Our short-term forecasts for wind, solar, demand, and prices rely on two external weather data providers:

Both providers run four forecasts per day — 00, 06, 12, and 18 — across two model types :

For more details on our raw weather data coverage, see here

Our price forecasting combines machine learning and time series analysis, with a strong foundation in fundamental market drivers such as production, consumption, fuel prices, and interconnector capacity.

To ensure transparency and avoid the “black box effect”, our models operate in two sequential steps:

We first convert raw weather data (temperature, wind speed, solar radiation, cloud cover, etc.) into forecasts of demand, wind production, solar production

This stage captures the physical supply-demand fundamentals that will later influence price formation.

The outputs in MW from the fundamental models are then fed into our machine learning-based price models to forecast 15-minute day-ahead electricity prices (€/MWh, £/MWh for GB). This approach combines fundamental analysis with historical market behavior, effectively capturing the complex interplay of physical system dynamics and structural price signals by combining several key parameters, such as:

From the changes of these different variables, our model is able to predict for each of the future quarter hours a balance between supply and demand on the merit order for the wholesale market.

We use a single coupled machine learning model that predicts all 40+ interconnected bidding-zone prices simultaneously - this structure helps ensure greater coherence across the European day-ahead market. Our model is retrained daily, using an expanding historical window from 2015 to the most recent actuals.

Once finalized, forecast files are not modified further and are made available via FTP and API access.

On top of the day-ahead forecasts, we apply a “sawtooth correction” that refines the 15-minute dynamics without impacting the hourly performance of the model.
This additional layer learns recurring patterns and adjusts the baseline forecast accordingly, leading to more accurate 15-minute predictions.
The corrector retrains on short rolling windows so it adapts quickly as market behavior changes. When sawtooth effects fade, it naturally winds down to avoid over-adjustment.

This enhancement improves 15-minute forecast accuracy - reducing mean absolute error, preserving hourly averages, and avoiding bias or overshooting.

Please note that all short-term forecasts are generated at 15-minute granularity. Hourly values are derived as the average of the four corresponding quarter-hour intervals.

We ingests weather runs from ECMWF and NOAA as soon as files are released.

Delivery times vary by model and run. The times below represent the median, 90th percentile (p90), and 95th percentile (p95) of the availability of the full forecast (i.e. when the last forecasted hour becomes available).

Notes:

API

FTP

For FTP access, connect to:
​Host: calc1.cor-e.fr
​Port (FTP): 21
​Port (SFTP): 2222

Directory paths :

File naming conventions :

Note :