Online retailers are using increasingly sophisticated technology to adapt their prices in response to changes in competitors’ prices and consumer demand in real time. One of these technologies is pricing algorithms which allow for more fast-paced, frequent price changes.

Pricing algorithms enable businesses to instantly determine and set prices at a level that will maximise profits beyond that which is obtainable through the usual interaction of supply and demand.

In 2017, we asked the question ‘can robots collude?’ We considered how the increased capacity to process large amounts of data to execute price changes raised concerns about compliance and enforcement of competition laws relating to price fixing arrangements and extreme forms of price discrimination between buyers.

Since then, the risk of pricing algorithms has increased. Algorithmic pricing systems have become much more common in digital marketplaces, with research indicating that more than one-third of all sellers on Amazon Marketplace using pricing algorithms. Yet many of the questions about prevention, detection and liability for algorithmic collusion have gone unanswered – including most fundamentally, when and if collusion is involved at all.

A recent study by Rocher, Tournier and de Montjoye (2023) suggests that adversarial strategies, where one retailer manipulates their pricing strategically to impact its competitors who also use their own pricing algorithm, can lead to increased prices and increased profits for both firms in a market. This study found that “increasing use of sophisticated algorithms to set prices on online markets creates incentives for competitors to decode strategies, detect weaknesses and exploit them”.

As Rocher, Tournier and de Montjoye (behind a pay wall, but well worth buying the full article) explain, an attacking algorithm progressively learns the pricing strategy of its competitors over time and then uses this knowledge to maximise its own profits by anticipating the reactions of its competitors and applying the best response to maximise not only their immediate profit but also future profits by selling products on the market at optimised prices.

This kind of machine-learning algorithmic interaction may fall outside of the scope of current competition laws as the main actors are algorithms, not individuals, and the conduct lacks the usual hallmarks of cartel or collusive conduct such as an agreement or clandestine conversations between competitors. But is that a problem - is this just a case of high speed, high accuracy ‘price discovery and response’ by one firm of another’s pricing? Will these AI-assisted price wars mean consumers end up paying less or more?

How pricing algorithms work

Pricing algorithms input various types of data, such as product inventory, product demand, consumer characteristics and competitors’ prices, to output a price for a product. This price can be presented as a recommendation to a human decision-maker or being automatically implemented in the online marketplace.

Pricing algorithms can also set dynamic rules, such as to match the second lowest-priced competitor’s price or to set the price $0.05 lower than the highest-priced competing product.

These rules can be fixed or adapted over time as a result of machine learning. Autonomous learning protocols present a risk that pricing algorithms will learn that coordinating with competitors’ algorithms of competitors is the optimal strategy to maximise profit and that by following these rules, pricing algorithms may lead to collusion and price fixing between competitors.

While the notion of offering lower prices and price match guarantees is not necessarily new, the use of algorithms has increased the speed and scale of these price comparisons:

• Speed: By using automated pricing algorithms, analysis of real-time pricing data is able to inform pricing decisions almost instantaneously. Compared to traditional comparative techniques performed manually by a human, algorithmic pricing creates efficiencies by significantly reducing time and labour costs associated with employees observing and analysing competitor data and manually setting prices. Analysis by pricing algorithms can also be more nuanced as it can take into account fluctuations in price throughout different times of day as algorithms can process information in regular intervals with high frequency – potentially daily or hourly.
• Scale: In a concentrated market, the market structure makes it is possible for humans to easily observe competitors' prices and to engage in parallel pricing behaviour. Classic real-world examples include airlines and petrol stations. By using automated pricing algorithms, this process can be repeated in any market and aggregate a broad range of data across large numbers of competitors and vast time periods. The capacity to overcome structural characteristics of an industry means that through algorithmic pricing, any market is more prone to collusion.
• Anticipate and shape pricing behaviour: humans may be ‘taken by surprise’ when a competitor moves its prices. If viewed over the longer terms of past pricing conduct, the competitor’s pricing movements may not necessarily be a surprise. However, pricing algorithms can also go beyond merely monitoring competitors’ prices, but both the amount of data and their processing capabilities, algorithms can ‘see’ relationships which humans cannot, allowing them to anticipate competitors’ prices. But they can go further still – pricing algorithms can incorporate adversarial techniques to try to ‘discipline’ competitors (also operating a pricing algorithm) through a reward or punishment structure. Calvano et al. (2020) observed that pricing algorithms can tacitly learn to consistently charge supra-competitive prices and learn punishment mechanisms against players trying to defect. As these pricing algorithms become the norm, competitors are forced to choose between intervening and colluding. The reward-punishment scheme incentivises competitors to revise their own pricing algorithms to collaborate and coordinate their prices at the expense of consumers.

When does comparison become collusion?

Traditionally, there has been a distinction between ‘conscious parallelism’ and ‘collusion’.

Conscious parallelism occurs when companies independently adopt a common course of conduct, without any agreement or communication with competitors. This behaviour is not of itself unlawful or considered to be collusion because it lacks an intention to collude on prices.

Collusion refers to competitors coordinating or agreeing to work together with the objective of raising profits. This can occur explicitly (eg. by having conversations or signing written agreements) or tacitly (eg. by using the same algorithm to determine prices or react to changing market conditions).

Pure tacit algorithmic collusion is unlikely to be captured by the existing Australian competition law framework, as discussed below.

Price signalling vs price fixing

Observing a competitors publicly available prices and making unilateral pricing decisions will not fall foul of the current competition law regime. However, when all competitors are using similar algorithms, there can be coordinated effects such that everyone can instantly coordinate their prices at the same time. Even if algorithms cause higher prices than would otherwise exist in the market, this can still be legal conscious parallelism provided there is no agreement between sellers and their algorithms. In a sense, the pricing algorithm is just better at seeing and responding unilaterally to competitors’ pricing moves than the ‘human eye and mind’.
But the human limitations in detecting and responding to competitors’ pricing result in a key ingredient of successful competitive markets. Uncertainty amongst competitors about price is one of the key characteristics which creates competition and innovation. If algorithms remove this uncertainty by enabling competitors to have prior knowledge of how their competitors will react to changes in price, should the resulting more effective co-ordination or degree of ‘parallelism’ than could be achieved by humans be considered anti-competitive and if so, how would this be accommodated within existing theories of collusion? Like much in AI, there is not a simple, single answer.

Algorithmic pricing models in action

Rocher, Tournier and de Montjoye (2023) explore the need for economists and regulators to pay attention to algorithmic pricing models to how pricing algorithms, often with a degree of opaqueness and subtly, impact the competitiveness of online markets, sometimes with beneficial impacts on consumers and sometimes with harmful impacts. Their article sets out different scenarios which may emerge in two-firm markets and three-firm markets involving pricing algorithms:

• Adversarial competition: If one company uses a pricing algorithm to set it prices just under its competitor, the competitor will quickly notice that their profits have changed. This sends a clear signal to the competitor who could then revise their pricing strategy. If competitors iteratively revise and retrain their algorithms to attack, then profits will slowly decrease overall. In this scenario, the high profits initially obtained by one company are not necessarily sustainable. This suggests that pricing algorithms could ‘tear down’ the ability of a powerful competitor to price independently of its competitors faster and more effectively than ‘human-based’ pricing behaviour amongst competitors.
• Adversarial collusion: If one company uses a pricing algorithm to set its prices based on a competitor’s prices, it may communicate that if competitors’ algorithms do not follow the higher price, then the company’s prices will be lowered as a punishment. By disincentivising competitors’ intervention, pricing algorithms can instead unilaterally increase both its profits and the profits of competitors. Algorithms programmed to optimise their profits will continue to choose actions which produce a more successful result based on its understanding of the market conditions and will ultimately choose collusion as it leads to an outcome with symmetric and supra-competitive profits. This suggests more subtle, difficult to discern and ‘stickier’ price co-ordination between competitors than achievable by ‘human-based’ pricing behaviour.

Example:

Company A knows that its competitor Company B is using a pricing algorithm that is programmed to set its prices $1 below its competitor's price. Assume the competitive price without any algorithmic intervention is $10.

Scenario 1 - Adversarial competition: Company A sets its price at $10 and Company B instantly responds by setting its price at $9. Company A is now losing market share even though it has set its price at the competitive price and so is incentivised to lower its prices in response to $9. However, because Company B is following the same algorithm as before, its prices will automatically be reduced again.

Scenario 2 - Adversarial collusion: Company A will set its prices at $15 knowing that will maximise its own profits, even if Company B will set its price at $14 and also receive a profit. Company A has been forced to set higher prices  given the threat of Company B's algorithm undercutting Company A.

The adversarial competition scenario maximizes the attacker’s profits while the adversarial collusion scenario maximizes the profits of all agents.

Tesauro and Kephart similarly found that when reinforcement learning algorithms are used by multiple competitors in the same market, pricing algorithms learn to increase profitability and damp out or eliminate cyclic price ‘wars’.

Other forms of sophisticated pricing algorithms or arrangements could also lead to higher prices, such as:

• personalised pricing, where an algorithm will adapt or discriminate prices based on certain consumer characteristics or their willingness to pay. This would raise prices for some consumers and create secret discounts for others and be harder to detect when prices are adapted to specific consumer profiles.
• price cycling, where an algorithm will continuously lower prices until prices get so low that they automatically reset at a higher price; and
• hub and spoke arrangements, where competitors are using the same third-party algorithm supplier, the supplier may play a role and have an incentive to raise prices above the competitive level or set up cross platform parity agreements which facilitate collusion between the retailers.

Competition law risks

Some of the risks posed by algorithmic pricing have been considered before by competition lawyers and regulators in Australia and abroad. However, issues unique to the online retail marketplace introduce new questions such as determining intent and finding evidence of an agreement or understanding between algorithms.

Cartel conduct and concerted practices

There are two ways that anti-competitive algorithmic pricing could be captured under the Competition and Consumer Act 2010 (Cth) (CCA) are:

• cartel conduct: where competitors agree, or attempt to agree, to a contract, arrangement or understanding with the purpose, effect or likely effect of fixing or controlling prices; or
• concerted practices: where competitors do not reach a contract, arrangement or understand, but engage in cooperative conduct with the purpose, effect or likely effect of substantially lessening competition in a market.

Both cartel conduct and concerted practices require some sort of proof of direct or indirect conduct to show that competitors have not acted independently from each other. This means regulators focus on whether there has been a ‘meeting of the minds’ between competitors. But what happens when the decisions are partly made by pricing algorithms reacting to each other?

ACCC inquiries into online marketplaces

The ACCC considered the impact of algorithmic pricing in the Digital Platform Services Inquiry (DPSI) Interim Report ‘General online retail marketplaces’ released in March 2022 and found that algorithmic pricing increased the risk of collusion as a result of:

1. increased transparency of up-to date pricing data;
2. the enhanced ability to monitor and retaliate in real time; and
3. the availability of mechanisms which automatically respond and adapt with limited need for (slower and less reliable) human engagement.

The ACCC also considered the increased risk of price gouging in rapidly changing market conditions where demand suddenly outstrips supply and could lead to unrealistic and excessive prices. So far, it has been up to the retailers rather than regulators to monitor and address this risk. For example, eBay Australia has a policy against price gouging which requires sellers to offer reasonable prices for items that are considered essential.

Competition law enforcement

While not all algorithmic pricing will be actionable, what will be the yard stick? Identifying the requisite intention, detecting collusion and establishing liability present hurdles to reducing adversarial collusion and ensuring compliance with competition laws:

• Intention: Intention is not always objectively determinable, even for humans engaging in anticompetitive arrangements. The sequential pricing behaviour could be traced through an algorithm’s data trail, pattern recognition programming and analysis patterns once an algorithm detects what a competitor’s algorithm is doing. But is this ‘intention’? Intention of the programmers could be inferred from rules that lead to higher prices or punish for undercutting. However, an algorithm’s objective to maximize profits could be consistent with both pro-competitive and anti-competitive conduct. Patterns of detection/analysis/response in processing data by an algorithm are just that.
• Detection: There is limited capacity for regulators, and even researchers, to investigate potential collusion without direct access to data and algorithms. Algorithmic pricing strategies are often designed to be difficult to detect or unnoticeable. Algorithms contain commercial proprietary information and are unlikely to be provided voluntarily or are difficult to analyse in isolation of a broader automated system. In any event, AI programmers candidly acknowledge that they don’t quite understand how AI often works.
• Liability: Another issue is who should be held liable for anticompetitive conduct enacted by algorithms without human intervention. Human developers may code the software with fixed rules or machine learning protocols that do not necessarily instruct the algorithm to collude, but merely instructed the algorithm to implement a pricing strategy to optimise profits. As algorithms self-learn and anticipate market conditions, collusive schemes can be created and implemented automatically and autonomously. Should the responsibility be on companies to monitor these inputs and outputs rather than simply following its pricing blindly?

Global competition law concerns

We can also take guidance from the policy approaches in other jurisdictions, which depend on the awareness and agreement of the competitors.

• United Kingdom: The Competition and Markets Authority (CMA) released an economic research paper in 2018 on pricing algorithms and certain risk-factors. The CMA was primarily concerned about hub and spoke arrangements as presenting the most immediate risk because all that this required was for companies to knowingly use the same algorithm as their competitors and be engaging in potentially collusive behaviour. The CMA has also set up a specific Digital Markets Unit (DMU) to assess and quantify harm caused to digital markets, including the possibility of ‘autonomous tacit collusion’, whereby pricing algorithms learn to collude without requiring other information sharing or existing coordination (but this may overlook the warning from Rocher, Tournier and de Montjoye that pricing behaviour by pricing algorithms needs to be understood in all its subtle forms, as discussed above).
• European Union: Although Article 101 of the Treaty on the Functioning of the European Union (TFEU) clearly states that collusion through agreements, decisions by associations of undertaking or concerted practices are prohibited, the Court of Justice of the European Union (CJEU) has historically afforded considerable leeway to companies unilaterally setting prices if there is no collusion. However, algorithmic collusion can be viewed through the lens of concerted practices if it can be proven that the competitor became aware that its algorithm was being manipulated (so human knowledge is still needed beyond the firms’ algorithms interacting with each other in a punishment reward mode).
• United States: The Department of Justice (DOJ) has suggested that algorithmic collusion will not be illegal if there is no agreement among the participants. The DOJ has advised that when analysing whether conduct constitutes collusion, regulators should put to one side the fact that algorithms are involved and focus on the existence of any communications which rise to the level of an agreement. When algorithmic pricing models learn to collude autonomously, there is no agreement and no form of communication (this was, in effect, an application of standard pre-algorithmic thinking on the differences between collusion and parallelism).

But in February 2023, the United States Department of Justice Antitrust Division announced a policy shift to address information exchanges facilitated by algorithms and the Principal Deputy Assistant Attorney General Doha Mekki explained:

“… high-speed, complex algorithms can ingest massive quantities of ‘stale,’ ‘aggregated’ data from buyers and sellers to glean insights about the strategies of a competitor.. … Where competitors adopt the same pricing algorithms, our concern is only heightened. Several studies have shown that these algorithms can lead to tacit or express collusion in the marketplace, potentially resulting in higher prices, or at a minimum, a softening of competition.”

Are there public benefits to algorithmic pricing?

As consumers feel the pinch of the cost of living crisis, could pricing algorithms bring down prices overall and facilitate a positive economic effect for consumers?

In some instances, pricing algorithms may actually benefit the competitive process if they are used to compete for sales in online marketplaces.

Generally, price transparency will be a good thing as it leads to more favourable prices for consumers and improves consumers’ decision-making power by providing access to more comparative information. When consumers search and compare products online, price is often one of the dominant factors in a consumer’s decision-making process. If businesses are seeking to boost their online sales by competing on price, then dynamic pricing algorithms can help certain sellers stand out and connect consumers with the cheapest available products.

Considered from a theory of harm point of view, each algorithm is trying to find the equilibrium price. If all firms are trading at the equilibrium price with sufficient quantity to meet consumer demand, this is also maximising surplus. In the DSPI Report, the ACCC acknowledged that “dynamic pricing algorithms can reduce business costs for sellers and increase price competition to the benefit of consumers”.

One view of algorithmic pricing is that it arms competitors with more market data and is simply another tool to influence market dynamics. Automatic adjustments to price will not overhaul all existing market forces which still hold significant influence such as consumer demand, brand loyalty and offline retailers.

That said, as Rocher, Tournier and de Montjoye point out, pricing algorithms can work just as powerfully, quickly and subtly in achieving higher prices for competitors to the disadvantage of consumers.

Where from here?

While the inherent limitations of human ‘eyes and minds’ in detecting and responding to pricing behaviour in markets provide a practical buffer or uncertainty on which antitrust and competition regulators could implicitly rely, how do we address the potentially accelerated ‘good’ and ‘bad’ of algorithmic pricing? Broadly, the options are:

1. Leave well enough alone: Existing competition law distinctions between collusion and parallelism continue to hold true – and it should make no difference that pricing algorithms are just faster and better at detecting and responding unilaterally to competitors’ pricing;
2. Redefining existing competition laws: Expanding the definition of intention or collusion to capture automated learning processes and algorithms which have the effect or likely effect of resulting in collusion. By updating the current approach to the elements of cartel conduct and concerted practices, there is a greater likelihood of responding to actual or potential algorithmic collusion in online marketplaces. But if the dividing line between collusion and parallel conduct is shifted to embrace learned co-ordinated behaviour between algorithms, what is that new standard and how can it avoid capturing instances of legitimately attacking behaviour between algorithms?
3. Discrete regulatory or legislative framework for algorithms: Introducing a new regulatory and legislative framework specifically for algorithms could offer a more focused approach to auditing, investigating, and enforcing competition law, provided it is flexible enough to adapt as technology evolves. This could enable regulation the anticompetitive features of algorithms without requiring detailed knowledge of precise algorithms or setting the competitive price levels. For example, a legal obligation for companies to program algorithms in a way that will prevent them from setting oligopolistic prices. But again, if the search for some agreed behaviour between competitors is abandoned in favour of capturing unilateral behaviour, what is the relevant pricing standard (oligopolistic pricing by what standard?)

Read more: Adversarial competition and collusion in algorithmic markets