Problem

Farmers globally face high-stakes decisions regarding planting and investments, heavily reliant on uncertain future weather conditions, particularly the timing of monsoon onset in tropical regions. A significant challenge is the lack of readily available local weather forecasts with sufficiently long lead times, which are crucial for effective agricultural planning. Furthermore, designing useful forecasts for a diverse population of farmers—each with unique land characteristics, resources, and risk tolerances—is complex, as a one-size-fits-all approach is ineffective. Conventional forecast evaluation often uses static climatology baselines, which are unrealistic as farmers' expectations about weather events like monsoon onset dynamically evolve throughout a season.

Method

The researchers developed a decision-theory framework, guided by Blackwell's Informativeness Theorem, to design forecasts that empower heterogeneous farmers to make their own decisions. This framework emphasizes providing probabilistic forecasts over deterministic ones and incorporating information farmers already possess into the baseline. A key component is the "evolving-expectations" statistical model, which uses Bayesian inference to dynamically update the probability distribution of future monsoon onset dates based on historical data and the observation that onset has not yet occurred. This model is then blended with forecasts from two artificial intelligence weather prediction (AIWP) models, Google's NGCM and ECMWF's AIFS. The blending is achieved using a multinomial logistic regression that combines the probabilistic information from the evolving-expectations model with AIWP rainfall forecasts, dynamically weighting their contributions by lead time. The monsoon onset definition used is a modified Moron-Robertson definition, focusing on rainfall sequences after April 1 and avoiding prolonged dry spells.

Results

The blended model consistently outperformed both static climatology and the evolving-expectations model across various metrics, including Brier Score, Ranked Probability Score (RPS), and Area Under the Receiver Operating Characteristic (ROC) curve (AUC), over cross-validation periods (2000–2024), hold-out periods (1965–1978), and the 2025 dissemination period. For instance, during 2000–2024, the blended model showed a 5–10% improvement in Brier Score and 20–25% in RPS compared to static climatology. Notably, in 2025, a year with an atypical monsoon onset, the blended model maintained its accuracy, resulting in a substantial 20% improvement in Brier Score and RPS relative to the poorly performing climatological baseline. The skill of the blended model was highest at a one-week lead time and remained positive out to four weeks. This system was successfully deployed operationally in 2025, delivering subseasonal monsoon onset forecasts weekly to 38 million Indian farmers.

Implications

This research provides a replicable pathway for developing effective climate adaptation tools for vulnerable populations globally, particularly in monsoon regions. The decision-theory framework highlights the necessity of probabilistic forecasts, the incorporation of farmers' prior knowledge, and the measurement of decision changes rather than just outcomes. The blended AI/statistical approach effectively leverages the strengths of multiple models, addressing limitations like AIWP models' shorter-range skill for full agronomic onset criteria. Future work could focus on increasing forecast resolution (from weekly to daily) and enhancing spatial granularity. The study also underscores the importance of using evolving climatological baselines in forecast evaluation to ensure that reported skill reflects true added value for users.