Adithya. ← All work
PROJECT 03 · LIVE · FORECAST DESK

Windhedge

Shows the UK wind forecast that was really available when each day-ahead order was due, and how much safety margin to plan for.

FastAPIPythonReactViteDockerElexon API
Visit live ↗ Live · Web app
windhedge.vercel.app/desk
Screenshot of the Windhedge desk: controls, headline numbers and the forecast-vs-actual wind chart for January 2026
THE STORY

Don’t guess the gust, buy what you must.

A UK energy company buys electricity on the day-ahead market, so orders for tomorrow go in today. Wind is a big and changeable part of supply, so the buyer needs to know what the forecast said at the moment the order was due, not what a later, better forecast says.

Windhedge uses public Elexon data to show exactly that, measures how wrong those forecasts were in January 2026, and works out a safety margin to plan for. It doesn’t build a prediction model of its own: it displays and evaluates the forecasts that already exist.

84%
of January hours had less wind than the day-ahead forecast
~1,700 MW
average over-forecast on decision day
88%
of unseen hours covered by the 90% cushion
99%
of those hours caught short with no cushion

Elexon WINDFOR forecasts against actual FUELHH output, January 2026. Rough figures from one month, so treat them as indicative.

HOW IT WORKS

From public data to a margin.

SourceDataBackendInterfaceOutputPackaging
Elexon BMRS API
FUELHH actuals + WINDFOR forecasts
Clean + quality checks
Parquet, committed with the repo
FastAPI backend
matching rule, metrics, cushion
React desk
controls, chart, readouts
Safety margin
how much wind to plan for
Docker + Vercel
one command locally, live on the web
downloaded once
loaded at start-up
JSON /api
shows
the dashes show which way the data flows ↑
KEY FEATURES

What it does.

01

Forecast as the buyer saw it

Picks the newest forecast published before the 09:50 UTC day-ahead deadline, never a later, better one that didn’t exist yet.

02

Honest accuracy numbers

Measures miss, lean and wobble across all 744 hours of January 2026, by hours ahead, wind level and time of day.

03

A safety cushion

Says how much wind to plan as missing, at 80%, 90% or 95% caution, learned from past forecast errors.

04

Back-tested on unseen days

The cushion is learned from 1–24 January and tested on 25–31 January. The per-wind-level version failed, and the desk labels it that way.

05

An interactive desk

Sliders for hours ahead and caution level, a date range, a forecast-vs-actual chart and CSV export.

06

Containerised and tested

FastAPI and React run together with docker compose up, with pytest, Vitest, linting and CI on every push.