AI enters the monsoon forecast

The 2026 monsoon was the first in which the India Meteorological Department launched two AI-enabled monsoon forecast models operationally, using a multi-model ensemble approach that draws on global climate models from multiple international research centres. It was a quiet institutional change with large consequences, because the monsoon forecast shapes sowing decisions, water releases, food prices, and relief planning for hundreds of millions of people. A forecast is not weather, but for a farmer deciding when to sow, it is the closest thing to it.

The season gave the new tools a mixed report card. IMD's April forecast put seasonal rainfall at 92 per cent of the long period average, with a five-percentage-point model error, and its late-May update lowered the estimate to 90 per cent, with a four-point error. The observed outcome was 87 per cent, inside the forecast error limits, and IMD reported that its seasonal forecast was accurate within those limits and that its El Niño and Indian Ocean Dipole outlooks were broadly consistent with what was observed. The Kerala onset forecast did not verify: the forecast date was 26 May with a four-day window, while the observed onset was 4 June. Forecast skill therefore differed between the national seasonal total and the onset date, and both results deserve to be reported together.

The deeper challenge is resolution. A national or even regional forecast cannot tell a district whether its dry spell will last ten days or thirty, and it was district-level dry spells, such as 67 rainless days out of the first 102 in Sakri taluka of Dhule district, with a continuous dry period of 22 days, that damaged crops in 2026 even where seasonal totals looked closer to normal. Researchers argue that reanalysis and global datasets, and finer-resolution data such as kilometre-scale grids, are needed to improve temperature and humidity information, which is critical to measuring heat stress as well as rain.

The lesson for forecast users is statistical literacy. A forecast with an error band is not wrong because the observed value differs from the central estimate; it is wrong only if the outcome falls outside the band, or if it is systematically biased. The 2026 season, in which the seasonal total verified within error but the onset date did not, is a useful classroom example of that distinction.