This presentation by the OECD Competition Division was made during the discussion “Algorithms and collusion” held at the 127th meeting of the OECD Competition Committee on 23 June 2017. More papers and presentations on the topic can be found out at oe.cd/1-0.

1. Antonio Capobianco (antonio.capobianco@oecd.org)
Pedro Gonzaga (pedro.gonzaga@oecd.org)
Anita Nyesö (anita.nyeso@oecd.org)
The opinions expressed and arguments employed herein are those of the authors and do not
necessarily reflect the official views of the OECD or OECD member countries.

2. 2

3. • Roundtable on Collusion and Algorithms
– Discussion by a panel of experts:
• Ariel Ezrachi
• Michal Gal
• Avigdor Gal
– Background note by
the OECD Secretariat
3
www.oecd.org/competition/algorithms-and-collusion.htm
Short URL: oe.cd/1-0

4. 1. Algorithms: concepts and applications
2. Risk of algorithmic collusion
3. Challenges for competition law enforcement
4. Regulation of algorithms

5. 5

6. 6
“An algorithm is an unambiguous, precise, list of simple operations applied
mechanically and systematically to a set of tokens or objects. (…) The initial
state of the tokens is the input; the final state is the output.”
Wilson and Keil (1999)
Instructions
Manual
• Start by doing this
• Then of course do that.
• Yeah, that sounds good.

7. • Plain language
• Diagrams
• Voice instructions
• Computer codes
– Automatic
– Fast processing
– Complex calculation
7
Digitalisation
Adoption of
computer
algorithms
Increased
productivity

8. • Artificial intelligence
– Detailed algorithms that mimic human intelligence
• Machine learning
– Algorithms that iteratively learn from data
• Deep learning
– Artificial neural networks that replicate the activity of human
neurons…
8

9. 9
ML requires manual features engineering, while in DL
feature engineering is automatic…
Input
Feature
Extractor
Features
Traditional
ML
Algorithm
Output
Input Deep Learning Algorithm Output

10. 10
Business Consumers Government
Predictive analytics
Process
optimisation
Increase
supplier power
Consumer
information
Decision-making
optimisation
Increase
buyer power
Crime detection
Determine fines
and sentences
Positive impact on static and
dynamic efficiency

11. 11
Predictive analytics
Optimisation of
business processes
Supply-chain
optimisation
Target ads
Recommendations Product
Customisation
Dynamic pricing
Price differentiation
Fraud prevention
Risk
management
Product innovation

12. 12
• Melanoma cancer detection
• Detect cancerous cells in brain
surgery
• Dispersion of dust in drilling
• Response of buildings to earthquakes
• Prediction of traffic conditions
• Buy and sell stocks
• Predict corporate bankruptcy
• Detect and classify deep-sea animals
• Estimate concentrations of chlorophyll
• Colorize black and white images
• Add sound to silent movies

13. 13

14. 14
Algorithmic collusion consists in any form of anti-
competitive agreement or coordination among
competing firms that is facilitated or
implemented through means
of automated systems.

15. 15
Relevant factors for collusion
How algorithms affect
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 –
+ positive impact; – negative impact; 0 neutral impact; ± ambiguous impact.

16. 16
Result: 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.
Intuition: If markets are transparent and companies react
instantaneously to any deviation, the payoff from deviation is
zero.
ICC: 𝑒−𝑟𝑡 𝜋 𝑀. 𝑑𝑡
∞
0
≥ 𝑒−𝑟𝑡 𝜋 𝐷. 𝑑𝑡
𝑇+𝐿
0
+ 𝑒−𝑟𝑡 𝜋 𝐶. 𝑑𝑡
∞
𝑇+𝐿

17. • However, even if collusion is
an equilibrium strategy, firms
may fail to coordinate…
• Cartel umbrella effect:
– In industries with many players, each firm has an
incentive not to participate in the agreement in the
first place in order to benefit from the “cartel
umbrella”.
17
Structural measures might still be effective…

18. • Any collusive arrangement requires:
– A meeting of minds
– A structure to implement and govern firms’
interaction:
• Common policy
• Monitoring
• Punishment mechanism
18

19. • Some algorithms may eliminate the need
of explicit communication during the
initiation and implementation stages:
– Monitoring algorithms
– Parallel algorithms
– Signalling algorithms
– Self-learning algorithms
19

20. 20
C: Monitoring algorithm
Description : Collect and process
information from competitors to
monitor their compliance and,
eventually, to punish deviations.
Legend :
/𝑝 <collusive price>
/𝑝𝑖 <price set by firm i>

21. 21
C: Parallel algorithm
Description : Coordinate a common policy or parallel behaviour,
for instance by programming prices to follow a leader
Legend :
/𝑝 <collusive price>
/𝑝𝑖 <price set by firm i>

22. 22
C: Signalling algorithm
Description : to disclose and
disseminate information in order to
announce an intention to collude or
negotiate a common policy
Legend :
/𝑠 <tentative signal>
/𝑠𝑖 <signal sent by firm i>

23. 23
C: Self-learning algorithm
Description : maximise profits
while recognising mutual
interdependency and readapting
behaviour to the actions of other
market players
…

24. 24
Algorithms
Change market
characteristics
Govern collusive structures
Transparency
High frequency
trading
Signal & negotiate
common policy
Coordinate
common policy
Monitor & punish
Optimise joint
profits
Increase likelihood of
collusion
Replace explicit
communication
Tacit
collusion

25. 25

26. • IF algorithms are a tool used to implement or
amplify an illegal conduct:
– Traditional antitrust tools apply
• IF algorithms create new anti-competitive
behaviours not covered by antitrust rules:
– Can existing antitrust tools be adapted?
– Should we reconsider fundamental
antitrust concepts?
26

27. Market studies &
investigations
• Obtain empirical
evidence of
algorithmic collusion
• Identify problematic
markets and sectors
• Define appropriate
measures
Ex-ante merger
control
• Reconsider the
threshold of
intervention
• Evaluate the impact
of transactions on
market transparency
and high frequency
trading
• Account for multi-
market contacts in
conglomerate
mergers
Commitments &
remedies
• Design remedies to
prevent the use of
algorithms as
facilitating practices
• Apply “notice-and-
take-down”
processes
• Introduce auditing
mechanisms for
algorithms?
27

28. Algorithms
Tacit
collusion
LiabilityAgreement
28

29. • Competition rules do not forbid collusive outcomes,
but only the means to achieve collusion
• Establishing an infringement of competition law
requires:
– Evidence of parallel conduct AND
– “Plus factors” (communication, information
exchanges, signalling…)
29
01100011 01101111 01101100 01101100 01110101
01110011 01101001 01101111 01101110 00100001

30. Should jurisdictions review their approach towards tacit collusion, by
adjusting their criteria for an infringement?
30
% Example of law enforcement algorithm
if {price correlation > 0.9999
and companies use:
dynamic pricing algorithm
or scraping algorithm
or third party data centre
or machine learning}
conduct = infringement
else
conduct = no infringement
end
evidence of
parallel conduct
possible
“plus factors”

31. • Identifying an “agreement” is a prerequisite to enforce
the law against collusion
• The concept of “agreement” is usually broadly defined, in
order to ensure a wide reach of competition rules
– In the EU the term “agreement” involves simultaneously:
• A common will
• Some form of manifestation
– In the US is involves solely a
common will or a “meeting of minds”
31

32. • In practice, existing concepts provide little guidance and
courts rely on explicit communication to prove an
agreement…
• Should legislators create a more clear definition of
agreement, in order to:
– Reduce legal uncertainty
– Account for more subtle behaviours, such as “meetings of
algorithms”?
32

33. (…) 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)
33
time
price

34. Offer Acceptance
1 Firm intermittently sets a higher price for
brief seconds (costless signal)
Competitor increases price to the value
signalled
2 Firm programs algorithm to mimic the price
of a leader
The leader, recognising this behaviour,
increases the price
3 Firm publicly releases a pricing algorithm Competitor downloads and executes the
same pricing algorithm
4 Firm programs an anti-competitive price to
be triggered whenever the competitor’s
price is below a threshold
Recognising the algorithm, the
competitor always keeps the price above
the threshold
5 Firm uses ML algorithm to maximise joint
profits (for instance, by accounting for the
spillovers on competitors’ profits)
Competitor reacts with the same strategy
34
Can a “meeting of algorithms” amount to an anti-competitive agreement?

35. • Weak link between the agent (algorithm) and the
principal (human being)
• Defining a benchmark for illegality requires assessing
whether any illegal action could have been anticipated or
predetermined
35
The challenges that automated systems
create are very real. (…) So as competition
enforcers, we need to make it very clear that
companies can’t escape responsibility for
collusion by hiding behind a computer
program.
Margrethe Vestager (2017)

36. • Who is liable for the decisions and actions of
algorithms?
– Creators
• Programmers
• Third party centres
– Users
• Managers
• Commercials
– Benefiters
• Shareholders
• Managers
36
Creators
Users
Benefiters

37. 37

38. 38
• The use of automated computer systems to organise
and select relevant information affects fundamental
structures of the society…
“(…) 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.”
Domingos (2017)

39. 39
• “Echo chambers”
• Product recommendations
• Content-control software
• Feedback scores
• Rankings of search engines’ results
• Target ads
• Price discrimination
• Collection of data protected by IP rights

40. • Can the risks of algorithmic selection be eroded through
“algorithmic competition”?
40
Imperfect Information
Lack of algorithmic
transparency
Algorithms as trade
secrets
Complexity of program
codes
Barriers to entry
Scale economies of IT
infrastructures
Scope economies of
datasets
Network economies in
online platforms
Spill-overs
Nature of knowledge as a
public good
Spill-overs of a variety of
information

41. 41
Algorithmic
risks
Market
failures
Regulatory
intervention?
Competitive
Impact
Can market regulation prevent algorithmic collusion?

42. 42
Market
solutions
Self-
organisation
Self-
regulation
Co-
regulation
State
intervention
WHO?
Online companies operate at the interface of many laws
enforced by different agencies:
Privacy law Data protection
Competition law Consumer protection
Transparency law
IPR

43. • New FTC Office of Technology Research and Investigation
responsible for studying algorithmic transparency
• EU Commissioner Vestager’s statement advocating for
compliance by design with data protection and antitrust
laws
• German Chancellor Merkel’s public statement:
43
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.

44. Principles
Awareness
Access &
redress
Account-
ability
Explanation
Data
provenance
Auditability
Validation
& testing
44

45. • Public disclosure of algorithms may reduce incentives for
investment and innovation
• Disclosing a complex program code may not suffice as a
transparency measure
• Transparency and accountability are challenging when
decisions are taken autonomously by the algorithm
• Enforcement cost of reviewing and supervising algorithms
45
Risk that algorithmic
transparency
facilitates further
algorithmic collusion

46. • Extreme forms of algorithmic collusion enabled by deep
learning may be hard to prevent through competition law
• As a result, some regulatory interventions to prevent
algorithmic collusion might be considered in the future:
46
Price
regulation
Market
design
Algorithm
design
Risk of competitive
impact

47. 47
• Regulation objective:
– To restrain the ability of firms to set high prices, regardless of
whether they are achieved through explicit collusion, tacit
coordination or other means
• Barriers to competition:
– Reduces incentives to innovate or to supply high-quality products
– Creates a focal point for collusion
– Creates barriers to new market entrants

48. 48
• Regulation objective:
– To make digital markets less prone to collusion, by reducing
transparency and frequency of interaction
• Restrictions on information disclosures
• Systems of secret discounts
• Enforcement of time lags to implement price adjustments
• Barriers to competition:
– Limits the information available to consumers
– Restricts the ability of firms to adjust strategies fast and efficiently

49. 49
• Regulation objective:
– To enforce programmers to comply with competition principles in
the design of algorithms:
• Restrictions on the features / market information that the algorithm
can use (e.g. recent price changes)
• Restrictions on the objective function (e.g. joint profit maximisation)
• Barriers to competition:
– Limits the ability of firms to adjust strategies efficiently
– Raises entry costs by forcing firms to comply by design