The use of algorithmic trading in the power sector has been steadily increasing over recent years and, as more data is made available to inform decision making, this is only set to grow. This approach has taken longer to establish in power than other sectors due to disparate and unfriendly data over different platforms. This has meant it has been difficult to create efficient algorithms and train them, so power trading has historically been reserved for human traders. However, recent improvements in data availability have meant that algorithmic trading is now taking off in the power market.

There are a number of electricity markets where algorithmic trading can be effective, one of the largest being the within-day market. High levels of liquidity, a large number of interdependent factors, and a tendency for volatile prices make this a prime market for algorithmically driven strategies. If the wind suddenly blows more, or a power station develops a fault, this can cause a significant price change and present opportunities that can drive a business case. These volatile swings are only set to grow as we see increasing levels of renewable generation on the system on the route to net zero.

If you are able to predict the direction of the system and the cashout price, you can position yourself to benefit from particularly high or low prices

The most volatile of prices available is known as the ‘cashout’ price, which shows a much greater spread than the day-ahead price (see below). The cashout price represents the cost of balancing the system in a given half-hour. For example, if the wind blows less than expected, National Grid must turn other units up to compensate. In times of severe imbalance, units are likely to set high prices to be turned up knowing that they will be needed. The cost of these actions is then reflected in the cashout price and is charged back to those units that caused or helped the imbalance in the first place. For units that were ‘short’ in this situation, i.e. they generated less than expected, they will be charged the high cashout price. For units that were ‘long’ and generated too much, they actually helped the net imbalance problem, so will be paid this lucrative price.

So if you are able to predict the direction of the system and the cashout price, you can position yourself to benefit from particularly high or low prices. 

In order to provide traders with the information they need to trade the intraday market effectively, LCP built Enact. Enact cleans data from a huge range of sources and brings it into one single platform. Moreover, Enact provides forecasts for demand, wind, Net Imbalance Volume (NIV), and cashout prices to help algorithmic trading. To do this we developed a cutting-edge proprietary algorithm to forecast the NIV of the system for the next 6 hours, on a real-time basis. Rather than rolling forward the errors in the published NIV forecasts (a so-called ‘top-down’ approach), we derive imbalance fundamentally as a combination of every factor affecting it, e.g. wind level, demand, plant outages (a ‘bottom-up’ approach). Certain aspects of this forecast can be modeled through machine learning techniques while others require expert knowledge of the domain to understand issues around minimum run times, plant outages, etc. As this calculation is live, if any market movements occur (e.g. a plant declares an engine is faulty, a European interconnector begins importing, the solar forecast changes, etc) our view of imbalance will update. The result is a forecast that is 74% accurate in predicting system direction 30 minutes ahead of delivery. When combined with our cashout price forecast this allows algorithms to make decisions based on this information. 

To demonstrate this, we built a trading bot to test the accuracy of our NIV and cashout price forecasts in real-time, by comparing our cashout price forecast against the price being traded on the EPEX intraday market 30 minutes before delivery. The bot chooses to simulate a buy or sell action of 1MWh and then rebalance through cashout when the period closes. The bot makes a profit when it correctly predicts which direction it should position itself (long or short), and makes a loss if it makes an incorrect decision. LCP’s trading bot has consistently made an overall profit since it was launched earlier this year. In October 2020 it made a profit of £1.53/MWh by simulating 1200 separate trades. The charts below show this profit cumulatively over each trade and the distribution of profits made on each trade.

With substantial opportunities already available through algorithmic trading, an increase in the number of participants looking to enhance their business cases through efficient trading and the quality of data sources improving all the time, we expect to see the use of algorithmic trading to continue to grow.