Algotrading on Power Markets vs. Human Decision-Making
Posted in   Energyblog   on  June 23, 2023 by  admin0
WITH THE ASTONISHING CAPABILITIES OF CHATGPT, THE DEBATE AROUND THE AUTOMATION OF POWER TRADING (ALGOTRADING) HAS GATHERED STEAM ONCE AGAIN. IN THIS ARTICLE WE EXPLAIN OUR VIEW, NAMELY THAT ALGOTRADING IS ESTABLISHED AND IT IS A QUESTION OF TIME UNTIL MOST TRADING ACTIVITY WILL BE AUTOMATED AND EXECUTED THROUGH BOTS. HOWEVER, IT REMAINS TO BE SEEN HOW AUTOMATION WILL OCCUR, WHO IS BEST PLACED TO EXECUTE IT AND WHETHER THERE IS STILL A PLACE FOR HUMAN DECISION MAKERS IN THE MARKET.


Asking the machine

Clearly this article can only start one way, namely by asking Chat GPT what it thinks. So this is what I prompted:

"write a short blog post with the title "Short-term power trading – have the machines taken over?". Refer to German Power markets on EPEX Spot, Intraday Trading and the relevance of algorithmic trading vs human based decision making."

And this is what I got:

In recent years, the energy industry has seen a significant shift towards algorithmic trading in short-term power markets. The German Power markets on EPEX Spot, for instance, have been dominated by high-frequency trading algorithms, which can execute trades within fractions of a second.
The rise of algorithmic trading has been driven by several factors. First, the increasing availability of real-time market data has made it easier for algorithms to analyze market conditions and identify profitable trading opportunities. Second, advances in machine learning and artificial intelligence have enabled algorithms to learn from historical data and improve their performance over time.
However, the relevance of algorithmic trading versus human-based decision making in short-term power trading is a topic of ongoing debate. While algorithms can process vast amounts of data quickly and efficiently, they lack the human intuition and creativity that can sometimes lead to unexpected market outcomes. On the other hand, human traders may be prone to bias, emotion, and cognitive limitations that can affect their decision making.
Despite these potential drawbacks, the evidence suggests that algorithmic trading is here to stay. Intraday trading, for example, has seen a sharp increase in algorithmic trading activity in recent years, with many market participants using algorithms to optimize their trading strategies and achieve better results.
Ultimately, the key to success in short-term power trading lies in finding the right balance between human expertise and algorithmic efficiency. By combining the strengths of both approaches, traders can gain a deeper understanding of market dynamics and make more informed trading decisions. While the machines may have taken over, the role of human traders in short-term power trading remains essential for navigating complex market conditions and achieving long-term profitability.


Basically, we could end this exercise here, but with some human pride in play, I wish to expand on this view.


The Status Quo: Automation is already the new Industry Standard

Automation on short-term power markets is a well-established practice and not at all a new development. According to EPEX Spot, 84% of trades on its intraday continuous market were completed via API and only 16% via the manual trading front end Comtrader while 60% of volumes where API traded vs. 40% on Comtrader. The difference is due to the fact that Bots (i.e. algorithms) tend to have smaller order sizes than humans using Comtrader and their fingers. In Germany, arguably the most liquid and advanced intraday market, the ratio for volumes was already at 75% API vs. 25% Comtrader. 

This metric is not perfect, as a trade via an API does not necessarily mean that it was executed by a bot, however it can safely be assumed that most of them are. The trend here is very clear and it heavily leans towards automation on continuous markets. However, automation does not at all mean high technical sophistication, let alone trading based on artifical intelligence (AI). 

Automated Execution vs. Autonomous Decision Making

The terminology in this discussion has not been settled, which tends to cause quite a bit of confusion. People speak of automated trading, algorithmic trading (algotrading), autonomous trading, AI based decision making, machine based trading, just to name a few and people mean very different things when they are using these terms. So I will try to distinguish between automated execution and autonomous decision making to get a bit of structure into our thinking. 

Automated Execution in Power Trading is here to stay

Automated execution on short term markets is arguably a large part of all automated trading on EPEX Spot. Automated execution is a tool for a human decision maker to access the market more efficiently. A human intraday trader has to manage 24 hourly and 96 quarter-hourly products with constantly changing bids and offers around very different price levels within each day and trading shift. Her renewable position may change every minute or so, which means she may have to input/change 120 bids and offers every minute. Clearly this is no task for a human. Therefore, most traders use algorithms that automatically execute some/ or most of their orders. The standard automated execution is a position-closer with a set of price limits that try to manage a constantly changing renewable power position. In my view, these bots are nothing more than a tool that is clearly highly useful but in itself not very impressive. The complexity of position closers varies from the complexity of a simple screwdriver to that of a very flexible assembly robot, however in principle they don't do much more than executing the pre-programmed wishes of a person. 

Even when it gets to the automated trading of proprietary strategies or the execution of trading strategies for flexible assets such as batteries, bots are very important, indispensable even, however they fail to impress on any fundamental level. Most bots that trade prop strategies execute trades based on a logic they have been given by a human in some way, such as:

"Buy hour 12, if and( day = sunday, renewable production < 20 GW, demand > 40 GW, price < 60 EUR/MWh)." (just to make sure: no, this is not a strategy that will make you money). 

Automated Execution of Power Trades is not all that fancy

Calibrating a good execution then becomes more sophisticated and fun, when you start thinking about order sizes, timing, bid/offer positioning and various other conditions. Some automated trading cockpits have become so varied and complicated that most of the sales experts will not be able to explain what their functionalities do in practice. 

Also, the hyped bots for trading batteries are usually not the rocket ships they tend to be made out as. Most of them are simple spread-searchers that trade in and out of positions locking in the biggest spread between products with some attached risk-profile. Bear in mind that I am not making an argument against automation. No human could manage to trade a portfolio of say 10 different batteries with different cycling restrictions and optimize them across 120 intraday products with constantly changing prices. This is clearly a job for machines. Bots don't get tired, they don't make calculating mistakes, they can take in vast amount of information simultaneously and they don't ask for salaries (yet). Nonetheless, I am saying that most of these machines are not all that impressive. It will and has replaced a lot of the clicking intraday traders such as myself had to do in the past, but it has not replaced the decision making itself. 

The really impressive application that would and to a large extend will replace human traders is autonomous decision making coupled with automated execution. Autonomous decision making means that I give an AI a task such as "close my position in a way that maximizes revenues over time" and let it take over from there. Even better: "Trade in and out of position with a limit at X MW and a maximum Var of Y EUR and maximize profits". I would feed this AI lots of historic data including prices, trades, order books, weather data, grid data, fuel costs etc. and then just let it train through the past to make decisions in the future. I could also have it watch a profitable human trader and have it train according to her decisions. Having seen AI such as ChatGPT or midjourney do its magic, I am sure that this kind of AI will do its replacing at one time or another, however, I will explain why it may be more difficult than many enthusiasts claim. 

In August 2023, I was invited to talk about algos and bots in short-term power trading by Montel's Energy Podcast. If you're interested in a deep dive on the topic, here you go...

Click the button below to load the content from www.linkedin.com.

Load content

The Hurdles to Algotrading in Autonomous Decision Making

Over the last years, we have tried to use machine-learning for various use cases and have run into problems that are common in the application of this technology. The main problems we have discovered are incomplete and faulty data, changing market environments, outlier market situations, technical errors and a lack of expertise. 

Incomplete and faulty Data

Incomplete and faulty data deserves a post on its own, however I encourage the volues and montels of the world to keep improving their services to make the market machine-learning-ready. However, at this stage, most data analysis problems involve 20% of analysis and 80% of data cleaning and until this is solved, machine learning won't have it easy to find enough useful material to train with.

The Markets be changing

Changing market environments and outliers are probably the biggest problems for autonomous decision making at the moment. I believe that autonomous algotrading can be better compared to autonomous driving than to AI applications such as Large Language Models (LLMs) such as ChatGPT. Chat GPT can train on a stable set of data and then apply it to changing questions. And it is impressive at that. However, in short-term power trading your learning set of data is not at all stable and the world around you changes with each choice you make, such as in driving a vehicle. The Quarterhour 13q4 today is a very different product than the product 13q4 yesterday as technical conditions, demand and supply are different and worst of all: the market has learned from what happened in 13q4 yesterday and may carry this bias into this day. Also, it is difficult for an AI to understand that if gas prices moved 10 EUR within a day, which they did during the recent crisis, the price levels from yesterday are pretty much useless to predict prices today. I am aware that all the Quants are laughing now, because you can fix these issues, but each fix makes the model so much more complex and difficult to manage. In our experience, markets go irrationally crazy at least once a week, and every experienced human trader will tell you that it is not difficult to play out the bots (or algotraders) in these situations.

Handling technical Failures

An underestimated issue for autonomous decision making and unsupervised  algotrading are technical failures. Streaming live data and making decisions based on them is complex enough if everything goes well. But often it does not: Markets have technical issues and break down, websites for essential input data crash, scrapers fail, data pipelines break, trade collection malfunctions. All these processes are annoying for humans but we are quite good and flexible in finding workarounds and solving problems. Any of these issues would be catastrophic for an AI handling a large unsupervised position as it may not be able to pick up the phone and find a way out.

Resources and Expertise

The last major issue I see is simply a lack of expertise and resources. While I see many a machine-learning course on CVs, I believe that the best people in this field are still to be found in Silicon Valley solving other problems than European Power markets. In comparison to the resources going into the Meta-Verse or autonomous driving, no one has seriously attempted implementing AI for short term trading yet, which may be the biggest reason that this market is not yet dominated by one outlying player. If you are that person, feel free to send us an application 😉

If we come back to the autonomous driving metaphor, think about the resources that have gone into solving this exceptionally difficult issue and it still remains somewhat unsolved. Now as a society, we allow every 18 year old after some basic training to drive a car. Trust me, the hurdles to become a trader on power markets are a lot higher. So if we cannot yet solve the decision making problem of an average 18-year old with the use of immense resources, how long will it take to fully replace a sophisticated human decision maker while having invested a fraction of that? 

We believe in Bots and Humans

When it comes to risk management and prop trading, humans still have a role to play. When it comes to managing renewable portfolios or to trading small scale flexibility from batteries and other assets, we believe in automation to a very large degree. To us building bots is not something fancy as it has been part of our everyday practice for the last 5 years. However, we still see situations every single week where the market turns a little crazy and most of the bots are out of their depth. These are the situations where we push the manual-override button and we go into infighting mode. Very much to the financial advantage of our customers as this allows us to offer lower prices on managing their portfolio and higher revenues when it comes to managing their flexibility. If you need help with these issues, you can always rent-a-trader.

We still believe in the interplay of human market experts that have fundamental price views, will securely deal with unforeseen situations and can compose bots that reflect these views in elegant ways. The day that AI takes over most of short-term trading will surely arrive, but human traders are here to fight for another day. Or decade. Time will tell. 


Tags

Algotrading, Bot, Power Trading, Trader


You may also like

{"email":"Email address invalid","url":"Website address invalid","required":"Required field missing"}
>