Hugging Face's logo

Datasets:
judicialmind
/
BharatWater-Bench

Tasks:
Time Series Forecasting
Modalities:
Text
Formats:
parquet
Languages:
English
Size:
1M - 10M
Tags:
water
reservoirs
rainfall
soil-moisture
india
time-series
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

This dataset is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 licence (CC BY-NC-SA 4.0): research, academic, personal and other non-commercial use is permitted with attribution; adaptations must be shared under the same licence; commercial use is not permitted. The original public source data remains under the Government Open Data License - India (GODL-India). Commercial use: contact surya@judicialmind.ai or hello@judicialmind.ai.

Log in or Sign Up to review the conditions and access this dataset content.

BharatWater-Bench

Licence: CC BY-NC-SA 4.0 — research and non-commercial use only; derivatives must be shared under the same licence. Commercial use: contact surya@judicialmind.ai or hello@judicialmind.ai

Access is approved automatically when the email you enter matches your Hugging Face account email and belongs to the organisation you name; other requests are reviewed manually.

A probabilistic forecasting benchmark for India's water: daily live storage and fill of the reservoirs monitored by the Central Water Commission, daily and weekly district rainfall, and daily district soil moisture. Each track holds out its last year as the test period and forecasts every series from rolling monthly origins, 7, 30 and 90 days ahead (4 and 13 weeks for weekly rainfall), with nine quantiles. The release ships 2,410 cleaned series, the task definitions, the forecasts of 7 models and an evaluate.py that recomputes the leaderboard from those forecasts.

  • At 7 days, the lowest overall WQL is BharatWater-Chronos (bharatwater_chronos) (0.2631); 1.9% below chronos2, 16.3% below lightgbm.
  • At 30 days, the lowest overall WQL is BharatWater-Chronos (bharatwater_chronos) (0.2928); 4.0% below chronos2, 17.6% below lightgbm.
  • At 90 days, the lowest overall WQL is BharatWater-Chronos (bharatwater_chronos) (0.3331); 11.0% below chronos2, 14.5% below climatology.
  • BharatWater-Chronos has the lowest WQL in 10 of 14 track x horizon tasks.

Coverage: 153 CWC reservoirs (daily live storage and % of capacity, 1991-01-01 to 2026-04-14), 702 districts of daily rainfall (weekly: 670), 733 districts of daily soil moisture (2018-07-24 to 2025-06-30)

Headline results

Overall = mean of the four daily tracks (reservoir_storage, reservoir_pct, rain_daily, soil_moisture), each averaged over its test origins. Lower WQL and MASE are better; a calibrated 80% interval covers 0.80.

h (days) model WQL MASE cov_80
7 BharatWater-Chronos (bharatwater_chronos) 0.2631 0.9030 0.7458
7 chronos2 0.2682 0.9531 0.7502
7 timesfm25 0.2840 0.9308 0.7552
7 lightgbm 0.3141 1059.4330 0.7780
7 climatology 0.3925 4.5752 0.4339
7 naive 0.6665 1.0683 0.6352
7 seasonal_naive 0.8247 4.7287 0.5994
30 BharatWater-Chronos (bharatwater_chronos) 0.2928 1.4518 0.7330
30 chronos2 0.3050 1.5879 0.7470
30 timesfm25 0.3458 1.5877 0.7499
30 lightgbm 0.3552 809.5556 0.7461
30 climatology 0.3948 4.5696 0.4457
30 seasonal_naive 1.0563 4.6995 0.6030
30 naive 1.0884 1.7537 0.6235
90 BharatWater-Chronos (bharatwater_chronos) 0.3331 20.2052 0.7283
90 chronos2 0.3744 20.8066 0.7286
90 climatology 0.3895 22.3129 0.4526
90 lightgbm 0.4444 1105.8990 0.7378
90 timesfm25 0.4525 20.8013 0.7111
90 seasonal_naive 1.1338 22.4545 0.6173
90 naive 1.7081 21.0180 0.5782

Per track (WQL and MASE by horizon):

Reservoir live storage (reservoir_storage)

model WQL h=7 WQL h=30 WQL h=90 MASE h=7 MASE h=30 MASE h=90
BharatWater-Chronos (bharatwater_chronos) 0.0289 0.0511 0.0949 0.9176 1.8934 39.1006
timesfm25 0.0307 0.0637 0.1530 0.8497 1.9528 39.8558
chronos2 0.0298 0.0620 0.1586 0.9245 2.0149 39.9879
naive 0.0387 0.0860 0.1950 0.8994 2.0670 40.0600
climatology 0.2199 0.2181 0.1979 7.9440 7.9061 43.2191
seasonal_naive 0.2175 0.2215 0.2007 7.9841 7.9740 43.2915
lightgbm 0.1059 0.1475 0.2486 4222.7047 3221.7563 4366.6128

Reservoir fill (reservoir_pct)

model WQL h=7 WQL h=30 WQL h=90 MASE h=7 MASE h=30 MASE h=90
BharatWater-Chronos (bharatwater_chronos) 0.0365 0.0678 0.1183 0.9198 1.8973 39.4647
timesfm25 0.0367 0.0741 0.1542 0.8522 1.9566 40.2249
chronos2 0.0380 0.0753 0.1598 0.9272 2.0189 40.3563
lightgbm 0.0674 0.1002 0.1845 13.0184 14.1751 54.1326
naive 0.0419 0.0862 0.1853 0.9026 2.0720 40.4300
seasonal_naive 0.2755 0.2680 0.2536 8.0215 8.0016 43.6762
climatology 0.2865 0.2770 0.2579 7.9811 7.9450 43.6160

District rainfall (daily) (rain_daily)

model WQL h=7 WQL h=30 WQL h=90 MASE h=7 MASE h=30 MASE h=90
climatology 0.9186 0.9235 0.9246 0.8826 0.7670 0.7456
chronos2 0.8940 0.9300 0.9478 0.8603 0.7432 0.7308
BharatWater-Chronos (bharatwater_chronos) 0.9008 0.9358 0.9665 0.8490 0.7444 0.7347
lightgbm 0.9643 1.0161 1.1081 0.9020 0.7911 0.7771
timesfm25 0.9332 1.0619 1.2195 0.8778 0.7558 0.7472
seasonal_naive 2.6475 3.5624 3.8930 1.3091 1.1150 1.1172
naive 2.4484 3.9991 6.1657 1.3436 1.2523 1.3118

District soil moisture (soil_moisture)

model WQL h=7 WQL h=30 WQL h=90 MASE h=7 MASE h=30 MASE h=90
BharatWater-Chronos (bharatwater_chronos) 0.0861 0.1167 0.1526 0.9257 1.2722 1.5207
climatology 0.1450 0.1605 0.1776 1.4929 1.6601 1.6711
seasonal_naive 0.1584 0.1735 0.1877 1.6001 1.7074 1.7331
chronos2 0.1109 0.1527 0.2315 1.1005 1.5745 2.1516
lightgbm 0.1189 0.1570 0.2363 1.1071 1.5000 2.0735
timesfm25 0.1356 0.1836 0.2834 1.1435 1.6856 2.3773
naive 0.1368 0.1825 0.2863 1.1277 1.6233 2.2703

District rainfall (weekly) (rain_weekly)

model WQL h=4 WQL h=13 MASE h=4 MASE h=13
climatology 0.9168 0.9679 0.8744 0.7229
BharatWater-Chronos (bharatwater_chronos) 0.8606 1.0220 0.8183 0.8197
chronos2 0.9850 1.0694 0.8058 0.7324
timesfm25 2.0543 2.3165 1.2428 1.1577
seasonal_naive 2.3373 2.6994 1.1420 1.0241
lightgbm 4.3331 3.2157 1.4338 1.2396
naive 3.3698 6.1483 1.5082 1.9195

The full table with 80% coverage and window counts is in leaderboard.md / leaderboard.json and in the BharatWater-Bench-Leaderboard Space.

Tasks

track what unit freq series observations span test cutoff origins horizons
reservoir_storage daily live storage of each reservoir monitored by the Central Water Commission BCM (billion cubic metres) daily 153 664,854 1991-01-01 to 2026-04-14 2025-05-01 9 (2025-05-01 to 2026-01-01) 7, 30, 90
reservoir_pct daily live storage as a percentage of live capacity, per reservoir % of live capacity at full reservoir level daily 152 663,717 1991-01-01 to 2026-04-14 2025-05-01 9 (2025-05-01 to 2026-01-01) 7, 30, 90
rain_daily daily rainfall per district mm/day daily 702 1,454,423 2018-01-01 to 2026-05-04 2025-05-01 9 (2025-05-01 to 2026-01-01) 7, 30, 90
soil_moisture daily soil moisture at 15 cm depth per district volumetric soil moisture at 15 cm (%) daily 733 1,744,379 2018-07-24 to 2025-06-30 2024-07-01 9 (2024-07-01 to 2025-03-01) 7, 30, 90
rain_weekly weekly (Monday-start) rainfall totals per district mm/week weekly 670 124,897 2018-01-15 to 2026-04-20 2025-05-05 9 (2025-05-05 to 2025-12-15) 4, 13

Training data for any model = observations strictly before the track's test cutoff; every origin lies on or after the cutoff. A window needs at least 8 observed steps in the context and one observed step in the horizon. Models get the 512 days (156 weeks) before the origin as context.

Quickstart 

import pandas as pd
from huggingface_hub import snapshot_download

root = snapshot_download("judicialmind/BharatWater-Bench", repo_type="dataset")
s = pd.read_parquet(f"{root}/data/reservoir_storage/series.parquet")
print(s.head())

import sys; sys.path.insert(0, root)
from evaluate import evaluate
lb = evaluate(root, models=["seasonal_naive", "chronos2"], only_tracks=["reservoir_storage"])
print(lb.round(4))

To score your own model, write forecasts/<track>/<your_model>/<origin>.parquet (columns series_id, origin, h, q0.1 ... q0.9) for the origins in data/<track>/splits.json and run python evaluate.py --models <your_model>. python evaluate.py --check recomputes the shipped leaderboard.

Files

path content
data/<track>/series.parquet long series table: series_id, ts, target and static columns
data/<track>/splits.json cutoff, origins, horizons, frequency, seasonality
data/<track>/tasks.json one task per origin x horizon (fev-style)
windows/<track>.parquet scored windows (series_id, origin) with their MASE scale
forecasts/<track>/<model>/<origin>.parquet quantile forecasts of every model (315 files)
leaderboard.json, leaderboard.md scores per track x model x horizon, and the overall table
evaluate.py, requirements.txt scoring code (numpy, pandas, pyarrow)

Metrics

  • WQL: mean over the quantiles 0.1-0.9 of 2 x summed pinball loss / summed |y|, pooled over the windows of an origin x horizon task, then averaged over origins.
  • MASE: mean absolute error of the median divided by the in-context mean absolute seasonal difference (mase_season of splits.json; lag 1 when no seasonal pair exists), averaged over windows and origins.
  • cov_80: share of observed values inside the 0.1-0.9 interval.

Models

model description
bharatwater_chronos BharatWater-Chronos: Chronos-2 fine-tuned on train-period water series
chronos2 amazon/chronos-2, zero-shot
timesfm25 google/timesfm-2.5-200m, zero-shot
lightgbm global LightGBM quantile model on lagged values
climatology same calendar window in earlier years
seasonal_naive same day (week) last year
naive last observed value

Source data

Derived from three resources of the Open Government Data Platform India (data.gov.in): the CWC daily reservoir level and storage bulletin, the NRSC district-wise daily soil moisture series and IMD district rainfall (gridded IMD rainfall averaged to districts, continued with the NDAP district daily rainfall series), all under the Government Open Data License - India (GODL-India). Cleaning: duplicate rows removed, values outside physical bounds (negative storage, storage far above live capacity, rainfall or soil moisture outside plausible ranges) flagged and left out of the series, series with too little history before the cutoff dropped.

Original data: published by Central Water Commission (CWC), Ministry of Jal Shakti (reservoir levels and storage); India Meteorological Department (IMD) and NITI Aayog's NDAP (district rainfall); National Remote Sensing Centre (NRSC), ISRO (district soil moisture) on the Open Government Data Platform India (data.gov.in) and, where marked, NITI Aayog's National Data and Analytics Platform (NDAP), under the Government Open Data License - India (GODL-India). The original public data remains under GODL-India; the CC BY-NC-SA 4.0 licence of this repository covers JudicialMind's compilation, derived fields, annotations, benchmarks and models.

Citation 

@misc{judicialmind2026bharatwaterbench,
  title={BharatWater-Bench: Forecasting India's Reservoirs, Rainfall and Soil Moisture},
  author={JudicialMind},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/judicialmind/BharatWater-Bench}}
}

License

CC BY-NC-SA 4.0 (see LICENSE): research, academic, personal and other non-commercial use, with attribution to JudicialMind and to the original data publishers; adaptations must be shared under the same licence. Commercial use requires a separate licence.

Commercial use: contact surya@judicialmind.ai or hello@judicialmind.ai.
Downloads last month
15
Total file size:
407 MB

Models trained or fine-tuned on judicialmind/BharatWater-Bench

Space using judicialmind/BharatWater-Bench 1

Collection including judicialmind/BharatWater-Bench