Every winter, Delhi blames the weather, the farm fires in Punjab and Haryana, and Diwali. They arrive together, so nobody can say how much each one adds. We used the years when they didn't line up to pull them apart.
The monitors behind these numbers pull air through white glass-fibre tape and leave a grey spot for each sample. The dirtier the air, the darker the spot. Each strip below is one season, July to June, as recorded at the US Embassy in New Delhi. Every year the tape darkens from late October and stays dark into January. The marigold ticks are Diwali. White gaps are days the monitor did not report.
Daily PM2.5 at the US Embassy in New Delhi from 2016 to 2026, one row per season. Every season darkens from late October to January.
Diwali follows the lunar calendar, so it moves across the burning season. In 2017 it fell on 19 October, before most farm fires. In 2020 it fell on 14 November, at their height. Delhi's firecracker rules changed four times in the same decade. And the timing of crop burning itself appears to have shifted. These changes let us compare like with like.
Diwali dates from 2016 to 2026 placed on an October to November axis, with the firecracker rules for each year.
For every monitor, a machine-learning model learns how PM2.5 responds to temperature, humidity, wind, rain and the height of the mixed layer of air near the ground. Re-running it hundreds of times with the weather shuffled gives the pollution each site would have had under typical weather.
Weather explains most of the rise from monsoon to winter in Kolkata, Lucknow or Mumbai: about 82% along the eastern Gangetic Plain and 81% outside it. In Punjab and Haryana, where paddy residue is burned, it explains only about 48%, and in Delhi-NCR, downwind, about 62%. Removing the weather narrows the winter gap between northern and other cities only from 45 to 33 µg/m³. Most of the north's winter problem is what is emitted, not the weather it is emitted into.
For each city, the winter minus monsoon PM2.5 gap before and after removing weather.
Mostly. With the weather taken out, Delhi's PM2.5 was 14% below its 2016–2019 level during the lockdown. From mid-2020 on it stayed about 11% below. But since then it has barely moved: down about 0.5% a year.
Delhi's monthly PM2.5 at the US Embassy, observed and weather-normalised, 2016 to 2026, with a step down in 2020.
We built a baseline for every hour: what PM2.5 should have been given that night's weather, the time of year and how much farm smoke the wind was carrying from Punjab and Haryana. The model never sees Diwali, and every hour is predicted by a version of it that never saw that week. Diwali's excess is what the air held above that baseline from 6 pm to 3 am.
In the CPCB network median, Diwali night added between 127 and 392 µg/m³ in every year with data. Among 211 ordinary autumn nights scored the same way, the 99th percentile was 193.
Hour by hour PM2.5 above the baseline around Diwali night, one line per year, coloured by the firecracker rules.
Mean Diwali-night excess was 192 µg/m³ in the two years without a bursting restriction, 303 with green crackers and 366 under a full ban. The two largest nights in the record came in 2021, under a full ban, and 2025, when green crackers were allowed again. The smallest came in 2017, when sales were suspended about ten days before an early Diwali. There are only a few Diwalis per regime, so this is not proof that bans cannot work, but it gives no sign that these did.
Diwali night excess PM2.5 for each year from 2016 to 2025, against the range of ordinary autumn nights.
The US Embassy monitor is the most-used reference for Delhi. On Diwali it often repeats its last reading and then stops reporting during the worst hours. Late-night hours go missing 13 times as often as on an ordinary night, so any Diwali average built on this monitor alone leaves out the peak.
Share of hours missing at the US Embassy monitor from 5 pm to 6 am, Diwali nights against ordinary nights.
Compared with the same hours a week earlier, Diwali evening air is finer and more sulphurous. The PM2.5 to PM10 ratio rises from 0.41 to 0.69, and SO₂ rises by 37%. The traffic gases go the other way: NO₂ and CO both fall slightly, as fewer people drive. Fine particles and sulphur up, traffic down: that is fireworks.
The timing follows local custom. Delhi peaks around midnight. Mumbai and Hyderabad peak between 8 and 11 pm. Chennai peaks on the evening before, when Tamil Nadu celebrates Deepavali.
Paired comparison of Diwali evening against reference evenings for PM2.5 to PM10 ratio, sulphur dioxide, carbon monoxide and nitrogen dioxide.
Median excess PM2.5 by hour around Diwali in Delhi, Kolkata, Mumbai, Hyderabad and Chennai.
Official progress on stubble burning is measured in satellite fire counts. In Punjab and Haryana, VIIRS detections in October and November fell 94% between 2021 and 2025. But the same satellite also passes at 1:30 am, and the share of fires it catches then rose from 0.2% to 4.7%. More fires are still burning in the middle of the night, which is what you would see if farmers now light them after the afternoon pass.
Afternoon-pass fire detections fall sharply after 2021 while the night-pass share rises from under one percent to several percent.
Delhi's weather-adjusted PM2.5 follows upwind fires over one to four days. From 2016 to 2021, 100 units of the fire index went with about 2.2% more PM2.5. In 2025 the same 100 units went with 25% more. The 2025 estimate is uncertain (95% interval 0 to 55%), but it is the pattern you would expect if each detected fire now stands for more burning.
Percent rise in Delhi PM2.5 per 100 fire index units by season, with confidence intervals; 2025 stands well above earlier years.
If burning moved to the evening, monitors inside Punjab and Haryana should see more evening pollution during the harvest. They don't, so far. But the open archive is missing 2023 and 2024, and 2025 had so little burning that a change in its timing would barely move a city monitor. We report this as a test that could not yet be run, not as evidence against the shift.
Harvest evening ratio relative to December for monitors in Punjab and Haryana by season.
Explaining the past is easy to overdo. So before Diwali 2026 we locked a forecast of the excess for each possible set of firecracker rules, published its checksum, and will score it after 8 November.
Explore any of the 105 monitors in the city explorer, read how every number is made in data and methods, or read the full paper.