Current performance and potential of AI

The concept of artificial intelligence is not new. As early as the 1950s, Alan Turing wondered whether a machine could “think”. Learning algorithms, which are much talked about, were used in the 20th century in the chess match between Deep Blue and Kasparov, for example. What really changes today is #1 the computing power and #2 the amount of data we have.

It is therefore quite logical that AI is the number one priority of all the giants of the Tech industry: Google, Facebook, Apple, Microsoft, Oracle, IBM, Tesla…
Technology, powered by our data, is a bit like their DNA, and even their business model for some.

At the 2018 World Economic Forum, Sundar Pichai (CEO of Google) said: “”AI is probably the most important thing humanity has ever worked on. I think of it as something more profound than electricity or fire.” The extract of the interview is here.

So, in 2021, where do we stand in terms of AI performance, specifically for the concrete applications that interest us: power market forecasts?
And what advances can be anticipated in the medium term (10/20 years) in this field?

Let’s take a look at the historical players in finance, for example. They do not lack data, processing information is even their core business. However, their advances in terms of AI are strangely little known, or communicated by their marketing and communication departments. In terms of fundamentals or price forecasts, no AI-based forecasts have been launched on the market by them.
Good for us, COR-e, that’s probably one of the reasons we exist.

Without naming our smaller direct competitors, we know of perhaps only one or two that actually use AI to model their forecasts. The others are on traditional or hybrid methodologies (if you take, for example, the previous day’s data to guess the next day). The former are manual and therefore extremely costly and time-consuming, while the latter are blinkered, since they ignore many historical factors.

Our bias is to be 100% AI for our forecasts. We do not “cheat” with the previous day’s data to soften our forecasts when they announce large deviations that could surprise. We optimise the algorithms and make decisions to include or exclude certain data or time ranges to improve performance, but it is the AI that calculates the forecast.

In terms of performance, here are some examples on DE and FR in 2019 and 2020:

• Our rate of good direction – which is probably the most important indication for a trader – on a full day spot oscillates between 55% and 60% on average over a year. On futures trends, it can be around 75%.
• Our average absolute error on the daily price calculated the day before (D-1) is below 2€ on FR and slightly above for DE in 2019 and 2020. This average is of course impacted by a few days where prices rise or fall, and where it will be more important to look at the direction to know whether to sell, buy or do nothing. 2021 is around €3, impacted by the exceptional rise in CO2 which has led us to re-optimise our models.

2 years of pure R&D were necessary before we started to market our first predictive models in 2019… Because, contrary to an illusion that we sometimes hear from certain startups, AI cannot be improvised in 3 lines of code.

In addition, our customers now enjoy 3 extremely important advantages:

• Flexibility in adding new datasets or in accepting a few lower quality datasets from time to time (very useful for some countries).
• The level of resources required to update the forecast calculation for all countries in the CWE area. We do it 20 times a day, we could do it every 5 minutes if necessary, by simply increasing our server resources.
• The speed of the calculations: between the reception of a new information, its processing and its restitution in a forecast (not only of the price, but also of the fundamentals), it often takes us less than 20 seconds.

Finally, to take a look at AI from a purely business perspective, we are still in the early stages of the profound changes that AI will bring to our society.