Exploratory studies explain the past. To show the method also works on data it has not seen, we published a forecast for Diwali 2026 before it happened, and we will score it in public afterwards.
The firecracker rules for Delhi in 2026 had not been announced when we locked the forecast, so we give one forecast per regime. Each is the average of past Diwalis under the same rules, with an 80% prediction interval.
| Rules in force | Forecast, µg/m³ | 80% interval | Based on |
|---|---|---|---|
| Full ban | 366 | 233 to 499 | 2020, 2021 |
| Green crackers, fixed hours | 303 | 177 to 428 | 2018, 2019, 2025 |
| No effective restriction | 192 | 58 to 325 | 2016, 2017 |
Forecast intervals for Diwali 2026 by firecracker regime, with past Diwalis shown as points.
Quantity. mean hourly (observed - BAU) PM2.5, 18:00 8 Nov to 02:59 9 Nov 2026 IST, missing hours carried forward from the last valid hour.
Where. CPCB Delhi network median, hours with 3 or more reporting sites (config/stations.csv).
Baseline. LightGBM on ERA5 weather, calendar and wind-aware fire index; Diwali windows excluded from training; out-of-fold prediction by ISO week (src/elevenwinters/analysis/partb.py).
Scoring. After 8 Nov 2026 the same code is rerun with data through 15 Nov 2026. The forecast for the regime actually in force counts as correct if the observed excess falls inside its interval.
The record is prereg/diwali-2026-forecast.json in the repository. Its SHA-256 checksum is 0b5f3a06349b1dd1311300029ef3446796338ac83adc0a3b78a3412ef0120a4f. Any later change to the file would change the checksum.