Algorithmic power trading is trading where software, not a person, decides the details of each order. It decides when to send it, at what price, how much, and what to do with it afterwards. On a power desk the software works from forecasts of weather, demand and prices. It is not the high-speed trading of stock markets. Its edge is that it never misses a deadline and never tires of re-checking a number.
This article is for developers, analysts and new traders who will build, run or work beside these systems. It assumes you know how the auctions work, and gate closure: the last deadlines before delivery. How power trading works covers those for Great Britain, and how European energy trading works for the Continent. Here we follow one British battery through one day of automated trading. Then we look at what stays with people, what has gone wrong, and what the rules ask.
What counts as algorithmic trading
The law gives a precise answer. REMIT is the EU's rulebook for wholesale energy markets. It calls it algorithmic trading when "a computer algorithm automatically determines individual parameters of orders", with little or no human intervention. The parameters include whether to send an order at all, its timing, its price, its size and how it is managed once sent.
A system that only routes orders, or only confirms them, does not count. ACER, the EU's energy market regulator, drew the line in an open letter in July 2024:
| In scope | Out of scope |
|---|---|
| A desk's own trading algorithms | A signal generator, where a trader checks each order before it is sent |
| An exchange's execution tools that slice or re-send orders with no person involved | Standard exchange order types, such as iceberg or stop orders |
| Trading robots bought from a vendor, however simple | Systems that only route, confirm or settle orders |
So a desk that rents a trading robot is trading algorithmically, even if it wrote no code.
Why power desks automate
There is too much to do, too often, for people to do by hand.
- More prices. Great Britain settles power in 48 half-hours a day. Since delivery on 1 October 2025, the Continent's shared day-ahead auction sets 96 quarter-hour prices a day, where it used to set 24.
- More trading close to delivery. EPEX SPOT's intraday markets traded 241 TWh (terawatt-hours; one is a million MWh) in 2025, up from 215 TWh in 2024. EPEX credits part of the growth to algorithmic trading. Nord Pool's intraday volume rose by more than 60% in the same year.
- Forecasts that keep moving. Renewables made 52% of the UK's electricity in 2025. Wind and solar output is known well only hours ahead, so a desk's position changes every time the forecast does.
- Batteries. The UK had 7.5 GW of grid-scale batteries at the end of 2025, 2.3 GW of it connected that year. A battery can buy and sell in every half-hour, and its owner wants it to. Ofgem notes that battery trading generally relies on algorithms.
Each of these multiplies the decisions a desk makes in a day. A trader can watch a few prices closely. Software can watch all of them, every few seconds, and act before each deadline.
The loop every trading algorithm runs
Strip away the detail and a power trading algorithm repeats one loop:
- Read. Take in the latest forecasts, prices, the state of the asset and the trades already done.
- Decide. Work out the position the desk wants for each delivery period, given what it now believes.
- Size. Turn the gap between that and the current position into orders: which product, how much, at what price limit.
- Check. Test each order against the desk's limits before it leaves the building.
- Send. Submit the orders to the exchange, before its deadline, through its interface for automated trading.
- Record. Read back what was filled, update the position and keep a record of why each order was sent.
Then it starts again. A desk running a battery might go round this loop every minute through the day. The hard part is rarely the maths in step 2. It is making every step correct when data arrives late, an order is rejected or a deadline moves.
One day for Northfield Battery
Northfield is an invented trading firm. Northfield Battery is its 100 MW battery in Great Britain. It stores up to 200 MWh, so it can run at full power for two hours. Its round-trip efficiency is 86%: of every 100 MWh it buys, it can sell back 86. The desk's rule for this battery is a position of no more than 100 MW either way in any half-hour.
Every price below is made up for the example. The delivery day is Wednesday 11 November 2026.
Tuesday morning: the plan
The algorithm reads Tuesday morning's price forecast for each of Wednesday's 48 half-hours. Prices are lowest overnight, when demand is low and the wind is blowing. They are highest between 17:00 and 19:00, when people come home.
Its plan is simple. Charge in the four cheapest half-hours. Sell in the four dearest. It can buy 200 MWh, which fills it. The losses come off on the way out: at 86% efficiency it can sell back only 172 MWh. That is 50 MWh in three evening half-hours and 22 MWh in the fourth. It puts the 22 MWh in the half-hour it expects to be cheapest, 17:00.
| Half-hour | Forecast, £/MWh | Plan |
|---|---|---|
| 02:00 | 60 | Buy 50 MWh |
| 02:30 | 58 | Buy 50 MWh |
| 03:00 | 57 | Buy 50 MWh |
| 03:30 | 59 | Buy 50 MWh |
| 17:00 | 105 | Sell 22 MWh |
| 17:30 | 112 | Sell 50 MWh |
| 18:00 | 116 | Sell 50 MWh |
| 18:30 | 108 | Sell 50 MWh |
Tuesday afternoon: the auction
The forecast is only a forecast, so the algorithm does not send orders that take any price. It puts a price limit on each. It will pay at most £80 for the overnight power, and sell the evening power for no less than £95.
Those two numbers protect the margin. Selling at £95 and getting back 86% means each megawatt-hour bought is worth £95 x 0.86 = £81.70 at worst. That is more than the £80 most it will pay. So whatever the auction does, a filled buy and a filled sell never lose money between them. Wear on the battery is a real cost too; a real desk would add it to the sum. Battery storage in Great Britain prices it, with the battery's other markets.
Before sending, the checks run. No half-hour may go beyond 100 MW. The battery cannot sell energy it will not have. No price may be outside a sensible range. The orders pass, and the algorithm sends them to the half-hourly day-ahead auction on EPEX SPOT before it closes at 15:30.
At 15:45 it reads back the results. Every order is filled:
| Half-hour | Auction price, £/MWh | Traded | Cash, £ |
|---|---|---|---|
| 02:00 | 61 | Bought 50 MWh | -3,050 |
| 02:30 | 59 | Bought 50 MWh | -2,950 |
| 03:00 | 58 | Bought 50 MWh | -2,900 |
| 03:30 | 62 | Bought 50 MWh | -3,100 |
| 17:00 | 108 | Sold 22 MWh | 2,376 |
| 17:30 | 115 | Sold 50 MWh | 5,750 |
| 18:00 | 118 | Sold 50 MWh | 5,900 |
| 18:30 | 104 | Sold 50 MWh | 5,200 |
| Total | 7,226 |
It records the trades and updates the position for all 48 half-hours. Then it passes the plan on, so the battery's operator can tell the grid what it intends to do.
Wednesday: the intraday market
Overnight the battery charges as planned. By Wednesday lunchtime the forecasts have moved. A cold evening is now expected to arrive earlier. On the continuous intraday market, 17:00 trades at £126 and 18:30 at £101.
The loop sees an opportunity. The battery holds the same energy either way, but it could sell more of it at 17:00 and less at 18:30. Moving 28 MWh does that:
- sell another 28 MWh for 17:00 at £126: +£3,528;
- buy back 28 MWh for 18:30 at £101: -£2,828.
That adds £700, with the same energy sold. The checks run again. At 17:00 the position becomes 50 MWh sold in the half-hour, which is 100 MW: exactly at the limit, which the desk's rule allows. The orders go in.
There are deadlines here too. In Great Britain the continuous market for a half-hour closes 15 minutes before it starts. Changes to the battery's physical plan must reach the grid an hour before each half-hour, at Gate Closure. The algorithm has to know both, for every half-hour, every day, including the two days a year when the clocks change.
Wednesday, 16:10: when the data stops
Now suppose that at 16:10 on Wednesday the battery trips and stops, and the feed that reports its state freezes at the last good reading. The algorithm's view says the battery is full and ready. In truth it can deliver nothing. Northfield has sold 172 MWh for the evening, as reshuffled at lunchtime, that it cannot produce. If nobody notices, it will pay the imbalance price on all of it.
A well-built algorithm does not trust data just because it arrived once. It checks how old each reading is. When the battery's data goes stale, it stops sending new orders for that asset and raises an alarm. A person then decides what to do: phone the site, confirm the fault, and buy back the evening's sales before each half-hour stops trading. The physical plan for 17:00 is already fixed, because Gate Closure passed at 16:00. The trades still count in settlement. That decision belongs to a person, because the situation is one the algorithm was not built for.
What stays human
Automation changes what people on a desk do. It does not remove them. These stay with people:
- The strategy. What the algorithm is trying to do, which markets it trades, and when it should stop.
- The limits. How much risk it may take: positions, order sizes, price ranges, losses in a day. The algorithm works inside them and never changes them.
- Changes. Every change to the code or its settings is tested and approved before it trades. A small change to a forecast can change every order that follows.
- The kill switch. One action that stops the algorithm sending orders and cancels the ones in the market. Someone must be allowed to press it, and know when to.
- The unexpected. Stale data, a rejected order, an exchange outage, a day nobody tested. The algorithm should notice, stop and tell someone, not guess.
The rule underneath is simple: the algorithm proposes and the desk's rules decide what reaches the market. Neither the algorithm nor the person who wrote it should be able to move the limits on their own.
When it goes wrong
Algorithms that do exactly what they were built to do can still break the rules. Two recent cases show the two ways. Neither is final yet.
Spain: robots built to jump the queue. In October 2024 Spain's energy regulator, the CNMC, fined Gesternova €6 million for market manipulation. It fined a second firm €1.5 million. The trading took place on Spain's continuous intraday market between 30 September and 30 December 2022.
Spanish orders reach the shared European order book through the exchange OMIE's own system. It checked every order message from every firm in one first-in, first-out queue. Trading in next-day contracts opened at 15:10. Under a rule in force until March 2023, orders waiting for later were released at about 22:00, in the order the queue had checked them. At that moment new export capacity to France opened, and French buyers were paying far more than Iberian prices. So a firm's place in the queue at 15:10 decided who sold to France first.
The firms' robots flooded the queue around 15:10 with messages they knew would be rejected: sent too early, too fast or too small. Rejected messages still took their turn, so competitors' orders were checked later. Gesternova sent 74,797 next-day orders, 36.5% of all such orders from 56 firms; 2.2% were accepted. The CNMC put its gain at about €6.83 million, above the €6 million legal maximum fine. The firms argued they were competing fairly for priority. The CNMC held that firms may compete for priority, but not by manipulating the system with false orders. It also held that manipulation needs no effect on price. The decisions are under appeal. Since 4 March 2023 OMIE has ordered its queue by the best price offered instead of first come, first served.
Nothing here was a bug. The robots did what they were built to do.
Finland: a forecast with no human check. On 23 November 2023 Kinect Energy placed a day-ahead sell order in Finland. It was far above what the firm could produce. Staff rolling out a new wind forecasting system had configured one asset wrongly, which multiplied its expected output a thousand times. ACER's account is that the wrong forecast passed into the trading system automatically, and nobody spotted it before the order went to the market.
The order averaged about 5,787 MW in each hour, at any price. Finland's day-ahead price averaged -€203.40/MWh and sat at the -€500 floor for ten hours. Without the order, a re-run of the auction put the price at around €48. In December 2025 Finland's Energy Authority decided that this was market manipulation, because the order gave a false picture of supply. It also found that Kinect's systems lacked the checks a market participant should have. In April 2026 it proposed that the Market Court impose a fine of about €9.25 million. The court has not yet ruled.
ACER's guidance says a firm is responsible for the behaviour of its algorithm, even when the harm comes from poor design. A forecast that flows straight into orders, with no check between them, is the desk's responsibility. A plausibility check on the forecast, or a cap on order size, is the kind of control that catches this.
What the rules ask
In the EU, REMIT Article 5a has applied since 7 May 2024. A firm that trades power or gas by algorithm must:
- have effective systems and risk controls, so its trading stays within limits and does not send erroneous orders or disrupt the market;
- make its systems resilient, with enough capacity, and keep them fully tested and properly monitored;
- have plans for when its systems fail;
- tell its national regulator and ACER that it trades by algorithm;
- keep records of its systems and controls for five years.
In May 2026 ACER reported that only 671 of the 21,169 firms registered under REMIT had declared algorithmic trading. It called that low. From 29 October 2027, the reports of algorithmic orders and trades on organised markets will also carry an Algorithm ID. Our guide to Article 5a covers the duties and records in full, and revised REMIT reporting covers the Algorithm ID.
In Great Britain, UK REMIT has no copy of Article 5a. Investment firms that trade power futures by algorithm must follow the financial markets' rules for algorithms, kill switch included. A firm whose futures trading is only a sideline to its main business can be exempt, but not if it trades at high frequency. Physical day-ahead and intraday power is outside those rules. The ban on manipulation applies to everyone, with or without an algorithm.
What the desk's systems have to do
Behind the loop sits a set of systems, and most of the work is in the joins between them. A desk that trades by algorithm needs to:
- take in forecasts, prices and asset data, and know how old each one is;
- hold one position for every product and every half-hour, which the algorithm, the traders and risk all read;
- run the checks before every order: limits, sizes, price ranges and whether the asset can deliver;
- connect to each exchange's interface for automated trading, and handle rejected orders, partial fills and lost replies;
- record every order with the reason it was sent and the version of the algorithm that sent it;
- stop cleanly when someone presses the kill switch, and show what is left in the market.
Jordan Dimov, who wrote this article, built a cloud platform for automated day-ahead power auctions at Limejump, on EPEX SPOT and N2EX (now Nord Pool). It covered everything from exchange connectivity and bid generation to submission and real-time positions. It was adopted by the whole floor. Choosing software for this? Our guide to energy trading software sets out the parts of a desk's stack and how to choose them.
Building or replacing this on your own desk? It is the work A115 does. See the systems work
What the full guide adds
This article is the short version. The full guide, Algorithmic Trading in Power Markets, builds a working trading algorithm step by step:
- Forecasting as code: a forecast as a range, not a number, and how the range changes the order;
- Sizing the trade: how much to buy when being short costs more than being long;
- Execution: slicing an order against a deadline, partial fills and orders that never come back;
- Trading between markets: spotting when two neighbouring prices have drifted too far apart;
- Backtesting: the three traps that make a strategy look better than it is, and the steps from shadow trading to full size;
- Controls: the checks before every order, the kill switch, and keeping duties apart.
The full guide is in the community on Skool, with the rest of the series, and it is updated regularly as the markets change. If you want to work on a trading desk, read it to understand the subject properly. To join, start at the Energy Trading Fundamentals course page.
Sources
- REMIT, Regulation (EU) No 1227/2011 as amended by Regulation (EU) 2024/1106: Article 2(18), the definition of algorithmic trading, and Article 5a.
- ACER, Open letter on algorithmic trading (30 July 2024): what is in and out of scope.
- ACER, Guidance on the application of REMIT, 6.1st edition (December 2024): a firm is responsible for the behaviour of its algorithm.
- Commission Implementing Regulation (EU) 2026/256: the Algorithm ID field, from 29 October 2027.
- CNMC, resolution SNC/DE/019/23 (3 October 2024) and press release (in Spanish): the Gesternova case.
- ACER, REMIT Quarterly Q1 2026 (the declarations of algorithmic trading; OMIE's queue; the appeals) and Q3 2026 (the Finnish case).
- Energy Authority of Finland, the breach decision (December 2025) and the proposed fine (April 2026).
- EPEX SPOT, Annual Power Trading Results 2025 (19 January 2026), and its Trading Brochure (February 2026): the GB auction times and intraday lead times.
- Nord Pool, yearly volumes 2025 (January 2026).
- Nord Pool, REMIT Best Practice, 3rd edition (April 2026): kill and suspend functions, and checks before each order.
- ACER, 30-minute intraday gate closure across EU borders (December 2025).
- DESNZ, Grid-scale battery storage statistics (Energy Trends, June 2026) and Digest of UK Energy Statistics 2026, Chapter 5.
- Ofgem, Repetitive re-trading by electricity storage in transmission constraint periods (June 2026): battery trading relies on algorithms.
- UK REMIT, Regulation (EU) No 1227/2011 as kept in UK law: it has no Article 5a.
- MiFID II, Directive 2014/65/EU, Article 2(1)(j): the exemption for trading that is ancillary to a main business, lost by high-frequency traders.
- FCA, PS25/24 (December 2025): the UK's ancillary activities test from 1 January 2027.