standard techniques usually used in the literature to
analyse the relevant factors for collusion (see, for instance, Ivaldi et al. 2003). However,
in order to incorporate the fact that algorithms allow companies to adapt their strategies
very fast, the inter-temporal stream of profits are represented in continuous time, in
opposition to the economic literature where collusion is commonly model in discrete
time.
Collusion is a super game Nash equilibrium if, at any moment of time, the present
discount value of the profits of an arbitrary firm under the collusive path is greater than or
equal to the present discount value of the profits that the firm would earn by deviating
from collusion. The value of the collusive path is given by:
𝑉 𝑀
= ∫ 𝑒−𝑟𝑡
𝜋 𝑡
𝑀
. 𝑑𝑡
∞
0
, (1)
where 𝜋 𝑡
𝑀
is the profit of the firm at moment t under collusion and r is the discount rate.
The definite integral represents the sum of all discounted profits from moment 0 up to
infinite.
On the other hand, the value obtained by a firm that deviates from the collusive path is:
𝑉 𝐷
= ∫ 𝑒−𝑟𝑡
𝜋 𝑡
𝐷
. 𝑑𝑡
𝑇+𝐿
0
+ ∫ 𝑒−𝑟𝑡
𝜋 𝑡
𝐶
. 𝑑𝑡
∞
𝑇+𝐿
, (2)
where 𝜋 𝑡
𝐷
is the profit from deviating from collusion, 𝜋 𝑡
𝐶
is the payoff under competition
(in other words, under the punishment phase), T is the number of periods that it takes for
other firms to detect a deviation and L is the time lag required to react to the deviation
(for instance, by introducing a price change). This way, T is a measure of the
transparency in the market and L is a measure of the velocity of response or of the
frequency of interaction. In equation (2), the first term of the right-hand side corresponds
to the one shot gain from deviation, which takes place between periods 0 and 𝑇 + 𝐿,
while the second term is the value of the competitive path that takes place after period
𝑇 + 𝐿.
Then, collusive is an equilibrium strategy if the following incentive compatibility
constraint (ICC) is fulfilled:
𝑉 𝑀
≥ 𝑉 𝐷
. (3)

66 │
Replacing 𝑉 𝑀
and 𝑉 𝐷
by equations (1) and (2) respectively:
∫ 𝑒−𝑟𝑡
𝜋 𝑡
𝑀
. 𝑑𝑡
∞
0
≥ ∫ 𝑒−𝑟𝑡
𝜋 𝑡
𝐷
. 𝑑𝑡
𝑇+𝐿
0
+ ∫ 𝑒−𝑟𝑡
𝜋 𝑡
𝐶
. 𝑑𝑡
∞
𝑇+𝐿
. (4)
For simplification, the profit functions are considered to be constant over time. However,
the same results could be easily demonstrated by considering that profits grow at a
constant rate g, as long as 𝑔 < 𝑟.
∫ 𝑒−𝑟𝑡
𝜋 𝑀
. 𝑑𝑡
∞
0
≥ ∫ 𝑒−𝑟𝑡
𝜋 𝐷
. 𝑑𝑡
𝑇+𝐿
0
+ ∫ 𝑒−𝑟𝑡
𝜋 𝐶
. 𝑑𝑡
∞
𝑇+𝐿
. (5)
Bringing the profits to outside the integrals:
𝜋 𝑀
∫ 𝑒−𝑟𝑡
. 𝑑𝑡
∞
0
≥ 𝜋 𝐷
∫ 𝑒−𝑟𝑡
. 𝑑𝑡
𝑇+𝐿
0
+ 𝜋 𝐶
∫ 𝑒−𝑟𝑡
. 𝑑𝑡
∞
𝑇+𝐿
. (6)
Taking into consideration that the indefinite integral ∫ 𝑒−𝑟𝑡
. 𝑑𝑡 is equal to
−𝑒−𝑟𝑡
𝑟
, the
defined integrals above can be solved as follows:
𝜋 𝑀
[lim
t→∞
(
−𝑒−𝑟𝑡
𝑟
) − (
−𝑒−𝑟×0
𝑟
)] ≥
≥ 𝜋 𝐷
[(
−𝑒−𝑟(𝑇+𝐿)
𝑟
) − (
−𝑒−𝑟×0
𝑟
)]
+ 𝜋 𝐶
[lim
t→∞
(
−𝑒−𝑟𝑡
𝑟
) − (
𝑟
)]. (7)
Solving the limits and simplifying some of the terms:
𝜋 𝑀
[0 − (−
1
𝑟
)]
≥ 𝜋 𝐷
[(
𝑟
) − (−
1
𝑟
)]
+ 𝜋 𝐶
[0 − (
𝑟
)]. (8)
Finally, using standard algebra, equation (8) can be simplified to the following equation:
𝜋 𝑀
≥ 𝜋 𝐷
− (𝜋 𝐷
− 𝜋 𝐶
). 𝑒−𝑟(𝑇+𝐿)
. (9)
As a preliminary result, it is possible to conclude from equation (9) that transparency and
velocity of interaction facilitate collusion. Indeed, when T and L fall, the right-hand side
of equation (9) also falls (notice that the profit under deviation is always greater than the
profit under competition), relaxing the incentive compatibility constrain and thus making
collusion more likely.
In order to see what happens when markets are perfectly transparent and firms are able to
adjust their strategies instantaneously, the limit of the right-hand side of equation (9) is
taken when T and L tend to zero:
𝜋 𝑀
≥ lim
𝑇→0
𝐿→0
(𝜋 𝐷
− (𝜋 𝐷
− 𝜋 𝐶
). 𝑒−𝑟(𝑇+𝐿)
). (10)

│ 67
Solving the limit:
𝜋 𝑀
≥ 𝜋 𝐷
− (𝜋 𝐷
− 𝜋 𝐶
) × 1. (11)
Finally, it follows directly that:
𝜋 𝑀
≥ 𝜋 𝐶
. (12)
Equation (12) completes the proof: when markets are perfectly transparent, so that firms
can immediately identify any deviation from collusion (𝑇 = 0), and the time lag of
response L is zero, collusion can always be sustained as an equilibrium strategy as long as
the profit of any firm under collusion is greater than under competition.

oe.cd/1-0
www.oecd.org/competition
The combination of big data with technologically advanced tools,
such as pricing algorithms, is changing the competitive landscape
in many markets and sectors. While this is producing benefits and
efficiencies, it is also raising concerns of possible anti-competitive
behaviour. This paper looks at whether algorithms can make tacit
collusion easier and discusses some of the challenges they present
for both competition law enforcement and market regulation.
This report contributes to the OECD Going Digital project which
provides policy makers with tools to help economies and
societies prosper in an increasingly digital and data-driven world.
For more information, visit www.oecd.org/going-digital.

Algorithms and-colllusion-competition-policy-in-the-digital-age

More Related Content

What's hot (15)

Similar to Algorithms and-colllusion-competition-policy-in-the-digital-age (20)

More from Giuliano Tavaroli (20)

Recently uploaded (20)

Algorithms and-colllusion-competition-policy-in-the-digital-age