Skip to content

Explore ensemble-based forecast confidence on the 16-day view (WeatherNext 2) #95

Description

@kellylford

Context

Open-Meteo now offers Google DeepMind WeatherNext 2 — a 64-member global AI ensemble (0.25°, 6-hourly, out to 15 days) via the Ensemble API endpoint (different from our forecast endpoint). It rides our existing paid Open-Meteo subscription — confirm no extra cost before building.

Most of this model is not for FastWeather: a 64-member ensemble is a meteorologist's tool, and a 64-line "spaghetti plot" is exactly the decoder-ring UI this app avoids (see WAM/simplicity principles). This issue is scoped to the one on-brand idea, not adopting the model wholesale.

The idea: an accessible confidence signal on the 16-day forecast

The value an ensemble adds over a single deterministic model is uncertainty — how much the members agree. Our 16-day forecast currently shows day 14/16 with the same visual authority as tomorrow, which quietly overstates certainty; single-model forecasts that far out are close to noise.

Use ensemble spread (Open-Meteo already exposes ensemble mean + standard deviation) to surface confidence in plain language, e.g.:

"Next 3 days: high confidence. Days 8–16: forecasts disagree — treat as a rough outlook."

This is the same principle behind the air-quality work (#94): never present a number as more certain than it is.

Hard scope guardrails

  • Text/label only — a high / mixed / low agreement signal per far-out day. No ensemble chart, no member lines, no probability UI.
  • Fully VoiceOver-first: the confidence reads as words, not color.
  • Supplement, not replacement — keep the current forecast source for the actual values; ensemble is consulted only for the agreement signal. (WeatherNext 2 is 6-hourly / 12-hourly updates, coarser than what we show, so it must not drive day-to-day values.)
  • Decide the confidence thresholds (what spread = low/mixed/high) and where the signal appears (16-day rows and/or day detail).

Open questions

  • Cost/quota impact of the Ensemble API endpoint on our plan.
  • Whether to derive confidence from WeatherNext 2 specifically or a blend of ensembles Open-Meteo offers.
  • Portability of the confidence logic to the web and Windows apps.

Priority

Backlog / exploration — explicitly not a detour from finishing the air-quality feature. Captured so the framing isn't lost.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions