Functional Architecture of European Electricity Trading Markets: Requirements for AI Supported Trading Systems under Regulatory Constraints

1 Introduction

• European electricity trading is a constrained coordination problem in which legal rules and system physics operate jointly under cross-zonal transfer limits, bidding-zone definitions, and balancing responsibilities.Dispatch feasibility depends on network conditions and market-design parameters.Source: §1 · pp. 1–2

2 Market design

• Forward, day-ahead, intraday, and balancing horizons are linked because positions transferred from one horizon constrain feasible actions in the next.Coupling algorithms apply common cross-zonal capacity inputs at gate closure.Source: §1 · p. 2
• Cross-border implementation faces interface friction from venue-specific access, order handling, fallback procedures, and synchronized coupling timelines despite legal harmonization.The market is also institutionally heterogeneous, with persistent capacity-availability and implementation-quality bottlenecks across jurisdictions.Source: §1 · p. 2, §2 · p. 3

1 Introduction

• Existing AI studies focus on price forecasting across horizons or reinforcement-learning bid construction mainly for day-ahead auctions, without integrating coupling, venue permissions, and compliance controls.The cited strands include Di Persio et al., 2025; Lago et al., 2021; O’Connor et al., 2025.Source: §1 · p. 2
• The paper proposes an operational architecture that maps legal obligations, market-coupling mechanics, and actor permissions into implementable requirements for AI-supported trading.Its research questions cover horizon interactions, binding decision-makers, governance constraints, decision-ready state variables, and auditable AI support.Source: §1 · p. 2

2 Market design

• Day-ahead coupling jointly matches NEMO order books with cross-zonal capacities, while intraday and balancing address forecast errors and residual physical deviations under tighter timing.Intraday trading updates commercial positions close to delivery; balancing resolves deviations remaining after market close.Source: §2 · p. 2

3 Actors and rules

• The rule hierarchy combines sector regulation, market-coupling codes, and exchange rulebooks, linking allocation horizons and balancing obligations.The 2024 reform retained marginal pricing and tightened long-term risk-allocation and consumer-protection instruments.Source: §3 · p. 3
• REMIT integrity and transparency duties and the MiFID II, MiFIR, and EMIR constraints on derivatives make compliance part of executable strategy design, not only post-trade processing.The cited rules require auditable records and constrain conduct, transparency, clearing, and reporting.Source: §3 · p. 3
• Cross-border day-ahead and intraday execution relies on coupled algorithms and synchronized gate windows, while interface boundaries create operational failure points.Distinct operator governance, product eligibility, and procedural timetables require explicit per-venue rules.Source: §3 · p. 3
• Executable actions depend on a permission matrix combining venue membership, product eligibility, gate-closure timing, balancing responsibility, clearing route, and surveillance obligations.Those constraints must be encoded during order construction because downstream repair can cause economic loss or regulatory exposure.Source: §3 · pp. 3–4
• The short-term exchange landscape is plural, and Table 1 compiles major EU exchanges with operator context.The venue map matters because product eligibility, operator governance, and procedural timetables require explicit per-venue encoding even under harmonized coupling frameworks.Source: §3 · p. 3, Table 1 · p. 4

4 Trading workflow

• Rolling portfolio-state construction maps updated load, generation, and availability forecasts against open positions to calculate residual volume risk by delivery period and bidding zone.Long-term transmission-right positions constrain feasible cross-zone hedge translation before day-ahead order construction.Source: §4 · p. 4
• Day-ahead matching jointly combines NEMO order books and cross-zonal capacities at gate closure, producing a coupled allocation rather than an exchange-isolated merit order.Both local bid quality and interzonal transfer limits determine clearing jointly.Source: §4 · p. 4
• Intraday trading repeatedly re-optimizes after day-ahead publication to absorb forecast errors and operational events under strict time priority.As delivery approaches, narrowing correction windows make execution latency and order-book depth binding decision variables.Source: §4 · p. 4
• Balancing settlement resolves residual physical deviations using measured imbalances and applicable pricing rules, with cross-border activation dependent on reserve products and platforms.The passage identifies aFRR and mFRR activation through PICASSO and MARI structures.Source: §4 · p. 4
• Post-trade processing reconciles confirmation, clearing, reporting, and surveillance into a timestamped audit trail linking forecast state, order intent, execution, and settlement exposure.REMIT applies to spot and cross-border wholesale activity; derivatives also face MiFID II, MiFIR, and EMIR requirements.Source: §4 · p. 5

5 Licensing and permissions

• Market entry requires separate legal-regulatory and technical or contractual permission stacks; combining them can cause onboarding failures and compliance gaps at go-live.The distinction is between authorization to operate in a regulatory perimeter and the ability to execute orders in particular markets.Source: §5 · p. 5

5.1 Legal and regulatory prerequisites

• REMIT registration and an ACER code through CEREMP are pre-trade conditions for reportable wholesale energy transactions.Automated execution adds notification and identifier obligations under the revised REMIT framework; power derivatives also bring MiFID II, MiFIR, and EMIR duties.Source: §5.1 · p. 5

5.2 System and venue permissions

• Legal status does not grant venue access: exchange participation, clearing admission, participant agreements, and technical onboarding are distinct requirements.The passage notes that EEX exchange participation and ECC clearing admission must both be in place before production trading.Source: §5.2 · p. 5
• The platform must apply independent legal-regulatory and venue-execution gates to each action candidate; only their intersection defines executable actions.The first tests entity-level admissibility, while the second tests market-, product-, and counterparty-level permissions.Source: §5.2 · p. 5

6 System requirements

• The operational state combines market status, portfolio, cross-zonal capacity, executable price and depth, risk, and compliance variables at decision time.These variables are represented by μ_t, ν_t, κ_t, π_t, ρ_t, and σ_t, respectively.Source: §6 · p. 5
• Residual exposure ξ_t,z is computed per interval and bidding zone and directly drives imbalance exposure across day-ahead, intraday, and balancing decisions.It combines forecast demand, controllable generation, translated forward and derivatives hedges, cleared market positions, and activated balancing volume.Source: §6 · p. 6
• The constrained objective combines execution cost, imbalance risk, and regulatory cost, with transfer-capacity and gate-closure constraints encoding coupling and timing rules.The formulation includes tail risk and prices imbalance risk; scheduled interconnector flow is bounded by available transfer capacity.Source: §6 · p. 6
• An order is releasable only when its executable-action gate is satisfied, linking venue access, permitted products, and active regulatory predicates to the control decision.For candidate i, Ω_i = 1 is required for release.Source: §6 · p. 6

7 Battery storage and temporal arbitrage

• European intraday markets regularly exhibit same-day spreads exceeding 100 EUR/MWh between off-peak surplus and peak scarcity hours.The passage attributes persistent temporal spreads to renewable output variability and merit-order pricing.Source: §7 · p. 6
• A battery captures temporal spreads by charging at low prices and discharging at high prices, net of efficiency losses and degradation cost.The example of a 100 MW / 200 MW h system buying at 20 EUR/MWh and selling at 140 EUR/MWh yields a stated gross margin of 24 000 EUR per cycle before adjustments.Source: §7 · p. 6
• Value stacking combines day-ahead arbitrage, intraday forecast-error correction, and balancing or frequency-response activation payments across horizons.The passage states that this combination can make battery returns exceed single-market projections by a wide margin.Source: §7 · p. 6
• Battery dispatch must jointly account for price expectations, correction opportunities, balancing probabilities, state of charge, and degradation because each commitment changes later feasible actions.Capacity committed to frequency response cannot simultaneously be used for intraday arbitrage; battery dispatch is identified as a primary validation case for the state vector and constrained objective.Source: §7 · pp. 6–7

8 AI architecture

• The policy maps operational state to executable actions while preserving risk and regulatory bounds, with expected utility penalized for tail risk and governance cost.The executable-action constraint links the policy to venue and rule feasibility.Source: §8 · p. 7
• Predictive distributions produce a coverage-controlled set for downstream decision and risk modules, with lower δ increasing forecast protection while reducing expected capture.The stated trade-off affects intraday repositioning aggressiveness.Source: §8 · p. 7
• Execution is fail-closed: an action is released only when profitability, risk, and compliance predicates are simultaneously satisfied, otherwise the null action preserves safety.The market-quality predicate combines depth and latency conditions, with an acceptance threshold.Source: §8 · p. 7
• The override controller switches to a conservative policy when distribution drift or calibration error exceeds supervisory limits.The drift statistic tracks state-distribution shift, while the calibration error and thresholds are used in supervision.Source: §8 · p. 7
• Each decision receives an immutable record of state, forecast, uncertainty set, executed action, feasibility-gate result, active policy, and residual exposure for line-by-line reconstruction.The record is presented as supporting surveillance expectations and documentation discipline aligned with EU AI governance requirements.Source: §8 · p. 7
• Validation compares policies across economic, risk, compliance, and latency metrics and an explicit stress set, accepting a policy only when it improves on baseline while preserving gate integrity and governance thresholds.The scenario basis covers normal operation, congestion, forecast shock, outage, and decoupling fallback events.Source: §8 · pp. 7–8

9 Discussion

• Prior forecasting and reinforcement-learning work improves local components but treats market layers as independent, whereas this architecture jointly models horizon coupling, cross-zonal capacity, and multi-venue permissions.A day-ahead bid’s feasibility depends on forward hedge state, intraday correction capacity, and balancing exposure.Source: §9 · p. 8
• The permission gate moves compliance from post-trade filtering into pre-trade construction, making venue, product, timing, and regulatory requirements hard optimization constraints.The fail-closed logic defaults to inaction when a constraint is violated.Source: §9 · p. 8
• The architecture operationalizes AI governance through auditability, uncertainty calibration, and override controls embedded as system components rather than documentation requirements.The audit record and drift-triggered override are described as executable logic that can be tested, monitored, and verified against regulatory thresholds.Source: §9 · p. 8

• European electricity trading is interpreted as a coupled control problem across forward, day-ahead, intraday, and balancing horizons, with outcomes shaped by capacity constraints, venue timing, and actor obligations.Source: §10 · p. 8
• REMIT, MiFID II, MiFIR, and EMIR map into pre-trade admissibility checks and post-trade accountability, while NEMO designation and passporting explain EU spot markets’ multi-venue topology.Source: §10 · p. 8
• The minimal decision architecture links state representation, residual exposure accounting, constrained optimization, uncertainty calibration, and fail-closed execution control.Source: §10 · p. 8
• The proposed operational framing models the market as one constrained system, encodes permissions before execution, and treats uncertainty and governance as state variables.Source: §10 · p. 8
• AI is proposed as a bounded decision component inside regulated market operation, rather than an autonomous layer outside institutional control.Source: §10 · p. 9

• The framework covers EU wholesale electricity trading but does not claim completeness for national implementation detail.Local market-design differences leave rule interpretation jurisdiction-sensitive at delivery-zone and venue levels.Source: §11 · p. 9
• The operational model abstracts from feed latency, revision-frequency asymmetry, and outage behavior across venues and operators.These factors can dominate theoretical edge in intraday correction windows and degrade policies calibrated on clean historical snapshots.Source: §11 · p. 9
• Evolving AI-governance, market-integrity, and financial-supervision obligations limit the temporal stability of fixed compliance encoding.Production deployment requires continuous legal versioning and rule-to-control trace updates, not one-time certification.Source: §11 · p. 9
• The quantitative formulation still requires empirical validation on synchronized multi-venue event streams with realistic transaction-cost, latency, and imbalance-penalty models.The next stage calls for out-of-sample stress-regime evaluation and controlled ablations of risk, calibration, and governance gates.Source: §11 · p. 9

• All authors contributed equally to the article.Source: Declarations · p. 12
• The research received no external funding, and the authors declare no conflicts of interest.Source: Declarations · p. 12
• Supporting data and code are available from the corresponding author on reasonable request.Source: Declarations · p. 12
• Generative AI tools were used for editorial work such as language editing and formatting, and the authors verified the content and remain responsible.Source: Declarations · p. 12
• The article is a research article published in Swissi AI Journal in 2026 under CC BY 4.0.Source: Declarations · p. 12