Using annotations for better forecasting

From our journey to build forecasting as a feature within FlexMeasures: Good forecasting models are usually the result of small iterations.

Here is one of those iterations: the image shows the improvement from telling your model during training (and in production) what days are public holidays.

Forecasts without and with annotations of ascension day — Forecasts without and with holiday annotations used in training (here: Ascension Day)

In this example, energy prices are much lower than on a normal Thursday afternoon. That's because it is Ascension Day's afternoon (and thus much lower energy demand).

The model we trained with knowledge what a holiday is catched this correctly.

Now, integrating a correct list of holidays for all regions of the world (or maybe Europe, to start with), is the next challenge :) But we already solved that, too.

In FlexMeasures, training a forecaster with holidays as regressor is shipping soon (in v1.0, now already on main).