How did we adapt our Spot forecast to the context?

We have always assumed this, especially as the situation is the same for all, it is complicated to model what will become of the Spot prices D+1 at 8:00 am UTC the day before, in a market context where gas prices vary strongly and in a very brutal way. For this forecast, which was our reference, the mean error (MAE) has nothing to do with our past standards.

However, our impression is that no one is doing much better. On the one hand, traditional models have great difficulty in following the high volatility of gas and its immediate consequences due to their inertia. On the other hand, models using AI find themselves with too little historical data to make the current context their new standard and thus regain performance.

Daily MAE in Euros of one of our old FR price models, calibrated to “the world before”:

Price models FR to D-1 based on EC OP weather forecast

It is quite visual: the 8:00 UTC model was completely devaluing in terms of performance. The 10:15 UTC model is still performing well, benefiting from more accurate gas prices for the next day.

Without revealing confidential elements, using statistical methods giving more weight to recent market variations, we managed to launch – for willing customers – an alternative model optimised on 10 September, more sensitive to peaks and volatility of other energies, to follow much more accurately the new shapes of the price curve.

Price models FR to D-1 based on EC OP weather forecast

The results are very interesting and the MAE much better. Each client will make its own opinion in comparison with its internal models and/or operational constraints, but we are in any case satisfied to have been able to react while fully preserving our 100% AI identity.

To go further, and at the request of some clients, we are currently working on the creation of an alternative directional (binary) model to predict whether the EPEX price will come out higher or lower than the morning market price.

Finally, we will soon integrate an error allocation module into our Spot dashboard, in order to help our clients better understand the strengths and weaknesses of each of our models (EC/GFS OP/ENS).

Extract from a series of models made by our web designer, Léa. The above figures are not realistic.

Do not hesitate to contact us to evaluate our new models!

‍