WEATHER · FORECAST OBSERVATORY
Live model output
Live model output · commit 24b7120 · 2026-07-05
Global daily-mean temperature · 211 countries

Forecasting the planet's temperature to within a quarter‑degree.

An end-to-end machine-learning pipeline that predicts global daily temperature, flags the outliers, and explains what drives them. Every figure is regenerated from source and stamped to the commit that produced it.

0.27°C RMSE
Best model · GradientBoosting
211
Countries scored
30days
Untouched holdout
7
Models benchmarked
Three findings, in order
01

Forecast

Daily-mean temperature across 211 countries, scored on a 30-day holdout.

02

Anomalies

Temperature outliers flagged by z-score and Isolation Forest, mapped by location.

How to read this

Two conventions recur across every section on this page.

  • [ FINAL ]

    A provenance chip names exactly what a number measures and how it was produced, so a claim is never color-only.

  • [ HOLDOUT ]

    The holdout is a final stretch of days none of the models trained on. Every forecast metric shown here is scored against readings the model had not yet seen.

03

Drivers

What drives air-quality (PM2.5) predictions

MODEL: PM2.5 · NOT TEMPERATURE

Humidity and visibility lead; temperature contributes less than most people expect.

§

Built to be reproduced

Every figure regenerates

The dashboard reads only the committed JSON contracts the pipeline emits. Nothing is typed in by hand, so a number cannot drift from the model that produced it.

Commit-stamped and honest

Every dataset carries its generation date and source commit. A metric that cannot be trusted yet is marked pending, with a stated reason, instead of padded.

Forecasting
GradientBoosting, LightGBM, ARIMA, SARIMA, Prophet
Detection
z-score, Isolation Forest
Explainability
SHAP feature attribution
Search
all-MiniLM-L6-v2 embeddings, offline cosine
Mapping
MapLibre GL
Charts
d3-scale, d3-shape
Serving
Python, FastAPI
Interface
Next.js, Astryx design system
WEATHER · FORECAST OBSERVATORY