Algorithmic collusion: From smokeroom to code lines? How collusion can evolve & are we prepared

tags
Competition Algorithmic Pricing

Notes

1. Introduction: Setting the Scene

NOTER_PAGE: (2 0.1003717472118959 . 0.2958860759493671)

growing resort to dynamic pricing by businesses and sellers.

NOTER_PAGE: (2 0.3250132766861391 . 0.4414556962025316)

Dynamic pricing, which is a type of price discrimination, refers to a set of pricing strategies implemented to increase profits and is facilitated by algorithms.

NOTER_PAGE: (2 0.5002655337227827 . 0.7824367088607594)

an increasing number of businesses rely on pricing algorithms, but use ones that focus on react- ing to other competitors’ pricing. Platforms such as Amazon’s facilitated this expansion.

NOTER_PAGE: (2 0.7636749867233138 . 0.23417721518987342)

60% of the third-party sellers on Amazon that use algorithmic pricing charged more than other sellers.

NOTER_PAGE: (2 0.8645778013807753 . 0.3662974683544304)

pricing algorithms can independently learn to not compete,

NOTER_PAGE: (4 0.11949017525225702 . 0.4699367088607595)

the most appropriate solutions may come from regu- lating the design of algorithms themselves.

NOTER_PAGE: (4 0.2851832182687201 . 0.3401898734177215)

more opaque forms of collusion, i.e. the sheltering of unlawful behaviour and liability from consequences, behind an algorithm.

NOTER_PAGE: (4 0.46574614976101963 . 0.15585443037974683)

2. Artificial Intelligence: the What and the Why

NOTER_PAGE: (4 0.6771109930961231 . 0.21518987341772153)

2.1 The Nomenclature: Defining the Terms

NOTER_PAGE: (4 0.7833244822092406 . 0.25079113924050633)

2.2 Algorithmic Business

NOTER_PAGE: (5 0.6303770578863516 . 0.3536392405063291)

2.2.1 Predictive Analytics & Business Processes Optimization

NOTER_PAGE: (5 0.8390865639936271 . 0.15031645569620253)

2.2.2 Algorithms Catered Towards Consumers & Algorithms Catered Towards Businesses

NOTER_PAGE: (6 0.27403080191184276 . 0.15427215189873417)

2.2.3 The Appeal of Algorithmic Business

NOTER_PAGE: (6 0.697822623473181 . 0.2602848101265823)

On the supply side, algorithms allow to gain efficiencies,

NOTER_PAGE: (6 0.9054699946893254 . 0.15268987341772153)

dynamic pricing algorithms have been recognized to improve market efficiency,

NOTER_PAGE: (7 0.11949017525225702 . 0.36075949367088606)

the ease with which consumers can compare offerings incentivizes businesses to innovate and promote competition.

NOTER_PAGE: (7 0.4020180562931492 . 0.27927215189873417)

surely we don't still believe this so late in the game

3. Too Good to Be True? A Looming Threat that Should Be Heeded

NOTER_PAGE: (7 0.48327137546468396 . 0.14794303797468356)

3.1 Debates on the Anti-Competitive Effects of Pricing Algorithms

NOTER_PAGE: (7 0.6500265533722782 . 0.17325949367088608)

use of algorithms will favour some form of collusion and make it harder to detect by regulators.

NOTER_PAGE: (7 0.7089750398300584 . 0.5340189873417721)

algorithms may learn to cooperate autonomously—and thus to independently achieve collusive outcomes

NOTER_PAGE: (7 0.8066914498141263 . 0.31170886075949367)

algorithmic pricing tends to reach supra-competitive prices,

NOTER_PAGE: (8 0.24004248539564524 . 0.569620253164557)

Algorithms, however, allow for better coor- dination and monitoring irrespectively of the number of players.

NOTER_PAGE: (8 0.43919277748274027 . 0.45332278481012656)

AI tools also reinforce market transparency and the frequency of interactions, which are elements that favour collusive behaviour.

NOTER_PAGE: (8 0.49229952203929894 . 0.22151898734177214)

3.2 Recent Robust Experimental Studies

NOTER_PAGE: (9 0.3138608603292618 . 0.2626582278481013)

when a competing algorithm lowers its prices—that is, when it would deviate from the rest—, it is met with punishment from the other algorithms until it realigns itself.

NOTER_PAGE: (9 0.7849176845459372 . 0.5783227848101266)

collusive outcomes are common and not dependent on a select few points.

NOTER_PAGE: (10 0.39564524694636216 . 0.6020569620253164)

In their experiment setting, the algorithms do not observe the prices of their competitors. Just like Calvano et al, they also find that in such a setting, collusion can materialize just as well.

NOTER_PAGE: (10 0.514073287307488 . 0.21281645569620253)

high informational value. When this is present, independent algorithms inadvertently correlate their prices,

NOTER_PAGE: (10 0.575677110993096 . 0.3939873417721519)

3.3 Empirical Evidence from the Case of the German Retail Gasoline Market

NOTER_PAGE: (10 0.7679235262878384 . 0.15268987341772153)

The near-perfect transparency of prices in the market studied, in part due to price disclosure regulations, make deviations from collusive behaviour easily detectable and punishable—thus, favouring an environment that will sustain the supra-competitive prices obtained.

NOTER_PAGE: (11 0.7296866702071162 . 0.1574367088607595)

the road to hell is paved with good intentions. the road to heaven is dirt

mean margins at the station level increased by 0.7 euro per litre, after the adoption of pricing algorithms, which represents a 9% increase.

NOTER_PAGE: (11 0.8672331386086032 . 0.22151898734177214)

€500 million per year increase in consumer expenditures for gasoline,

NOTER_PAGE: (11 0.9298990971853425 . 0.40348101265822783)

consistent with estimates of other researchers on the effects of coordination in the retail gasoline market.

NOTER_PAGE: (12 0.21986192246415293 . 0.5221518987341772)

margins only start increasing a year into the adoption of pricing algorithms in the market, which indicates that the algorithms are effectively learning tacit collusive strategies.

NOTER_PAGE: (12 0.3536909187466808 . 0.310126582278481)

4. A Canadian Legal Landscape … In Need of Landscaping?

NOTER_PAGE: (12 0.859798194370685 . 0.14477848101265822)

NOTER_PAGE: (13 0.2596919808815719 . 0.27294303797468356)

4.1.1 The Criminal Avenue: Sections 45 of the Competition Act

NOTER_PAGE: (13 0.2894317578332448 . 0.14082278481012658)

Under Section 45, there is no need to demonstrate that the object of the conspiracy was carried out or that the agreement had any effect on competition. Importantly, however, there must be an actual agreement for there to be a conspiracy, which requires “genuine intention”

NOTER_PAGE: (13 0.5687732342007434 . 0.15901898734177214)

conscious parallelism does not meet the threshold

NOTER_PAGE: (13 0.715878916622411 . 0.5498417721518988)

4.1.2 The Civil Avenue: Section 90.1 of the Competition Act

NOTER_PAGE: (14 0.370154009559214 . 0.15348101265822783)

Similar to Section 45, Section 90.1 requires a consensus between the parties and does not apply to conscientious parallelism. 126 However, this new test is “effects-based.” In that sense, an otherwise reviewable collabora- tion will be saved under Subsection 90.1(4) if it creates efficiencies.

NOTER_PAGE: (14 0.6664896441848114 . 0.15427215189873417)

4.1.3 The Tort Avenue: The Tort of Civil Conspiracy

NOTER_PAGE: (15 0.14976101964949548 . 0.20569620253164556)

4.2 Putting the Law to the Test

NOTER_PAGE: (15 0.44609665427509293 . 0.31645569620253167)

4.2.1 Competition Law & Pricing Algorithms as Tools to Execute Pre-Existing Explicit Collusion

NOTER_PAGE: (15 0.5592140201805629 . 0.17484177215189872)

With pricing data being more readily available, the use of pricing algo- rithms can facilitate explicit coordination.

NOTER_PAGE: (15 0.6133828996282528 . 0.16218354430379747)

agreement was executed using pricing algorithms that were specifically programmed to implement this cartel.

NOTER_PAGE: (15 0.8481147105682421 . 0.15822784810126583)

these scenarios all rely on AI to put into effect a previously agreed upon collu- sive pact. There, AI merely becomes a vector of their collusive intent.

NOTER_PAGE: (16 0.143919277748274 . 0.807753164556962)

4.2.2 Competition Law & Algorithmic Hub-and-Spoke Scenarios

NOTER_PAGE: (16 0.33244822092405735 . 0.13528481012658228)

businesses who indepen- dently acquire a pricing algorithm, knowing that their competitors use it too, may be subjected to a collusion inference

NOTER_PAGE: (16 0.5719596388741369 . 0.5632911392405063)

alleged that Uber facilitated a hub-and-spoke-structured col- lusion, whereby Uber conspired with the drivers—whom Uber has always argued are independent contractors—to use the company’s pricing algorithm to set the prices charged to the users.

NOTER_PAGE: (16 0.6887944768985661 . 0.2634493670886076)

there would be a “meeting of the minds” especially if “spoke A” agrees to implement a certain price policy if “spoke B” does it too.

NOTER_PAGE: (17 0.2580987785448752 . 0.12737341772151897)

CJEU stresses the importance of the spokes’ knowledge of the concerted practice.

NOTER_PAGE: (17 0.33191715347849177 . 0.6938291139240507)

Drawing upon the doctrine of willful blindness, it could be said that, in these circumstances, an intent to enter into an agreement can be inferred from the adoption of these pricing algorithms.

NOTER_PAGE: (17 0.4795539033457249 . 0.19145569620253164)

self-imposed ignorance should not be rewarded.

NOTER_PAGE: (17 0.6054168879447689 . 0.18829113924050633)

4.2.3 Competition Law & Pricing Algorithms Autonomously Colluding

NOTER_PAGE: (17 0.8348380244291025 . 0.15901898734177214)

could be argued that businesses that implement these pricing algorithms as part of their processes, knowingly assume the anti- competitive effects,

NOTER_PAGE: (18 0.2862453531598513 . 0.32832278481012656)

these same points that raise concerns in the literature and to regulators are the same points that draw these businesses to adopt the AI tools.

NOTER_PAGE: (18 0.3494423791821561 . 0.16693037974683544)

still presupposes that underneath, there is an agreement.

NOTER_PAGE: (18 0.5369091874668083 . 0.31566455696202533)

recognize that collaboration can be—and is—good for competition.

NOTER_PAGE: (18 0.7148167817312798 . 0.4509493670886076)

5. Conclusion: The Path Forward

NOTER_PAGE: (19 0.1009028146574615 . 0.30696202531645567)

point rather into the direction of regulating algorithms themselves. Maybe certain properties of pricing algorithms should be prohibited;164 changing the design to reward algorithms that cut process for instances,165 or making software designers liable for their algorithms’ designs are all paths forward

NOTER_PAGE: (19 0.6951672862453531 . 0.819620253164557)

Digisprudence