Skip to content
Power Pulse

04 — method

Methodology

A short process note — not a model catalogue. For what is running right now, see Models.

Research / portfolio demo. Not trading advice. Prices and fundamentals from ENTSO-E (and SMARD where used).

History currently loaded

9,601 h

2025-06-23 → 2026-07-28

Metrics

MAE · RMSE · Skill vs naive

Skill > 0 means better than seasonal naive (168h).

01 — process

Pipeline

  1. 01 — Market target

    Hourly day-ahead auction prices by bidding zone. The public site emphasises the next delivery curve (and a multi-day path on the forecast page).

  2. 02 — Data & storage

    Ingest via scheduled GitHub Actions into Supabase (raw → mart). Day-ahead load, wind and solar fundamentals feed residual-load style features when present.

  3. 03 — Information set

    At decision time T, only data published by T is allowed. Lags look backward; day-ahead fundamentals for future hours are valid exogenous inputs; same-day actuals for future hours are not.

  4. 04 — Train & score

    Baselines (seasonal naive, hour-of-week), classical (SARIMAX, Prophet), LightGBM multi-horizon, and LSTM. LightGBM is trained per zone; gallery models focus on the primary zone.

  5. 05 — Walk-forward

    Origins every day after a minimum history. Train or fit with data ≤ T, forecast the next horizon, score MAE / RMSE / skill vs seasonal naive. No shuffled K-fold.

  6. 06 — Publish

    Export small zone JSON packs and site indexes. The Astro site loads the primary zone fast and fetches other zones on demand. No ML in the browser.

ENTSO-E  (A44 prices · A65/A69 DA fundamentals)
   │
   ▼
GitHub Actions
   ingest → features → train → score → export
   │
   ▼
Supabase  (raw / mart / meta)
   │
   ▼
Astro site  ·  zone packs  ·  map console