Understanding our Prediction Series

Prediction Series are used to catalogue the different predictions we generate, where as the Predictions themselves are generated as Prediction Events.

Properties of a Prediction Series

Prediction Series are categorised using the following properties:

Property	Description
product	The different products in the market the series relates to, such as mFRR EAM Up or Imbalance
area	The geographical region, such as NO1 or FI
statistic	Defines something about the shape of the data in each prediction event. More on the different statistics below
unit_type	What it is that is being predicted, More on the different unit types below
unit	EUR, MW or N/A for series that are probabilities (unitless)
resolution	The period for which each event is valid for. More on this when we look at the Prediction Events

Unit Type

Price

The expected price.

Price Spread

The expected difference (spread) from the Spot price.

Direction

Which direction will be dominant.

Volume

The expected volume.

Statistics

Point

A single expected value for the corresponding Product and Unit Type.

Quantile

Probabilities the Product being predicted will fall into a certain range. For example, given the product Imbalance, and unit type of Price_Spread:

{
    "10": -9.918647732527685,
    "25": -5.9330300407128735,
    "50": -2.14651955171376,
    "75": 1.8668928923393722,
    "90": 8.326481400150248
}

• There is a 10% probability that the price will be 9.91€ below the spot price
• There is a 25% probability that the price will be 5.93€ below the spot price
• The 50% quantile (also referred to as median), means there is a 50% probability the price will be above (Spot price - 2.14)€, and a 50% probability the price will be below (Spot price - 2.14)€
• There is a 25% probability that the price will be 1.86€ above the spot price
• There is a 10% probability that the price will be 8.32€ above the spot price

Distribution

Our model predicts the probability of each possible direction - based on historical data. These probabilities reflect the empirical distribution observed in the data: that is, they represent how frequently each outcome occurred in the past under similar conditions.

In statistical terms, the empirical distribution is an estimate of the true underlying distribution that generated the data.

To ensure the predicted probabilities align well with real-world outcomes, the model is calibrated. Calibration means that if the model predicts, for example, a 70% chance of Up, then over many similar situations, Up should occur roughly 70% of the time.

For example, given the product Imbalance , and unit type of Direction

{
    "up": 0.2971276214864301,
    "down": 0.36712538740813155,
    "none": 0.33574699110543826
}

• There is a 36% probability that the dominant direction will be down
• The most likely outcome is the direction with the highest predicted probability.
• The spread between probabilities (how concentrated or even they are) indicates how confident the model is — and therefore how much risk is present.
• Comparing predicted probabilities to the historical (prior) distribution helps identify whether current conditions are unusual or typical.

Distribution Score

Unlike the empirical distribution, which estimates the cumulative distribution function of the real-world data, the balanced prior reflects a uniform/balanced distribution used during model training. For example, given the product Imbalance , and unit type of Direction:

{
    "up": 0.66,
    "down": 0.22,
    "none": 0.12
}

here, 66% for up would indicate a strong up signal as it is 2x the prior. The dominant direction can be easily identified by selecting the class with the largest probability. In this case, the dominant direction is up. This distribution does not describe how the data naturally occurs but is intentionally structured to reduce the effect of class imbalance and improve model learning across categories.

Conditional Index

Probabilities for different thresholds the series will exceed. For example, given the product Imbalance, and unit type of Price_Spread:

{
    "20": 0.06071186939588584,
    "50": 0.03407819229710775,
    "100": 0.026504537656296383,
    "-20": 0.07577408096156756,
    "-50": 0.012431568356134725,
    "-100": 0.005805317597058654
}

• "20" refers to the the probability the price will be 20€ higher than spot (~6%)
• "-50" is the probability the price will be 50€ below spot (~1.2%).