Research
158d agoRolling Validation Upsets Air Quality Forecasting
A new study shows that common static testing can overstate machine‑learning gains in air‑quality forecasting. Using daily PM10 data, researchers compared XGBoost and SARIMA to a persistence baseline under both static and rolling‑origin protocols.
The Signal
The rolling approach reversed the model rankings, underscoring the need for realistic, continuously updated evaluation worldwide.