1 Introduction↩︎

Every day, a day-ahead market operator decides how much energy has to be delivered for each of the twenty-four hours of the next day. The amount that will be needed depends on what households, factories, hospitals and shops will eventually draw from the grid, and nobody knows these numbers exactly in advance; for each hour tomorrow the actual total demand is random. The operator asks the consumers of the day-ahead market to report their demand, and schedules producers for tomorrow to match the total of all reports [1], [2]. The next day, because at any given time the energy supplied to the grid should balance the energy demanded, a real-time (balancing) market corrects, at delivery time, any deviation between the scheduled generation and the realized demand. In the context of this work, we use the term energy market to refer specifically to the sequential coupling of the day-ahead and real-time markets.

Although scheduling generation to match demand is the most sensible thing to do subject to the grid balance constraint, this arrangement relies on the assumption that the day-ahead market participants report their demands truthfully. The operator cannot independently measure how much each consumer believes to need and has to trust their reports. The same trust shows up in many other institutions. Online advertising exchanges ask advertisers what an impression is worth to them, and then allocate it accordingly [3]. Multi-tier supply chains ask each tier to forecast demand so that the next tier can plan [4]. Insurance underwriters ask applicants for risk information and use it to price the premium [5].

Our work was motivated by striking empirical observations. 1 is based on public market data fetched from [6] and reports a persistent gap between the aggregate demand reported by participants in the day-ahead market and the total demand actually realised the next day for three European countries on three different years.

For each country we pick a recent year and show two views. The top chart in each column gives, hour by hour, how often the volume sold in the day-ahead market fell 10% below the electricity people actually used that hour. The bottom chart shows, again hour by hour, how large that shortfall was on average – in megawatts on the left axis and as a share of the mean (actual) demand on the right axis. Blue bars mark hours where the market scheduled less than what was needed; dark-green bars mark hours where it scheduled more. The same pattern shows up everywhere: the gap is largest overnight and in the evening, and almost disappears around midday.

This gap cannot be attributed to prediction error: it is one-sided, it persists across long observation windows, and survives changes in market conditions and weather patterns. We argue that this gap reflects a strategic bias in the demand reports — there exist energy market participants that systematically under-report their demand whenever is profitable under the existing settlement rules.

The economic consequence is direct. The day-ahead market collects reports from many self-interested players to recover the latent true distribution of next-day demand and shapes the very data it is trying to learn from. If participants do not report their demand honestly, the biased day-ahead market drives the real-time market towards a non-zero expected imbalance, and the real-time market clears more than just the irreducible random fluctuations in actual consumption, causing the market outcome to drift away from its competitive equilibrium, thus reducing social welfare.

That said, and given that in both the U.S. and the European energy markets, the settlement rules are an evolving design objective (see 1.2), the central question of this paper – taking the demand reports as the strategic component – is the following:

Can one design a settlement layer so that the institution estimates the latent demand belief and hence recovers the competitive equilibrium of the underlying economy?

1.1 Our Contributions.↩︎

To answer this question, we start in 2 by first providing a unified institutional and formal model for the two-stage energy markets. This model is the common ground where all the subsequent results rest. Then, in 3 we identify and study the implementable incentives that explain the empirically observed strategic bias. While respecting institutional and informational constraints, we design in 4, a contrastive scoring-rule-based mechanism implemented by the day-ahead market operator and added to all consumer’s bill. 5 demonstrates our findings and evaluates our mechanism design via numerical simulations on synthetic and on real market data.

Our theoretical and empirical contributions are as follows:

Theoretical Contributions

1. Vanishing incentives to underreport for consumers (1). For any consumer who does not own real-time generation, the maximum benefit from under-reporting her demand shrinks toward a non-positive value as the size \(|I|\) of demand side of the day-ahead market grows, at a rate proportional to \(1/\sqrt {|I|}\). Honesty is therefore an approximate best response for consumers in large day-ahead markets.

2. Non-vanishing incentives to underreport for prosumers (2). For any prosumer — a consumer who owns a strictly positive share of real-time generation assets — the benefit from under-reporting is bounded below by a strictly positive constant that depends only on the size of her portfolio and the shape of the real-time supply stack. This constant does not depend on \(|I|\). Therefore, under the standard institution not everyone can be made approximately honest simultaneously, no matter how large the day-ahead market grows.

3. Scoring-rule restoration of incentive compatibility (3, 1 and 5). While respecting the existing institutional constraints, we modify the imbalance settlement by designing a contrastive scoring-rule-based payment that the day-ahead operator can compute from observable quantities (the reports, the realised total consumption, and historical data) and add to each participant’s bill. The extra payment is built from a strictly proper scoring rule and has a single tuning knob, the strong convexity parameter \(\mu\) of the associated strongly convex function. We prove that, by choosing \(\mu\) large enough, the prosumer’s benefit from lying can be driven strictly below the persistent constant of (2). Honesty is therefore restored for everyone.

   Turning \(\mu\) up deters lying more strongly, but it also charges honest participation more heavily when their true demand happens to differ from the population average. 5 in ¿sec:trade-off? characterises this trade-off by providing the value \(\mu^\star\) that balances the two effects and prove that, in a large market, both the residual incentive to lie and the residual cost paid by honest participants vanish at the same rate.

Empirical Contributions

1. Simulations on synthetic market data (5.1). We illustrate the three qualitative phenomena presented by our theoretical results in a constructed synthetic market that focuses only on the incentive question. The day-ahead and real-time supply stacks are fixed and every player holds a uniform belief over private demand. We focus on a single player with fixed true demand across market sizes so that the comparisons have a common baseline. In every experiment we evaluate the expected gain from under-reporting, with the expectation taken by Monte Carlo. The four experiments differ only in which mechanism is in force and which structural parameters are varied.

2. Evaluation on real market data (5.2) We demonstrate the performance of our mechanism by applying it to the strategic component that we infer from real market data. We work with hourly time series of low, medium and high voltage reported loads together with the corresponding realized demand. We rationalize the day-ahead market reports as subjective equilibria by means of the prosumers parameters and real-time generation capacity bounds that make the observed day-ahead reports approximately optimal. Having inferred the model’s parameters, we evaluate our scoring rule-based mechanism and demonstrate its effects.

1.2 Related Work↩︎

This paper contributes to several literature domains. We organize our discussion of the related work around three themes (i) the design of wholesale electricity markets and the institutional treatment of imbalance settlement, (ii) mechanism design, subjective beliefs and scoring rules, (iii) strategic machine learning.

Wholesale Electricity Market Design and Imbalance Settlement.

The design of competitive wholesale electricity markets has been a subject of study since deregulation in the 1990s. [1] is the first textbook treatment of the two-stage structure — day-ahead scheduling followed by real-time balancing (what we previously called the energy market)— that is the institutional setting of the present paper.

In Europe, the Third Energy Package (TEP) [7] and the most recent Clean Energy Package (CEP) [8] are each, a set of laws passed in 2009 and 2019, respectively. The former established the framework for the internal electricity market and the most recent one changed the rules for two things: balancing (the real-time process ensuring that supply equals demand) and imbalance settlement (the rule for charging participants’ deviations from schedule). In particular, Directive 2019/944 [9] from the CEP officially recognized two kinds of players that hadn’t really existed in EU electricity law before – active customers and aggregators.

1. An active customer is an energy market participant that isn’t just passively buying electricity from the grid. They can produce and sell their own electricity. Directive 2019/944 defines them in Article 2(8) and gives them concrete rights in Article 15 (e.g. the right to sell their excess electricity).

2. An aggregator is an energy market participant that bundles together lots of small players and offers their combined flexibility to the electricity market as if it were one big power plant. The Directive 2019/944 defines aggregators in Article 2(18) and (19) and sets the rules for how they operate in Article 17.

Article 5(1) of Regulation (EU) 2019/943 [10] establishes that all market participants are responsible for the imbalances they cause in the system. Since aggregators and active customers qualify as market participants under Article 2(25) of the same Regulation, they are subject to this obligation.

Although both active customers and aggregators qualify as market participants under Regulation (EU) 2019/943, only aggregators are realistically positioned to engage in cross-market exchange (e.g., buying energy in the day-ahead market and selling it closer to real time, via the balancing market – see Regulation (EU) 2019/943 [10]). Active customers may benefit indirectly from such strategies by participating through an aggregator (Directive (EU) 2019/944, Article 15(2)(a)).

That said, a central energy market entity of study in this paper is the market-active prosumer, or simply prosumer: an entity that both consumes and produces electricity, qualifies as a market participant under Article 2(25) of Regulation (EU) 2019/943, and engages in cross-market exchange between the day-ahead and balancing markets. The prosumer thus sits at the intersection of the active customer (Directive (EU) 2019/944) and the aggregator function described in the same Directive, and is subject to the balance responsibility imposed by Article 5(1) of Regulation (EU) 2019/943. The formal definition we adopt for the player’s economic role in our model is given in 2.1.

The EU regulatory framework signals that imbalance settlement is an active design objective. Regulation (EU) 2019/943 establishes the general principle of balance responsibility. On top of that, recital 12 explicitly characterizes imbalance pricing as a behavioral incentive mechanism, stating that the pricing method for balancing energy "should create positive incentives for market participants in keeping their own balance or helping to restore the system balance [...], thereby reducing system imbalances and costs to society," and that such pricing approaches "should strive for the economically efficient use of demand response and other balancing resources." This framing is reinforced by the Electricity Balancing Guideline [11], to which recital 12 directly refers (Articles 18, 30 and 32). Taken together, these provisions reveal that the EU treats the design of imbalance settlement as an open, behaviorally consequential lever for integrating new market actors such as prosumers, rather than as a fixed technical detail. This paper contributes directly to this open design problem.

Mechanism Design, Subjective Beliefs and Scoring Rules.

Mechanism design in its central insight, going back to Vickrey [12], Myerson [13], and a long line since, holds that honesty cannot be assumed – it has to be engineered. For mechanisms with money like energy markets, this often means designing a payment rule so that, for every participant, telling the truth is the most profitable thing to do. If lying is unprofitable, self-interest does the rest.

In our setting, participants report their demand for the next day. These reports can, in principle, be aggregated to reflect each player’s subjective belief about their consumption needs the following day. De Finetti [14] was the first who defined subjective belief axiomatically through betting: a person’s subjective belief about an event is the price at which this person would be indifferent to sell or buy a $1-contract written on that event. De Finetti’s coherence argument shows that consistent beliefs are precisely those that admit no sure-loss bet. However, the question of how to elicit a person’s belief is not answered in his work.

[15] independently, working in meteorology, likely without thinking about subjective probability foundations but rather about how to verify weather forecasters, was the first who proposed the quadratic scoring rule as an empirical tool. Two years later [16] introduced the logarithmic scoring rule. De Finetti [17] observed that Brier’s and Good’s scores had the strict propriety property of measuring – in the sense of eliciting – (true) subjective beliefs. This was a completely different insight from the question of forecasting skill: a strictly proper scoring rule controls the belief that a person communicates given what they know, and not the quality of what they know, which depends on their ability in information acquisition. Essentially, according to de Finetti, the scoring rule acts as an information-elicitation tool that filters out non-rational instincts that usually distort human judgment. By tying real-world consequences, the score aligns a person’s self-interest with honesty.

This insight was formalized into a general theory. [18] and [19] characterized the class of strictly proper scoring rules and established their equivalence to strongly convex functions defined over distributions, while [20] and [21] extended the framework to continuous and multivariate forecasts and provided the modern synthesis on which most applications rely.

Although in principle individual demand reports can be mapped to the player’s subjective belief about their consumption needs, scoring these beliefs requires knowing their actual consumption levels. Such information can be obtained only via metering which introduces measurement errors and it is often of questionable reliability. That said, any settlement modification is intended to be implemented by the day-ahead market operator. Accordingly, the design should be subject to the information available to the day-ahead market stage (see 4) which does not include the individual consumed energy levels.

For each player, our mechanism builds on two population-level subjective beliefs – one constructed from all reports including that player’s and one constructed without it, thus charging based on their marginal contribution to the day-ahead market’s forecasting ability. The first work where scoring-rule theory is applied to collective beliefs constructed from many players’ inputs is [22], [23] where a market scoring rule aggregates individual reports into a single distribution, and each player’s payoff depends on how their action moves that collective belief.

Our mechanism aggregates past demand reports and past actual consumption realizations, forms two market-level subjective beliefs for the real-time imbalance and scores each player based on their marginal contribution to the market’s forecasting ability of the imbalance. Our mechanism can be viewed as a scoring-rule formulation of ideas previously developed in [24] on truthful elicitation from strategic data sources, and [25] on marginal-contribution valuation. Casting these ideas in scoring-rule language allows us to establish approximate Bayesian strategy-proofness under the information constraints (see 4).

Strategic Machine Learning.

The problem we study is, at its core, a strategic statistical estimation problem: an estimator – here, the day-ahead market – is fitted to data submitted by self-interested players whose reports depend on how the estimator will use them. This places our work within the growing literature on strategic machine learning, which examines learning and estimation problems in which data are produced by strategic players rather than drawn from a fixed distribution of nature. Foundational contributions established the basics of strategy-proof regression and classification [26], [27], followed by the strategic classification framework of [28] and the performative prediction framework of [29]. Closest to our setting are the strategic-data-source framework of [24], which designs payments so that the resulting statistical estimate is optimal in expectation under truthful reporting, and the data valuation framework of [25], which uses leave-one-out marginal contributions as a fair measure of each data point’s value. We continue in this line, but apply its ideas to a domain not previously studied from this perspective – electricity imbalance settlement – in which the events being scored are endogenously defined by the real-time supply stack, a feature absent from both prior works.

2 Institutional Background and Model↩︎

Our analysis uses the simple timeline that is common in energy markets: (i) energy is scheduled ahead of delivery, and any real-time deviations between scheduled and actual consumption are corrected by the system operator and settled financially afterwards.

2.1 Key Actors and Terms↩︎

Terminology differs across regions. Therefore, before describing the institutional sequence, we briefly list the actors and terms that appear in it. We map EU terms (NEMO, TSO, BRP, ENTSO-E) to their closest U.S. counterparts (ISO/RTO, Balancing Authority, NERC).

ACTIVE PROSUMERS.

A prosumer is an entity that (a) both consumes and produces electricity, (b) qualifies as a market participant under Article 2(25) of Regulation (EU) 2019/943 [10], and (c) engages in cross-market exchange, in particular between the day-ahead and balancing markets. The prosumer thus sits at the intersection of the active customer defined in Directive (EU) 2019/944 [9] and the aggregator function described in the same Directive, and is subject to the balance responsibility imposed by Article 5(1) of Regulation (EU) 2019/943.

NEMO (EU) / ISO–RTO (US).

In the EU, a Nominated Electricity Market Operator (NEMO) is the entity designated to perform tasks related to single day-ahead or single intraday coupling [30].

In the U.S., the closest counterpart is the Independent System Operator (ISO) or Regional Transmission Organization (RTO), which operates the system and runs day-ahead and real-time wholesale markets [31], [32].

TSO (EU) / system operator and Balancing Authority (US).

A Transmission System Operator (TSO) is responsible for operating, maintaining, and (if needed) developing the transmission system in a given area [33].

In the U.S., the operational responsibility to keep supply and demand balanced in real time belongs to a Balancing Authority (BA) (often an ISO/RTO in organized-market regions) [32], [34].

ENTSO-E (EU) / NERC (US).

ENTSO-E is the European network that brings together TSOs to support secure and coordinated operation of Europe’s interconnected grid [35].

In North America, bulk power system reliability standards are developed and overseen through the North American Electric Reliability Corporation (NERC) under FERC oversight [36].

LSE:

Load Serving Entity. The retail-facing entity that is responsible for serving electricity to end-use consumers. An LSE forecasts consumption, purchases energy in wholesale markets (e.g., day-ahead), and is responsible for deviations between scheduled and actual consumption through imbalance settlement [37].

BSP:

Balancing Service Provider. In the EU, a market participant with reserve-providing units (or groups) able to provide balancing services to TSOs [38].

In U.S.ISO/RTO markets, the closest counterpart is a resource/market participant that offers ancillary services (e.g., regulation and operating reserves) to the ISO/RTO and is compensated according to the ancillary-service market rules when providing these reliability services [39], [40].

BRP (EU) / responsible market participant (US).

A Balance Responsible Party (BRP) is a market participant (or its representative) responsible for its imbalances in the electricity market [41].

In U.S. ISO/RTO markets, the ISO/RTO financially settles real-time deviations through its real-time market and settlement rules; the responsible party is the market participant (or scheduling entity) behind the schedule [31].

Buy/sell orders:

Instructions submitted to the market operator for a given delivery period. In EU terminology these are orders that state how much energy a participant wants to buy or sell for a given delivery period, and at what monetary rate (price) limit. [30]. A buy order specifies a quantity the participant is willing to purchase up to a maximum price; a sell order specifies a quantity the participant is willing to supply above a minimum price. The market clearing process aggregates these orders to determine the accepted schedules and the market price.

In U.S.ISO/RTO terminology they are typically called bids (to buy) and offers (to sell). A buy order/bid specifies a quantity and usually a maximum price; a sell order/offer specifies a quantity and usually a minimum price. The market clearing process matches bids and offers to determine the accepted schedules and the market price(s) [42], [43].

2.2 Institutional Sequence: From Schedules to Imbalances↩︎

We describe the institutional sequence at a level sufficient for incentive analysis. While the precise design differs across markets, the institutional functioning of energy markets can often be represented as the following sequence:
Day-Ahead Scheduling (DAM). Ahead of delivery, market participants submit buy/sell orders to the Day-Ahead Market (DAM) operated by the NEMO. Market operator selects accepted demand and generation schedules and computes uniform prices.
Intraday Adjustments (IDM). After DAM, participants may adjust schedules closer to delivery in the Intraday Market. For clarity we skip intraday adjustments and treat the DAM schedule as the final schedule for the day ahead.
Balancing and Activation (TSO layer). In real time, the TSO activates balancing resources to maintain system balance (generation equals demand). Balancing resources can be integrated cross-border through certain platforms (e.g. ENTSO-E).
Imbalance settlement (financial layer). After delivery, BRPs are charged/credited for deviations between scheduled and realized positions using an imbalance price derived from balancing actions (design-specific details).

2.3 Energy Mechanisms↩︎

Fix one delivery slot. An outcome of an energy market for this slot consists of: (i) an electricity allocation \(x\) and (ii) monetary transfers \(t\). A participant of an energy market has a private type that encodes their preference over the market outcomes. In energy markets, participants communicate their type directly to the mechanism and derive utility based on the market outcomes.

Energy markets are direct mechanisms with money and in this section, we define the sequential coupling of the day-ahead and real-time institutions explicitly as such.

2.3.1 The Day-Ahead Mechanism↩︎

The day-ahead mechanism is a direct mechanism with money executed by the NEMO.

Consumers. Individual households and businesses do not participate in the energy market themselves. Instead, an LSE bundles their total demand and buys the electricity for them. Therefore, within the day-ahead market, LSEs function as actual consumers.

Producers. Power plants act as the market’s producers. Rather than selling directly to the end-consumers, they submit supply orders to the day-ahead market based on their nominal generation capacity. Therefore, within this market, power plants function as the primary producers.

Each participant has a private type that encodes their preference over the mechanism’s outcomes, the latter consisting of allocations and transfers (see 1). The day-ahead market operates as a double auction, where both LSEs and producers simultaneously submit price-quantity orders (bids/offers) to the NEMO (ISO). Within the framework of mechanism design, these orders constitute the reported types of the participants.

Simple Orders. In this study, we focus on simple orders. Simple orders consist of straightforward price-quantity pairs thus yielding quasi-linear preferences (for more details about the different orders submitted in energy markets, see e.g. [2]).

Let \(I\) be the set of consumers and \(J\) the set of producers. Each buyer \(i\in I\) has a type \[\theta_i \equiv (v_i,\mathrm{d}_i)\in \mathbb{R}_{>0}\times \mathbb{R}_{\ge 0},\] where \(v_i\) is their marginal value, i.e. willingness to pay, measured in $/MWh and \(\mathrm{d}_i\) is \(i\)’s true demand (MWh) for the particular time slot. Given an allocation \(x_i\in[0,\mathrm{d}_i]\) and transfer \(t_i\in\mathbb{R}\), buyer \(i\) has quasi-linear utility \[U_i(x_i,t_i;\theta_i)=v_i x_i + t_i.\] Each producer \(j\in J\) is characterized by a type \[\phi_j\equiv(c_j,g_j)\in\mathbb{R}_{>0}\times \mathbb{R}_{\ge 0},\] where \(c_j\) is constant marginal cost ($/MWh) and \(g_j\) is the nominal generation capacity (MWh) for the hour. Given dispatch \(y_j\in[0,g_j]\) and transfer \(t_j\), producer \(j\) has quasi-linear utility \[U_j(y_j,t_j;\phi_j)=t_j - c_j y_j.\] Participants communicate their types to the mechanism by reporting their types to the NEMO. On the basis of these reports then, the NEMO determines allocations by solving a social welfare maximization program. On top of that, given these allocations, the NEMO assigns payments to the participants.

In what follows we define the allocation and payment rules for the day-ahead mechanism under the assumption of simple orders (linear utilities).

2.3.1.1 Day-Ahead Allocation Rule.

For a given reported profile \((\theta,\phi)\), where \(\theta = (v,\mathrm{d}) \in \mathbb{R}^{2|I|}\) collects buyers’ values and demands and \(\phi = (c,g) \in \mathbb{R}^{2|J|}\) collects sellers’ costs and nominal generation capacities, the NEMO’s role is to find the allocation of energy that creates the most value for society. The NEMO solves \[\label{eq:DA-LP} (x^{DA},y^{DA}) \in \arg\max _{(x,y) \in \mathbb{R}^{|I|}_+ \times \mathbb{R}^{|J|}_+} \left\{\; v^\top x - c^\top y\;:\;\; x \leq \mathrm{d},\; y \leq g,\; \sum_{i \in I} x_i = \sum_{j \in {J}} y_j \right\},\tag{1}\] where the inequalities are component-wise and the equality constraint ensures that delivered energy matches demanded energy.

Market Restriction. The welfare maximization program (1 ) can be solved greedily by matching the highest-value consumers with the lowest-cost producers while maintaining the power balance constraint.

Program (1 ) has non-empty and bounded feasible region and hence a primal optimal \((x^{\star},y^{\star})\) exists and the optimal value is finite. The equality constraint can be relaxed and because strong duality holds it guarantees the existence of a dual optimal \(p^{\star}\). The convex and piece-wise linear dual function reads: \[\begin{align} \label{eq:decentralized} g(p)=\sum_{i \in I} \;\max _{x_i \in[0, \mathrm{d}_i]}\left(v_i-p\right) x_i+\sum_{j \in J} \;\max _{y_j \in[0, g_j]}\left(p-c_j\right) y_j, \end{align}\tag{2}\] the best-responses to some feasible \(p^{\star}\) are \[\label{eq:best-responses} \begin{align} x_i^\star(p^\star) \in \arg\max_{x\in[0,\mathrm d_i]} (v_i - p^\star)\,x, \qquad y_j^\star(p^\star) \in \arg\max_{y\in[0,g_j]} (p^\star - c_j)\,y, \end{align}\tag{3}\] and \(p^{\star}\) is dual optimal if \[\begin{align} \label{eq:sub-gradient-dual} 0\in \partial g\left(p^\star\right)=\left\{\;\sum_{i \in I} x_i = \sum_{j \in {J}} y_j \;:\; y_j \in y_j^\star\left(p^\star\right), x_i \in x_i^\star\left(p^\star\right)\right\}. \end{align}\tag{4}\] Then, \((x^{\star},y^{\star}, p^{\star})\), where \(x^{\star}=x^{\star}(p^{\star})\) and \(y^{\star}=y^{\star}(p^{\star})\), is primal-dual optimal. According to (3 ), for some fixed dual optimal \(p^{\star}\) the set of maximizers of (1 ) is not a singleton and therefore, the allocations determined are in general multi-valued correspondences.

Because of this non-uniqueness, even small variations in reported types could shift market outcomes in an unknown manner. Then, a participant is not able to determine the consequence of their reports and subsequently the mechanism designer cannot understand incentives and hence establish desired incentive properties. To this end, we impose the following rather non-restrictive assumption over reported types.

We adopt 1 to ensure the uniqueness of the market-clearing outcome. While this excludes exact ties between orders, such cases are negligible (of measure zero). This restriction allows us to treat the allocation rule as a well-defined function rather than a multi-valued correspondence, providing the necessary foundation for our incentive analysis and mechanism design in 3 and 4, respectively.

Each maximization problem in (2 ) is solved independently by the participants with the dual variable \(p^{\star}\) driving their individual incentives so that the market clears, i.e. (4 ) holds. The dual optimal that allows players to act based only on their private information while still trading with one another, is called market-clearing price.

The allocations obtained by the market and the allocations assigned by the mechanism are in general different since the latter depend on the reported than the private types. However, if day-ahead mechanism market participants report their true types, then the two allocations coincide.

Fairness of Competitive Allocations. We refer to market allocations as competitive allocations. A well-known fact from microeconomics is that competitive allocations are fair in the sense of Pareto: everyone is allocated what they deserve (according to their private type). That being said, it makes sense to benchmark our mechanism against the corresponding competitive allocation. This of course translates to incentivizing truthful reporting of the market participants.

In what follows, we define the day-ahead payments. Because energy markets clear through uniform pricing¹, we need first to define the pricing rule – how the reported types determine the day-ahead price. If we wish our mechanism to be able to recover competitive allocations (and hence be Pareto fair), we have to make sure that the pricing rule spans market clearing prices. First, we introduce some notation

Merit-Order Notation. Let \((c_{1},g_{1}),\dots,(c_{|J|},g_{|J|})\) denote producers ordered by increasing cost: \(c_{1}<\cdots<c_{|J|}\). Define cumulative capacities \[G_0\equiv 0,\qquad G_j\equiv \sum_{l\le j} \;g_{l}\quad(j=1,\dots,|J|).\] Let \(G_{\vert J\vert}\) be the total generation capacity available in the day-ahead market and assume that it always suffices for covering the total schedule \(q^{\star}\equiv\sum_{i}x^{\star}_{i}\in[0,G_{|J|}]\). The marginal producer index is defined as the cheapest marginal cost that covers that total (see also 2) \[m(q^\star)\equiv \min\{l: q^\star\le G_l\}.\] Then, the cost of the marginal producer is \(c({q^\star})\) where we slightly abuse the notation by identifying the cost of the \(m-\)th producer (which is the marginal) \(c_{m(q^\star)}\), with the cost funtion at that quantity \(c({q^\star})\). In a similar way, we define the marginal consumer.

Let \((v_{1},d_{1}),\dots,(v_{|J|},d_{|J|})\) denote consumers ordered by decreasing willingness to pay : \(v_{1}>\cdots>v_{|I|}\). Define cumulative demands \[D_0\equiv 0,\qquad D_l\equiv \sum_{i\le l}\; \mathrm{d}_{i}\quad(l=1,\dots,|I|).\] Then, the marginal consumer is defined as \(n(q^\star)\equiv \max\big\{l: q^\star\le D_l\big\}\) and the valuation of the marginal consumer is \(v(q^{\star})\), where once again, by \(v(q^{\star})\) we mean \(v_{n(q^{\star})}\).

2.3.1.2 Day-Ahead Market Pricing Rule

Let \[\begin{align} \label{eq:cleared-quantity} q(\theta,\phi)\;\equiv\;\sum_{i}x^{DA}_{i}(\theta,\phi)\in [0,{G}_{|J|}], \end{align}\tag{6}\] be the total quantity demanded for tomorrow. We follow the uniform-pricing convention and select the day-ahead price as the marginal cost of the marginal producer (in merit order) supplying the cleared quantity: \[\label{eq:DA-price-rule} p^{DA}(\theta,\phi)\;\equiv\; c({q(\theta,\phi)}).\tag{7}\] According to 2, under 1 the marginal cost of the marginal producer can be realized as a competitive price only if strictly \[\begin{align} \label{eq:32cut-by-consumers} c(q^{\star}) > v(q^{\star}+1). \end{align}\tag{8}\] This condition is rather qualitative. It implies that the decline in surplus is driven by a reduction in valuations rather than an increase in costs. In particular, this is the case in 2 where the market clears with the inverse demand curve “cutting” the inverse supply curve (see 2). In addition, (8 ) implies that non-zero allocated consumers are fully allocated their demand. This fact will simplify our incentive analysis later on in 3.

Under 2 the pricing rule (7 ) supports competitive allocations and hence renders possible benchmarking the day-ahead mechanism against competitive equilibria. As a result, it establishes a formal baseline for evaluating and achieving fairness. Having established the pricing rule, we are ready to define the day-ahead payment rule.

2.3.1.3 Day-Ahead Payment Rule

Energy markets clear through a uniform price. The payments are \[\label{eq:DA-transfers} \begin{align} t_i^{DA}(\theta,\phi)&\;\equiv\; -\,p^{DA}(\theta,\phi)\,x_i^{DA}(\theta,\phi)\quad(i\in I), \\[5pt] t_j^{DA}(\theta,\phi) &\;\equiv\; +\,p^{DA}(\theta,\phi)\,y_j^{DA}(\theta,\phi)\quad(j\in J). \end{align}\tag{9}\]

Let \(x \equiv (x_i)_{i=1}^{|I|}\), \(y \equiv (y_j)_{j=1}^{|J|}\), and \(t = \big((t_i)_{i=1}^{|I|}, (t_j)_{j=1}^{|J|}\big)\) with \(t_i = p\,x_i\) and \(t_j = p\,y_j\), denote the outcomes of the day-ahead mechanism.

2.3.2 The Real-Time Mechanism↩︎

The day-ahead producers that are dispatched to cover the actual consumption in the day-ahead commit to supply \(y\) based on the scheduled demand \(x\) reported in the day-ahead market. However, the actual consumption \(D\) (MWh), demanded in real-time is uncertain. The difference between the total scheduled demand \(q(\theta,\phi)\) from the day-ahead market and the actual consumption places the grid in imbalance \[\begin{align} \label{system-imbalance} \rho(D,(\theta,\phi))\;\equiv\; D-q(\theta,\phi), \end{align}\tag{10}\] and must be balanced with an appropriate real-time supply adjustment. The role of the real-time (balancing) market is therefore to provide this required quantity.

To safeguard operational security around the day-ahead plan, the TSO procures reserve capacity in advance (FCR/FRR/RR) [44] for maintaining physical balance between generation and demand. Operationally, it does so by activating balancing resources (BSPs), the latter being the actors of the real-time market. A BSP can be either a producer, or market actor with assets capable of rapidly adjusting electricity production.

Each BSP \(j\in \overline{J}\) is characterized by a type \(\overline{\theta}_{j}\equiv (\overline{c}_{j},\overline{g}_{j})\in \mathbb{R}_{>0}\times \mathbb{R}_{\geq 0}\), where \(\overline{c}_k>0\) is the marginal cost and \(\overline{g}_k\ge 0\) the awarded capacity. For simplicity, we assume that the set of balancing resources awarded upward activation in the hour is the same with the BSPs.

2.3.2.1 Real-Time Market Allocation Rule.

Given required balancing energy \(\rho \geq 0\) and reported profile \(\overline{\theta} = (\overline{c}, \overline{g}) \in \mathbb{R}^{|\overline{J}|}_{\geq 0}\times \mathbb{R}^{|\overline{J}|}_{\geq 0}\) collecting BSPs’ costs and reserved capacities, the role of the TSO is to dispatch BSPs for balancing real-time imbalance at minimum cost. \[\label{eq:RT-LP} y^{RT} \in \arg\min_{\overline{y} \in \mathbb{R}^{|\overline{J}|}_+} \left\{\; \overline{c}^\top \overline{y} \;:\;\; \overline{y} \leq \overline{g},\; \sum_{j \in \overline{J}} \overline{y}_j = \rho \right\},\tag{11}\] where the inequality is component-wise and \(\sum \overline{y}_j = \rho\) is the balancing constraint ensuring that dispatched reserves clear the system imbalance. Let \(\overline{G} \equiv \sum \overline{g}_j\) denote the total reserved capacity. For a consumption realization \(D\), the LP 11 has a solution only if \(\overline{G}\) suffices for supplying the emerged system imbalance.

4 ensures that the RT allocation rule (11 ) remains well-defined. Failure to satisfy this condition poses a significant risk of grid instability; consequently, the TSO is required to secure real-time reserves to account for tail events. This approach trades grid safety with increased system costs, as the TSO must procure reserve capacity that may ultimately remain idle with high probability. Therefore, accurate consumption forecasting is vital for ensuring both grid reliability and market cost-effectiveness.

We can easily link the structure of the real-time allocation to the day-ahead allocation by modeling the TSO’s commitment to balancing the grid as a fictitious consumer with finite willingness to pay greater than any real-time cost, \(v_o > \overline{c}_j\) for all \(j \in \overline{J}\), and demand equal to the imbalance, \(\mathrm{d}_o \equiv \rho\) \[\label{eq:RT-LP-2} y^{RT} \in \arg\max_{(x_o, \overline{y}) \in \mathbb{R}_+ \times \mathbb{R}^{|\overline{J}|}_+} \left\{\; v_o\, x_o - \overline{c}^\top \overline{y} \;:\;\; x_o \leq \mathrm{d}_o,\; \overline{y} \leq \overline{g},\; \sum_{j \in \overline{J}} \overline{y}_j = x_o \right\}.\tag{12}\] Similarly to the discussion in 2.3.1.1, we impose the following assumption for ensuring uniqueness of the real-time allocation.

Given that for each consumption realization the real-time allocation rule is structurally the same with the day-ahead allocation rule, the definition of real-time competitive equilibria, and pricing rules are also follow directly from the discussion of 2.3.1.1. Accordingly, under 3, for each realized consumption \(D\), strong duality holds for (11 ) and there exists dual optimal \(\overline{p}^{\star}(D)\), that we call real-time market clearing price, that regulates real-time producers’ incentives so that the real-time system imbalance is supplied \[\begin{align} \overline{y}^{\star}_{j}\in \underset{\overline{y}\in [0,\overline{g}_{j}]}{\mathrm{argmax}} \;\;\overline{p}^{\star}(D)\;\overline{y}\;-\;\overline{c}_{j}\overline{y} \label{rt-gen-opt1} \end{align}\tag{13}\] and \[\begin{align} \rho\;=\;\sum\overline{y}^{\star}_{j}. \label{rt-gen-opt} \end{align}\tag{14}\] Since the TSO commits to delivering the day-ahead schedule, \(v_o>\overline{p}^{\star}(D)\) and hence, there exists large enough budget \(w_{o}>0\) such that for each \(D\), TSO is allocated the grid imbalance \(\rho(D,x^{\star})\), \[\begin{align} \overline{x}^{\star}_o(D)\; =\; \rho(D,x^{\star}),\quad w_{o}\;\geq\; \overline{p}^{\star}(D)\rho(x^{\star},D),\quad Q-\text{a.e.}\label{rt-budget} \end{align}\tag{15}\]

Once again, \(\left(x^{\star}_o(D),\overline{y}^{\star}(D),\overline{p}^{\star}(D)\right)\) are primal-dual optimal for (11 ). Moving forward, fairness for the real-time mechanism is feasible only if the real-time pricing rule spans real-time market clearing prices.

2.3.2.2 Real-Time Market Pricing Rule.

Order balancing units by increasing costs, with cumulative capacities \(\overline{G}_r=\sum_{\ell\le r}\overline{g}_{(\ell)}\), and define \(\overline{m}(\rho)\equiv\min\{r:\rho\le \overline{G}_r\}\). The real-time uniform pricing policy is defined as the marginal cost of the marginal unit: \[\label{eq:RT-price-rule} p^{RT}(\theta,\phi,\overline{\theta},D)\;\equiv \;\overline{c}{(\rho)},\tag{16}\] where again, we slightly abuse notation by writing \(\overline{c}{(\rho)}\) instead of \(\overline{c}_{(\overline{m}(\rho))}\). Similarly to (8 ) the marginal cost of marginal producer supports real-time competitive allocation only if \[\begin{align} \label{eq:RT-comp-price} \overline{c}(\rho)\;>\;v(\rho+1). \end{align}\tag{17}\] This condition can be easily satisfied assuming the existence of a second real-time consumer with zero valuation.

2.3.2.3 Real-Time Market Payment Rule.

On top of their operational role, TSO’s are granted financial role to act as a real-time settlement counterparty: they pay BSPs for the balancing energy they deliver and charge (or credit) Balance Responsible Parties (BRPs) for their deviations from schedule.

Let \(o_i(\theta,\phi)\equiv \mathrm{d}_i-x_i^{DA}(\theta,\phi)\) be BRP \(i\)’s deviation. With one-price settlement at the real-time price \(p^{RT}(D,\theta,\phi)\), BRPs pay their imbalance \[\label{eq:32BRPs-payment}t_i^{RT}(D,\theta,\phi)\;\equiv\;- \;p^{RT}(\theta,\phi,\overline{\theta},D)\;o_{i}(\theta,\phi)\;\mathbb{1}\left[x^{DA}_{i}(\theta,\phi)\neq 0\right],\tag{18}\] and BSPs receive \[\label{eq:32BSPs-payment} t_k^{RT}(D,\theta,\phi)\;\equiv\;+\; p^{RT}(\theta,\phi,\overline{\theta},D)\;y_k^{RT}(D,\theta,\phi).\tag{19}\] The real-time settlement process requires knowledge of \(\mathrm{d}_i\) in order to allocate imbalance charges to each BRP, and this information is typically provided via metering. A BRP can either be the participant itself or an appointed representative that takes on the financial responsibility for imbalances. Conceptually, BRPs serve as the TSO’s financial counterparties in the same way that buyers serve as the NEMO’s counterparties in the day-ahead market.

Let \(x_o(D)\), \(\overline{y}(D) \equiv (\overline{y}_j(D))_{j=1}^{|\overline{J}|}\), and \(\overline{t} \equiv \big(\overline{t}_o(D),\, (\overline{t}_j(D))_{j=1}^{|J|}\big)\) denote the real-time outcomes, where \(\overline{t}_o(D) = \overline{p}(D)\,x_o(D)\) and \(\overline{t}_j(D) = \overline{p}(D)\,\overline{y}_j(D)\).

The day-ahead market operator (NEMO) implements only the day-ahead allocation and uniform payments as a function of submitted bids. The imbalance settlement is implemented by the BRP/TSO layer, which observes metered consumption and applies an exogenous settlement rule. Incentives must therefore be analyzed in the composed mechanism that we now define:

Observe that the day-ahead and real-time markets form two sequential equilibria, where the real-time equilibrium is driven by the day-ahead schedule. Moving forward, it is important to characterize the defined institutions w.r.t.whether they preserve money.

Neither the day-ahead nor the real-time mechanism is ex-post budget balanced in isolation due to imbalance settlement. Nevertheless, the energy mechanism \(M^{E}(D)\) is ex-post budget balanced by construction since \[\label{energy-budget-balanced} \begin{align} &\sum_{i \in I} t_i^{D A}(\theta,\phi)+\sum_{j \in J} t_j^{D A}(\theta,\phi)+\sum_{i\in I} t_i^{RT}(\theta,\phi,\overline{\theta},D) \\[5pt] &= p^{RT}(\theta,\phi,\overline{\theta},D)\left(\;\sum_{j\in|\overline{J}|}\overline{y}_j^{RT}(\theta,\phi,\overline{\theta},D)-\sum_{i\in I}\delta_i(D,\theta,\phi)\right)\\[5pt] & +p^{D A}(\theta,\phi)\left(\;\sum_{j \in J} y^{D A}_j(\theta,\phi)-\sum_{i \in I} x^{D A}_i(\theta,\phi)\right)\\[5pt] &=0. \end{align}\tag{20}\]

3 Incentives↩︎

Market participants have preferences over the market outcomes, typically expressed as utility functions derived from their private types. On the basis of their private information, the publicly available institutional framework and their knowledge about the world and the others, they best respond to the induced game so as to influence the social choice for their own benefit.

As we already saw in 2, the response of the real-time mechanism, driven by the day-ahead schedule, determines energy costs to the final consumers and affects grid stability. In particular, any bias of the day-ahead schedule from the actual consumption increases imbalances which later on are bought by the end consumers via the imbalance settlement (18 ). Beyond financial damage, large biases threaten operational security by destabilizing the power grid.

3.1 Social Choice↩︎

As we saw in 1, empirical observations in European energy markets—and, in particular, inspection of consumption data from [6] reveal a systematic underestimation of realized demand in the day-ahead stage. While forecast errors are always present in practice, the severity and persistence of this underestimation motivates the hypothesis that day-ahead market consumers are strategically misreporting their demands.

To investigate this hypothesis, we assume that valuations, costs and nominal capacities are publicly available or at least, they can be inferred, and focus on the demands. Let \(\mathrm{d}\equiv(\mathrm{d}_i)_{i\in I}\) denote the vector of true (private) demands. Building on the discussion in 2.3.1.1, if day-ahead players were acting in isolation, the resulting day-ahead allocation would be Pareto fair relative to the private types and in the absence of strategic bias, the TSO would have to safeguard only for the inherent uncertainty of the actual consumption.

Consequently, Pareto fairness in the Day-Ahead stage is a prerequisite for minimizing real-time costs and preventing grid failures. Given that all types except demands are public, truthful reporting of demands in the day-ahead mechanism becomes the necessary and sufficient condition for Pareto fairness.

Given that the day-ahead and real-time institutions are linked, incentives for demand under-reporting in the day-ahead stage might be induced by preferences that extend beyond the day-ahead outcomes. Therefore, we benchmark the entire energy institution (see 6) against the corresponding market equilibria \[\begin{align} \label{eq:32energy-mech} \ifmmode\mathrm{M}\else M\fi^{E} : \mathrm{d} \longmapsto \Big(\, (x^{\star},\, y^{\star},\, t^{\star}),\; \big(x^{\star}_{o}(D),\, \overline{y}^{\star}(D),\, \overline{t}^{\star}(D)\big) \,\Big). \end{align}\tag{21}\]

Intuitively, when the day-ahead allocation is truthful, the real-time imbalance varies around zero, thus minimizing risk of grid instability, the real-time price and the overall energy costs. A nonzero average deviation means that the day-ahead plan inherits some strategic component (strategic bias). That said, observe that under public values, costs, and generation capacities, if \[\begin{align} \label{eq:da-benchmark} \ifmmode\mathrm{M}\else M\fi^{DA}~:~ \mathrm{d}\mapsto \left(x^{\star},\;y^{\star},\;t^{\star}\right), \end{align}\tag{22}\] is truthfully implemented, then so is \(\ifmmode\mathrm{M}\else M\fi^{RT}\) and hence \(\ifmmode\mathrm{M}\else M\fi^{E}\). Crucially, because existence and uniqueness conditions are independent of the demands \((\mathrm{d})\), we can analyze incentives without concern that varying the demand reports will compromize the well-definedness of (21 ).

3.2 Setup↩︎

Consumers in the day-ahead-mechanism privately observe demands from some distribution. They also maintain a belief about the others types \(\mathrm{d}_{-i}\) and actual consumption \(D\), usually formed by aggregating past profiles and states of the world.

Let \((\mathrm{d}_{-i}, D)\sim P_{i}\) be such a belief maintained player \(i\) and let \(P_{-i}\) denote the associated marginal w.r.t.\(\mathrm{d}_{-i}\). Then, for a given sampled private type, they decide based on that subjective belief and play a subjective equilibrium (see [45]–[47]).

Unlike Bayes-Nash, subjective equilibria are inefficient since there is no guarantee that honest forecasting will ever emerge. If, however, under the energy mechanism, all consumers were required to report their demands truthfully, and this requirement were common knowledge, then the induced game would admit a well-defined common prior.

In order the mechanism participants to report truthfully on the basis of their knowledge, mechanism should be such that they rationalize truth-telling –approximately– as their best response.

The implementable preferences over the mechanism’s outcomes span positive combinations of consumers in the day-ahead market, producers in the day-ahead market and producers in the real-time market. However, under the defined institutions (see 1 and 4), day-ahead producers’ components would constrain the demand under-reporting incentives and hence act against what we empirically observe in the European electricity markets. For that reason, we rule them out of the incentive study.

The incentive question studied in this section is whether, under only cross-market ownership (prosumers), the rational behavior under the defined institutions is compatible with the market equilibria (see 2).

3.3 Consumers↩︎

We now study the incentives of consumers in the energy institution. Typically, preferences over the day-ahead consumption outcomes are implemented by (LSEs), the latter being associated with their own clientele. Thus naturally, such entities are modeled as a single consumer. Consumer \(i\)’s preferences over outcomes \((x,t)\) of the energy mechanism are given by \[U_{i}(x,t)\;=\;v_{i}\;x_{i}+t_{i}.\nonumber\] Being aware of the institutional framework, each consumer affects the social choice by acting rationally and reporting their demand \(\mathrm{d}_{i}\in \mathbb{R}_{\geq 0}\) to the day-ahead institution so that their preferences are maximized in expectation. Let \(( \mathrm{d}^{\prime}_{i},\mathrm{d}_{-i})\in \mathbb{R}^{|I|}_{\geq 0}\) be the reported demand profile in the day-ahead mechanism. At that profile, under mechanism \(\ifmmode\mathrm{M}\else M\fi^E\), consumer \(i\in I\) derives net utility \[\label{consumers-utility}U_{i}\big(\mathrm{d}_{i},\;\ifmmode\mathrm{M}\else M\fi^E(\mathrm{d}^{\prime}_{i},\mathrm{d}_{-i})\big) = v_{i}\mathrm{d}_i+t^{DA}_i(\mathrm{d}^{\prime}_i,\mathrm{d}_{-i})+t_i^{RT}(D,\mathrm{d}^{\prime}_{i},\mathrm{d}_i,\mathrm{d}_{-i}),\tag{23}\] where, according to (9 ) \[t^{DA}_i(\mathrm{d}^{\prime}_i,\mathrm{d}_{-i})=-p^{DA}(\mathrm{d}^{\prime}_{i},\;\mathrm{d}_{-i})\;x^{DA}_{i}(\mathrm{d}^{\prime}_{i},\;\mathrm{d}_{-i}),\] and according to (18 ) \[\begin{align} t_i^{RT}(D,\mathrm{d}^{\prime}_{i},\mathrm{d}_i,\mathrm{d}_{-i}) = -p^{RT}\big(\rho(D,x^{DA}(\mathrm{d}^{\prime}_{i},\mathrm{d}_{-i}))\big)\big(\mathrm{d}_{i}-x^{DA}_{i}(\mathrm{d}^{\prime}_{i},\;\mathrm{d}_{-i})\big)\mathbb{1}\!\big[x^{DA}_{i}\neq 0\big].\nonumber \end{align}\]

The first term in (23 ) is consumer \(i\)’s utility from both the day-ahead and the real-time stage. The second term is the payment charged to the consumer in the day-ahead stage. If consume more than bought in the day-ahead stage, consumers have to buy the remainder at the real-time price. The third term in (23 ) expresses the payment for these imbalances as being defined by the real-time settlement (19 ). When \(i\in I\) is zero day-ahead-allocated the real-time settlement is not applied (see (18 )) and the net utility vanishes. To set the stage for our incentive analysis, given report \(\mathrm{d}^{\prime}_{i}\in[0,\mathrm{d}_i]\) let \[\begin{align} \label{eq:32dispatch32event} A_i\;\equiv\; \left\{\;\mathrm{d}_{-i}\in \mathbb{R}^{|I|-1}_{\geq 0}\;\;\Big|\;\; v_i>p^{DA}(\mathrm{d}^{\prime}_{i},\mathrm{d}_{-i})\right\}, \end{align}\tag{24}\] be the event within which \(i\) is allocated when reporting \(\mathrm{d}^{\prime}_i\). Let \(A_i=A(\mathrm{d}_i)\) be the event where \(i\) is allocated when reporting truthfully, and \(A^{\prime}_i=A(\mathrm{d}^{\prime}_i)\) the event where \(i\) is allocated when under-reporting.

Under 2 a consumer that is allocated in the day-ahead market is (fully) allocated their report. Therefore, \[\begin{align} \label{eq:da-alloc-under-assumption-1} x^{DA}_{i}(\mathrm{d}^{\prime}_{i},\mathrm{d}_{-i}) \;=\; \mathrm{d}^{\prime}_{i}\;\mathbb{1}_{A^{\prime}_i}(\mathrm{d}_{-i}). \end{align}\tag{25}\] If consumer \(i\) is allocated when reporting truthfully, so is when under-reports \[\begin{align} \label{eq:da-alloc-ind} \mathbb{1}_{A_i}(\mathrm{d}_{-i})=1\implies\mathbb{1}_{A^{\prime}_i}(\mathrm{d}_{-i})=1, \end{align}\tag{26}\] since due to merit order, under-reporting in the day-ahead mechanism can only drop the day-ahead price. In addition, observe that for some report \(\mathrm{d}^{\prime}_{i}\) \[\begin{align} \label{eq:32da-alloc-eq-ind} \mathbb{1}\left[x^{DA}_i(\mathrm{d}^{\prime}_{i},\mathrm{d}_{-i})\neq 0\right]=\mathbb{1}_{A^{\prime}_i}(\mathrm{d}_{-i}). \end{align}\tag{27}\] The point-wise gain obtained by under-reporting \(\mathrm{d}^{\prime}_{i}=\mathrm{d}_i-\delta_i\) is defined as \[\begin{align} \label{eq:32gain-under-reporting} G_{i}\big(\mathrm{d}_{i},\;\ifmmode\mathrm{M}\else M\fi^E(\mathrm{d}_i-\delta_i,\mathrm{d}_{-i})\big)\;\equiv\; U_i\big(\mathrm{d}_{i},\;\ifmmode\mathrm{M}\else M\fi^E(\mathrm{d}_i-\delta_i,\mathrm{d}_{-i})\big)-U_i\big(\mathrm{d}_{i},\;\ifmmode\mathrm{M}\else M\fi^E(\mathrm{d}_i,\mathrm{d}_{-i})\big). \end{align}\tag{28}\]

Given their true (private) demand, consumers decide about their report by maximizing their utility in expectation over all other reports and actual consumption. By plugging (25 ), (26 ), (27 ) into (28 ) we obtain:

\[\begin{align} G_{i}\big(\mathrm{d}_{i},\;\ifmmode\mathrm{M}\else M\fi^E(\mathrm{d}_i-\delta_i,\mathrm{d}_{-i})\big)&= \;\Big( p^{DA}(\mathrm{d}_i,\mathrm{d}_{-i})\mathbb{1}_{A_i}-p^{DA}(\mathrm{d}_i-\delta_i,\mathrm{d}_{-i})\mathbb{1}_{A^{\prime}_i} \Big)\mathrm{d}_i\nonumber\\[5pt] & + \Big( p^{DA}(\mathrm{d}_i-\delta_i,\mathrm{d}_{-i})-p^{RT}(\rho)\Big)\delta_i\mathbb{1}_{A^{\prime}_i},\nonumber \end{align}\label{eq:random}\tag{29}\] where \(\rho=\rho(\delta_i)= D - \sum_{l} x^{DA}_{l}\!\big(\mathrm{d}_i-\delta_i,\mathrm{d}_{-i}\big)\). At this point we assume that consumer \(i\) believes that she is certainly allocated.

Under 6, \(\Pr(A^{\prime}_i) = 1\). By taking expectations (29 ) reads

\[\begin{align} \label{eq:Delta-consumer-33} \mathbb{E}_{(D,\mathrm{d}_{-i})\sim P_i}\;G_{i}\big(\mathrm{d}_{i},\;\ifmmode\mathrm{M}\else M\fi^E(\mathrm{d}_i-\delta_i,\mathrm{d}_{-i})\big) &= \mathbb{E}\!\Big[ p^{DA}(\mathrm{d}_i,\mathrm{d}_{-i})-p^{DA}(\mathrm{d}_i-\delta_i,\mathrm{d}_{-i}) \Big]\mathrm{d}_i \nonumber\\[5pt] &+\mathbb{E}\!\Big[ p^{DA}(\mathrm{d}_i-\delta_i,\mathrm{d}_{-i}) -p^{RT}(\delta_i,\mathrm{d}_{-i}) \!\Big]\delta_i\; \nonumber\\[5pt] &\leq \mathbb{E}\!\Big[ -p^{DA}(\mathrm{d}_i-\delta_i,\mathrm{d}_{-i}) +p^{DA}(\mathrm{d}_i,\mathrm{d}_{-i}) \!\Big]\mathrm{d}_i \nonumber\\[5pt] &+\mathbb{E}\!\Big[ p^{DA}(\mathrm{d}_i,\mathrm{d}_{-i}) -p^{RT}(0,\mathrm{d}_{-i}) \!\Big]\delta_i, \end{align}\tag{30}\] where \(\mathbb{E}\!\left[\, p^{RT}\!\big(\rho(\delta_i)\big) \,\middle|\, \mathrm{d}_{-i} \right] = p^{RT}(\delta_i,\mathrm{d}_{-i}),\) and the inequality follows because the day-ahead price increases with the demand report of \(i\) and the real-time price is increasing with imbalance \(\rho\). According to the first term of (30 ), by requesting less energy, consumer \(i\) pulls the day-ahead price down. For handling the second term of (30 ) we state the following

Moving forward, because the inverse supply curve is piece-wise constant, if the reduction stays within the same step, the same producer remains marginal and the price does not change at all. Further, the price drops only if the reduced demand crosses a step boundary in the supply curve.

The following Lemma extends Lipschitz continuity to piece-wise constant inverse supply.

Empirical evidence suggest that the day-ahead market clears within a sub-range of costs where the Lipschitz part and the price gap are both very small.

3.3.1 Anti-Concentration of Aggregate Demand.↩︎

While Lemma 1 alone yields a deterministic bound on the gain from the day-ahead price drop, this bound can be further tightened by accounting for the fact that the price change occurs only when the deviation causes total demand to cross a supply step boundary. Whether or not this crossing occurs depends on where the aggregate demand of other players \(\sum_{j\neq i}\mathrm{d}_j\), happens to fall relative to the step boundaries of the supply curve. If player \(i\)’s beliefs \(P_{-i}\) are sufficiently spread out —that is, if no small interval of possible demand values is assigned too much probability—then the chance of such a crossing is small, and the expected gain shrinks accordingly.

As mentioned earlier in 3.2, players form their beliefs about others’ total demand from historical market data. The condition we require is that these beliefs do not concentrate excessively in any narrow interval. This is a rather mild requirement: it holds whenever the historical demand data is not pathologically concentrated at a single point and it is directly verifiable from past data.

The appropriate tool for formalizing this condition is the Lévy concentration function, a classical object in probability theory introduced by Lévy [48] and developed systematically by Petrov [49].

For a random variable \(X\) and a width \(\delta \geq 0\), the Lévy concentration function is defined as \[\begin{align} \label{eq:levy} Q(X,\,\delta) \;\equiv\; \sup_{x \in \mathbb{R}}\; \Pr\big(X \in [x,\; x+\delta]\big). \end{align}\tag{31}\] It measures the maximum probability that \(X\) places on any interval of width \(\delta\). A small value of \(Q(X,\delta)\) means that the distribution of \(X\) is spread out and does not concentrate too much mass in any narrow region. As we show below, the Lévy concentration function of the aggregate demand \(\sum_{j\neq i}\mathrm{d}_j\) is small whenever enough individual players exhibit variability in their bids. This follows from the Kolmogorov–Rogozin inequality, which we now recall.

In our setting, \(\xi_j = \mathrm{d}_j\) for each \(j\neq i\), and the sum \(\sum_{j\neq i}\mathrm{d}_j\) is the aggregate demand of all players other than \(i\). The quantity \(Q_j(l) = \sup_x \Pr(\mathrm{d}_j \in [x,x+l])\) measures how concentrated player \(j\)’s bid is: it equals one if and only if player \(j\) always bids within the same interval of width \(l\), and is strictly less than one whenever player \(j\)’s bid varies beyond that width. The quantity \(1-Q_j(l)\) captures how much variability player \(j\) contributes, and the sum \(s = \sum_{j\neq i}(1-Q_j(l))\) aggregates this variability across all other players. 4 then says that the more players exhibit variable bidding behavior, the more spread out the aggregate demand is, and the smaller the Lévy concentration function becomes.

This motivates the following assumption, which is a primitive condition on individual consumers’ reporting behavior.

8 says that there exist at least \(m\) consumers that do not always report within the same narrow interval of width \(l\): each such consumer’s report varies by more than \(l\) across market realizations. The constant \(\kappa\) quantifies the minimum variability each of these players exhibits. This is a weak and empirically verifiable condition: it fails only if nearly all players report the exact same quantity in every market period, which would mean there is essentially no demand uncertainty. In any realistic energy market variability in weather, and individual consumption ensures that many consumers’ demands fluctuate, so both \(m\) and \(\kappa\) are bounded away from zero.

9 says that the share of players with variable demand does not vanish as the market grows. It holds whenever the sources of demand variability — weather, economic activity, individual consumption patterns — affect a stable fraction of the population.

We are now ready to state the first result of this section.

3.3.2 Empirical Evidence for Small Cost Gaps at Market Clearing.↩︎

Recall from 2.3.1 that in day-ahead electricity markets, the clearing price is set by the marginal producer—the most expensive unit needed to meet total demand. Empirical studies show that, for a large percent of operating hours, this marginal producer is a natural gas-fired power plant. In European day-ahead auctions, Zakeri et al. [51] find that gas-fired plants set the wholesale price approximately 80% of the time  [51] and a European Commission study confirms similar findings [52].

In the US, the PJM day-ahead market clears using cost-based offers from the same power plants [53], and the Monitoring Analytics State of the Market Report shows that fuel costs—mainly gas—constitute the dominant component of day-ahead clearing prices [54].

When gas is the marginal fuel, the cost gaps between consecutive producers in the merit order are small, because all gas-fired units share the same fuel price and differ only in their heat rates—a measure of how efficiently they convert fuel to electricity [55].

In a large market such as PJM, the combined-cycle gas power plants contain a large number of units spanning a narrow range of heat rates [56], [57] across tens of thousands of megawatts of capacity [58]. Because these units share the same fuel and differ only slightly in efficiency, their marginal costs are tightly grouped, and the cost gap between any two consecutive units in the merit order is negligibly small relative to the price level.

According to 1, consumers have no incentive to under-report in a large market. Next, we study a subset of day-ahead market consumers — namely, consumers who own generation assets in the balancing market and can produce energy in real time. We call these consumers prosumers (see also Section 2.1). As we will see, under the same market parameters and conditions stated in 1 (and 1), prosumers’ gain from under-reporting is lower-bounded by a strictly positive constant. That is, there exists a subset of players in \(I\) whose gain from under-reporting does not go to zero as the market grows (see 2).

3.4 Cross-Market Ownership & Prosumers↩︎

Consumers might also own one or more generation assets in the real-time market. We refer to such actors that buy energy in the day-ahead market while owning shares on production (generation) units in the real-time market as prosumers (see 2.1).

Let \((\alpha_{i,1},\ldots,\alpha_{i,|\overline{J}|})\) be the convex combination of shares for prosumer \(i\in I\) over generation assets. Then, prosumers’ preferences over the energy mechanism’s outcomes \((x,t,\overline{y},\overline{t})\) read: \[\begin{align} U^{\alpha}_{i}(x,t,\overline{y},\overline{t})= U_{i}(x,t)+\sum_{l \in \overline{J}}\;\alpha_{il} \;\left(\;\overline{t}_{l}-\overline{c}_{l}\;\overline{y}_{l}\;\right). \end{align}\] Each prosumer then, influences the social choice by acting rationally and reporting their demand \(\mathrm{d}_{i}\in \mathbb{R}_{\geq 0}\) to the day-ahead institution so that her preferences are maximized. Let \(( \mathrm{d}^{\prime}_{i},\mathrm{d}_{-i})\in \mathbb{R}^{|I|}_{\geq 0}\) be the reported demand cap profile in the day-ahead mechanism. At that profile, under mechanism \(\ifmmode\mathrm{M}\else M\fi^{E}\), prosumer \(i\in I\) derives utility \[\begin{align} U^{\alpha}_{i}\big( \mathrm{d}_{i},\;\ifmmode\mathrm{M}\else M\fi^{E}(\mathrm{d}^{\prime}_{i},\;\mathrm{d}_{-i})\big)&\equiv U_{i}\big( \mathrm{d}_{i},\;\ifmmode\mathrm{M}\else M\fi^{E}(\mathrm{d}^{\prime}_{i},\;\mathrm{d}_{-i})\big)+C_{i}\big(D;\;\mathrm{d}^{\prime}_{i},\;\mathrm{d}_{-i}\big), \label{prosumers-utility} \end{align}\tag{37}\] where \[\begin{align} \label{shares} C_{i}\big(D;\;\mathrm{d}^{\prime}_{i},\;\mathrm{d}_{-i}\big)\equiv \mathbb{1}\left[x^{DA}(\mathrm{d}^{\prime}_{i},\mathrm{d}_{-i}){\neq} 0\right]\sum_{l\in\overline{J}}\alpha_{il}\big({p}^{RT}(D, \mathrm{d}^{\prime}_{i},\mathrm{d}_{-i})-\overline{c}_{l}\big)y^{RT}_{l}(D,\mathrm{d}^{\prime}_{i},\mathrm{d}_{-i}\bigl). \end{align}\tag{38}\] Under the day-ahead and real-time institutions described in 2, by under-reporting their day-ahead demand, a prosumer lowers the net day-ahead schedule thus increasing both the likelihood and the magnitude of upward real-time imbalances in such a way so that prosumer’s generation asset gets activated and offers real-time revenues.

Due to the piecewise-constant structure of the real-time inverse supply curve (see 2), the resulting real-time payoff is a discrete random variable. Its realized value is determined by the specific capacity bin \[\ifmmode\mathrm{B}\else B\fi_{j}\equiv [\overline{G}_j,\;\overline{G}_{j+1}),\] into which the aggregate system imbalance materializes.

Formally, Lemma 2 expresses the second term in (37 ) as a state-contingent balancing-energy payoff whose availability and size are supported by maintaining shares over the reserved generation capacity in real time.

Prosumers are a subset of consumers. Our goal then is to identify an interpretable and feasible (w.r.t.1, 4, and 8) market instance such that prosumers have incentives to under-report. This would then prove that \(\mathrm{M}^{E}\) is not \(\varepsilon-\)BIC.

To begin with, according to (40 ), the energy mechanism partitions the real-time capacity axis into mutually exclusive imbalance states (bins). Each prosumer then is provided for free by the energy mechanism a portfolio of Arrow–Debreu-style contracts: the contract for state \(\ifmmode\mathrm{B}\else B\fi_j\) pays \(w_{i,j}\) if \(\rho\) lands in \(\ifmmode\mathrm{B}\else B\fi_j\) and zero otherwise. This is a flaw of the energy mechanism.

Recall 6 and define the probabilities \[P^j(\delta_i) \;\equiv\; \mathbb{E}\;\Big[\mathrm{Pr}(\ifmmode\mathrm{B}\else B\fi_j-\delta_i\mid \mathrm{d}_{-i})\Big], \qquad P^j(0) \;\equiv\; \mathbb{E}\;\Big[\mathrm{Pr}(\ifmmode\mathrm{B}\else B\fi_j\mid \mathrm{d}_{-i})\Big],\] and the probability shift \(\Delta P(\delta_i)\equiv P^j(\delta_i)-P^j(0)\). Under 6, prosumer’s expected gain from under-reporting \(\mathrm d'_i = \mathrm{d}_i - \delta_i\) reads \[\begin{align} \label{eq:dyo} &\mathbb{E}_{\;\mathrm{d}_{-i}\sim P_{-i}} G^{\alpha}_{i}\big(\mathrm{d}_{i},\;\ifmmode\mathrm{M}\else M\fi^E(\mathrm{d}_i-\delta_i,\mathrm{d}_{-i})\big)= \mathbb{E}_{\;\mathrm{d}_{-i}\sim P_{-i}} G_{i}\big(\mathrm{d}_{i},\;\ifmmode\mathrm{M}\else M\fi^E(\mathrm{d}_i-\delta_i,\mathrm{d}_{-i})\big)\; + \big\langle w_i,\,\Delta P(\delta_i)\big\rangle.\nonumber \end{align}\tag{41}\] Let \(\overline{\beta}_i\equiv \max_{\delta_i \in [0,\mathrm{d}_i]}\mathbb{E}p^{RT}(\delta_i,\mathrm{d}_{-i})\) be the the maximum expected real-time price that player \(i\) can influence by fully under-reporting, and let \(\overline{\beta}_i \mathrm{d}_i\) the corresponding cost. Given the fact that the day-ahead price is increasing w.r.t.the demand report, equation (30 ) yields \[\begin{align} \mathbb{E}_{\;\mathrm{d}_{-i}\sim P_{-i}}\;G_{i}\big(\mathrm{d}_{i},\;\ifmmode\mathrm{M}\else M\fi^E(\mathrm{d}_i-\delta_i,\mathrm{d}_{-i})\big) \;\geq\;-\overline{\beta}_i \mathrm{d}_i \end{align}\] Boundary-crossing decomposition. Let \(\rho_0\geq 0\) be the real-time imbalance when prosumer \(i\) reports truthfully (\(\delta_i=0\)), with continuous CMF \(F(x)\equiv\mathrm{Pr}(\rho_0\leq x)\), where \(F(x)=0\) for \(x\leq 0\). Then \[\label{eq:decomp} P^j(\delta_i) \;=\; F(G_{j+1}-\delta_i) - F(G_{j}-\delta_i),\tag{42}\] and the mass crossing (from the left) boundary \(\overline{G}_j\) when under-reporting by \(\delta_i\) is \[\label{eq:inflow} T_{j}(\delta_i) \;\equiv\; \mathrm{Pr}\!\left(\overline{G}_j - \delta_i \leq \rho_0 < \overline{G}_k\right) \;=\; F(\overline{G}_k) - F(\overline{G}_k - \delta_i).\tag{43}\]

Dropping the index \(i\) from \(w_i\) and writing \(\Delta w_j\equiv w_j - w_{j-1}\), and using the boundary conditions \(T_1(\delta_i)=T_{|\overline{J}|}(\delta_i)=0\) (no mass crosses \(0\) from the left, and no mass crosses \(G_{|\overline{J}|}\)), equations (42 )–(43 ) yield the gap-weighted inflow formula: \[\label{eq:weighted95boundary95decomp} \big\langle w,\,\Delta P(\delta_i)\big\rangle \;=\; \sum_{j=2}^{J-1}\Delta w_j\,T_j(\delta_i).\tag{44}\]

Unimodal profits. Suppose the profit vector is unimodal across bins, i.e.there exists \(j_p\) such that \[\label{eq:unimodal95w} w_1 \;\leq\; \cdots \;\leq\; w_{j_p} \;\geq\; \cdots \;\geq\; w_{J-1}.\tag{45}\] Then \(\Delta w_j\geq 0\) for \(j\leq j_p\) and \(\Delta w_j\leq 0\) for \(j\geq j_p+1\), and (44 ) splits into \[\label{eq:split95good95bad95boundaries} \big\langle w,\,\Delta P(\delta_i)\big\rangle \;=\; \underbrace{\sum_{j=2}^{j_p}\Delta w_j\,T_j(\delta_i)}_{\text{uphill crossings}} \;-\; \underbrace{\sum_{j=j_p+1}^{|\overline{J}|-1}|\Delta w_j|\,T_j(\delta_i)}_{\text{downhill crossings}}.\nonumber\tag{46}\] Under-reporting is therefore profitable at \(\delta_i\) when the mass crossing profit-increasing boundaries dominates both the mass crossing profit-decreasing boundaries and the baseline cost \(\beta_i\delta_i\).

Unimodal mass function. We assume that the probability mass function is unimodal. Under this assumption the boundary-crossing masses \(j\mapsto T_j(\delta_i)\) are themselves unimodal in \(j\) for every fixed \(\delta_i\in(0,\mathrm{d}_i]\): as \(\overline{G}_j\) moves toward the high-probability region of \(\rho_0\), the width-\(\delta_i\) window mass \(T_j(\delta_i)=\mathrm{Pr}([\overline{G}_j-\delta_i,\overline{G}_j))\) first increases and then decreases.

Combined with unimodal profits (45 ), this gives clear intuition: profitable under-reporting requires choosing \(\delta_i\) so that the (typically localised) boundary crossings with the largest \(T_j(\delta_i)\) fall on the uphill side \(j\leq j_p\), while crossings on the downhill side \(j\geq j_p+1\) remain small.

Single-asset special case. Suppose prosumer \(i\) holds exactly one real-time generation asset on bin \(j_0\), with payoff \(w\equiv w_{j_0}\). Then \[\mathbb{E}_{\;\mathrm{d}_{-i}\sim P_{-i}} G^{\alpha}_{i}\big(\mathrm{d}_{i},\;\ifmmode\mathrm{M}\else M\fi^E(\mathrm{d}_i-\delta_i,\mathrm{d}_{-i})\big) \geq w\!\left(T_{j_0}(\delta_i)-T_{j_0+1}(\delta_i)\right).\] If moreover \(j_0=|\overline{J}|-1\) (the previous from rightmost bin), then \(T_{|\overline{J}|}(\delta_i)=0\) and \[\begin{align} \label{eq:prosumer-gain} \mathbb{E}_{\;\mathrm{d}_{-i}\sim P_{-i}} G^{\alpha}_{i}\big(\mathrm{d}_{i},\;\ifmmode\mathrm{M}\else M\fi^E(\mathrm{d}_i-\delta_i,\mathrm{d}_{-i})\big) \;\geq\; w\,T_{|\overline{J}|-1}(\delta_i) . \end{align}\tag{47}\]

The term \(\mathrm{d}_i(w\,q_i - \overline{\beta}_i)\) is the prosumer’s net gain from under-reporting, scaled by their demand. The quantity \(w\,q_i\) is the real-time profit rate: the revenue the prosumer earns per unit of demand by pushing the imbalance across the boundary of their generation bin, where \(w\) is the cost gap between consecutive bins times the nominal capacity of the prosumer’s asset and \(q_i\) is the rate at which boundary-crossing scenarios are activated. The quantity \(\overline{\beta}_i\) is the real-time price cost: the most the prosumer pays per unit of demand for the energy that spills into real-time due to the under-report. The difference \(w\,q_i - \overline{\beta}_i\) is the net rate: what the prosumer earns from the real-time dispatch manipulation minus what they pay for the spilled energy. The lower bound \(\varepsilon^{\alpha}_i\) is positive when \(w\,q_i > \overline{\beta}_i\), meaning the prosumer earns more from manipulating the real-time dispatch than they pay for the resulting real-time energy cost.

1 and 2 are contradictory in the following sense. 1 states that for any tolerance \(\varepsilon > 0\), the gain from misreporting for every consumer falls below \(\varepsilon\) once the market is large enough. 2 states that a prosumer — who is a player in \(I\) — has gain at least \(\varepsilon^{\alpha}_i > 0\) regardless of how large the market is. In particular, choosing \(\varepsilon = \varepsilon^{\alpha}_i/2\) in  1: for \(|I|\) large enough, every consumer’s gain is below \(\varepsilon^{\alpha}_i/2\), yet the prosumer’s gain remains at least \(\varepsilon^{\alpha}_i > \varepsilon^{\alpha}_i/2\). Therefore, the mechanism \(\ifmmode\mathrm{M}\else M\fi^E\) cannot be \(\varepsilon\)-BIC with vanishing \(\varepsilon\) for all players in \(I\).

Below is the proof of 2:

4 Mechanism Design↩︎

We design a mechanism for the two-stage energy institution (see 6) consisting of a day-ahead scheduling stage and a real-time balancing stage. The allocation rules in both stages are fixed and reflect the existing regulatory framework. Uniform pricing is also taken as given. In 3.1 we saw that under public costs, values and nominal capacities, if the day-ahead institution is truthfully implemented, then so is the energy mechanism. Our design problem then concerns the settlement rule faced by demand-side participants at the day-ahead stage.

The strategic component of the day-ahead schedule can destabilize the grid but also increase the energy costs for consumers. In 3 we saw that under certain conditions, the current imbalance settlement is unable to incentivize prosumers in reporting their demand truthfully.

Moreover, we identify an additional major concern with the existing imbalance settlement framework. Its reliance on metering introduces vulnerabilities to manipulation and measurement errors, raising questions about its trustworthiness.

To this end, we propose an extended imbalance settlement mechanism \[\begin{align} t^{DA,S}_{i}(\cdot)\;\equiv \; t^{DA}_{i}(\cdot)+\Pi_i(\cdot),\label{new-settlement}\quad i\in I, \end{align}\tag{49}\] implemented by the day-ahead market, where \(\Pi_{i}(\cdot)\) is chosen from an appropriate class \(C\) of mechanisms characterized by properties we define in 4.3. We have the following definition.

On the basis of 9, our mechanism is obtained as the solution to the following optimization problem \[\mathrm{M}^{S\star}(\cdot) \in \underset{\mathrm{M}^S(\cdot)\, \in\, C}{\arg\min}\; \left\{\, \mathbb{E}\left|\rho\big(\mathrm{d}^{\prime}_1,\ldots,\mathrm{d}^{\prime}_{|I|}\big)\right|^2 \;:\;\; \mathrm{d}^{\prime}_i \in \underset{x\in [0,\mathrm{d}_i]}{\arg\max}\; \mathbb{E}_{\sim P_{-i}}\, U^\alpha_i\!\left(\mathrm{d}_i,\mathrm{M}^{S}(x,\mathrm{d}_{-i})\right),\; \forall i \in I \,\right\}. \nonumber\] In what follows, we characterize the set \(C\) by defining the specific properties this mechanism must satisfy.

4.1 Information Constraints↩︎

The proposed settlement modification is intended to be implemented by the day-ahead market operator (NEMO). Accordingly, we explicitly restrict the information available to the designer to what is institutionally observable at the day-ahead stage and after delivery.

At the time settlement is computed, the designer observes:

1. the full profile of day-ahead reports \(\mathrm{d}=(\mathrm{d}_i)_{i\in I}\);

2. the day-ahead clearing outcome \(x^{DA}(\mathrm d)\) and the cleared total \(\sum_{i\in I}x_i^{DA}(\mathrm d)\);

3. realized system total consumption \(D\) for the delivery hour;

4. public historical data consisting of past day-ahead reports and realized system consumptions.

The designer does not observe realized consumption at the level of individual day-ahead participants or Balance Responsible Party (BRP) portfolios, nor any settlement allocation of realized consumption to individual participants.

While individual BRP-level allocated volumes are calculated and used for imbalance settlement by the TSO (or a nationally assigned settlement entity), they are outside the defined scope of NEMO tasks (which concerns day-ahead/intraday coupling and clearing/settlement of the resulting trades) and are subject to professional secrecy where they are treated as confidential information under network codes: [38] [30].

4.2 Scoring Rule-Based Mechanisms↩︎

The inputs to the designer’s problem are the information available to the NEMO, the structure of prosumer’s incentives and the publicly available market parameters. The market participants should be aware of the incentive changes applied by the NEMO.

Recalling our earlier notation, \(\ifmmode\mathrm{B}\else B\fi_j\) refers to the \(j\)th interval formed by the cumulative real-time generation capacities, \([\overline{G}_j, \overline{G}_{j+1})\). With \(\mathrm{d}^{\prime}_{i}\) we denote the report of the \(i\) prosumer in the day-ahead mechanism and with \(\mathrm{d}_{-i}\) the demands of the rest of the players. The real-time imbalance is given by \(\rho=D-\sum_{i}x^{DA}_{i}(\mathrm{d}^{\prime}_{i},\mathrm{d}_{-i})\), where \(D\) is the random (true) total consumption and \(w_i=(w_{ij})_{j\in\overline{J}}\) is the real-time profit vector of prosumer \(i \in I\).

Let \(\mathbb{1}(\mathrm{d}^{\prime}_{i},\mathrm{d}_{-i},D)\) denote the vector with components the indicator functions \(\mathbb{1}(\rho\in \ifmmode\mathrm{B}\else B\fi_j)\),  \(j\in \overline{J}\). Then, on the basis of 2, at reported profile \((\mathrm{d}_i,\mathrm{d}_{-i})\), prosumer’s utility reads \[\begin{align} U^{\alpha}_{i}\big( \mathrm{d}_{i},\;\ifmmode\mathrm{M}\else M\fi^{E}(\mathrm{d}^{\prime}_{i},\;\mathrm{d}_{-i})\big) = U_{i}\big( \mathrm{d}_{i},\;\ifmmode\mathrm{M}\else M\fi^{E}(\mathrm{d}^{\prime}_{i},\;\mathrm{d}_{-i})\big)+ \big\langle w_i,\; \mathbb{1}(\mathrm{d}^{\prime}_{i},\mathrm{d}_{-i},D)\big\rangle\;\mathbb{1}_{A_i}(\mathrm{d}_{-i}). \end{align}\] The term \(\mathbb{1}_{A_i}(\mathrm{d}_{-i})\) encodes the allocation event of \(i\in I\) when reporting \(\mathrm{d}_i\) (see (24 )), though, under 6 it can be omitted and hence its precise mechanics are not important to the following discussion.

Information used by the designer. Prosumer \(i\in I\) reports \(\mathrm{d}^{\prime}_i\) so as to maximize their expected utility. The designer observes others’ reports too. Thus, when conditioned on the others’ reports, prosumer \(i\) submits \(\mathrm{d}^{\prime}_{i} \in [0,\mathrm{d}_i]\) and shares their belief \[\begin{align} \ifmmode\mathrm{P}\else P\fi( \mathrm{d}^{\prime}_{i},\mathrm{d}_{-i})\equiv \mathbb{E}\left[\mathbb{1}( \mathrm{d}^{\prime}_{i},\mathrm{d}_{-i},D)\big|\;\mathrm{d}_{-i}\right] \in \Delta^{|\overline{J}|},\label{eq:playeri-belief} \end{align}\tag{50}\] about the imbalance with the mechanism. In expectation over the true total consumption given the others submissions, prosumer \(i\) reports \(\mathrm{d}_i\) so as to maximize \[\begin{align} \label{eq:pro-in} \mathbb{E}_{\sim D|\mathrm{d}_{-i}}U^{\alpha}_i( \mathrm{d}_i,\mathrm{d}_{-i},D)\equiv \mathbb{E}_{\sim D|\mathrm{d}_{-i}} U_i(\mathrm{d}^{\prime}_{i},\mathrm{d}_{-i},D)+\left\langle w_i,\;P( \mathrm{d}^{\prime}_{i},\mathrm{d}_{-i})\right\rangle. \end{align}\tag{51}\] According to (51 ), prosumer’s incentives are forced by their profit vector and there is no reason to expect them to report their belief truthfully.

The work of [18] is one of the earliest that formally links the value of information to the convexity of a function, laying the groundwork for Bregman divergences. Then, [19] shows that every strictly proper scoring rule is expressed via the supporting hyperplanes of a strongly convex function.

That said, on the basis of the available information, the designer would like to penalize prosumers’ incentives via a mechanism constructed from a family \(\phi(\cdot,\;\boldsymbol{\gamma}): \Delta^{|\overline{J}|}\to \mathbb{R}\) of smooth, \(\mu(\gamma)\)-strongly convex (with respect to \(\Vert \cdot \Vert_{p}\) for some \(p \geq 1\)) real-valued functions parameterized by \(\boldsymbol{\gamma} \in \Gamma_{\phi}\).

Let \(S_{\phi(\gamma)}\) be the corresponding scoring rule. The vector \(\gamma\) controls the modulus of strong convexity of the function. Following 10, a scoring rule is maximized if and only if the player submits their subjective belief truthfully and hence it can be deployed as a properly \(\gamma-\)tuned penalty to incentivize the player to report truthfully.

Although the NEMO has access to the player’s past reports, which could potentially be used to form a probabilistic forecast the player, this is not the case for their true reports, as indicated by the observability constraints (see 4.1). Hence, a scoring-rule penalty that evaluates a player’s report solely on the basis of their forecast cannot be constructed.

What can be done instead is to score (50 ) against a distribution at zero that expresses the true (competitive) imbalance (see 3.1). However, by doing so, even if player \(i\) report truthfully, (50 ) could be still far from zero because of the others \(-i\).

We overcome this difficulty be designing a contrastive scoring rule penalty as follows:

4.2.1 Contrastive Scoring Rule-Based Payment↩︎

Given the past reports of player \(i\), \(\mathrm{d}_i\), the past reports of the others (available information see 4.1) \(\mathrm{d}_{-i}\), and the past consumption realizations \(D\), let \(\lambda=\frac{|I|-1}{|I|}\) and \(D_{-i}\equiv \lambda D\) be an approximation of the actual consumption associated with all players except player \(i\). This approximation becomes more accurate whenever players’ true demands are close to each other. Let \(\rho_{-i}\equiv D_{-i}-\sum_{j\neq i}\mathrm{d}_{j}\) be the imbalance associated with players \(-i\) only and \(\mathbb{1}_{-i}(\mathrm{d}_{-i})\) the vector with components the indicator functions \(\mathbb{1}(\rho_{-i}\in \ifmmode\mathrm{B}\else B\fi_{j})\), \(j\in \overline{J}\).

On the basis of \(\rho_{-i}\) the NEMO can form the probability mass function (PMF) about how the rest of the players, \(-i\in I\), predict their corresponding imbalance. Let \[\begin{align} \mathrm{P}_{-i}(\mathrm{d}_{-i})\equiv \mathbb{E}\big[ \mathbb{1}_{-i}(\mathrm{d}_{-i};\;\lambda)\big| \mathrm{d}_{-i} \big] \label{eq:32all-but-i} \end{align}\tag{52}\] be this PMF which is either communicated to or can be easily obtained by player \(i\) again, based on past total consumption realizations, past reports and the coefficient \(\lambda\) which is communicated from the mechanism to the player.

Then, both forecasts are evaluated against a PMF centered at zero or more precisely, against a PMF for which all the mass is concentrated, in the first bin (one-hot vector with the first entry being the non-zero entry). Let \(Q\) be this PMF. The designer scores both (50 ) and (52 ) against \(Q\) and forms the contrastive loss \[\begin{align} \label{incentive-structure-scoring-rules-0} D_{\gamma_i}\big(\mathrm{d}^{\prime}_{i},\;\mathrm{d}_{-i}\big)\;\equiv\; S_{\phi( \boldsymbol{\gamma}_i)}\big(\ifmmode\mathrm{P}\else P\fi_{-i}(\mathrm{d}_{-i}),\;Q\big)-S_{\phi(\boldsymbol{\gamma}_i)}\big(P(\mathrm{d}^{\prime}_{i},\mathrm{d}_{-i}),\; Q\big). \end{align}\tag{53}\] Although the actual total consumption is observable, it is not needed to forming the above difference. The mechanism announces to all consumers (including prosumers) of the day-ahead market the following scoring-rule-based payment \[\begin{align} \label{incentive-structure-scoring-rules} \Pi_{\gamma_i}\big(\mathrm{d}^{\prime}_{i},\;\mathrm{d}_{-i}\big)\equiv \big[ D_{\gamma_i}\big(\mathrm{d}^{\prime}_{i},\;\mathrm{d}_{-i}\big) \big]_{+}+\zeta_i, \end{align}\tag{54}\] that acts as a penalty charged to player \(i\in I\) when reporting \(\mathrm{d}^{\prime}_{i}\) while others report \(\mathrm{d}_{-i}\). Here, \(\zeta_i\) is a real number and by \([\cdot]_{+}\) we mean positive part.

According to (54 ), whenever the forecast constructed from all players performs poorly relative to \(Q\), while the forecast based on all players except \(i\) accurately predicts their share of the imbalance (demand), the discrepancy (53 ) is positive and is attributed to player \(i\) ’s report. In this case, player \(i\) is deemed responsible for the degradation in predictive accuracy and should therefore be heavily penalized. In the case where the player contributes more to forecast accuracy than the rest of the players, expression (53 ) is negative, and the player receives payment equal to \(\zeta_i\).

Prosumer’s utility at profile \((\mathrm{d}^{\prime}_{i},\mathrm{d}_{-i})\) under mechanism \(\ifmmode\mathrm{M}\else M\fi^S\) reads \[U_i^\alpha\big(\mathrm{d}_i, \ifmmode\mathrm{M}\else M\fi^{S}(\mathrm{d}^{\prime}_{i},\;\mathrm{d}_{-i})\big)\equiv U_i^\alpha\big(\mathrm{d}_i, \ifmmode\mathrm{M}\else M\fi^{E}(\mathrm{d}^{\prime}_{i},\;\mathrm{d}_{-i})\big)-\Pi_{\gamma_i}(\mathrm{d}^{\prime}_{i},\mathrm{d}_{-i}),\] where \(\alpha\geq 0\).

Bregman Divergence Properties. We recall two standard properties of Bregman divergences that will be used latter when we prove the Theorem of this work (see 3).

At this point, we assume that \(P_i( \mathrm{d}_i,\mathrm{d}_{-i})\) lies between \(Q\) and \(P_i(\mathrm{d}^{\prime}_{i},\mathrm{d}_{-i})\) in the sense that the player’s more honest report is closer to the truth \(Q\) than the manipulated report.

This says that the direct divergence from \(Q\) to \(P(\mathrm{d}^{\prime}_{i},\mathrm{d}_{-i}))\) is at least as large as the two-step path \(Q \to P(\mathrm{d}_i,\mathrm{d}_{-i}) \to P(\mathrm{d}^{\prime}_{i},\mathrm{d}_{-i}))\). In particular, it implies \(D_\phi(Q\|P(\mathrm{d}^{\prime}_{i},\mathrm{d}_{-i}))) \geq D_\phi(Q\|P(\mathrm{d}_i,\mathrm{d}_{-i}))\): the manipulated report \(P(\mathrm{d}^{\prime}_{i},\mathrm{d}_{-i}))\) incurs a strictly larger penalty than the more honest report \(P(\mathrm{d}_i,\mathrm{d}_{-i})\).

Observe that  10 is equivalent, via the three-point identity (see 4), to: \[\big\langle \;\nabla \phi(P(\mathrm{d}_i,\mathrm{d}_{-i})) - \nabla \phi(P(\mathrm{d}^{\prime}_{i},\mathrm{d}_{-i})),\, Q - P(\mathrm{d}_i,\mathrm{d}_{-i}) \;\big\rangle \;\geq\; 0,\] which says that the direction from \(P(\mathrm{d}_i,\mathrm{d}_{-i})\) to \(P(\mathrm{d}^{\prime}_{i},\mathrm{d}_{-i})\) in gradient space is opposed to the direction from \(P(\mathrm{d}_i,\mathrm{d}_{-i})\) toward truth \(Q\).

4.3 Objectives↩︎

Having designed the class our mechanism \(\ifmmode\mathrm{M}\else M\fi^S\) we are now in a position to state the objectives we want it to satisfy.

4.3.1 Approximate Bayessian Incentive Compatibility↩︎

As we already mentioned in 3.1, (21 ) is truthfully implemented if, under public valuations, costs and nominal capacities, the day-ahead mechanism is truthfully implemented. Therefore, our focus is to truthfully implement (22 ). Our implementation is such that, given everyone’s beliefs about uncertainty, the players’ play leads the mechanism to the competitive benchmark. Hence, our objective is \(\varepsilon\)-BIC: a single \(\varepsilon\) that bounds the gain from misreporting for every player in the (demand side of the) market.

As established by 1 and 2, the standard mechanism \(\ifmmode\mathrm{M}\else M\fi^E\) achieves \(\varepsilon-\)BIC for consumers with vanishing \(\varepsilon\) as the market grows, but fails for prosumers whose gain is bounded below by \(\varepsilon^{\alpha}_i > 0\), which as we saw, is independent of the day-ahead market size.

The following Theorem shows that the scoring rule mechanism \(\ifmmode\mathrm{M}\else M\fi^{S}\) reduces the prosumer’s gain to at most \(\varepsilon(\vert I \vert)+\varepsilon^S\) where \(\varepsilon(|I|)\) vanishes with market size and \(\varepsilon^S\) is a design constant controlled by the strong convexity parameter \(\mu(\gamma)\) and can be set less than the lower bound \(\varepsilon^\alpha_i\).

3 provides a uniform bound on the gain from misreporting for all players under the penalized mechanism \(\ifmmode\mathrm{M}\else M\fi^S\). For consumers, the bound is \(\varepsilon(|I|)\), which vanishes with market size — the same as under \(\ifmmode\mathrm{M}\else M\fi^E\). For prosumers, the bound has an additional term \(\|w_i\|_q^2/(2\mu)\) that reflects the real-time profit. This term does not vanish with market size, but the designer controls it through the strong convexity parameter \(\mu\). A natural question is whether \(\mu\) can be chosen large enough to make this term smaller than the prosumer’s gain under the original mechanism \(\ifmmode\mathrm{M}\else M\fi^E\). We first prove  3 and then state a Corollary (see  1), which confirms that such a choice is always possible.

3 establishes that the prosumer’s gain under \(\ifmmode\mathrm{M}\else M\fi^S\) is at most \(\varepsilon(|I|) + \|w_i\|_q^2/(2\mu)\).  2 established that prosumer’s gain under \(\ifmmode\mathrm{M}\else M\fi^E\) can be at least \(\varepsilon^{\alpha}_i > 0\). The penalty is effective if the upper bound under the new mechanism is strictly below the lower bound under the old mechanism. We now state that the designer can always achieve this.

Of course, \(\mu\) can be increased further to infinity but as we are going to see in the next section, \(\mu\) poses a trade-off between eliminating prosumer’s incentives to under-reporting and a large penalty when the same prosumer happens to report truthfully.

4.3.2 Approximate Penalty Guarantee for Honest Participants↩︎

Ideally, we would like the mechanism to impose zero expected penalty on any player who reports truthfully: \[\begin{align} \label{eq:tax-free-objective} \mathbb{E}_{\mathrm{d}_{-i}\sim P_{-i}}\;\Pi_{\gamma_i}\big(\mathrm{d}_{i},\;\mathrm{d}_{-i}\big) \;=\; 0, \end{align}\tag{60}\] for all true demands \(\mathrm{d}_i\) and all \(i\in I\). That is, an honest participant should expect zero penalty in expectation over the others bid. To achieve (60 ) exactly, the designer would set \(\zeta_i = -\mathbb{E}_{\mathrm{d}_{-i}}\big[D_{\gamma_i}(\mathrm{d}_i, \mathrm{d}_{-i})\big]_+\), which offsets the expected penalty at the truthful report. However, this requires knowledge of the true demand \(\mathrm{d}_i\), which is private information. If the designer knew \(\mathrm{d}_i\), the contrastive penalty structure would be unnecessary — one could simply score each player’s report against a belief constructed from the true demands. The entire motivation for the contrastive nature of the penalty mechanism is precisely that \(\mathrm{d}_i\) is not observed (see 4.1).

We resolve this issue with a simple approximation. The designer does not observe individual true demands, but does observe the total realized demand \(D = \sum_{i \in I} \mathrm{d}_i\) ex post. In a large market with comparable consumers, the average demand \(D/|I|\) is a reasonable proxy for each player’s true demand. Then, the designer sets the offset as \[\begin{align} \label{eq:zeta-approx} \zeta_i \;=\; -\mathbb{E}_{\mathrm{d}_{-i}\sim P_{-i}}\big[D_{\gamma_i}(D/|I|,\; \mathrm{d}_{-i})\big]_+, \end{align}\tag{61}\] which is computable only from observable quantities. Under this choice, the expected penalty at the truthful report \(\mathrm{d}_i\) is no longer exactly zero, but approximately zero: \[\begin{align} \label{eq:approx-tax-free} \mathbb{E}_{\mathrm{d}_{-i}\sim P_{-i}}\;\Pi_{\gamma_i}\big(\mathrm{d}_{i},\;\mathrm{d}_{-i}\big) \;=\; \mathbb{E}_{\mathrm{d}_{-i}}\big[D_{\gamma_i}(\mathrm{d}_i, \mathrm{d}_{-i})\big]_+ \;-\; \mathbb{E}_{\mathrm{d}_{-i}}\big[D_{\gamma_i}(D/|I|, \mathrm{d}_{-i})\big]_+. \end{align}\tag{62}\]

The penalty trade-off for honest participants.

From 3, the prosumer’s gain from misreporting is at most \(\varepsilon(|I|) + \|w_i\|_q^2/(2\mu)\), which decreases as \(\mu\) grows. From 5, the penalty at truthful reporting is at most \(\kappa_{\phi}\cdot\mu\cdot\mathbb{E}\|P - \tilde{P}\|_1\), which increases as \(\mu\) grows. The two bounds move in opposite directions: a large \(\mu\) eliminates the prosumer’s incentives, but when the same prosumer happens to report truthfully, the large \(\mu\) amplifies the penalty they pay for honest participation. We now study this tradeoff and characterize the value of \(\mu\) that best balances the two.

Consider a fixed prosumer \(i\) with profit vector \(w_i\) and a fixed penalty parameter \(\mu\). The same \(\mu\) governs the mechanism regardless of whether the prosumer underreports or reports truthfully.

The tradeoff. Define the following shorthands: \[\begin{align} R(\mu) \;\equiv\; \frac{\|w_i\|_q^2}{2\mu}, \qquad T(\mu) \;\equiv\; \kappa_{\phi}\cdot\mu\cdot\mathbb{E}\big\|P - \tilde{P}\big\|_1, \end{align}\] where \(\tilde{P} = P(D/|I|, \mathrm{d}_{-i})\). Here \(R(\mu)\) is the residual gain from under-reporting (excluding \(\varepsilon(|I|)\)) and \(T(\mu)\) is the penalty at truthful reporting. The function \(R\) is decreasing in \(\mu\) and \(T\) is increasing in \(\mu\). The designer faces a tradeoff: increasing \(\mu\) tightens the bound on under-reporting but loosens the bound on truthful reporting. The value of \(\mu\) that balances the two objectives is obtained by setting \(R(\mu) = T(\mu)\). Solving for \(\mu^{\star}\) yields \[\begin{align} \label{eq:optimal-mu} \mu^{\star} \;=\; \frac{\|w_i\|_q}{\sqrt{2\,\kappa_{\phi}\cdot\mathbb{E}\big\|P - \tilde{P}\big\|_1}}. \end{align}\tag{68}\] At this value, the residual gain from misreporting and the penalty at truthful reporting are both equal to \[\begin{align} \label{eq:common-value} R(\mu^{\star}) \;=\; T(\mu^{\star}) \;=\; \|w_i\|_q\;\sqrt{\frac{\kappa_{\phi}\cdot\mathbb{E}\big\|P - \tilde{P}\big\|_1}{2}}. \end{align}\tag{69}\]

In summary, no fixed \(\mu\) can achieve both objectives perfectly: deterring under-reporting requires large \(\mu\), protecting truthful reporting requires small \(\mu\). The optimal \(\mu^{\star}\) balances the two, and the resulting common bound vanishes in large markets.

6pt

Large market regime. The common value (69 ) is the product of two factors: the prosumer’s profit norm \(\|w_i\|_q\) and \(\sqrt{\kappa_{\phi}\cdot\mathbb{E}\|P - \tilde{P}\|_1/2}\). The first is a fixed property of the prosumer. The second vanishes in the large market regime where players are comparable in size, because \(D/|I| \to \mathrm{d}_i\) implies \(\mathbb{E}\|P - \tilde{P}\|_1 \to 0\). Therefore, at the optimal \(\mu^{\star}\), both the gain from under-reporting and the penalty at truthful reporting vanish as the day-ahead market grows.

4.3.3 Budget Balance.↩︎

Baseline day-ahead and real-time energy settlement are jointly budget balanced by construction under uniform pricing and physical balance constraints (see (20 )). The introduction of the settlement adjustment \(\Pi_i\) generally generates additional revenue that needs to be balanced.

The TSOs are allowed to develop an additional settlement mechanism separate from imbalance settlement to recover procurement costs of balancing capacity, administraitive costs, and other balancing-related costs (subject to RNA approval) [61]. On top of that, any net financial outcome must be passed on to network users (so extra revenue must ultimately be returned / offset somewhere) [61].

Thus, to preserve feasibility of the overall institution, we impose budget balance through a neutrality account. Specifically, we introduce an account \(0\) with transfer \[t_0\equiv \sum_{i\in I}\Pi_i\] and require the accounting identity \[\begin{align} \label{o3} \sum_{i\in I}t_i^{DA,\text{new}}(\mathrm{d}) +\sum_{j\in\overline{J}}t_k^{RT}(D,\mathrm{d}) +t_0(\mathrm{d})=0 \end{align}\tag{70}\] to hold for every realization-type profile \((D,\mathrm{d})\).

5 Numerical Results↩︎

5.1 Simulation Setup and Synthetic Results↩︎

We isolate the three qualitative phenomena presented in Sections 3-4 in a constructed synthetic market that focuses only on the incentive question.

5.1.1 Simulation Setup↩︎

The day-ahead and real-time merit-order supply curves are fixed. Every player holds a uniform belief over private demand of common width \(l\). We focus on player \(i\in I\) with true demand \(\mathrm{d}_i\), fixed across market sizes. This allows the comparisons to be made at a common baseline.

In every experiment we evaluate, on a grid of candidate reports \(\mathrm{d}^{\prime}_i\in [0,\mathrm{d}_i]\), the expected gain from under-reporting, \(\mathbb{E}_{\mathrm{d}_{-i}\sim P_{-i}} \left[ G_i(\mathrm{d}_i,\;\ifmmode\mathrm{M}\else M\fi(\mathrm{d}^{\prime}_i,\mathrm{d}_{-i})) \right],\) with the expectation taken by Monte Carlo. The four experiments differ only in which mechanism \(\ifmmode\mathrm{M}\else M\fi\) is in force and which structural parameters are varied. Each subsequent experiment adds one ingredient on top of the previous one, so that the figures read progressively.

Units. Throughout, energy quantities (\(\mathrm{d}_i\), \(\mathrm{d}'_i\), \(D\), \(\overline{g}_j\)) are in \(\mathrm{MWh}\) delivered over the one-hour delivery period, and prices (\(p^{DA}\), \(p^{RT}\), \(v_i\), \(c_j\), \(\overline{c}_j\)) are in $/\(\mathrm{MWh}\). Each term of the per-period utility is therefore a product \(\$/\mathrm{MWh} \times \mathrm{MWh} = \$\), which we read as a per-hour dollar amount, i.e.\(\$/h\). Thus the gain and expected gain inherits the same units. That said, all payoff and gain axes in the figures below are reported in \(\$/h\).

5.1.2 Synthetic Results↩︎

1 — vanishing consumer gain. We consider a family of energy markets indexed by the number of consumers \(|I|\). All players are consumers with \(\alpha_j=0\) for all \(j\in \overline{J}\). The demand of player \(i\in I\), \(\mathrm{d}_i\) is held fixed and the day-ahead nominal generation capacities are rescaled across market sizes so that the truthful clearing always lands on a common merit-order supply point. Recall (30 ) under 6. The intuition is as follows: For an under-report to be profitable, it must shift the day-ahead clearing price which requires \(\sum_{j\neq i}\mathrm{d}_j\) to fall within a strip of width \(\mathrm{d}_i-\mathrm{d}^{\prime}_i\) around a supply-curve step. 1 bounds the probability of that event by \(O(1/\sqrt{|I|})\), so the expected gain inherits the same rate.

3 confirms the predicted decay. The left panel shows that the sign of the gain is driven by the under-report cost: every curve rises toward zero from below as \(\mathrm{d}^{\prime}_i\) tends to \(\mathrm{d}_i\) (\(\mathrm{d}^{\prime}_i/\mathrm{d}_i\to 1\)), and the family shifts upward as \(|I|\) grows. The right panel removes the report dimension by taking for each \(|I|\), the maximum of the gain over non-truthful reports and comparing it against the theoretical bound reference. For small markets the maximum is positive but small: a narrow interior band of under-reports still yields profit because the probability of shifting the clearing price is high enough to beat the residual. As \(|I|\) grows, that band shrinks at rate \(1/\sqrt{|I|}\) and eventually disappears for \(|I|=50\); the curve crosses below the reference and into the negative region. Intuitively, a single consumer’s ability to move the market price decreases with scale, and once it has decreased enough, the residual forces every misreport into strict loss.

2 — persistence of the prosumer gain. We repeat the previous experiment, but now player \(i\) is endowed with an ownership vector \(\alpha_i\) on the simplex of real-time producers and with nominal capacities \(\overline{g}\). The remaining players stay consumers with \(\alpha_j=0\). The gain under \(\ifmmode\mathrm{M}\else M\fi^{E}\) conists of the previous consumer term and an real-time-share term (see (38 )). The explanation is that, by under-reporting, the prosumer pushes a positive imbalance into real time. Once that imbalance crosses the next capacity threshold of the real-time merit order, her own asset is activated and collects the cost gap between it and the preceding unit. This profit depends only on the real-time supply curve. The convex combination \(\alpha_i\), and the nominal generation capacities does not vanish with \(|I|\). In every market size, one therefore expects a band of interior under-reports in which the real-time-share channel dominates the under-report cost and yields a strictly positive gain.

4 reproduces 2 visually. The right panel contrasts the two maxima: the consumer curve drops as before, while the prosumer curve stabilizes at a strictly positive level. The left panel explains why: even for large \(|I|\) there is a band of under-reports, located close to truthful (in a neighborhood of \(\mathrm{d}^{\prime}_i/\mathrm{d}_i=0.89\)), on which the prosumer’s real-time-share revenue increases the under-report cost and produces a strictly positive gain. The height of that hump depends on the real-time supply curve, generation capacities and \(\alpha_i\), and hence, it does not shrink with the market size. No single tolerance \(\varepsilon\) can serve consumers and prosumers simultaneouly once \(|I|\) is large enough, and \(\ifmmode\mathrm{M}\else M\fi^{E}\) therefore fails to be \(\varepsilon-\)BIC with vanishing \(\varepsilon\) over the whole player set.

3 — scoring rule restoration. We reuse the market of the previous experiments and augment the day-ahead settlement with the contrastive scoring-rule payment \(\Pi_{\gamma_i}\) of 4. The payment is parameterized by the strong convexity parameter \(\mu\) of the underlying potential function. We sweep \(\mu\) over a set of values while holding all other primitives constant. The prosumer’s gain under the extended mechanism \(\ifmmode\mathrm{M}\else M\fi^S\) equals her gain under \(\ifmmode\mathrm{M}\else M\fi^E\) minus the scoring-rule payment (54 ).

5 shows how the penalty closes the prosumer gap. The interior hump of 4 which was responsible for the persistent prosumer advantage under \(\ifmmode\mathrm{M}\else M\fi^E\), is pushed into the negative region once \(\mu\) is large enough. The right panel gives the quantitative story: the family of green curves traces the 3 bound \(\varepsilon(|I|)+\Vert w_i\Vert^{2}_2/(2\mu)\). The first term decays with \(|I|\) at rate \(1/\sqrt{|I|}\) established for consumer, and the second term is a free parameter that the designer can shrink by increasing \(\mu\). For large \(\mu\), the prosumer bound coincides with the consumer bound, restoring a uniform \(\varepsilon-\)BIC guarantee over the full player set.

Incentive deterence vs. cost of honest reporting. If under the value of \(\mu\) such that prosumers’ gain drops to the consumer bound (5), the prosumer plays honestly under mechanism \(\ifmmode\mathrm{M}\else M\fi^S\), what does it cost them and how much they give up when under-reporting?

6 makes this comparison at every penalty strength a designer might consider, on the same synthetic market that we used in 1-3. The horizontal axis is the penalty strength \(\mu\) which varies from \(0\) up to \(150\) – that is the range that matters for this market. Let \(U^{0}\) be the prosumer’s average payoff (measured in \(\$/h\)) on this market when they report their true demand under mechanism \(\ifmmode\mathrm{M}\else M\fi^{E}\).

It makes sense to use \(U^{0}\) as a reference since it expresses what the prosumer would have earned if the penalty had never been introduced. Hence, any cost or gain the penalty creates is naturally read against it. On top of that, \(U^0\) adapts to the size of the market: if we double every price, every demand, and every producer capacity, \(U^0\) doubles too, and so do the penalty and the under-reporting gain. The percentage of \(U^0\) stays the same, so the numbers we report applies to a market that is either ten times larger or ten times smaller. Lastly, \(U^0\) is the quantity the prosumer actually cares about. The penalty strength \(\mu\) is also in dollars per hour, but it is a setting the designer picks. Comparing \(\mu\) to \(U^0\) via the percentage axis is what makes the figure say something concrete for the prosumer.

5.2 Empirical Evaluation on Real-Market Data↩︎

This section demonstrates the performance of our mechanism by applying it to the strategic component that we infer from real market data. In particular, we work with hourly time series of low, medium and high voltage reported loads (interpreted as three players’ day-ahead reports) found in [62] together with the corresponding realized demand, extracted from an ENTSO-E-based dataset [63] and preprocessed to retain only under-estimated hours in which the aggregate report does not exceed actual total consumption.

On this set we rationalize the day-ahead market reports as subjective equilibria by means of the prosumers valuations, demands, real-time share vectors and real-time nominal capacity bounds that make the observed day-ahead reports approximately optimal with respect to a finite set of breakpoint deviations.

We formulate this inverse problem as a convex feasibility program, relaxed with slacks and recast it as a regularized convex optimization program (inner level), while the nominal real-time capacities are treated as outer-level decision variables tuned by a derivative-free covariance matrix evolution strategy (CMA-ES) search rather than by a large mixed-integer formulation.

We detail the key modeling and algorithmic parameters (valuation floor, slack penalties, regularization weights, quantile supports, and solver choices), and finally present numerical results that compare observed and predicted best-response reports, day-ahead/real-time prices, imbalances, and player utilities under the calibrated model, with and without the effect of our mechanism.

5.2.1 Time Resolution, Units, and Sample Period↩︎

Day-ahead time unit (MTU) in our sample. In our empirical sample (prior to October 2025), the coupled EU day-ahead market (SDAC) clears at an hourly MTU, i.e.a delivery day is partitioned into 24 hourly MTUs. SDAC transitioned from hourly to 15-minute MTUs on trading day 30 September 2025 for delivery day 1 October 2025; this occurs after our sample period. [64], [65]

Imbalance settlement time unit (ISP). At the EU level, EBGL sets the harmonisation objective for a 15-minute ISP. [66]

Model period and units. Because our dataset reports realized load only at the hourly level, we conduct the theoretical and empirical analysis at the hourly resolution. We study a single delivery hour and suppress the hour index whenever no confusion arises. Let \(\tau>0\) denote the duration of one model period in hours; in the baseline specification, \(\tau=1\). All quantity variables are energy quantities in MWh delivered over that hour. For example, \(100\) MW sustained over 15 minutes corresponds to \(25\) MWh, while \(100\) MW sustained over an hour corresponds to \(100\) MWh.

5.2.2 Data and Pre-Processing↩︎

The dataset combines day-ahead reported quantities at three aggregation levels: low-voltage LV, medium-voltage MV, high-voltage HV, and realized total load (actual total system demand) \(D\). In the implementation, these quantities are fetched from an annual ENTSO-E [63] "Total Load" time series and a preprocessed day-ahead market dataset containing the voltage levels forecasts [67]. The data are stored in an array of shape \((\text{days}, 4, 24)\) containing LV, MV, HV, and actual total consumption for each hour of the day, from which we extract a single hour across all days. For that fixed hour, each selected slot \(t\) refers to a day of the year. We define three "players" corresponding to the LV, MV, and HV aggregates, and denote the collection \(\mathrm{d}_{i}^t\) as the observed day-ahead report of player \(i\) at day \(t\), for \(i \in {I}\), \(|I|=3\). The realized total consumption is denoted \(D^t\). We restrict attention to under-estimated days where the reported day-ahead quantities do not exceed the realized average demand. Specifically, we retain only those tuples \(\left(\mathrm{d}_1^t, \mathrm{d}_2^t, \mathrm{d}_3^t, D^t\right)\) that satisfy \(\sum_{i}\mathrm{d}^{t}_{i}\leq D^{t}\) and discard all remaining days. This ensures that the implied day-ahead schedule is physically feasible without upward curtailment and simplifies the real-time modeling: imbalances are always nonnegative, \(\rho^t\geq 0\) and therefore, the real-time market always operates in a "short" regime that activates positive real-time capacities. After this filtering step, we obtain a dataset \(\mathrm{dataset}=\left\{\left(\mathrm{d}_1^t, \mathrm{d}_2^t, \mathrm{d}_3^t, D^t\right)\right\}_{t \in {T}}\), where \({T}\) indexes the remaining days for the chosen time slot.

5.2.3 Inverse Best-Response Model↩︎

Fixing a particular time slot of the day-ahead, we are provided with a sequence of reported profiles and total consumptions \(\{(\mathrm{d}^{t}_{1},\ldots,\mathrm{d}^{t}_{|I|},D^{t})\}^{T}_{t=1}\) over different days \(t\). Such sequences are publicly available by e.g., the Hellenic Energy Exchange [62] and the ENTSO-E Transparency Platform [63]. We assume that each player decides upon the same historical horizon \(H>0\) and maintains beliefs \(P^{t}_{-i}\), and \(Q\). For each \(t> H\), these histograms are constructed by aggregating previously observed samples over a rolling window of length equal to the horizon \(H\), \((\mathrm{d}^{\tau}_{i},\ldots,\mathrm{d}^{\tau}_{|I|},D^{\tau})\), \(\tau=t-H,\ldots, t-1\).

Given observed (from the data) profiles \(\{(\mathrm{d}^{t}_{1},\ldots,\mathrm{d}^{t}_{|I|}),D^{t}\}^{T}_{t=H+1}\), we want to determine, for each player and each day, those valuations, demands and incentive strength (convex combinations) that rationalize the players’ reports as subjective equilibria of \(G^{t}\). Let \(\boldsymbol{v}^{t}=[v^{t}_{1},\ldots v^{t}_{|I|}]\), \(\boldsymbol{\mathrm{d}}^{t}=[\mathrm{d}^{t}_{1},\ldots \mathrm{d}^{t}_{|I|}]\), \(\boldsymbol{A}^{t}=[\boldsymbol{\alpha}^{t}_{1};\ldots;\boldsymbol{\alpha}^{t}_{|I|}]\), where \(\boldsymbol{\alpha}^{t}_{i}=[\alpha^{t}_{i,1},\ldots,\alpha^{t}_{i,|\overline{J}|}]\) is the row vector of shares for player \(i\in I\). Lastly, \(\overline{\boldsymbol{g}}^{t}=[g^{t}_{1},\ldots,g^{t}_{|\overline{J}|}]\) is the vector of nominal capacities of the real time market. The problem we are interested in solving is: For each \(t=H+1,\ldots, T\) \[\begin{align} & \boldsymbol{ Find } \qquad\qquad\left(\boldsymbol{\mathrm{d}}^t, \boldsymbol{v}^t, \boldsymbol{A}^t, \boldsymbol{\overline{g}}^{t}\right) \\ & \boldsymbol{ subject to :} \qquad \mathrm{d}_{i}^t \in \arg \max _{x \in[0, \mathrm{d}_i^t]} U_i^t\left(x, ; v_i^t, \boldsymbol{\alpha}_i^t,\overline{\boldsymbol{g}}^{t}\right), \\ & \qquad\qquad\qquad\quad \sum_{i} \mathrm{d}^{t}_{i} = D^t, \\ & \qquad\qquad\qquad\quad \sum_{j\in\overline{J}} \alpha^{t}_{i,j} = 1,\\ & \qquad\qquad\qquad\quad v^{t}_{i}\geq 0, \\ & \qquad\qquad\qquad\quad \overline{\boldsymbol{\alpha}}^{t}\geq 0, \\ & \qquad\qquad\qquad\quad \overline{\boldsymbol{g}}^{t}\geq 0, \end{align}\quad \forall i\in I.\label{ibr}\tag{71}\] In practice of course, the dominance of the observed report is approximated by enforcing it on a finite grid of deviations \(A_{i,t}\) around the observation.

The feasibility problem (71 ) is infinite-dimensional: the best-response conditions involve expectations with respect to empirical distributions and an \(\arg\max\) over a continuum of actions \(x\in[0,\mathrm{d}_i^t]\), and the real-time capacities \(\overline{\boldsymbol{g}}^{t}\) themselves are additional unknowns. In order to implement it numerically we proceed in two steps:

1. for fixed real-time capacities \(\overline{\boldsymbol{g}}^{t}\) we build a finite set of breakpoint deviations \(A_{i,t}\) and formulate an inner convex program that enforces approximate best-response inequalities and infers \((\boldsymbol{\mathrm{d}}^t,\boldsymbol{v}^t,\boldsymbol{A}^t)\);

2. treating the capacities \(\overline{\boldsymbol{g}}^{t}\) as outer variables, we tune them so that the model’s forward best responses match the observed reports. This outer problem is tackled with a derivative-free CMA-ES search rather than a large mixed-integer program.

5.2.3.1 Inner Level: Finite Deviations and Linear Inequalities.

Fix a finite set of calibration days \({T}^{\mathrm{fit}}\subset\{H+1,\dots,T\}\) and assume that the real-time capacities \(\overline{\boldsymbol{g}}^{t}\) for \(t\in{T}^{\mathrm{fit}}\) are given. For each such day \(t\) and each player \(i\in I\), we construct a finite deviation set \(A_{i,t}\subset [0,D^t]\) which contains the endpoints \(0\) and \(D^t\), day-ahead and real-time breakpoints (values of \(x\) where the day-ahead marginal unit or the real-time marginal genset changes, computed from the day-ahead/real-time stacks and the empirical supports of \(P^t_{-i}\) and \(Q^t\)), and the observed report \(\mathrm{d}_{i}^t\) (used to form utility differences but not treated as a deviation).

For each \((i,t)\) and each deviation \(x\in A_{i,t}\), we compare the expected utility of the observed action \(\mathrm{d}_{i}^t\) and the deviation \(x\), with \(U_i^t(\cdot)\) as in (37 ). The approximate best-response condition is: \[U_i^t( \mathrm{d}_{i}^t; v_i^t,\boldsymbol{\alpha}_i^t,\overline{\boldsymbol{g}}^{t}) \geq U_i^t(x; v_i^t,\boldsymbol{\alpha}_i^t,\overline{\boldsymbol{g}}^{t}) - s_{i,t}(x), \qquad\forall x\in A_{i,t}, \label{eq:br-ineq-raw}\tag{72}\] for some slack \(s_{i,t}(x)\geq 0\) that allows for small violations and model error. By construction, and for fixed capacities and empirical distributions, \(U_i^t(x)\) is affine in the unknowns \((v_i^t,\boldsymbol{\alpha}_i^t)\) and the real-time term is a weighted sum \(\boldsymbol{\alpha}_i^{t\top}\text{(expected RT reward vector at x)}\). Hence there exist coefficients \[A_{i,t}(x)\in\mathbb{R},\quad \boldsymbol{B}_{i,t}(x)\in\mathbb{R}^{|\overline{J}|},\quad b_{i,t}(x)\in\mathbb{R},\] computed from the data and \(\overline{\boldsymbol{g}}^{t}\), such that \[U_i^t( \mathrm{d}^{\prime t}_{i}) - U_i^t(x) = A_{i,t}(x) v_i^t + \boldsymbol{B}_{i,t}(x)^{\!\top}\boldsymbol{\alpha}_i^t - b_{i,t}(x).\] Substituting this into (72 ) yields the linear constraint \[A_{i,t}(x) v_i^t + \boldsymbol{B}_{i,t}(x)^{\!\top}\boldsymbol{\alpha}_i^t \geq b_{i,t}(x) - s_{i,t}(x), \qquad \forall x\in A_{i,t},~\forall i\in I,~\forall t\in{T}^{\mathrm{fit}}. \label{eq:br-ineq-linear}\tag{73}\] We retain the feasibility constraints from (71 ) in a slightly strengthened form: \[\begin{align} &\sum_{i\in I} \mathrm{d}_i^t = D^t, \qquad \mathrm{d}_i^t \geq {\mathrm{d}}_{i}^t, \qquad \mathrm{d}_i^t\geq 0, \quad\forall\, t\in T^{\mathrm{fit}},\tag{74}\\ &\sum_{j\in\overline{J}} \alpha_{i,j}^t = 1, \qquad \alpha_{i,j}^t\geq 0, \quad\forall\, i\in I,~\forall\, t\in T^{\mathrm{fit}},\tag{75}\\ &v_i^t \geq v_{\min}, \quad\forall\, i\in I,~\forall\, t\in T^{\mathrm{fit}},\tag{76} \end{align}\] where \(v_{\min}\) is a valuation floor chosen strictly above the largest marginal cost in the day-ahead stack.

5.2.3.2 Inner Regularized Convex Program.

To select a particular solution and keep the inferred parameters economically reasonable, we minimize the total slack together with small quadratic regularization terms. Introducing positive weights \(\gamma_{\mathrm{slack}},\gamma_{\alpha},\gamma_{v}>0, \gamma_{\mathrm{d}}>0\), we solve the following inner inverse best-response problem: \[\label{eq:inner-opt} \begin{align} \min_{\substack{v_i^t,\mathrm{d}_i^t,\boldsymbol{\alpha}_i^t\\ s_{i,t}(x)\geq 0}} \quad & \sum_{i\in I} \biggl[ \gamma_{\mathrm{slack}}\sum_{x\in A_{i,t}} s_{i,t}(x) + \gamma_{\alpha} \|\boldsymbol{\alpha}_i^t\|_2^2 + \gamma_{v}(v_i^t)^2 + \gamma_{\mathrm{d}} (\mathrm{d}_i^t)^2\bigr) \biggr] \\[2mm] \text{s.t.}\quad & \text{BR inequalities (\ref{eq:br-ineq-linear}) for all }i,t,x,\\ & \text{demand constraints (\ref{eq:cap-constraints}), share constraints (\ref{eq:alpha-constraints}),}\\ &\text{and valuation floor (\ref{eq:vfloor}).} \end{align}\tag{77}\] All constraints in (77 ) are affine and the objective is a strictly convex quadratic function. Thus, for fixed capacities \(\overline{\boldsymbol{g}}^{t}\) and empirical distributions \((P^t_{-i},Q^t)\), problem (77 ) is a convex quadratic program which we solve with off-the-shelf solvers (OSQP, with ECOS and SCS as fallbacks), obtaining inferred primitives \((\widehat{v}^{t}_{i},\widehat{\mathrm{d}}^{t}_{i},\widehat{\alpha}_{i}^{t})\) for each \((i,t)\in I\times{T}^{\mathrm{fit}}\).

5.2.3.3 Outer Level: Calibration of Real-Time Capacities.

In the inner problem, we treated the real-time capacities \(\overline{\boldsymbol{g}}^{t}\) as fixed parameters. The second step is to adjust these capacities so that the model’s predicted best-response bids reproduce the observed ones. Given a candidate collection of real-time capacities \(\overline{\boldsymbol{g}} = \bigl(\overline{\boldsymbol{g}}^{t}\bigr)_{t\in{T}^{\mathrm{fit}}}\), we keep the inner primitives \((\widehat v_i^t,\widehat{\mathrm{d}}_i^t,\widehat{\boldsymbol{\alpha}}_i^t)\) fixed and simulate the subjective games \(G^t\) forward. For each day \(t\) and player \(i\) we compute a soft (quantal) best response \(\mathrm{d}_{i}^{t,\mathrm{BR}}(\overline{\boldsymbol{g}}^t)\) by evaluating \(U_i^t(x)\) on a refined breakpoint grid and applying a softmax over utilities: \[\mathrm{d}_{i}^{t,\mathrm{BR}}(\overline{\boldsymbol{g}}^t) = \sum\nolimits_{x\in A^{\mathrm{sim}}_{i,t}} \pi_i^t(x;\overline{\boldsymbol{g}}^t) x, \qquad \pi_i^t(x;\overline{\boldsymbol{g}}^t) = \frac{\exp\bigl(U_i^t(x; v_i^t,{\boldsymbol{\alpha}}_i^t,\overline{\boldsymbol{g}}^t)/\tau\bigr)}{\sum_{\tilde{x}\in A^{\mathrm{sim}}_{i,t}}\exp\bigl(U_i^t(\tilde{x};\widehat v_i^t,\widehat{\boldsymbol{\alpha}}_i^t,\overline{\boldsymbol{g}}^t)/\tau\bigr)},\] where \(A^{\mathrm{sim}}_{i,t}\) is a simulation breakpoint set (constructed in the same spirit as \(A_{i,t}\)), and \(\tau>0\) is a temperature parameter. As \(\tau\to 0\) the soft best response approaches the hard \(\operatorname{argmax}\), whereas larger \(\tau\) yields smoother responses and better numerical robustness.

We then measure the fit between the model and the data by the mean squared deviation between observed and predicted bids: \[\label{eq:outer-loss} L(\overline{\boldsymbol{g}}) := \frac{1}{|I|} \sum\nolimits_{t\in{T}^{\mathrm{fit}}} \sum\nolimits_{i\in I} \bigl( \mathrm{d}_{i}^{t,\mathrm{BR}}(\overline{\boldsymbol{g}}^t) - \mathrm{d}_{i}^t \bigr)^2.\tag{78}\] The outer problem is then \[\label{eq:outer-opt} \begin{align} \min_{\overline{\boldsymbol{g}}}\quad & L(\overline{\boldsymbol{g}})\\ \text{s.t.}\quad & \underline g_{j}^t \le \overline{g}_j^t \le \overline{g}_{j,\max}^t, \qquad \forall j\in\overline{J},~\forall t\in{T}^{\mathrm{fit}}, \end{align}\tag{79}\] where \(\underline g_j^t>0\) and \(\overline{g}_{j,\max}^t\) are simple box bounds (in the implementation, they are built from a heuristic “suggested” real-time capacity based on the empirical distribution of positive imbalances).

The mapping \(\overline{\boldsymbol{g}}\mapsto L(\overline{\boldsymbol{g}})\) is evaluated via simulation and is nonconvex and nonsmooth. We therefore solve (79 ) with a covariance-matrix adaptation evolution strategy (CMA-ES), acting on the flattened vector of capacities. In the numerical experiments we use a two-stage procedure: a first “cheap” stage with a very short belief horizon (used to warm-start the capacities), followed by a second stage with a longer horizon and refined CMA-ES search.

5.2.4 Results and Discussion↩︎

In this section we answer, slot by slot and player by player, the following question: if we keep the same fitted primitives and switch from the mechanism \(\ifmmode\mathrm{M}\else M\fi^E\) to mechanism \(\ifmmode\mathrm{M}\else M\fi^S\), does the best response move towards truthfulness? If yes, by how much, at what cost, and on which slots? Under \(\ifmmode\mathrm{M}\else M\fi^S\) each player maximizes the same \(\ifmmode\mathrm{M}\else M\fi^E\) utility minus the contrastive penalty \(\Pi_{\gamma_i}\) defined in (54 ).

The first thing to look at is what the players actually report. 7 shows that under \(\ifmmode\mathrm{M}\else M\fi^E\) we see the observed report \(\mathrm{d}_i^{\,t}\) and the fitted true demand \(\widehat{\mathrm{d}}_i^{\,t}\) sitting strictly above it on most slots. That vertical gap is the empirical evidence of strategic under-reporting: the fitted player knows their true demand is \(\widehat{\mathrm{d}}_i^{\,t}\), but, given the real-time profit channel, it is cheaper for them to drop the day-ahead report down to \(\mathrm{d}_i^{\,t}\).

Moving forward, we set the strong-convexity modulus per player to \(\boldsymbol{\mu}=(10^{5},10^{5},10^{6})\), mapped to \(\boldsymbol{\gamma}_i\) via Table 1 under the log potential (logarithmic scoring rule). The spread is asymmetric: a large penalty for LV and MV, and a strong one for HV. This way a single plot shows, what “penalty in between”, and “penalty on” look like on the same real data.

8 shows that under \(\ifmmode\mathrm{M}\else M\fi^S\), on exactly those gapped slots, the best response lifts off the observed curve and travels towards the fitted cap. The mechanism makes the profitable under-report expensive enough that the player prefers to increase their report. On slots where the two \(\ifmmode\mathrm{M}\else M\fi^E\) curves were already close, \(\ifmmode\mathrm{M}\else M\fi^S\) changes nothing, which is the correct behavior of a mechanism that is supposed to activate only when there is something to correct.

8pt

The summary in 2 tells the same story. Under both mechanisms the day-ahead market clears inside the same generation capacity bin, and hence the day-ahead clearing price does not change: the larger reports under \(\ifmmode\mathrm{M}\else M\fi^S\) consume more but remain inside the same bin without crossing the upper step of the bin. The real-time market behaves differently. Under \(\ifmmode\mathrm{M}\else M\fi^E\) the residual that the TSO must cover in real time is large enough to activate the most expensive producer on the real-time supply stack, while under \(\ifmmode\mathrm{M}\else M\fi^S\) that residual is much smaller and either the cheapest real-time marginal cost is activated or no real-time producer is dispatched at all and the real-time clearing price collapses correspondingly. Both effects are consistent with what the theory predicts.

9 completes the picture by showing the real-time imbalance, i.e.how much demand had to be absorbed in the real-time market. Under \(\ifmmode\mathrm{M}\else M\fi^E\) this quantity is positive: that is the residual real-time profit channel 2 tells us the mechanism is meant to close. Under \(\ifmmode\mathrm{M}\else M\fi^S\) the same slots show a substantially smaller residual. Price and imbalance curves show that the mechanism does what it is supposed to do overall.

Lastly, the purpose of 10 is to illustrate when mechanism \(\ifmmode\mathrm{M}\else M\fi^{S}\) activates and how much cost it incurs. In order to ease the comparison the top row of 10 reproduces the content of 7. The two rows at the bottom of 10 check the property that matters most for a well-posed mechanism: when a player is honest, the mechanism does not charge them. For each player we plot the per-slot penalty, and we look at what happens on the slots where the observed report already agrees with the fitted truthful demand (so the player is, in effect, honest) versus the slots where the fit detected a meaningful under-report (so the player is not honest). On the honest slots the penalty is flat and near zero across all three players. On the dishonest slots the penalty activates based on how strong the player’s \(\mu\) is: player 1 (LV, \(\mu=10^{5}\)), player 2 (HV, \(\mu=10^{6}\)), and player 3 (MV, \(\mu=10^5\)). The last row shows for each player, how the realized payoff diminishes under the equilibrium profiles played w.r.t.mechanism \(\ifmmode\mathrm{M}\else M\fi^S\) for each time slot. In short, the plots show that dishonest play is what turns the penalty on, and honest play keeps it off and how the payoffs reduce under our mechanism design.

6 Conclusion↩︎

We studied a strategic estimation problem in which an institution—the day-ahead market—aggregates reports from self-interested players in order to recover a latent population-level object: the true distribution of next-day demand. Classical statistical estimation treats data as drawn from nature; in our setting the data are reports chosen by players who respond to the very rules that will process them, and the estimator is therefore endogenous to its own data-generating process.

Empirically, this circularity has a visible signature. Using publicly available data from three European wholesale markets, we documented a persistent, one-sided gap between aggregate day-ahead reports and realized consumption. The gap survives long observation windows and changes in conditions, ruling out prediction noise as the explanation. We argued that it is best understood as the equilibrium response of strategic prosumers to the prevailing settlement rules.

Our theoretical results explain the asymmetry of this response. For players participating on the demand side only—consumers in our model—the gain from misreporting shrinks at rate \(1/\sqrt{|I|}\) in the size \(|I|\) of the day-ahead market, so honesty is an approximate best response in large day-ahead markets. For consumers who also hold real-time generation—prosumers—the gain is bounded below by a strictly positive constant that depends only on the prosumer’s portfolio and the real-time supply stack, and does not vanish as the day-ahead market grows. The two-stage institution therefore cannot be made approximately truthful for everyone simultaneously.

To close this gap we introduced an imbalance-settlement mechanism built on a strictly proper, leave-one-out contrastive scoring rule. The auctioneer constructs two population-level subjective beliefs about the real-time outcome—one using the player’s report and one omitting it—and charges each player according to her marginal contribution to the fidelity of the market’s belief. The mechanism has a single tuning parameter \(\mu\), the strong-convexity parameter of the strongly convex function associated with the scoring rule, which the designer can set large enough to dominate the prosumer’s persistent strategic gain. Increasing \(\mu\) also charges honest players whose private demand happens to lie far from the population, and we characterized the value \(\mu^\star\) that balances these two effects: in a large market the residual incentive to misreport and the residual cost paid by honest participants vanish at the same rate. The mechanism is informationally light: the auctioneer needs nothing beyond what is already observable in the day-ahead stage, and the existing allocation and pricing rules are left untouched—only the settlement transfer is modified.

We then tested the mechanism on real day-ahead and real-time data from a European wholesale market. Once the penalty is in place, reports rise toward true demand, more energy clears in the day-ahead stage, and the real-time imbalance shrinks by roughly an order of magnitude. In our model, the day-ahead clearing price is unaffected because the larger reports remain within the same step of the supply curve; the real-time clearing price however drops sharply because the smaller imbalance no longer requires the most expensive producers to be activated. The empirical picture matches the theory.

References↩︎