AVV. ROBERTO CORNETTA
ABSTRACT
Italy’s wholesale electricity market has changed profoundly in a few years. Until 2024, demand paid a single national price and Terna procured dispatching services in a market largely reserved for large dispatchable plants. Today zonal prices also apply to consumers, the Integrated Dispatching Code has opened balancing to aggregators, storage and units of any size, the day-ahead market is coupled with the European market at fifteen-minute resolution, and a growing share of bids is generated by algorithms.
These reforms may mitigate the risks that marked the previous phase, namely local markets with very few operators and weak price signals, as illustrated by the capacity-withholding cases decided by the Italian Competition Authority in Campania and Brindisi. They leave open a new risk: tacit collusion among learning algorithms that, interacting in repeated, frequent and transparent auctions, learn to sustain supra-competitive prices without any agreement.
The article examines to what extent this risk can be addressed with existing tools. Article 101 TFEU may apply to algorithms that exchange signals, whereas for autonomously learning algorithms the case-law test of the ‘only plausible explanation’ makes concertation very hard to prove; Article 102 TFEU and Regulation (EU) No 1227/2011 on wholesale energy market integrity and transparency (REMIT) offer alternative routes but face limits of their own. The problem thus appears to be one of proof rather than of a legal gap. Using law and economics, and drawing on the North American experience, it compares three responses: liability of the firm deploying the algorithm, market-design rules, and algorithmic surveillance by the regulator. It finally outlines possible lines of intervention based on documentation duties for bidding algorithms and on the use of the market operator’s public data for early detection, which would make existing rules applicable.
Contents: 1. Introduction. – 2. The Italian electricity market before the reforms. – 3. The 2025–2026 reforms. – 4. Tacit collusion between algorithms. – 5. The legal framework and its gaps. – 6. Economic analysis of the possible responses. – 7. The North American experience. – 8. Conclusions.
1. Introduction
A growing share of bids on European electricity markets is no longer formulated by human operators but by algorithms: according to analyses cited by the European Union Agency for the Cooperation of Energy Regulators (ACER), on the EPEX electricity exchange trades executed by automated systems rose from 44 per cent of volumes in 2020 to 70 per cent in 2024.[1] In the same years the Italian market changed its rules: from 2025 demand pays zonal prices and balancing is open to operators of any size, and since October of the same year the day-ahead market has run on fifteen-minute intervals. The analysis that follows concerns tacit collusion between algorithms in wholesale and balancing electricity markets, with particular reference to the conditions that may favour it and to the legal instruments available to counter it. As will be seen, the crux appears to lie less in the absence of rules than in the difficulty of proving their breach: in a market such as the electricity market, parallel bidding almost always admits an explanation other than concertation, and proof to the contrary depends on the availability of information on how the algorithms work.
2. The Italian electricity market before the reforms
The organised electricity market came into being with the electricity exchange run by the Gestore dei mercati energetici (GME), operational since 2004 in implementation of Legislative Decree No 79 of 16 March 1999. Its core is the day-ahead market: for each hour of the following day, producers submit sell offers indicating quantity and minimum price, and buyers submit purchase offers; the GME ranks the offers and sets the price at the point where demand and supply meet. The price is uniform, in the sense that all accepted offers receive the marginal price of their own zone. The territory is in fact divided into market zones separated by transmission constraints, so that, when interconnection capacity between zones is saturated, zonal prices diverge. Until 2020 there were six zones, alongside four limited production poles, that is, areas with few plants and constraints of their own;[2] since 2021 the configuration has been revised.
The demand side followed a peculiar rule: whereas producers received the price of their own zone, buyers paid a single national price (Prezzo Unico Nazionale, PUN), equal to the average of zonal prices weighted by the quantities purchased.[3] The PUN attenuated territorial differentials for consumers, but removed from demand any local price signal.
The day-ahead market was followed by the intraday markets, in which operators adjust their positions, and by the dispatching services market (Mercato per il Servizio di Dispacciamento, MSD), in which Terna purchases the resources needed to resolve congestion, build reserve margins and balance the system in real time. Three features of this market are relevant for present purposes. Terna operates there as the sole buyer, and the related costs are passed on to consumers through the bill. Accepted offers are remunerated at the price indicated by each operator (pay-as-bid), not at the marginal price. Participation, finally, was reserved to enabled units, largely large programmable plants required to offer their available capacity; demand and aggregated virtual units entered only with the pilot projects launched from 2017.[4] Moreover, the selection of offers does not follow a simple merit order: Terna solves a complex, zone-based optimisation problem, gives priority to resolving congestion and is not required to give reasons for the choice of individual offers.[5] For operators, anticipating when they will be indispensable is therefore an informational advantage, which grows with knowledge of the grid and of competitors.
Italian dispatching is, moreover, of the central type. It is Terna that establishes the production schedule of each plant, and operators submit offers referring to single units, not to their fleet as a whole. In most European countries, by contrast, self-dispatching applies: each operator decides how to use its own plants and communicates to the grid operator only the overall position of its portfolio. The central model is thus a minority one in Europe; besides Italy it is adopted by Greece, Spain, Poland and Ireland.[6] Since 2009, finally, the GME has published the individual offers submitted on the spot markets starting from the seventh day after the session, whereas previously the wait was twelve months.[7]
The fact that the two markets close one after the other creates a precise incentive. A producer who knows that it is indispensable to Terna in a given area may forgo selling its energy on the day-ahead market, offering it at a price too high to be accepted or not offering it at all: it knows that Terna will have to use its plants anyway in the dispatching services market, where it will be remunerated at the price it itself indicated. This is what is termed capacity withholding, which is called economic when the offer is submitted but at an out-of-market price, and physical when it is entirely absent. The remuneration criterion reinforces this incentive. In a uniform-price auction, such as the day-ahead one, every operator receives the marginal price irrespective of its own offer; in an auction in which each is paid the price it bid, as in dispatching, the operator must instead estimate the highest price that will still be accepted. A mechanism of this kind tends as a rule to contain prices, but encourages strategic bidding, detached from actual production costs.[8]
These incentives have not remained a theoretical hypothesis, since two cases decided by the Italian Competition Authority (AGCM) show that capacity withholding has been practised in the Italian market: in the first through an agreement between undertakings, in the second through a tacit alignment between operators.
The first case concerns a collusive practice established in Campania.[9] Between May and October 2010, on Sundays and public holidays, the four production units in operation in the area, belonging to three companies, submitted on the day-ahead market offers destined not to be accepted, and then took turns in the dispatching services market, where the average price came to exceed the day-ahead price by two and a half times. The cartel, which had escaped market monitoring, emerged only thanks to a complaint by one of the participants; the AGCM found the existence of a concerted practice and fined the undertakings involved.
The second case concerned the Brindisi area.[10] Between April and June 2016 the main operator in the area had withheld capacity on the day-ahead market, thereby increasing Terna’s recourse to the dispatching services market, and a second operator had adopted similar conduct. According to estimates, the conduct contributed to an increase of about €320 million, compared with the previous year, in the costs borne by Terna for dispatching.[11] Since no communications between the undertakings were found, the AGCM did not allege an agreement but opened two separate proceedings for abuse of a dominant position: the one against Enel ended with the adoption of corrective measures, the one against Sorgenia without a finding of infringement, because the undertaking was not considered dominant.
The comparison between the two cases, the one based on an explicit agreement, the other on a tacit alignment, highlights some critical points. In Campania the undertakings had reached an agreement, and it was precisely the existence of communications that enabled the AGCM to strike at the cartel as a restrictive agreement. At Brindisi communications were lacking, and the Authority was able to intervene only through the prohibition of abuse of a dominant position, which presupposes a dominant undertaking; for this reason the second operator, although it had aligned itself, went unpunished. It follows that tacit coordination between non-dominant operators risks escaping every prohibition, and this is exactly the type of coordination that bidding algorithms can produce without any contact between the undertakings. The two cases also offer two data points useful for economic analysis. The first concerns the probability of detection: in Campania the cartel emerged because of a participant’s complaint, not as a result of monitoring, although the offers were public. The second concerns the extent of the harm: at Brindisi about two months of conduct were estimated to have contributed to an increase in costs of about €320 million, passed on to consumers through the bill. Conduct with high harm and a low probability of detection is the case in which economic analysis would recommend investing in detection rather than in higher penalties.
3. The 2025–2026 reforms
Article 13 of Legislative Decree No 210 of 8 November 2021, as amended by Article 19 of Decree-Law No 181 of 9 December 2023, has done away with the single national price. From 1 January 2025 purchase offers on the day-ahead market are valued at zonal prices, and the GME calculates a reference price, the ‘PUN Index GME’, equal to the average of zonal prices weighted by the quantities purchased, used as an index in contracts; a transitional equalisation mechanism offsets the differences.[12] The declared objective is to give weight to local price signals in a system with growing shares of non-programmable renewable sources.
The deepest reform concerns dispatching. The Integrated Text on Electricity Dispatching (Testo integrato del dispacciamento elettrico, TIDE), approved by the Italian Regulatory Authority for Energy, Networks and Environment (ARERA) by Resolution 345/2023/R/eel, entered into force on 1 January 2025 in phases: a transitional phase until 31 January 2026 and a consolidation phase, from 1 February 2026, with near-complete implementation.[13] Four novelties are relevant to this analysis. The old dispatching services market gives way to the Integrated Scheduling Process and to the balancing and redispatching market, both still run by Terna. Responsibility for imbalances is attributed to a dedicated figure, the Balance Responsible Party.[14] There is no longer a minimum power threshold for participating in the services, so that aggregators, storage facilities, virtual units and demand itself may enter. National dispatching, finally, is coordinated with the European platforms for the exchange of balancing products.
Two innovations of European origin are, moreover, relevant to this analysis. The first concerns the day-ahead market. National markets no longer operate in isolation: they are coupled in a single European auction, resolved by a single algorithm that determines together the prices of the different zones and the use of cross-border interconnections. From delivery day 1 October 2025, moreover, the trading unit changed from one hour to fifteen minutes in all bidding zones.[15] Every day the same operators therefore face one another no longer in 24 auctions but in 96; as will be seen in section 4, this frequency of interaction is relevant to the risk of coordination between algorithms.
The second innovation concerns the regulation of algorithmic trading. Regulation (EU) No 1227/2011 on wholesale energy market integrity and transparency (REMIT), as amended by Regulation (EU) 2024/1106,[16] regards as algorithmic the trading in which a computer program automatically determines the parameters of the orders (whether to submit them, when, at what price and for what quantity) with limited or no human intervention. The new Article 5a requires those who engage in it to have in place effective systems and risk controls, thresholds that prevent the sending of erroneous orders and measures to deal with malfunctions, and to subject their systems to testing and monitoring. It also requires them to notify the activity to the national authority and to ACER. The national authority may request a description of the algorithmic strategies, parameters and controls adopted, which operators must retain for five years. Operators registered in Italy notify ARERA directly, and the notification is also valid vis-à-vis ACER.[17] From 29 October 2027, finally, transactions reported to ACER will have to indicate the identification code of the algorithm used.
The spread of the phenomenon nevertheless remains difficult to measure. The figure on the EPEX exchange cited at the outset is from a journalistic source, and ACER does not derive it from its own records. A more telling figure is another: of the 21,169 operators registered in the European register, only 671 have declared that they engage in algorithmic trading, a number which ACER itself considers at odds with the widespread use observed in market surveillance.[18] For Italy no public data are available.
None of these reforms was adopted to counter collusion. They may, however, attenuate the risks of the previous phase: a larger number of participants in balancing makes the local markets in which the Campania and Brindisi cases arose less narrow; integration into the European platforms widens the geographic market for balancing energy; zonal prices give back to demand the local signals that the PUN cancelled. Other elements remain unchanged: the sequential closing of the markets, central dispatching with unit-by-unit offers, the publication of offers after seven days. Some, finally, add new risks. With 96 daily auctions, interactions between the same operators multiply, and storage facilities and aggregators, the new entrants favoured by the TIDE, mostly bid through optimisation software. The result is a more open market, but also one better suited to mutual learning between algorithms.
The characteristics that expose the Italian market are not, moreover, common to the whole of Europe. The day-ahead market is by now single, but dispatching remains national. In Germany and France self-dispatching applies; Germany, moreover, forms a single price zone with Luxembourg and manages internal congestion through redispatching largely regulated on the basis of costs rather than market offers. The result is fewer narrow local markets and less public information referring to individual plants. Spain is instead the closest term of comparison: it shares central dispatching with Italy, and its electricity market has been the subject of studies that identified in it signs of an alternation between phases of tacit collusion and price wars among the main operators.[19]
4. Tacit collusion between algorithms
It is useful first of all to distinguish, with the scholarship, two categories of algorithms. The first execute the sequence of operations for which they were programmed, so that their outcome is always predictable; these include monitoring algorithms, which check competitors’ prices, and signalling algorithms, which indicate a price and verify its reception. The second, known as self-learning algorithms, receive only the objective to be attained and choose the strategy by trial and error, evaluating the effects of their own decisions and correcting themselves. Their outcome, Manzini observes, “is not predictable ex ante, because it is determined by the self-learning process which, once triggered, goes beyond the control of the programmer”.[20] The studies recalled below concern this second category.
A reinforcement-learning algorithm receives no instructions on how to behave: it tries different strategies, observes the results and reinforces those that yield the most. In a repeated auction it can thus discover that it does not pay to lower its price, because the competing algorithms have also learned to react to every price cut with a price cut of their own. The punishment strategy is the heart of the mechanism: if one algorithm lowers its price to gain share, the others respond for a few periods with even lower prices, until the price cut is no longer profitable, and then all return to the high level. No one programmed this sequence, which emerges because, in the simulations, it is the one that maximises profit over time. The result is a cartel price without any agreement, either between the undertakings or between the programs.
This outcome has been obtained, in simulation, with relatively simple algorithms, which arrive at prices above the competitive level and sustain them with punishment strategies learned on their own,[21] and similar results have been obtained in contexts where prices are set sequentially.[22] Empirical confirmation comes from the German retail fuel market, where margins increased only in the local markets in which both competitors had adopted pricing algorithms.[23] Legal scholarship has noted the consequence: antitrust law strikes at the agreement, and in collusion between algorithms the agreement is missing.[24]
The conditions that favour this outcome are typical of electricity markets: few sellers, auctions repeated every day with the same counterparties, demand that reacts little to price and published results. Specific studies are still few, but concordant. As early as the late 2000s a power exchange was simulated with producers learning by reinforcement, showing that the form of the auction affects collusive outcomes: payment at the bid price attenuates them when demand is low, whereas when demand is high the difference disappears.[25] Other studies have observed apparently collusive outcomes between algorithms managing storage plants.[26]
A recent contribution, conducted on a simulated electricity market, helps clarify the terms of the problem.[27] The authors entrusted the bids of four power producers to programs that learn by trial and error, without communicating with one another and knowing only their own results and the public prices, and observed their behaviour over numerous market sessions. When the grid presents bottlenecks, prices settle at levels significantly above competitive ones. In some scenarios the programs also display the typical traits of a tacit agreement: they react to the producer that lowers its offers by reducing their own prices in turn for a limited period, and then return to the previous levels. In others, by contrast, equally high prices are explained solely by the location of the plants in the grid, without any coordination. For the jurist the finding is relevant in two respects: an outcome analogous to that of a cartel is produced without any contact between the undertakings, that is, without the presupposition on which the prohibition of agreements rests; and the proof of such coordination cannot be based on the level of prices, but requires analysis of behaviour over time, such as the reaction of competitors to whoever deviates. The same authors, moreover, present their results as a warning signal rather than as conclusive proof.
Four characteristics, already described, make the risk concrete in the Italian market. The first is repeated and frequent interaction: the same units face one another every day, now over 96 periods, in the day-ahead market and then in balancing. The second is the presence of narrow local markets, because central dispatching, with offers referring to single plants, and grid constraints create areas in which very few units compete, as at Brindisi. The third is the publicity of data: publication of offers after seven days allows the regulator and scholars to uncover cartels, but also allows an algorithm to learn competitors’ strategies, and the relationship between transparency and coordination is, moreover, one of the recurring themes of studies on electricity markets.[28] The fourth is the sequential closing of the two markets: the incentive to withhold capacity on the day-ahead market in order to profit in balancing exists even without collusion, and an algorithm that optimises across both markets may discover this strategy on its own, just as two algorithms may discover it together. The second and third characteristics distinguish the Italian model from European self-dispatching systems; the legal problem, examined in the next section, therefore arises in Italy probably with a certain urgency.
5. The legal framework and its gaps
The first instrument is the prohibition of anticompetitive agreements in Article 101 TFEU, which catches agreements and concerted practices. To assess its reach with respect to algorithms, scholars distinguish three scenarios.[29] In the first, algorithms merely execute or make more effective a collusion decided by people, for example by monitoring competitors’ prices or automating retaliation: here Article 101 applies without strain, because the agreement remains attributable to the undertakings, and the use of the technology may at most be relevant to the gravity of the infringement. In the second, algorithms interact with one another, exchanging signals or communicating directly, and give rise to forms of coordination functionally equivalent to an agreement; here too the phenomenon can be brought within the prohibition through an evolutionary interpretation of the meeting of minds and the principles for attributing conduct to the undertaking.[30] In the third, finally, algorithms act autonomously but interdependently, adapting their decisions to competitors’ moves without agreements, exchanges of information or relevant contacts between the undertakings: this is algorithmic tacit collusion, which remains for the time being, like traditional tacit collusion, outside the scope of Article 101.
The boundary between the scenarios depends on the notion of concerted practice, which the case law understands broadly. It requires neither a manifestation of reciprocal will nor an actual plan, but is realised through contacts, direct or indirect, capable of influencing the conduct of competitors or of revealing one’s own; the administrative case law describes it as a category that is “instrumental, operating on the evidentiary plane” in aid of establishing a prohibited agreement.[31] At the same time, the case law recognises each undertaking’s right to adapt intelligently to the conduct, present or anticipated, of its competitors. Scholars have long debated whether this approach is adequate for oligopolistic markets: some authors consider that tacit collusion, as a voluntary and rational choice of undertakings, should be treated like an agreement, others that an undertaking cannot be forbidden to take competitors into account.[32] Algorithms make the question more acute, because they make tacit collusion faster and more stable, and possible even in markets that do not display the classic characteristics of oligopoly.[33]
The Court of Justice has extended the notion of concerted practice to a message sent, through the system of a common platform, to all the travel agencies that used it, presuming that they were aware of it.[34] According to scholars, applied to monitoring algorithms this case law entails a rebuttable presumption (iuris tantum) of participation, by tacit assent, in a concerted practice by each undertaking that adheres to a system employing such algorithms to influence prices.[35] In that case, however, there was a common vehicle, imposed from above; with distinct algorithms, developed by different undertakings and trained on the public data of the market, even this is lacking. The Commission’s guidelines on horizontal co-operation agreements adopt a consistent approach: they warn that algorithms may become an instrument of collusion, but presume lawful the independent use of an algorithm based on public data.[36] In July 2025 the Commission announced several investigations into algorithmic pricing, without indicating the sectors; as far as is known, none concerns electricity markets.[37]
Two other lines of authority deserve attention. For the Court, even a single contact between competitors suffices, if capable of eliminating uncertainty about their future conduct.[38] The Commission further held that public announcements of price increases, of no utility to customers, could function as signals between competitors, and closed the proceedings by accepting the undertakings’ commitments.[39] The offers published by the GME are not announcements addressed to competitors, but they share one trait with them: they make others’ conduct observable.
On this terrain sits the thesis, upheld in Italian scholarship, that signalling algorithms amount to a concerted practice. An algorithm that repeatedly indicates a price, observes competitors’ reactions and confirms or modifies it according to whether they align or not “puts the undertakings in virtual but effective communication, thereby enabling them to reach a veritable ‘meeting of minds’”.[40] Competitors’ adjustment is formally an autonomous reaction, but in substance, since it follows price indications rendered intelligible by their constant repetition, it represents “the acceptance of a proposal”, and hence the conclusion of an agreement. The conduct would be comparable to tacit participation in a cartel, like passive attendance at meetings at which prices are agreed, which the Court of Justice regards as a source of liability when the undertaking does not publicly distance itself from them.[41] The same scholarship, by contrast, rules out extending Article 101 to every autonomous alignment in a context of interdependence, because it would make uncertain the boundary between lawful and unlawful conduct.[42]
The thesis must reckon with a precedent pointing the other way. In the Wood Pulp judgment, the Court of Justice had excluded that price announcements addressed to users constituted a concerted practice, because, when each undertaking makes them, it cannot be sure of the future conduct of the others, and the announcements therefore do not reduce reciprocal uncertainty.[43] More recent practice has, however, scaled back that precedent. The Commission’s 2023 guidelines admit that even unilateral disclosure or an indirect exchange of information may constitute a concerted practice, and that information may circulate indirectly through third parties, among which they expressly mention “an algorithm”; they further warn that public announcements of future pricing intentions, especially in markets with few competitors and high barriers to entry, may serve to implement or monitor a collusive agreement.[44]
A recent application comes, in the energy sector, from Italian case law. In the ‘bio component’ case concerning fuels, that is, the part of the price that covers biofuel blending obligations, the AGCM found a concerted practice among the main oil companies: one of them announced in a trade magazine the value of its own bio component and subsequent increases, and the others aligned with it, with almost perfect convergence between 2020 and 2023. The Regional Administrative Court for Lazio (TAR Lazio) upheld the decision, holding that the announcements, together with the alignment and mutual monitoring, had created an “artificial” transparency, useful for verifying adherence to the cartel, and that the undertakings’ failure to distance themselves publicly from the announcements had to be read as acceptance.[45] The case confirms, in a market with a homogeneous product and few operators, the logic that scholars attribute to signalling algorithms: a repeated public signal, followed by a consistent adjustment by competitors, can amount to an indirect contact.
In the electricity market the thesis finds at least two possible fields of application. The first is the continuous-trading intraday markets, in which offers are visible to operators while trading is under way. The second comprises the day-ahead and dispatching markets, in which individual offers become public after seven days: deferred signals, but repeated every day for each trading period. In both cases the publicity of offers, besides supplying algorithms with learning material, could constitute the channel of the ‘indirect contacts’ on which the concerted practice rests. What in the bio component case passed through a trade magazine could, in the electricity market, pass through the data published by the GME and through the algorithms that process them.
For self-learning algorithms the position changes. They are not programmed to send signals and await the response, but merely analyse competitors’ offers and react to them, so that, according to the same scholarship, “it is difficult to see room for the application of Article 101”.[46] The decisive criterion then becomes the one enunciated by the Court of Justice in the Wood Pulp case and adopted by Italian administrative case law: parallel conduct may be considered proof of concertation only where concertation is the sole plausible explanation for it, because Article 101 TFEU does not deprive economic operators of the right to react intelligently to the known or anticipated conduct of their competitors. It must therefore be ascertained whether the parallelism can be explained otherwise, having regard to the nature of the products, the number and size of the undertakings and the volume of the market.[47]
Applied to the electricity market, the criterion produces a paradoxical result. The characteristics that make tacit collusion likely, namely a perfectly homogeneous product, few operators in areas subject to grid constraints and demand that reacts little to price, are the same ones that supply an alternative explanation for parallelism: in a market so structured, the alignment of offers can always be presented as a rational reaction to its conditions. With algorithms the paradox is accentuated. A program built to maximise profit and able to learn from competitors’ published offers supplies, by its very nature, precisely that alternative explanation; if parallelism alone were the available proof, tacit collusion between algorithms would remain almost always unprovable.
The same case law nevertheless points to a way out. Alternative explanations for concertation must be examined not in their abstract plausibility but in the concrete context and against the documentation obtained, and cease to be credible when contemporaneous documents contradict them; studies prepared after the event do not, moreover, have the same probative force.[48] In the case of algorithms, the contemporaneous documents are the code, the objectives assigned to the program, the data with which it was trained and the records of individual offers: precisely the information capable of confirming or excluding that the algorithm merely adapted to the market. The availability of this documentation thus becomes the condition for the criterion of the sole plausible explanation to be applicable, and it is on it that the lines of intervention indicated in section 7 are founded.
Italian administrative case law shows, moreover, a tendency to soften the rigour of the criterion. The Consiglio di Stato has oscillated between the requirement that the accusatory hypothesis be the only one capable of explaining the evidence gathered, or at least the most credible as against alternatives lacking support, and the sufficiency of serious, precise and concordant indicia assessed as a whole.[49] In the bio component case the TAR Lazio went further, stating that, when the Authority demonstrates anomalous economic behaviour, such as uniformity and parallelism of conduct, “it is permissible to presume the existence of concertation, placing on the undertakings the burden of providing an alternative and plausible explanation”.[50] The formula is stricter than that of European case law, according to which, when the authority relies on parallelism alone, it is for the authority to exclude alternative explanations, and it must be read in the light of the specific case, in which the Authority also had contacts and internal communications at its disposal. For collusion between algorithms, however, it is precisely the allocation of the burden of proof that decides the outcome: a regime that shifts it onto the undertakings only when documentation on the algorithms is missing or reveals their collusive character, as envisaged in section 7, appears more compatible with the presumption of innocence.
The prohibition of abuse of a dominant position in Article 102 TFEU, applied in the Brindisi case, in turn presupposes a dominant position. In a market with several operators there remains the notion of collective dominant position, which the General Court (at the time the Court of First Instance) made subject to three conditions: sufficient transparency to know others’ conduct, a retaliation mechanism that makes deviation unstable, and the absence of reactions by competitors and customers capable of calling the outcome into question.[51] A market of repeated auctions, with published offers and algorithms that punish price cuts, satisfies the first two conditions almost by definition. The limit lies elsewhere: once the collective position is established, an abuse is needed, and the typical abuse of a tacit oligopoly is the excessive price, a cause of action which the case law applies with great caution.[52]
It is on this very terrain that scholars identify the most promising route for tacit collusion between self-learning algorithms. Their widespread use in a market could constitute one of the ‘correlative factors’ which, according to the Court of Justice, enable legally independent undertakings to anticipate one another’s conduct and induce them to align it, so as to constitute a collective dominant position; and since algorithms increase the transparency even of markets with many operators, the collective position could arise even outside an oligopolistic structure in the strict sense.[53] The abuse remains to be proved. In the electricity sector, however, the abuse does not necessarily coincide with the excessive price: capacity withholding has been challenged as a possible abuse of a dominant position in the German electricity market, in proceedings closed with commitments,[54] and is conduct more easily observable, in bidding data, than a price above the competitive level. Areas with few plants, such as that of Brindisi, are thus the natural terrain for applying Article 102 to tacit collusion between algorithms. There remain, finally, the instruments with which competition law traditionally addresses tacit collusion preventively, such as merger control, which assesses the risk of coordinated effects among the undertakings that remain on the market.[55]
REMIT offers the legal basis closest to the problem. The Regulation prohibits market manipulation (Article 5) and includes in it transactions that fix or attempt to fix the price of a wholesale energy product at an artificial level, unless the author shows that it acted for legitimate reasons (Article 2(2)). ACER’s guidance brings capacity withholding within it, distinguishing physical withholding, which consists in not offering available capacity, from economic withholding, which consists in offering it at prices so high that it is not selected.[56] The latter is precisely the conduct that an algorithm can learn: it does not require hiding anything, but only raising offers when convenient. Unlike the prohibition of agreements, the provision does not require an agreement and, unlike the US model examined below, it is not built on deception.
Two difficulties remain, however. The first is demonstrating that the price is artificial when every offer, taken alone, is a rational response to market conditions. The second concerns attribution: the result arises from the interaction of several algorithms, whereas the rule reasons in terms of individual conduct. Article 5a does not fill the gap, because it governs the orderly conduct of trading and not its outcome. At the enforcement level, REMIT leaves to Member States the regulation of sanctions, which must be effective, dissuasive and proportionate (Article 18), and in Italy competence lies with ARERA; neither the European nor the national rule, however, provides a specific evidentiary regime for conduct carried out through algorithms.
The institutional framework adds a further difficulty. The AGCM applies Articles 101 and 102 TFEU and Law No 287 of 10 October 1990; ARERA applies REMIT and regulates dispatching; the GME supervises the markets it manages; Terna writes, with ARERA’s approval, the balancing rules. Tacit collusion between algorithms falls at the meeting point of these competences: too little concerted for antitrust, too little individual for REMIT, too sophisticated for compliance checks on individual offers.
One may ask whether Regulation (EU) 2024/1689 on artificial intelligence could fill this gap, at least in part.[57] The answer appears to be negative, for a reason bound up with the logic of the Regulation itself, which distinguishes artificial intelligence systems according to the risk they present and reserves the most incisive obligations for those classified as ‘high-risk’. For the electricity sector, Annex III, point 2, includes in this category systems intended to be used as safety components in the management and operation of the supply of electricity, that is, systems whose malfunction could compromise the physical integrity of the grid and the continuity of service. The provision thus protects the security of the electricity system, not the proper functioning of competition. An algorithm that formulates offers on the market does not govern the physical operation of the grid: its error may translate into an economic loss for the undertaking or a higher price for consumers, but not into a danger to security of supply. It is therefore reasonable to consider that, as a rule, it remains outside the category of high-risk systems, although on this point there are, for now, no precedents or official indications.
It follows that the obligations the Regulation provides for high-risk systems do not apply to bidding algorithms, among them the drawing up of technical documentation describing the functioning of the system and the data used to develop it (Article 11) and the automatic recording of events during its use (Article 12). Yet this is precisely the information an authority would most need in order to establish coordination between algorithms: to know what the program observes when it decides, with what data it was trained, what offers it formulated and on the basis of what elements. As early as 2017 the European Parliament had called for the most advanced systems to be equipped with a ‘black box’ capable of recording the data of every operation and the logical steps of their decisions.[58]
6. Economic analysis of the possible responses
Having identified the gap, one must ask which response is preferable. The economic analysis of law offers two criteria for this purpose, which it is useful to set out before examining the individual options. The first concerns the information needed by whoever must apply the rule: a rule that asks the judge or the authority to establish, case by case, what conduct was due works well if that information is easy to obtain, but becomes costly and error-prone when it is not. The second criterion concerns the party that can avoid the harm at the least sacrifice: according to a by now classic approach, it is on that party that obligations and liability should be concentrated, because it is the one that can prevent the harm most cheaply.[59]
These criteria guide two choices that recur in every regulation of harmful conduct. The first is between liability and preventive regulation. One may intervene after the harm has occurred, attributing its cost to whoever caused it, or beforehand, laying down rules of conduct whose observance is monitored by an authority: the first solution is preferable when the actor knows the risks of its own activity better than the authority does, the second when the harm is hard to establish after the fact or the party causing it might not be able to compensate it.[60] The second choice is between precise rules and general clauses. A precise rule, such as a speed limit, demands effort from whoever writes it but is simple to apply; a general clause, such as the duty to drive prudently, is simple to write but requires an assessment of the concrete case each time. Precise rules are suitable when the conduct to be regulated is numerous and similar, general clauses when it is rare and varied.[61] In the light of these criteria, three possible responses are examined below: the liability of the undertaking that deploys the algorithm, market design rules and algorithmic surveillance by the regulator.
a) Liability of the party deploying the algorithm
The first response consists in attributing the collusive outcome to the undertaking that uses the algorithm, and it may take two forms, each with its own difficulty. With a fault-based criterion, the judge or the authority would have to establish what care the undertaking should have exercised in designing and training the program: information that nobody possesses today, because there are no shared standards on how a ‘diligent’ bidding algorithm should be built. With strict liability, which disregards fault and looks only at the result, the undertaking would bear the cost of the harm its own algorithm helps to produce, and would therefore have an interest in preventing it. To measure that harm, however, one must know at what price the market would have settled under competitive conditions, and that price, by definition, is not observed: it must be estimated, with all the attendant uncertainties. The proposal of ‘outcome visibility’ moves in this direction, according to which undertakings that benefit from the use of automated systems should answer for the outcomes those systems generate, regardless of whether they are attributable to coordination in the traditional sense.[62]
On the plane of attribution, moreover, existing competition law offers a fixed point. The undertaking that employs a self-learning algorithm capable of pursuing a price equilibrium above the competitive one, according to Manzini, “knowingly assumes and accepts the risk that the software it uses infringes antitrust law”, nor can it plead the unpredictability of its choices, because it could ask the programmer to exclude collusive options or, if that were not possible, refrain from using it.[63] The relationship between the undertaking and the algorithm would be governed, by analogy, by the rules on principal and commercial agent, which make the two a single economic entity responsible for the infringement, so that the ‘veil of the algorithm’ would be, at present, “undeniably pierceable”.[64] The problem, in this perspective, is not so much to whom to attribute the wrong as whether there is a wrong and how to prove it.
The difficulty of the fault criterion could be partly overcome by establishing in advance what diligence requires. Before deploying an algorithm, the undertaking could be required to test it in a simulated market, in which it interacts with other algorithms, on the model of the experiments recalled in section 4, in order to verify that it does not tend to converge on prices above the competitive level; during use, it could be required to retain the information on which each offer is based. Fault would then consist in the omission of verifiable tests and records, and not in a far more uncertain judgment on the correctness of the model. In a similar direction goes the proposal to require by law that algorithms be programmed so as to prevent them from setting oligopolistic prices (‘algorithms by design’).[65]
A deeper problem remains, however, concerning incentives more than attribution. When the collusive outcome arises from the interaction of several algorithms, none of them, taken alone, is its cause. Economic analysis knows this problem as that of harm produced jointly by several parties: if each answers only for a share of the harm, each has an insufficient incentive to take the necessary precautions, and no rule that merely apportions the harm among those responsible succeeds in correcting this distortion.[66] To this is added the opposite risk, that of an excessively severe regime: the same algorithms that may favour collusion also improve the forecasts and flexibility of the electricity system, and an overly broad liability could induce undertakings to forgo them, sacrificing real benefits. To avoid both drawbacks, some scholars propose striking not at the outcome but at certain characteristics of the algorithm, such as the capacity to condition its own offers on competitors’ moves in order to ‘punish’ those who lower prices.[67]
b) Market design
The second response looks not at individual undertakings but at the rules of the market. Its advantage is evident: it applies to all operators and does not require establishing who did what. The instruments are known from the experience of electricity markets. Caps may be set on offers; it may be provided that, when an operator is indispensable to meet demand in a given area, its offers above a cost-based reference value are automatically replaced by that value; data published on offers may be delayed or aggregated; and, finally, energy and balancing services may be purchased in a single market, thereby eliminating the incentive to withhold capacity in one in order to profit in the other.
Each of these instruments has, however, a cost. Automatic substitution of offers presupposes that the regulator knows the costs of each plant and updates them over time: information that is costly to collect, and whose erroneous estimation may compress prices that would legitimately be high in periods of scarcity. Joint purchase of energy and services would require a structural reform, remote from the European model. As for transparency, reducing it hinders the learning of algorithms, but also makes it harder to detect cartels and for those intending to enter the market to assess it.
On this last point there is nonetheless room for intervention. Studies on electricity markets examine what information guides operators’ bidding decisions, and confirm its weight.[68] The seven-day delay with which the GME publishes offers is not a natural datum but a regulatory choice: lengthening it, or publishing only aggregate data for local markets in which very few units operate, would reduce the information available to algorithms without depriving the regulator of the complete data, which it would continue to receive in confidential form. The choice also has legal significance: if published offers can function as signals between algorithms, the way in which they are made public affects the very possibility of the ‘indirect contacts’ on which the concerted practice rests being established (section 5).
c) Algorithmic surveillance
The third response uses artificial intelligence against itself: instead of regulating the algorithms that bid, it entrusts to an algorithm the task of monitoring them. A recent study applied machine-learning techniques to the GME’s public data relating to the Campania and Brindisi cases, succeeding in distinguishing collusive market sessions from competitive ones in about 97 per cent of cases, with a false-alarm rate below 6 per cent and a rate of undetected collusion below 2 per cent.[69] The authors present the result as a proof of concept, obtained on only two cases and hence in need of confirmation on broader data.
The advantage of surveillance is understood in the light of a classic argument of economic analysis. A rational undertaking refrains from unlawful conduct when the sanction it expects to suffer exceeds the gain it would derive, and that expected sanction depends not only on the amount of the penalty but also on the probability of being detected.[70] If the probability of detection is low, as the Campania case suggests, in which the cartel emerged only thanks to a participant’s complaint, a deterrent effect would require very high sanctions, which European and national law limit for reasons of proportionality. Increasing the probability of detection through algorithmic surveillance makes it possible instead to obtain the same effect with moderate sanctions.
On the legal plane, however, a statistical flag is not equivalent to proof. It can justify a request for information or an inspection, not a sanctioning decision, and must be capable of being explained to the undertaking concerned, which requires understandable and verifiable methods. It can, however, contribute, as an indication, to the overall picture on which the case law admits that proof of concertation may be based, which in most cases “must be inferred from a number of coincidences and indicia” and is to be assessed according to a canon of ‘narrative congruence’.[71] Statistical tools, moreover, have already entered adjudication: in the bio component case the TAR Lazio itself calculated the correlation coefficients between the undertakings’ price increases and regarded them as a “decisive endorsement” of the parallelism.[72] The example also shows the opposite risk: in the presence of a common cost shock, such as an increase in a regulatory obligation, a high correlation may be entirely natural, and this is why indicators capable of distinguishing coordination from a common reaction to an external event are needed. One must also take account of the errors that every tool of this kind inevitably makes: it may flag as collusive sessions that are not, with costs for the undertakings investigated and for the authority, or let collusive sessions pass, with costs for consumers. In the data of the cited study both errors are contained, but these are two cases in which collusion was frequent; in a market in which it is rare, even a low percentage of false alarms may produce many unfounded flags compared with well-founded ones, for the simple reason that competitive sessions are far more numerous. The flag must therefore open a phase of verification, not close it.
d) Who deploys the algorithm and who supplies it
The choice of response depends, finally, on who actually controls the algorithm. Many operators, especially among new entrants, do not develop their bidding programs in-house but turn to specialised suppliers; ACER itself includes among the algorithms subject to notification those supplied by third parties.[73] These programs are as a rule adapted to the needs of the individual client, but the adaptation mostly concerns parameters, such as the characteristics of the plants or risk appetite, while the logic with which the program observes the market and reacts to its signals tends to remain common.
When several competitors rely on the same supplier, the risk of coordination may grow for three reasons. Algorithms that react in the same way to the same signals become more predictable to one another. The supplier might train or refine its models with the data of all clients, including information that competitors could not otherwise know. And, if its remuneration depends on clients’ results, it has an interest in none of them competing too aggressively. Economic studies show that, under these conditions, a common supplier can contribute to raising prices even in the absence of an agreement among clients,[74] according to a scenario that, in different forms, emerged in the common platform case examined by the Court of Justice in Eturas and in those of shared pricing software discussed in the United States (section 7). If instead the program is truly custom-built and clients’ data remain separate, the case reverts to that of independent algorithms examined earlier; which of the two hypotheses prevails in the Italian market is a question of fact, depending on the degree of concentration of suppliers and on their contractual practices. Some scholars have judged it “implausible” that several undertakings adopt, through autonomous decisions, the same price-setting software;[75] in energy markets, however, recourse to external suppliers of trading algorithms is a phenomenon recorded by ACER itself, so that the hypothesis does not appear remote.
When coordination passes through a common supplier, the party on whom it is best to place obligations also changes. The criterion recalled at the beginning of this section suggests concentrating them on whoever can prevent the harm at the least sacrifice, and here that party is the supplier: it is the one that knows how the program works, that decides how the data collected from clients are used and that can prevent one client’s information from serving to formulate another’s offers, whereas the undertaking that merely uses the software has limited control over the model and often does not know its details. The approach is not new: the Artificial Intelligence Regulation allocates obligations precisely by distinguishing between whoever provides a system and whoever uses it.[76] Existing competition law confirms the usefulness of this distinction, but also shows its limit. According to the Court of Justice, an undertaking is liable for a concerted practice carried out through an independent service provider only if it was aware of its anticompetitive objectives or could reasonably have foreseen them;[77] the supplier, in turn, may be held liable as a facilitator of an agreement even if it does not operate on the relevant market,[78] but only if an agreement between clients exists. Scholars have translated this case law into a three-part test: the supplier is liable if it was aware of the undertakings’ unlawful conduct or could reasonably have foreseen it, accepting the risk, if it intended to contribute to their common objectives and if it contributed to them in a significant and non-accessory way; and whoever supplies monitoring or pricing software to a group of competitors “will hardly be able to claim that it could not reasonably have foreseen the unlawful conduct”.[79] In tacit collusion between algorithms both conditions risk being absent: clients may be unaware of what the program does, and there is no agreement to facilitate. To prevent no one from being liable, an efficient regime should therefore address both parties, with different obligations: documentation for whoever deploys the algorithm, separation of client data for whoever supplies it.
e) Summary
| Response | Information required | Attribution problem | Main risk | Current legal basis |
| Liability of the undertaking | Due care or competitive price | High: joint outcome of several algorithms | Excessive deterrence | Uncertain: Article 101 TFEU for signalling algorithms only; Article 102 TFEU and REMIT to be proved |
| Market design | Reference costs of plants | None: general rule | Errors in reference values, reduced transparency | ARERA regulation and Terna rules |
| Algorithmic surveillance | Bidding data already public | Low: identifies sessions and operators | False positives | REMIT supervisory powers |
In terms of the distinction recalled at the beginning of this section, automatic mitigation of offers is a precise rule, whereas the prohibition of manipulation is a general clause; and when the conduct to be regulated is numerous and similar, like the repeated offers of an electricity market, economic analysis tends to prefer the former. No option, however, suffices on its own: liability without good detection remains on paper, surveillance without a sanctionable cause of action produces flags without follow-up, and market design, by itself, does not distinguish those who coordinate from those who do not.
7. The North American experience
a) The market
In the United States there are seven organised electricity markets (PJM, MISO, ERCOT, CAISO, SPP, NYISO and ISO New England), run by regional operators under the supervision of the Federal Energy Regulatory Commission (FERC); the exception is Texas, whose market is largely outside FERC’s jurisdiction and supervised by the state commission. Two differences matter with respect to Italy. The price is formed by grid node, not by zone. Energy and reserves are purchased in the same market, with joint optimisation, and not in markets closed in sequence, so that the incentive to withhold capacity in order to profit in balancing is reduced as a result.[80] Canada offers a different model: Alberta has an ‘energy-only’ market, supervised by an autonomous surveillance authority, the Market Surveillance Administrator.
b) Preventive remedies
The first pillar is the independent monitor. After FERC’s Order 719 of 2008, every regional operator must have a monitor independent of its management, charged with identifying market design flaws, violations and abuses of market power, and reporting them to FERC.[81] The institutional structure varies, since some monitors are offices within the operator and others external companies, as in PJM; in every case they must be independent of the operator’s management, have access to all bidding data and may propose changes to market rules.
The second pillar is automatic mitigation. In several markets a conduct test compares each offer with a cost-based reference price and an impact test measures its effect on price; if both tests are triggered, the operator replaces the offer with the reference one.[82] California uses a structural test, which mitigates offers when three suppliers are jointly indispensable, and Order 831 of 2016 finally sets offer caps, which can be exceeded only with verified costs.[83] These instruments look to the result, not to intent: they therefore neutralise an excessive offer under conditions of market power whoever formulated it, person or algorithm, without any need to prove an agreement. They have, however, a cost, because they require cost-based reference prices for each plant, which monitors calculate and update through continuous work, and they may compress legitimately high prices in periods of scarcity; the US debate on the calibration of the three-pivotal-suppliers test shows how delicate the choice of thresholds is.
c) Sanctions and case law
After the Enron case, the Energy Policy Act of 2005 gave FERC a general power against manipulation. The precedent closest to the subject of this work concerns JP Morgan Ventures Energy: in 2013, on a referral from the California and MISO monitors, FERC found twelve manipulative bidding strategies, aimed at obtaining make-whole payments for plants usually out of the market, and the company paid $410 million between penalty and disgorgement of profits.[84] The logic is that of the Italian cases: exploiting the balancing remuneration rules. The case is not isolated. In 2017 FERC approved an $81.8 million settlement with GDF Suez Energy Marketing for bidding strategies formally compliant with market rules but held to be manipulative, confirming its determination to pursue so-called gaming of the rules.[85]
The US cause of action has, however, a structural limit. A federal court has stressed that fraud is a necessary element of the violation,[86] and an algorithm that learns to collude without deceiving anyone hardly falls within it. The Powhatan case, in turn, poses the basic question of this study. FERC admitted that the transactions complained of did not violate market rules, and the defence argued that the flaw lay in market design, which should be corrected rather than sanctioning the operator;[87] on 11 February 2020 the Fourth Circuit Court of Appeals rejected the statute-of-limitations objection, allowing FERC to proceed.[88]
In Canada, in the TransAlta case, the company had scheduled the outages of its coal plants in periods of high demand in order to push prices up. The authority held that direct proof of subjective intent was not necessary, it being possible to presume that the company intended the consequences of its actions, and TransAlta paid C$56 million.[89] It is a criterion of attribution transferable to algorithms. Alberta also offers an example of a conscious choice in the opposite sense: the Market Surveillance Administrator’s guidelines on offer behaviour did not prohibit unilateral economic withholding, tolerated as a form of remuneration of investments in a market without capacity payments, and in 2017, with the move to a capacity market announced, the authority opened a consultation to revoke them.[90] Tolerance of market power was thus a choice of dynamic efficiency, revised when the market design changed.
d) The limits of antitrust
In the United States the main antitrust instrument is the Sherman Act, section 1 of which prohibits agreements that restrain trade, with a function analogous to that of Article 101 TFEU. Unlike the European system, much of the enforcement of US antitrust law is entrusted to private actions: whoever has suffered harm from a cartel can sue and obtain compensation equal to three times the damage. In the wholesale electricity market, however, this route is almost precluded by a principle of case-law origin, the so-called filed rate doctrine. When a price has been filed with FERC, or is in any event subject to its control, ordinary courts cannot review it, nor find in a damages action that it was unlawfully high: the assessment belongs to FERC alone. The reason is to prevent courts from overlapping with the regulator and different customers from ending up, in fact, paying different prices for the same service.
Federal courts have applied this principle precisely to the events of the California electricity crisis of 2000–2001. In a case in which a public body that had bought energy at very high prices sought compensation from the seller, the Ninth Circuit Court of Appeals rejected the claim, observing that for claims of this kind the only competent forum is FERC;[91] the same court reaffirmed the principle in 2017, in an action against a JP Morgan group company concerning the California market.[92] It follows that in the US electricity sector protection against price manipulation passes almost exclusively through the regulator, and this is also why the FERC instruments described above have assumed such importance.
Outside the energy sector, US courts have begun to rule on pricing algorithms, and their decisions help in understanding where the line between lawful conduct and agreement runs. In a case in which customers accused the large Las Vegas hotels of having coordinated room prices by all using the same software, which suggested the rates to apply, the Ninth Circuit Court of Appeals ruled out, on 15 August 2025, the existence of an agreement: each hotel had adopted the software independently, the suggested prices were not binding and the program did not make one hotel’s confidential data available to the others.[93] It is, so far, the first decision of an appellate court on the subject. The outcome of the RealPage affair is different: RealPage is a company that supplies property owners with software for setting rents and which the Department of Justice accused of using, to formulate its recommendations, the non-public data supplied by competing owners: in November 2025 the parties proposed a settlement that would prohibit RealPage from using those data to set rents and would limit the use of the most recent data to train the model.[94] Read together, the two cases indicate that in the US experience the critical point is not the use of the same software by several competitors, but the circulation, through the software, of confidential information. It is a criterion transferable to the case of common suppliers of bidding algorithms discussed in section 6.
e) Pointers for the Italian market
The comparison between the European, and in particular Italian, experience and the North American one leads to a partly unexpected conclusion: faced with tacit collusion between algorithms, the two systems encounter similar limits. In both, the prohibition of agreements presupposes an agreement or a form of concertation between undertakings, which in these cases is hard to demonstrate. In the United States, moreover, the prohibition of energy market manipulation requires that the operator have deceived the market, whereas an algorithm that learns to keep prices high deceives no one; European REMIT does not require deception, but in turn encounters difficulties in demonstrating that a price is ‘artificial’ and in attributing to a single undertaking a result produced by the interaction of several algorithms (section 5).
The two systems differ instead in their preventive instruments. US markets have an automatic mitigation based on outcome: when an operator is indispensable and its offer exceeds by a significant margin a reference value calculated on costs, the offer is replaced with that value, with no need to ascertain why it was formulated or whether coordination lies behind it. Supervision is then entrusted to independent monitors, who have continuous access to all bidding data and may propose changes to market rules. In Italy there is an institution with a partly similar logic, the regime of plants essential for the security of the system, which constrains the offers of indispensable plants identified by Terna;[95] it applies, however, to a list of plants established in advance and does not intervene automatically on individual offers, while supervision remains split among several authorities. And this is probably the most useful lesson that the North American experience can offer the Italian market.
The comparison suggests that none of the instruments examined is sufficient on its own, but that several instruments, combined, can each compensate for the limits of the others. Without claiming completeness, and with the caution demanded by a phenomenon still little observed, some lines of intervention can be indicated, largely achievable with existing regulatory instruments. The first concerns information. REMIT already allows the national authority to ask those who engage in algorithmic trading for a description of their strategies and the controls adopted; it would be a matter of specifying its content, indicating the objectives the algorithm pursues, the information it observes (in particular whether it uses competitors’ published offers), the training data and period, the constraints imposed and possibly the results of the simulated-market tests recalled in section 6(a). In financial markets analogous obligations are already provided for by European rules on algorithmic trading,[96] and the cost for undertakings would be contained, while the benefit would be information that no authority possesses today. It would also constitute the contemporaneous document against which to verify in concrete terms the alternative explanation of adaptation to the market, according to the criterion followed by the case law (section 5).
The second line concerns surveillance. ARERA, in collaboration with the GME, could systematically apply to offers the statistical indicators of collusion already tested on the Campania and Brindisi cases, starting from the narrowest local markets; a flag would not suffice to found a sanction, but would allow the algorithm’s documentation to be requested and verification to be concentrated where the risk appears greatest. The third concerns proof. When a statistical flag concerns a narrow market and the documentation is missing, incomplete or reveals an algorithm that conditions its offers on competitors’ moves, it could be held that it is for the undertaking to demonstrate the economic reasons for its offers. REMIT already offers a hook in this sense, since it excludes manipulation when the author shows that it acted for legitimate reasons; such a rule would avoid having to prove an agreement that does not exist and would place the burden of proof on whoever possesses the necessary information. It would be, moreover, a specific application of the criterion, already followed by administrative case law, whereby alternative explanations for concertation lose credibility when they find no support in contemporaneous documents. Part of the administrative case law goes further, admitting that, once anomalous parallelism is demonstrated, it is for the undertakings to provide a plausible alternative explanation (section 5); the solution proposed here, which makes the shifting of the burden conditional on the absence or content of the documentation, appears more compatible with the presumption of innocence. The fourth, finally, takes up the North American experience: in areas where few operators are indispensable, balancing offers that exceed by a significant margin a cost-based reference value could be replaced with that value, extending the logic that already inspires the regime of essential plants. An instrument of this kind operates on the result and does not require establishing who, or what, produced it, and for this reason it is particularly suited to a phenomenon in which attribution is difficult.
These lines lend themselves to a number of objections, which deserve an answer. The first concerns trade secrecy, since the description of algorithms touches the technological assets of undertakings. The risk nevertheless appears contained: REMIT already imposes professional secrecy on the authorities that receive such information (Article 17), and the documentation would be delivered only on request, without any form of publication. The second objection concerns costs for smaller operators, such as the aggregators of small resources that the TIDE sought to attract into balancing; the obligations could be graduated according to the capacity offered, just as the notifications under Article 5a already apply only to those who engage in algorithmic trading and not to every market participant. The third objection is the most serious, and concerns the risk that shifting the burden of proof turns a statistical flag into a presumption of guilt. For this reason the hypothesis put forward here limits it to narrow markets and makes it conditional on a second element, missing or revealing documentation; it would in any case be a rebuttable presumption, which the undertaking can overcome by demonstrating the economic reasons for its offers.
The same lines could, finally, guide undertakings’ compliance programmes. Administrative case law recognises a mitigating value in a compliance programme only if it is effective, that is, supported by concrete management commitment, an adequate risk assessment and monitoring and audit systems;[97] for an undertaking that bids through algorithms, an effective programme should extend to the design, training and control of the algorithms themselves.
An institutional condition remains. The lines of intervention indicated involve ARERA, the GME, Terna and the AGCM, that is, precisely the authorities among which tacit collusion between algorithms risks being dispersed today. A dedicated cooperation protocol, with systematic exchange of flags and a common forum for data analysis, could prevent each from waiting for the others to take the initiative. This too is a lesson from the North American monitors: the effectiveness of supervision seems to depend less on formal powers than on continuous access to data and on the competence of those who analyse them.
8. Conclusions
The 2025–2026 reforms have made the Italian electricity market more open: more operators in balancing, zonal prices for demand too, closer integration with the European market, auctions on fifteen-minute intervals. In this way they may have attenuated the structural conditions from which the Campania and Brindisi cases arose. They do not, however, seem to have considered the risk that emerges with the spread of bidding algorithms, that of a tacit collusion that occurs without any agreement between undertakings.
Current law does not appear fully equipped to address it. The prohibition of agreements may apply to algorithms that exchange signals, but for those that learn autonomously concertation is very difficult to prove, because parallel bidding almost always admits the alternative explanation of adaptation to the market; the prohibition of abuse of a dominant position requires a dominant undertaking and an abuse which, in the electricity sector, may take the more easily observable form of capacity withholding; REMIT offers the basis closest to the problem, but reasons in terms of individual conduct and regulates algorithmic trading above all from the standpoint of the orderly conduct of trading. The North American experience shows that not even broad powers against manipulation suffice to fill the gap when the cause of action is built on deception, and suggests that the most effective instrument may be an automatic mitigation based on outcome.
The problem therefore appears, more than one of a gap in the legislation, to be one of proof: the existing rules can operate if the authorities have the information needed to verify in concrete terms the alternative explanation offered by undertakings. The economic analysis of law invites us, rather than seeking a single answer, to combine different instruments, each of which compensates for the limits of the others. In this perspective the characteristics of the Italian market take on a double significance: offers referring to single plants, narrow zones and the publication of data could expose it to the problem before other European markets, but those same data also offer the tools to address it.
At least three directions of research remain open. The first is empirical, and consists in applying to the GME’s post-reform bidding data the detection tools already tested on the Campania and Brindisi cases. The second is comparative, and concerns comparison with the Spanish market, which shares central dispatching with the Italian one. The third is institutional: it would be necessary to measure the degree of concentration of the market for bidding software suppliers, on which depends whether the risk resembles more the spontaneous convergence between algorithms or coordination through a common supplier.
Notes
[1]ACER, REMIT Quarterly, No 44, first quarter 2026, p. 10, which in turn cites an analysis published by the Financial Times; https://www.acer.europa.eu/sites/default/files/REMIT/REMIT%20Reports%20and%20Recommendations/REMIT%20Quarterly/ACER-REMIT-Quarterly-Q1-2026.pdf.
[2]J. Proz, M. Huber, I. Marin, Machine learning for detecting collusion and capacity withholding in wholesale electricity markets, in Energy Economics, 2026, vol. 162, art. 109557, https://doi.org/10.1016/j.eneco.2026.109557.
[3]Autorità per l’energia elettrica e il gas (Authority for Electricity and Gas), Resolution of 15 July 2004.
[4]ARERA, Resolution 300/2017/R/eel.
[5]J. Proz, M. Huber, I. Marin, op. cit., 168.
[6]Document published on energypress.gr on 11 February 2022, https://energypress.gr/sites/default/files/media/i-316303-11.02.22.pdf; see also Universidad Pontificia Comillas, Integrating the central-dispatch markets into the European balancing platforms: main lessons learnt and challenges, https://repositorio.comillas.edu/items/edcac903-6d22-4fd2-bc5c-2cd72c058c7a.
[7]GME, press release of 19 May 2009, https://gme.mercatoelettrico.org/Portals/0/Documents/en-US/20090519CS_EN.pdf.
[8]J. Proz, M. Huber, I. Marin, op. cit., which cite on this point OECD, Data screening tools for competition investigations, 2022. On the relationship between auction format and the sustainability of tacit collusion see N. Fabra, Tacit Collusion in Repeated Auctions: Uniform Versus Discriminatory, in Journal of Industrial Economics, 2003, vol. 51, No 3, p. 271 ff.
[9]AGCM, decision No 23623 of 2012, case I736, Repower Italia – Prezzo dispacciamento energia elettrica Centro Sud.
[10]AGCM, decision No 26562 of 2017, case A498A, Enel – Prezzi servizi di dispacciamento area Brindisi, and decision No 26563 of 2017, case A498B, Sorgenia – Prezzi servizi di dispacciamento area Brindisi.
[11]J. Proz, M. Huber, I. Marin, op. cit.
[12]Ministerial Decree of 18 April 2024; ARERA, consultation document 194/2024/R/eel, https://arera.it/schede-tecniche/dettaglio/superamento-del-prezzo-unico-nazionale-e-relativa-componente-perequativa-orientamenti-dellautorita.
[13]ARERA, Resolution 8/2026/R/eel, https://www.arera.it/fileadmin/allegati/docs/26/8-2026-R-eel.pdf.
[14]The Balance Responsible Party (BRP) is the party responsible for imbalances, that is, for the difference between the energy that an operator has committed to inject or withdraw on the basis of market trades and the energy actually injected or withdrawn. The figure derives from Union law: Article 5 of Regulation (EU) 2019/943 provides that every market participant is responsible for the imbalances it causes in the system, either assuming directly the role of BRP or contractually delegating that responsibility to a BRP of its choice, and Regulation (EU) 2017/2195 completes the rules. The BRP communicates to Terna the schedules of its portfolio of production or consumption units, is the holder of the dispatching contract and pays, or receives, the imbalance charges, calculated on fifteen-minute intervals. It is to be distinguished from the Balancing Service Provider (BSP), which supplies balancing services to Terna. Until 31 December 2024 the two roles were combined in the ‘dispatching user’ (utente del dispacciamento); the TIDE allows their separation from the consolidation phase, 1 February 2026, with exceptions, among them that concerning essential plants under the cost-reimbursement regime.
[15]APG, press release of 30 September 2025 on the launch of the fifteen-minute trading unit in day-ahead market coupling, https://markt.apg.at/en/news-press/sdac-go-live-of-15-minute-market-time-unit-in-day-ahead-market-coupling-successful-as-of-september-30-2025/.
[16]Regulation (EU) No 1227/2011 of the European Parliament and of the Council of 25 October 2011 on wholesale energy market integrity and transparency, as amended by Regulation (EU) 2024/1106 of the European Parliament and of the Council of 11 April 2024.
[17]ACER, REMIT Quarterly, No 44, cited above, p. 12; ARERA, communication to operators on notifications pursuant to Article 5a of Regulation (EU) 2024/1106, https://www.arera.it/comunicati-operatore/dettaglio/notifiche-ad-arera-e-acer-ai-sensi-dellarticolo-5-bis-del-regolamento-ue-20241106-del-parlamento-europeo-e-del-consiglio-dell11-aprile-2024-che-modifica-i-regolamenti-ue-n-12272011-e-ue-2019942.
[18]ACER, REMIT Quarterly, No 44, cited above, p. 12.
[19]N. Fabra, J. Toro, Price Wars and Collusion in the Spanish Electricity Market, in International Journal of Industrial Organization, 2005, vol. 23, No 3-4, p. 155 ff.
[20]P. Manzini, Algoritmi collusivi e diritto antitrust europeo [Collusive algorithms and European antitrust law], in Mercato Concorrenza Regole, 2019, No 1, p. 163 ff., esp. p. 165, doi 10.1434/94242.
[21]E. Calvano, G. Calzolari, V. Denicolò, S. Pastorello, Artificial Intelligence, Algorithmic Pricing, and Collusion, in American Economic Review, 2020, vol. 110, No 10, p. 3267 ff.
[22]T. Klein, Autonomous Algorithmic Collusion: Q-Learning under Sequential Pricing, in RAND Journal of Economics, 2021, vol. 52, No 3, p. 538 ff.
[23]S. Assad, R. Clark, D. Ershov, L. Xu, Algorithmic Pricing and Competition: Empirical Evidence from the German Retail Gasoline Market, in Journal of Political Economy, 2024, vol. 132, No 3, p. 723 ff.
[24]A. Ezrachi, M.E. Stucke, Virtual Competition. The Promise and Perils of the Algorithm-Driven Economy, Cambridge (Mass.)–London, Harvard University Press, 2016, in particular the scenario that the authors call ‘Digital Eye’, in which algorithms capable of learning reach a collusive equilibrium on their own; J.E. Harrington Jr., Developing Competition Law for Collusion by Autonomous Artificial Agents, in Journal of Competition Law & Economics, 2018, vol. 14, No 3, p. 331 ff., according to whom collusion carried out by programs that choose pricing rules without human intervention does not violate section 1 of the Sherman Act.
[25]E. Guerci, S. Ivaldi, Learning Agents in an Artificial Power Exchange: Tacit Collusion, Market Power and Efficiency of Two Double-Auction Mechanisms, in Computational Economics, 2008, vol. 32, https://doi.org/10.1007/s10614-008-9127-5
[26]I. Abada, X. Lambin, Artificial Intelligence: Can Seemingly Collusive Outcomes Be Avoided?, in Management Science, 2023, vol. 69, No 9.
[27]J. Seredyński, G. Tsaousoglou, AI agents in Algorithmic Electricity Markets: On the Emergence of Tacit Collusion, preprint, arXiv:2608.26896, 2026, https://arxiv.org/abs/2608.26896.
[28]D.P. Brown, D.O. Cajueiro, A. Eckert, D. Silveira, Evaluating the role of information disclosure on bidding behavior in wholesale electricity markets, in Energy Economics, 2025, vol. 146, art. 108505.
[29]V. Caforio, Il nuovo oligopolio [The new oligopoly], Università Bocconi, 23 February 2026, https://www.unibocconi.it/it/news/il-nuovo-oligopolio; P. Manzini, op. cit., pp. 165–166; A. Ezrachi, M.E. Stucke, op. cit.
[30]A recent episode, unrelated to energy markets, shows how far coordination can go between artificial intelligence agents not instructed to coordinate. In July 2026, during an internal evaluation of the cyber capabilities of its models, some 1,200 OpenAI agents communicated through an unauthorised bulletin board, organising collective projects to circumvent the evaluation system, and some 700 took part in an attack on Hugging Face’s infrastructure without any human having ordered it; according to the company, the behaviours at the origin of the episode had been reinforced during training. See METR, Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident, 26 August OpenAI, OpenAI – Hugging Face Incident Technical Report, 26 August 2026.
[31]Consiglio di Stato, Section VI, 19 December 2019, No 8584, para 3.1; Court of Justice, 16 December 1975, Joined Cases 40/73 and others, Suiker Unie and Others v Commission, paras 173–174; Court of Justice, 8 July 1999, C-49/92 P, Commission v Anic Partecipazioni.
[32]D.F. Turner, The Definition of Agreement under the Sherman Act: Conscious Parallelism and Refusals to Deal, in Harvard Law Review, 1962, vol. 75, p. 655 ff.; R.A. Posner, Oligopoly and the Antitrust Laws: A Suggested Approach, in Stanford Law Review, 1969, vol. 21, p. 1562 ff.; L. Kaplow, Competition Policy and Price Fixing, Princeton, Princeton University Press, 2013; M.S. Gal, Algorithms as Illegal Agreements, in Berkeley Technology Law Journal, 2019, vol. 34, p. 67 ff. In the United States the Supreme Court has held that tacit collusion is not in itself unlawful: Brooke Group Ltd. v. Brown & Williamson Tobacco Corp., 509 U.S. 209, 227 (1993).
[33]V. Caforio, op. cit.; P. Manzini, op. cit., p. 170, according to whom signalling algorithms induce coordination “even in markets characterised by the presence of numerous undertakings”.
[34]Court of Justice, 21 January 2016, C-74/14, Eturas and Others.
[35]P. Manzini, op. cit., pp. 168–169.
[36]European Commission, Guidelines on the applicability of Article 101 of the Treaty on the Functioning of the European Union to horizontal co-operation agreements, OJ C 259, 21 July 2023, para 379; see also Algorithmic collusion: the next generation of price fixing?, in Financier Worldwide, https://financierworldwide.com/algorithmic-collusion-the-next-generation-of-price-fixing.
[37]Freshfields, EU steps up on algorithmic pricing cartels, 2025, https://www.freshfields.com/en/our-thinking/blogs/risk-and-compliance/eu-steps-up-on-algorithmic-pricing-cartels-joining-the-us-and-other-jurisdiction-102kt76.
[38]Court of Justice, 4 June 2009, C-8/08, T-Mobile Netherlands and Others.
[39]European Commission, decision of 7 July 2016, case AT.39850, Container Shipping.
[40]P. Manzini, op. cit., pp. 171–172.
[41]Court of Justice, 22 October 2015, C-194/14 P, AC-Treuhand v Commission, para 31; Court of Justice, 28 June 2005, Joined Cases C-189/02 P and others, Dansk Rørindustri and Others v Commission, paras 142–143; see P. Manzini, op. cit., p. 172.
[42]P. Manzini, op. cit., p. 171, which cites to similar effect R.A. Posner, Review of Kaplow, Competition Policy and Price Fixing, in Antitrust Law Journal, 2013–2014, vol. 79, p. 761 ff.
[43]Court of Justice, Ahlström Osakeyhtiö and Others v Commission, cited above.
[44]European Commission, Guidelines on the applicability of Article 101 of the Treaty on the Functioning of the European Union to horizontal co-operation agreements, cited above, paras 368, 396 and 400.
[45]TAR Lazio, Rome, Section I, judgment, appeal No 14337/2025, K.P.I. S.p.A. v AGCM (public hearing of 15 April 2026), dismissing the appeal against AGCM, decision of 23 September 2025, No 31673, case I864, which cites the horizontal guidelines, paras 397 and 410, and Court of First Instance, 12 July 2001, Joined Cases T-202/98 and others, Tate & Lyle and Others v Commission.
[46]P. Manzini, op. cit., p. 173.
[47]Court of Justice, 31 March 1993, Joined Cases C-89/85, C-104/85, C-114/85, C-116/85, C-117/85 and C-125/85 to C-129/85, Ahlström Osakeyhtiö and Others v Commission (‘Wood Pulp II’), ECLI:EU:C:1993:120, in ECR, 1993, p. I-1307, para 71; to the same effect Court of First Instance, 29 June 1995, T-36/91, ICI v Commission, in ECR, 1995, p. II-1847; General Court, 12 April 2013, T-442/08, CISAC v Commission, which distinguishes the case in which the authority relies on parallelism alone, in which it suffices for the undertakings to show circumstances that make an alternative explanation plausible, from that in which it has evidence of concertation; General Court, T-1026/23, Laudamotion v Commission, ECLI:EU:T:2025:928; in Italy Consiglio di Stato, Section VI, 19 December 2019, No 8584, cited above, para 3.2, and Consiglio di Stato, Section VI, 6 October 2020, No 5885. More recently the AGCM has reiterated that mere price parallelism, even if accompanied by closely timed announcements of increases, does not suffice to demonstrate an agreement in the absence of contacts between the parties: see Freshfields, Newsletter giuridica di concorrenza e regolamentazione [Competition and regulation legal newsletter], 19 January 2026, https://www.freshfields.com/en/our-thinking/newsletters/2026/01/newsletter-giuridica-di-concorrenza-e-regolamentazione—19-gennaio-2026
[48]Consiglio di Stato, No 8584/2019, cited above, para 7, which cites Consiglio di Stato, Section VI, 12 October 2017, No 4733, and, on the lesser probative value of documents drawn up after the events, General Court, 15 July 2015, T-423/10.
[49]Consiglio di Stato, Section VI, 9 May 2022, No 3570; Consiglio di Stato, Section VI, 20 January 2023, No 688; Consiglio di Stato, Section VI, 25 July 2023, No 7273.
[50]TAR Lazio, Rome, Section I, judgment, appeal No 14337/2025, K.P.I. S.p.A. v AGCM (public hearing of 15 April 2026), dismissing the appeal against AGCM, decision of 23 September 2025, No 31673, case I864, which also cites the presumption of innocence under Article 48 of the Charter of Fundamental Rights of the European Union and Article 6(2) ECHR, holding it compatible with rebuttable presumptions based on rules of experience.
[51]Court of First Instance, 6 June 2002, T-342/99, Airtours v Commission, para 62; on the notion of collective dominant position see also Court of Justice, 16 March 2000, Joined Cases C-395/96 P and C-396/96 P, Compagnie Maritime Belge Transports and Others v Commission.
[52]Court of Justice, 14 February 1978, 27/76, United Brands v Commission.
[53]P. Manzini, op. cit., pp. 173–176; Court of Justice, 10 July 2008, C-413/06 P, Bertelsmann and Sony Corporation of America v Impala, paras 119–123; Court of First Instance, 26 January 2005, T-193/02, Piau v Commission, para 111.
[54]European Commission, decision of 26 November 2008, cases COMP/39.388 and COMP/39.389, E.ON; see T. Duso, F. Szücs, V. Böckers, Abuse of dominance and antitrust enforcement in the German electricity market, in Energy Economics, 2020, vol. 92, art. 104936.
[55]V. Caforio, op. cit.; the Airtours judgment, cited above, was delivered precisely in a merger case.
[56]ACER, Guidance on the application of Regulation (EU) No 1227/2011, 6th ed., 2024.
[57]Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence.
[58]Resolution of the European Parliament of 16 February 2017 with recommendations to the Commission on civil law rules on robotics (2015/2103(INL)), point 12; see P. Manzini, op. cit., p. 178, note 34.
[59]G. Calabresi, The Costs of Accidents. A Legal and Economic Analysis, New Haven, Yale University Press, 1970.
[60]S. Shavell, Liability for Harm versus Regulation of Safety, in Journal of Legal Studies, 1984, vol. 13, No 2, p. 357 ff.
[61]L. Kaplow, Rules versus Standards: An Economic Analysis, in Duke Law Journal, 1992, vol. 42, No 3, p. 557 ff.
[62]V. Caforio, Il nuovo oligopolio, cited above.
[63]P. Manzini, op. cit., pp. 176–178.
[64]P. Manzini, op. cit., pp. 178–179, which cites Court of First Instance, 15 September 2005, T-325/01, DaimlerChrysler v Commission, paras 86–88, and Court of Justice, 21 July 2016, C-542/14, VM Remonts and Others, paras 23–24.
[65]V. Caforio, Algorithmic Tacit Collusion: A Regulatory Approach, 2023, https://boa.unimib.it/handle/10281/464538 who by contrast rules out modifying the notions of agreement and concerted practice so as to include algorithmic tacit collusion within them.
[66]B. Holmström, Moral Hazard in Teams, in Bell Journal of Economics, 1982, vol. 13, No 2, p. 324 ff.
[67]J.E. Harrington Jr., Developing Competition Law for Collusion by Autonomous Artificial Agents, cited above.
[68]D.P. Brown, D.O. Cajueiro, A. Eckert, D. Silveira, op. cit.
[69]J. Proz, M. Huber, I. Marin, op. cit.
[70]G.S. Becker, Crime and Punishment: An Economic Approach, in Journal of Political Economy, 1968, vol. 76, No 2, p. 169 ff.
[71]Consiglio di Stato, No 8584/2019, cited above, paras 5 and 12, which cite Consiglio di Stato, 20 February 2017, No 740, and Consiglio di Stato, 18 May 2015, No 2514.
[72]TAR Lazio, Section I, appeal No 14337/2025, cited above.
[73]ACER, REMIT Quarterly, No 44, cited above, p. 12.
[74]J.E. Harrington Jr., The Effect of Outsourcing Pricing Algorithms on Market Competition, in Management Science, 2022, vol. 68, No 9.
[75]P. Manzini, op. cit., p. 170, note 15.
[76]Reg. (EU) 2024/1689, Article 3, points 3 and 4, which define respectively the provider and the deployer.
[77]Court of Justice, 21 July 2016, C-542/14, VM Remonts and Others.
[78]Court of Justice, 22 October 2015, C-194/14 P, AC-Treuhand v Commission.
[79]P. Manzini, op. cit., pp. 179–181, which also cites General Court, 10 November 2017, T-180/15, Icap and Others v Commission, para 180.
[80]J. Proz, M. Huber, I. Marin, op. cit.
[81]FERC, Order No. 719, Wholesale Competition in Regions with Organized Electric Markets, 2008; CAISO, presentation on market monitoring, November 2018, https://www.caiso.com/Documents/Presentation-EIMRegionalIssuesForum-Nov282018.pdf.
[82]U.S. Department of Energy, Review of Market Monitoring Activities at U.S. Independent System Operators, https://www.energy.gov/oe/articles/review-market-monitoring-activities-us-independent-system-operators.
[83]FERC, Order No. 831, Offer Caps in Markets Operated by Regional Transmission Organizations and Independent System Operators, 2016.
[84]FERC, press release of 30 July 2013 on the JP Morgan Ventures Energy case, https://mondovisione.com/media-and-resources/news/federal-energy-regulatory-commission-jp-morgan-unit-agree-to-410-million-in-pe-2013730/.
[85]Akin Gump, Big Settlement Shows FERC’s Pursuit of Market Gaming, 2017, https://akingump.com/a/web/55777/Big-Settlement-Shows-FERC-s-Pursuit-Of-Market-Gaming.pdf.
[86]FERC v. Silkman (D. Mass.); see National Law Review, https://natlawreview.com/node/61788/printable/print.
[87]Analysis: FERC moves against company for energy market manipulation, in Renewable Energy World, https://www.renewableenergyworld.com/energy-business/analysis-ferc-moves-against-company-for-energy-market-manipulation/.
[88]FERC v. Powhatan Energy Fund, LLC (4th Cir., 11 February 2020); see Greenberg Traurig, Fourth Circuit Rejects Statute of Limitations Challenge to FERC Electricity Market Manipulation Suit, February 2020, https://gtlaw.com/it/insights/2020/2/fourth-circuit-rejects-statute-of-limitations-challenge-to-ferc-electricity-market-manipulation-suit.
[89]Alberta Utilities Commission, decision 3110-D01-2015; see Market Manipulation in Alberta: TransAlta Pays $56 Million, in Energy Regulation Quarterly, https://energyregulationquarterly.ca/case-comments/market-manipulation-in-alberta-transalta-pays-56-million.
[90]N. Bankes, The End of Economic Withholding in Alberta’s Electricity Market?, in ABlawg, 30 March 2017, https://ablawg.ca/2017/03/30/end-of-economic-withholding-in-albertas-electricity-market/.
[91]Public Utility District No. 1 of Grays Harbor County v. IDACORP, 379 F.3d 641 (9th Cir. 2004).
[92]U.S. Court of Appeals for the Ninth Circuit, memorandum of 12 May 2017, No. 15-56697, https://cdn.ca9.uscourts.gov/datastore/memoranda/2017/05/12/15-56697.pdf.
[93]Gibson v. Cendyn Group (9th Cir., 15 August 2025); see Baker McKenzie, United States: Ninth Circuit on Algorithmic Pricing, August 2025, https://www.bakermckenzie.com/en/insight/publications/alerts/2025/08/united-states-ninth-circuit-on-algorithmic-pricing.
[94]United States v. RealPage, Inc.; see Morrison Foerster, (Still) All About Algorithms: Antitrust Lessons from the Last Year and What Lies Ahead in 2026, January 2026, https://www.mofo.com/resources/insights/260126-lessons-questions-antitrust-scrutiny-algorithms.
[95]ARERA, Resolution No 111/06 as subsequently amended.
[96]Commission Delegated Regulation (EU) 2017/589 of 19 July 2016 supplementing Directive 2014/65/EU of the European Parliament and of the Council with regard to regulatory technical standards specifying the organisational requirements of investment firms engaged in algorithmic trading.
[97]Consiglio di Stato, 21 December 2017, Nos 5997 and 5998, cited in Consiglio di Stato, No 8584/2019, cited above, para 14.5.
