qhjqhj00 monsoon-onset-forecast-eval
Decision-oriented benchmarking to transform AI weather forecast access: Application to the Indian monsoon — Masiwal et al. (2026) (arXiv:2602.03767, 2026)
What this evaluates
Evaluates the ability of AI and traditional NWP models to forecast the local monsoon onset date over India, specifically tailored for agricultural decision-making in the Central Maharashtra Zone. It tests long-range subseasonal precipitation forecasting and event detection under operationally realistic initialization constraints.
Datasets
- IMD Gridded Rainfall — total 124; splits: recent_test (6), extended (20), all_available (60), common (18); repo https://github.com/envfluids/monsoon-benchmark
Metrics
onset_forecast(primary) — range: days- Compares the model-predicted monsoon onset date against the observed onset date derived from the modified Moron-Robertson index. The observed onset is defined as the first day of the first 5-day wet spell (≥1 mm/day) occurring after the climatological median Kerala Onset Date (June 2). Skill is measured by the absolute error in days or binary correctness of onset timing per grid cell and year.
Input / output format
Input: Global atmospheric state fields (ERA5 or NCEP/ECMWF analysis) at 00 UTC on initialization dates (Mondays and Thursdays starting in May), plus climatological sea surface temperature and sea ice boundary conditions for models requiring them.
Output: Daily global precipitation forecasts (and derived onset dates) up to 46 days lead time, regridded to a common 4°×4° land-masked grid.
Scoring recipe
def compute_onset_forecast(forecast_precip, obs_precip):
# 1. Derive observed onset from modified Moron-Robertson index
obs_onset = find_first_wet_spell_after(obs_precip, start_date='June-02', min_days=5, threshold_mm=1)
# 2. Derive predicted onset from forecast precipitation
pred_onset = find_first_wet_spell_after(forecast_precip, start_date='June-02', min_days=5, threshold_mm=1)
# 3. Calculate absolute error in days
error_days = abs((pred_onset - obs_onset).days)
return error_days
Common pitfalls
- Using the original Moron-Robertson criterion (2) requires verifying a 10-day dry spell up to 35 days ahead, which exceeds subseasonal forecast skill; this benchmark uses the MOK filter instead.
- Evaluating at native model resolution introduces ocean bias; all forecasts and observations must be conservatively remapped to a 4° grid with a land-sea mask applied.
- Initialization dates must strictly follow the twice-weekly (Mon/Thu) schedule to ensure fair comparison with the IFS baseline hindcasts.
Evidence (verbatim from paper)
To enable the development of actionable onset forecasts several weeks in advance, we adopt a modified Moron-Robertson onset definition anchored to the median climatological onset date over Kerala, which reduces the dependence on long-lead post-onset precipitation forecasts (criterion (2) above). IMD defines the onset of monsoon over Kerala (MOK), based on the fulfillment of specific criteria, including large-scale atmospheric circulation patterns, regional rainfall thresholds, and outgoing longwave radiation. The MOK is typically followed by a northwestward advancement of the monsoonal rains. This typical inland northward progression of monsoon onset following MOK can be leveraged as an alternative indicator to minimize premature “false start” detections, without necessitating reliance on the full Moron-Robertson onset definition. Most of these false starts are observed in the northwestern semi-arid region of India and often occur in early May. Thus, we modify this onset definition such that the first wet spell after MOK is considered as the onset date over a region, regardless of whether or not it is followed by dry spells. However, in operational forecasting, the actual MOK date
Citation
@misc{masiwal2026decision,
title={Decision-oriented benchmarking to transform AI weather forecast access: Application to the Indian monsoon},
author={Masiwal et al. (2026)},
year={2026},
note={arXiv:2602.03767}
}
- arXiv: 2602.03767