AirQualityBench: Realistic Global Air Quality Forecasting Benchmark
AirQualityBench is an innovative global standard designed to assess air quality forecasting models in realistic scenarios. It tackles the shortcomings of existing evaluation methods that depend on preprocessed, regional datasets, which often exclude or artificially fill in missing data. This benchmark incorporates hourly data from 3,720 monitoring stations worldwide from 2021 to 2025, focusing on six primary pollutants. By maintaining the original observation masks from providers and considering missing data as part of the forecasting challenge, it reports errors based on valid future observations after converting back to physical concentration scales. This methodology highlights issues like uneven global coverage, structured gaps in data, varying pollutant scales, and the costs associated with deployment.
Key facts
- AirQualityBench is a global multi-pollutant benchmark for air quality forecasting.
- It includes hourly observations from 3,720 monitoring stations.
- Data covers the period 2021–2025.
- It covers six major pollutants.
- The benchmark preserves provider-native observation masks.
- Missingness is treated as part of the forecasting problem.
- Errors are reported on valid future observations after inverse transformation.
- It addresses uneven global coverage, structured missingness, heterogeneous pollutant scales, and deployment cost.
Entities
—