How India forecasts the monsoon

When a forecast says India will have a below-normal monsoon, what does that tell you about the rain where you live? It depends on what the forecast covers. The rainfall total for the country, the date the monsoon reaches Kerala and the weather over the next few days are different questions. To understand the forecasts, keep those questions separate and ask what evidence supports each answer.

A forecast has a calendar

The India Meteorological Department, or IMD, under the Ministry of Earth Sciences, issues its first forecast in April for June to September rainfall averaged over India. An update follows by the end of May. These estimates cover the national seasonal total. For rain in a particular town on a particular day, you need a local forecast.

The process continues after the monsoon begins. Forecasts near the end of June, July and August cover the following month. The late-July release also includes an outlook for August and September together. IMD also forecasts seven-day averages for the next four weeks and updates these on Thursdays. Daily short- to medium-range forecasts provide more immediate guidance. Read the issue date and the period covered as carefully as the headline.

Two ways of estimating the future

IMD's revised strategy, introduced in 2021, combines statistical and dynamical forecasting. Statistical methods look for predictors whose past values have a stable relationship with what is being forecast. They use that relationship to estimate a future value. This assumes that the relationship will hold in the future, though it may change.

Dynamical methods use numerical models to simulate atmospheric and oceanic conditions. Mathematical equations describe physical processes and calculate how the system changes over time. Models represent the real system imperfectly.

Combining models produces an ensemble. In its documented statistical ensemble method, IMD assessed models with different combinations of predictors by how well they predicted past outcomes. It then combined selected models. The dynamical Multi-Model Ensemble draws on coupled global climate models from different prediction and research centres, including IMD's Monsoon Mission Climate Forecasting System. MoES says these dynamical forecasts are updated monthly. An ensemble can contain statistical or dynamical models. Its accuracy still needs to be checked.

What the Long Period Average means

The Long Period Average, or LPA, is a benchmark: rainfall for a specified region and period, averaged over many years. A forecast expressed as a percentage of LPA compares the expected amount with that benchmark, rather than giving the chance of rain.

For India's June to September rainfall, IMD introduced the 1971 to 2020 normal from the 2022 monsoon. The value is 868.6 mm, rounded to 87 cm in seasonal bulletins. This normal uses 50 years of observations and is updated once a decade. Use it for the national seasonal total; annual and regional rainfall need their own benchmarks.

The April 2026 national forecast used five categories: deficient below 90% of LPA, below normal at 90 to 95%, normal at 96 to 104%, above normal at 105 to 110%, and excess above 110%. Use those bands for that national seasonal forecast. Local rainfall categories have their own definitions.

Reading the uncertainty

On 13 April 2026, IMD forecast national seasonal rainfall at 92% of LPA, with a model error of ±5% of LPA. On 29 May, it updated the estimate to 90%, with a model error of ±4% of LPA. Both figures were predictions when issued; observed rainfall is measured later.

The error figures describe uncertainty in rainfall relative to the benchmark, rather than the chance that the forecast is correct. Outcomes can fall outside these limits. MoES identifies several sources of uncertainty: how El Niño-Southern Oscillation and the Indian Ocean Dipole develop, variation within the season, and imperfect modelling of monsoon low-pressure systems and their rain. The result depends on more than a single climate driver.

When has the monsoon reached Kerala?

Kerala's normal onset date is 1 June, with a standard deviation of about seven days. The arrival date varies from year to year. IMD has issued operational onset forecasts since 2005, using a statistical model with six predictors and a stated model error of ±4 days. The standard deviation describes past variation; the model error describes forecast uncertainty.

To declare that the monsoon has arrived, IMD checks observations. It monitors the monsoon through surface and upper-air observations, satellites, radars, weather-chart analysis and model guidance. Its Kerala rainfall criterion requires, after 10 May, at least 60% of the available stations in a designated 14-station network to record 2.5 mm or more on two consecutive days. Onset can be declared on the second day only if wind and radiation criteria also agree.

How accurate have the forecasts been?

MoES's February 2026 verification table compared national seasonal forecasts with actual rainfall for 2023, 2024 and 2025. In 2023, the forecast was 96 ±4% of LPA and actual rainfall was 95%. In both 2024 and 2025, the forecast was 106 ±4% and actual rainfall was 108%. All three actual totals were within the forecast limits.

MoES reported an average absolute error of 1.9% of LPA across those years. The rounded annual figures in the table give a slightly different average. This measure describes error in the national seasonal amount; local daily forecasts need a separate assessment.

A longer comparison also shows improvement. In April 2025, MoES reported that average absolute error had fallen by about 21% in 2007 to 2024 compared with 1989 to 2006. That is a relative reduction in error across two 18-year periods. The comparison leaves open how much each model contributed.

What 2026 shows about judging success

IMD's end-of-season report recorded national June to September rainfall in 2026 at 759.4 mm, or a rounded 87% of LPA. That rounded percentage fell within both the April and May forecast ranges. Yet 87% belongs to the deficient category, while the forecast headline had said below normal. Actual rainfall can fall within the forecast range while belonging to a different category.

The onset forecast missed its window. IMD's 15 May bulletin had predicted arrival over Kerala on 26 May, with a model error of ±4 days. The observed onset was 4 June, and IMD judged the forecast incorrect. It identified this as the second such miss since operational onset forecasts began, after 2015. The seasonal total and onset date therefore need separate assessments.

How to read the next bulletin

Start with the place, period and quantity being forecast. Check the rainfall benchmark, the uncertainty around the estimate and whether a later update is available. When judging performance, compare the forecast with observations of the same thing. For tomorrow's local weather, use a local daily forecast.

To explore how changing inputs changes a model's readouts, Learnacy Labs offers browser-based interactive experiments with a guided Understand, Explore and Prove flow. These provide a way to explore that general idea; they are not presented as reproductions of IMD's forecasting system.