IndiaS2S-Bench: A Comprehensive Benchmark of Subseasonal to Seasonal Forecasting over India

Subseasonal forecasting over India is not a single ranking problem: usefulness changes with lead, season, region, target, metric, and verification data. We present IndiaS2S-Bench, a unified comparison of archived operational numerical-weather-prediction (NWP) forecasts and retrospective ERA5-initialized AI/hybrid forecasts. A frozen 1.5° protocol scores 517 precipitation and 516 2-metre-temperature starts during 2020--2024 at Weeks 1--6. All learned configurations are evaluated after their documented training cutoff, so the benchmark is out of training sample but not wholly untouched or operationally equivalent. A fixed six-system mean has the highest Week-1 ACC for both targets; by Week 6, FuXi-S2S has the highest precipitation ACC and ECMWF the highest individual temperature ACC. RMSE, MAE, and native fair-CRPS produce different orderings, and the precipitation mean beats deterministic climatology only at Weeks 1--2. IMD gauge verification agrees at early lead but changes weak long-lead rainfall ordering. The benchmark’s contribution is therefore a practical selection map: which configuration fits a target, lead, region, season, and forecast objective---not a claim that AI universally beats physics.