Skip to content
EPF Docs

v1.0 — Initial Release

Date: February 12, 2026

What Shipped

The first production release of EPF with:

  • Single model: HistGradientBoosting regressor from scikit-learn with MSE loss
  • Two predictor modes: Recursive (default) and direct multi-horizon (H1-H8 groups, opt-in via --direct)
  • Hourly resolution: 168 hourly predictions (7 days ahead) from 23:00 UTC origin
  • ~20 basic features: Price lags (1h, 24h, 48h, 168h), rolling averages, day-of-week, hour-of-day
  • FastAPI backend: REST API with SQLite storage
  • React dashboard: Interactive charts with Recharts, dark theme

Architecture

REE/ESIOS API → Data Collection → Feature Engineering → HistGBT → API → React Dashboard
  • Data source: 16 REE indicators via ESIOS API
  • Database: SQLite with ree_hourly and predictions tables
  • Training: 5-fold TimeSeriesSplit cross-validation
  • Deployment: GCP e2-micro instance with Nginx

Limitations

  • Single model — no ensemble diversity or bias detection
  • No weather or commodity features
  • No uncertainty quantification (no confidence intervals)
  • Origin at 23:00 UTC meant short-horizon predictions were already known from OMIE
  • Basic temporal encoding (integer hour/day) lost cyclical structure