Author: Derek Michalski, Chief Editor.
European power trading is becoming a higher-frequency optimisation problem. The growth of intermittent generation has increased forecast and imbalance exposure, while BESS has introduced new sources of flexibility that can be monetised across energy, balancing and ancillary-service markets. Cross-zonal trading is increasingly shaped by flow-based capacity and congestion, while intraday liquidity and shorter decision windows are placing greater demands on trading desks. At the same time, algorithmic participation is increasing the speed at which market information is incorporated into prices.
The result is a trading environment in which the value of an individual forecast increasingly depends on what happens next: how the position is revalued, whether flexibility is deployed, which market offers the highest expected return and whether the trade can be executed before the opportunity disappears.
This is where the next application of AI is emerging. The focus is shifting from improving individual forecasts to optimising positions across markets, assets and time horizons — and ultimately allowing automated systems to make and execute trading decisions within defined risk parameters.
Forecasting is only the starting point
Trading organisations already have access to sophisticated price, weather, load and renewable-generation forecasts. The differentiating problem is increasingly what the desk does with them.
A revised wind forecast does not automatically determine a trade. The optimal response depends on the existing portfolio position, time to delivery, intraday liquidity, expected balancing prices, transaction costs and the probability of further forecast revisions. The same forecast can therefore produce different decisions for different portfolios.
The relevant AI problem is conditional optimisation: given the current information set and portfolio state, what is the expected value of changing the position, in which market, at what volume and at what price?
That requires probabilistic scenarios rather than a single price forecast, combined with portfolio constraints and execution parameters. The commercial KPI consequently shifts from forecast accuracy alone towards incremental risk-adjusted P&L, imbalance-cost reduction, capture-price improvement and execution quality.
The most accurate forecast is not necessarily the most valuable trading model.
Intraday is where the decision problem becomes acute
The strongest near-term applications sit in intraday and balancing, where the information set changes continuously and the time available to respond is limited.
A renewable portfolio can move materially out of position following a weather update. A plant outage can alter both exposure and expected market prices. A change in interconnector capacity can invalidate a cross-zonal position. A balancing-price signal can change the value of retaining flexibility.
The challenge is not simply to react faster. Excessive reaction creates turnover, market impact and potentially unnecessary exposure. The system needs to distinguish a meaningful change in expected portfolio value from information that does not justify execution.
This makes probabilistic forecasting and sequential optimisation more relevant than point forecasts. The trading system needs to assess the distribution of possible outcomes, the probability of further revisions and the opportunity cost of remaining exposed.
The objective becomes continuous position management under uncertainty: trade, wait, hedge, rebalance or retain optionality according to expected risk-adjusted value.
BESS exposes the limits of conventional automation
BESS provides the clearest example of why AI is moving from forecasting towards decision-making.
Storage can generate revenue from day-ahead and intraday arbitrage, balancing and ancillary services, but available capacity is finite. Every dispatch decision changes the opportunity set for subsequent intervals. A battery discharged into the intraday market cannot simultaneously provide the same capacity to the balancing market.
The central optimisation is therefore the shadow price of flexibility.
What is the marginal value of another MWh of state of charge, or another MW of available capacity, given expected opportunities across all markets?
That calculation needs to incorporate price distributions, state of charge, degradation, round-trip efficiency, cycling limits, ancillary-service prices, balancing expectations and liquidity. It also needs to account for the probability that market conditions will change before the next decision point.
AI can continuously recalculate this optimisation rather than relying on a fixed dispatch strategy.
The problem becomes even more relevant for portfolios combining BESS with renewables and PPAs. Storage can hedge renewable imbalance, absorb deviations from contracted schedules or change the economics of merchant positions. The optimal battery strategy is therefore a function of the portfolio, not simply the battery’s own price signal.
Portfolio optimisation is the larger opportunity
The same problem exists across the wider trading book.
An IPP may combine wind, solar, BESS, PPAs and merchant exposure. A utility may add retail load, conventional generation, structured contracts and cross-zonal positions. These exposures interact continuously.
Optimising them independently can therefore destroy value. A battery may be worth more as a hedge against renewable imbalance than as an arbitrage asset. A PPA obligation can alter the marginal value of generation. An ancillary-service commitment can constrain the energy-market position. Cross-zonal capacity can change the value of holding flexibility in one market rather than another.
The AI opportunity is to treat these exposures as a single constrained stochastic optimisation problem.
That requires an accurate portfolio state. Market prices alone are insufficient. The decision engine needs current positions, contractual obligations, asset availability, BESS state of charge, nominations, risk limits, collateral constraints and execution status.
This makes the integration of AI with ETRM, EMS, market-access and risk systems a front-office issue. An AI model operating on incomplete portfolio data can produce a technically optimal answer to the wrong problem.
Cross-zonal trading adds the network constraint
European power trading increasingly requires the network to be part of the trading model.
Cross-zonal spreads cannot be evaluated independently of transmission capacity, congestion, flow-based parameters, outages and generation patterns. A historical spread can disappear because the underlying transmission conditions have changed.
Market coupling and the development of flow-based allocation and intraday market design increase the complexity of the optimisation problem. For AI, the opportunity is therefore not simply identifying statistical relationships between bidding zones, but evaluating market state, available capacity, portfolio exposure and expected congestion simultaneously.
A system capable of identifying a €20/MWh spread but unable to determine whether sufficient capacity exists to monetise it is a signal generator, not an autonomous trading system.
Algorithmic execution is only halfway there
Algorithmic execution is already established across European wholesale energy markets. The next step is different.
An execution algorithm receives an instruction and determines how to execute it. An autonomous system could determine whether to trade, which market to use, how much exposure to take and which strategy to deploy.
That moves AI from execution into portfolio decision-making.
An autonomous system could continuously monitor the portfolio, reassess market conditions, generate scenarios, select an optimisation strategy and execute within predefined constraints. The technical components already exist separately; the challenge is connecting them into a robust decision loop.
The commercial attraction is continuous portfolio management across markets without requiring a trader to initiate every decision cycle. The risk is that an adaptive system can behave differently from a deterministic execution algorithm when market conditions move outside its training regime.
The new risk: model convergence
AI also creates a market-structure problem.
If trading organisations use similar weather data, price histories, fundamental inputs and model architectures, their strategies can become increasingly correlated. In relatively thin intraday or balancing markets, simultaneous reactions can amplify price movements.
The concern is therefore not only erroneous orders. It is feedback between adaptive trading systems and the market signals on which they operate.
A model may identify a statistically robust relationship in historical data. Once enough automated participants act on that relationship, the resulting market behaviour can change the relationship itself.
This requires continuous monitoring of strategy drift, signal crowding, execution impact, model confidence and deviations from expected behaviour. Back-testing alone is insufficient when the strategy can influence the environment in which it operates.
Data architecture becomes trading infrastructure
The limiting factor for AI deployment may be the trading architecture rather than the model.
Market data, ETRM positions, asset telemetry, weather forecasts, PPA obligations, risk limits and execution status are commonly distributed across different systems. Autonomous decision-making requires a consistent and current representation of the portfolio.
The required architecture is increasingly a closed loop:
market and fundamental data → portfolio state → probabilistic scenarios → optimisation → execution → post-trade attribution → model feedback.
Each component affects the economics. Poor data quality can invalidate the optimisation. Weak execution can destroy the value identified by the model. Poor post-trade attribution prevents the desk from distinguishing forecast skill from optimisation and execution performance.
The defensible advantage therefore lies less in the individual model than in the integration of data, portfolio intelligence, optimisation and execution.
Autonomy needs hard boundaries
The move from recommendation to autonomous execution changes the governance requirement.
A trader can challenge a recommendation. An autonomous system can place the order before a human sees it.
Trading organisations therefore need to reconstruct the portfolio state, inputs, constraints and decision logic behind material trades. They also need deterministic controls around adaptive systems: position limits, price and volume limits, asset constraints, execution controls, kill switches and escalation thresholds.
The AI should optimise within those boundaries. It should not define them.
This becomes particularly important as European regulators increase scrutiny of algorithmic trading and market surveillance under REMIT, including more granular identification of algorithmic activity.
The next trading edge
The competitive advantage will not come simply from predicting the market better than everyone else. Increasingly similar market data and increasingly accessible AI models will make that difficult to sustain.
The advantage will come from making better decisions when everyone has access to similar information.
For one organisation, that may mean reducing renewable imbalance costs. For another, maximising BESS revenue stacking. For another, improving cross-zonal optimisation or automated intraday execution.
The relevant test is commercial: lower imbalance costs, higher capture rates, better BESS utilisation, lower execution costs or improved risk-adjusted returns.
The transition is therefore not from human forecasting to machine forecasting. It is from forecasting prices to optimising positions, from optimisation to execution, and ultimately towards autonomous trading decisions.
The next trading edge will belong to organisations capable of determining continuously what position to hold, where to hold it, when to change it and how much risk to take — while keeping those decisions within physical, liquidity and regulatory constraints.
