Forecasting temperature across 211 countries

The brief looked deceptively simple: forecast average temperature for 211 countries. The obvious first move — one Prophet model per country — is also the one that quietly wastes most of the signal in the data.

Why one model per country is a trap

Fitting 211 independent models feels safe because each one only has to learn one place. But temperature is not 211 unrelated problems. Neighboring countries share seasons; the northern and southern hemispheres are mirror images; coastal and continental climates rhyme. A per-country model throws all of that shared structure away and re-learns seasonality from scratch on whatever short, noisy history each country happens to have.

It is also an operational headache: 211 models to retrain, monitor, and version.

A single global model

Instead, one model saw every country at once, with the country identity and a few static geographic features (latitude, hemisphere, a coarse climate zone) as inputs. That let the model borrow strength — a data-rich country’s clean seasonal signal informs a data-poor neighbor’s forecast.

The result: average error of 0.19°C, about 75% lower than the tuned per-country Prophet baseline, from one artifact instead of 211.

• One training run, one model to monitor.
• New countries get a reasonable forecast on day one, before they have much history, because the global patterns already apply.
• Seasonality is learned once and shared, not re-estimated 211 times.

The takeaway

Before reaching for “a model per entity,” ask what the entities share. When they share a lot — and geography almost always does — a single model that is told which entity it is looking at will usually beat a fleet of specialists, and it is far less to operate. Details are in the case studies.