Prediction Output

90% CI: 90% of actual prices should fall within [lower_90, upper_90]
50% CI: 50% of actual prices should fall within [lower_50, upper_50]

Prediction Record Structure

Each prediction row represents a single price forecast for a specific target hour, generated by a specific model on a specific date.

Field	Example	Description
`prediction_date`	`2026-02-26`	When the forecast was generated
`target_date`	`2026-02-27`	Date being predicted
`target_hour`	`14`	Hour (0-23)
`target_minute`	`0`	Minute (0/15/30/45 for 15-min, always 0 for hourly)
`model_name`	`ensemble_dayahead`	Model identifier
`run_mode`	`dayahead`	Product type
`predicted_price`	`52.3`	Point forecast (EUR/MWh)
`actual_price`	`48.7`	Actual price (backfilled after the fact)
`model_version`	`2026-02-26`	Model training date
`prediction_lower_90`	`35.1`	90% CI lower bound (5th percentile)
`prediction_upper_90`	`68.5`	90% CI upper bound (95th percentile)
`prediction_lower_50`	`44.2`	50% CI lower bound (25th percentile)
`prediction_upper_50`	`60.4`	50% CI upper bound (75th percentile)

Model names encode the model type, approach, and context:

Pattern	Example	Meaning
`{type}`	`ensemble`	Base hourly model
`{type}_{approach}`	`ensemble_expanded`	15-min approach variant
`{type}_backtest`	`ensemble_backtest`	Walk-forward backtest result
`naive_persistence`	—	Naive baseline (today = tomorrow)
`naive_weekly`	—	Naive baseline (this week = next week)

Model types: histgb, lightgbm, xgboost, ensemble

Intervals are computed using split conformal prediction with asymmetric signed residuals, calibrated per horizon bucket:

Intervals are asymmetric — the upper bound may extend further than the lower bound, reflecting the right-skewed nature of electricity price errors.