The challenge uses historical observations from the Central Pollution Control Board (CPCB), collected through India’s Continuous Ambient Air Quality Monitoring Stations (CAAQMS).

Participants will predict hourly PM2.5 concentrations (µg/m³) at four monitoring stations in Mumbai. Predictions will be evaluated against new observations as they become available.

Training data updates

Train.csv will be updated on 31 October, 5 November and 9 November, with a final update scheduled for 10 November at 09:00 IST. Participants may use newly released observations to retrain their models and update predictions.

Test and submissions

Test.csv and SampleSubmission.csv contain the same 2,592 IDs throughout the competition. IDs, station-hours, row count and row order remain fixed.

Every submission must contain the columns ID and Target, with a non-negative PM2.5 prediction for every ID, in the original row order.

Public leaderboard updates

At launch on 26 October, the public score will use observations for 22–23 October, excluded from the launch Train.csv. This window remains active until the 31 October update. Subsequent updates will use:

• 31 October update: observations for 27–28 October.
• 5 November update: observations for 1–2 November.
• 9 November update: observations for 5–6 November.

Only the current two-day window contributes to the public score. Each update replaces the previous scoring window; earlier periods no longer contribute. The test files remain unchanged, and participants may continue submitting updated predictions until the final submission deadline.

All dates are in 2026, using IST (UTC+05:30). Data and leaderboard updates depend on measurement availability.

Starter Notebook

Get started with the data, train a simple Random Forest model, and create your first submission.

Explore Additional Datasets

Looking to improve your air quality forecasts? Explore ClimateVerse’s Google Air Quality Forecasting Challenge page to discover additional climate and environmental datasets that may help inform your modelling approach.