Datasets:

judicialmind
/

HMIS-Bench

---
license: cc-by-nc-sa-4.0
pretty_name: HMIS-Bench
language:
- en
task_categories:
- time-series-forecasting
tags:
- health
- india
- hmis
- time-series
- forecasting
- benchmark
- probabilistic-forecasting
size_categories:
- 1M<n<10M
configs:
- config_name: bench
  data_files: series/bench.parquet
  default: true
- config_name: all
  data_files: series/all.parquet
- config_name: district_extended
  data_files: series/district_extended.parquet
extra_gated_prompt: '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.'
extra_gated_fields:
  Name: text
  Email: text
  Organization: text
  I accept the CC BY-NC-SA licence terms (research and non-commercial use only): checkbox
extra_gated_button_content: Accept and request access
---
# HMIS-Bench

<!-- jm:top -->
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.
<!-- /jm:top -->

A probabilistic forecasting benchmark built from India's Health Management Information System (HMIS): monthly
counts that public and private health facilities report for each sub-district, April 2020 to March 2022, for 48
headline indicators of maternal health, deliveries, immunisation, family planning, disease and outpatient
services. The task is to forecast each series 1, 2 and 3 months ahead from six rolling origins (August 2021 to
January 2022) with nine quantiles. Four states are held out entirely, so models can be scored on sub-districts
they never saw. The release ships the series, the task definitions, the forecasts of nine models and an
`evaluate.py` that recomputes the leaderboard from those forecasts exactly.

## Quickstart

```python
import pandas as pd
# after `huggingface-cli download judicialmind/HMIS-Bench --repo-type dataset --local-dir HMIS-Bench`
bench = pd.read_parquet("HMIS-Bench/series/bench.parquet")      # or "hf://datasets/judicialmind/HMIS-Bench/series/bench.parquet"
print(bench.series_id.nunique(), "series,", len(bench), "rows")
print(bench.loc[bench.series_id == "hmis|10|188|1112|1.1|total", ["ts", "target"]].head(4).to_string(index=False))
```

```
100000 series, 2364616 rows
        ts  target
2020-04-01     0.0
2020-05-01  1721.0
2020-06-01  1207.0
2020-07-01  1255.0
```

With the `datasets` library: `load_dataset("judicialmind/HMIS-Bench", "bench", split="train")` (configs `bench`,
`all`, `district_extended`).

### Score the shipped forecasts

```bash
cd HMIS-Bench
pip install -r requirements.txt      # numpy, pandas, pyarrow
python evaluate.py --check          # recompute every leaderboard row from forecasts/ (about 2 minutes, 5 GB RAM)
```

```
 h               model    wql   mase  cov_80  win_rate  n_windows
 1               tirex 0.1435 0.8348  0.8278    0.5175     585323
 1            chronos2 0.1458 0.8380  0.8267       NaN     585323
 ...
checked 1377 of 1377 leaderboard rows against leaderboard.json: 0 mismatching values
```

### Evaluate your own model

1. For each origin in `splits.json`, get the benchmark windows: `make_windows(root, origin)` in `evaluate.py`
   returns `series_id`, a 36-month `context` (oldest first, NaN = month not reported) and the 3 observed
   `future` values. A window exists when the series has at least 8 reported months in the context and at least
   one reported month in the 3 months from the origin.
2. Forecast 3 months with the quantiles 0.1, 0.2, ..., 0.9 and write `forecasts/<your_model>/<origin>.parquet`
   with columns `series_id, origin, h, ts, q0.1 ... q0.9, mean, target` (one row per series and step `h` = 1..3;
   `target` = the observed value from `future`).
3. Run `python evaluate.py --models chronos2,<your_model>` (add the other models to compare on the same windows).

A simple example (last value ± a widening normal band):

```python
import os, sys
import numpy as np, pandas as pd
sys.path.insert(0, "HMIS-Bench")
from evaluate import make_windows, QCOLS
origin = "2021-08-01"                                                    # repeat for each origin in splits.json
w = make_windows("HMIS-Bench", origin)
last = np.array([c[np.isfinite(c)][-1] for c in w.context])             # last reported value
spread = np.array([np.nanstd(c) for c in w.context])
z = np.array([-1.2816, -0.8416, -0.5244, -0.2533, 0, 0.2533, 0.5244, 0.8416, 1.2816])
rows = []
for h in (1, 2, 3):
    q = np.maximum(last[:, None] + np.sqrt(h) * spread[:, None] * z[None, :], 0)
    df = pd.DataFrame(q, columns=QCOLS)
    df.insert(0, "series_id", w.series_id); df.insert(1, "origin", pd.Timestamp(origin).date())
    df.insert(2, "h", h); df.insert(3, "ts", (pd.Timestamp(origin) + pd.DateOffset(months=h - 1)).date())
    df["mean"] = df["q0.5"]; df["target"] = [f[h - 1] for f in w.future]
    rows.append(df)
os.makedirs("HMIS-Bench/forecasts/my_naive", exist_ok=True)
pd.concat(rows).to_parquet(f"HMIS-Bench/forecasts/my_naive/{origin}.parquet")
```

```bash
cd HMIS-Bench && python evaluate.py --models chronos2,my_naive    # scored on the windows both models cover
```

```
 h    model    wql   mase  cov_80  win_rate  n_windows
 1 chronos2 0.1624 0.9946  0.7843       NaN      96211
 1 my_naive 0.1673 1.0557  0.7489    0.5065      96211
 ...
```

## Files

| file | rows | content |
|---|---|---|
| `series/bench.parquet` | 2,364,616 | the benchmark: 100,000 sub-district series, 24 months each at most |
| `series/all.parquet` | 5,840,115 | every headline series: 264,835 series, 5,622 sub-districts, 704 districts, 33 States/UTs |
| `series/district_extended.parquet` | 809,458 | 24,260 district series of 38 indicators, April 2010 to March 2022 (district sheets FY2010-11 to FY2012-13 plus sub-district sums FY2020-21 to FY2021-22) |
| `headline_indicators.parquet` | 48 | headline name, HMIS indicator code and text |
| `indicators.parquet` | 320 | every HMIS indicator of the sub-district file with coverage statistics |
| `legacy_alignment.parquet` | 255 | how each older district-sheet item maps to a current HMIS indicator |
| `splits.json`, `tasks.json` | | origins, horizons, held-out states; one task per origin × horizon (fev-style) |
| `forecasts/<model>/<origin>.parquet` | | forecasts of the nine leaderboard models, 6 origins each |
| `leaderboard.json`, `leaderboard.parquet` | 1,377 | every model × horizon × slice score |
| `evaluate.py`, `requirements.txt` | | the scoring code (numpy + pandas) and its pinned dependencies |

## Schema

`series/bench.parquet` and `series/all.parquet` (one row per series and month):

| column | type | meaning |
|---|---|---|
| `series_id` | string | `hmis|<state_lgd>|<district_lgd>|<subdistrict>|<indicator>|total` |
| `ts` | date | month (first day) |
| `target` | float | reported count, public + private facilities |
| `state_lgd`, `district_lgd`, `subdistrict` | int | LGD codes of the State/UT, district and sub-district |
| `state_name`, `district_name`, `subdistrict_name` | string | names |
| `indicator`, `indicator_name` | string | HMIS indicator code (2017 format) and text |
| `headline` | string | short name of the headline indicator, e.g. `anc_registered`, `penta3`, `opd_allopathic` |
| `type` | string | `total` |
| `cov_public`, `cov_private`, `cov_rural`, `cov_urban` | float | the same count split by facility type and area (covariates) |

`series/district_extended.parquet`: `series_id` (`hmisd|<state_lgd>|<district_lgd>|<indicator>|total`), `ts`,
`target`, `state_lgd`, `district_lgd`, `indicator`, `type`, `origin_src` (`legacy` = FY2010-13 district sheets,
`subdistrict_sum` = FY2020-22 sub-district values summed to the district).

`indicators.parquet`: `indicator`, `indicator_name`, `section`, `n_rows`, `n_subdistricts`, `n_months`, `total`,
`zero_share` (share of reported zeros), `identity_fail_share` (share of cells where public + private differs from
rural + urban).

`legacy_alignment.parquet`: `legacy_name` (item text in the older sheets), `indicator`, `indicator_name`, `score`
(text similarity), `method` (`exact`, `jaccard` or `none` = not aligned).

Forecast files: `series_id`, `origin`, `h`, `ts`, `q0.1` ... `q0.9`, `mean`, `target` (see above).

## Benchmark

- **Series**: 100,000 sub-district × headline-indicator series, drawn in equal numbers from every
  (headline indicator, State/UT) stratum (smaller strata are taken whole) among the series with at least 18
  reported months and a non-zero total.
- **Origins**: 2021-08-01, 2021-09-01, 2021-10-01, 2021-11-01, 2021-12-01, 2022-01-01. Each forecast uses only
  months before the origin.
- **Horizons**: 1, 2 and 3 months; quantiles 0.1 to 0.9.
- **Held-out states**: Uttar Pradesh (9), Bihar (10), Odisha (21) and Andhra Pradesh (28): 18,555 of the 100,000
  series. Slice `heldout` = these states, `indist` = the rest.
- **Metrics**: WQL (weighted quantile loss, pooled over a task), MASE (seasonal period 12, falling back to lag 1
  when no lag-12 pair exists in the context), `cov_80` (empirical coverage of the 0.1-0.9 interval) and win rate
  against zero-shot Chronos-2 (share of windows with lower MASE). Scores are averaged over the 6 origins, on the
  599,714 windows that every model forecast.

## Results

All series (lower WQL and MASE are better; a calibrated 80% interval covers 0.80):

| h | model | WQL | MASE | cov_80 | win rate vs chronos2 | windows |
|---|---|---|---|---|---|---|
| 1 | `tirex` | 0.1435 | 0.8348 | 0.8278 | 0.5175 | 585,323 |
| 1 | `chronos2` | 0.1458 | 0.8380 | 0.8267 | – | 585,323 |
| 1 | `timesfm25` | 0.1459 | 0.8378 | 0.7568 | 0.5193 | 585,323 |
| 1 | `lightgbm` | 0.1473 | 0.8349 | 0.7989 | 0.5070 | 585,323 |
| 1 | `naive` | 0.1540 | 0.9081 | 0.7441 | 0.4833 | 585,323 |
| 1 | `chronos2_ft_hmis_s0` | 0.1553 | 0.8224 | 0.8019 | 0.4985 | 585,323 |
| 1 | `chronos2_ft_hmis_s1` | 0.1565 | 0.8249 | 0.8191 | 0.4954 | 585,323 |
| 1 | `ets_lite` | 0.1593 | 0.8854 | 0.6756 | 0.4603 | 585,323 |
| 1 | `seasonal_naive` | 0.3161 | 1.2229 | 0.5726 | 0.3744 | 585,323 |
| 2 | `lightgbm` | 0.1574 | 0.8917 | 0.7959 | 0.5050 | 587,203 |
| 2 | `tirex` | 0.1575 | 0.8893 | 0.8404 | 0.5222 | 587,203 |
| 2 | `timesfm25` | 0.1582 | 0.8944 | 0.7521 | 0.5217 | 587,203 |
| 2 | `chronos2` | 0.1594 | 0.8949 | 0.8363 | – | 587,203 |
| 2 | `chronos2_ft_hmis_s0` | 0.1671 | 0.8738 | 0.8013 | 0.5073 | 587,203 |
| 2 | `chronos2_ft_hmis_s1` | 0.1682 | 0.8760 | 0.8184 | 0.5028 | 587,203 |
| 2 | `naive` | 0.1709 | 0.9823 | 0.7683 | 0.4674 | 587,203 |
| 2 | `ets_lite` | 0.1775 | 0.9563 | 0.7102 | 0.4439 | 587,203 |
| 2 | `seasonal_naive` | 0.3073 | 1.2408 | 0.5766 | 0.3490 | 587,203 |
| 3 | `lightgbm` | 0.1634 | 0.9326 | 0.7923 | 0.5099 | 587,740 |
| 3 | `timesfm25` | 0.1662 | 0.9359 | 0.7472 | 0.5247 | 587,740 |
| 3 | `tirex` | 0.1665 | 0.9291 | 0.8476 | 0.5305 | 587,740 |
| 3 | `chronos2` | 0.1683 | 0.9376 | 0.8409 | – | 587,740 |
| 3 | `chronos2_ft_hmis_s0` | 0.1770 | 0.9153 | 0.8022 | 0.5094 | 587,740 |
| 3 | `chronos2_ft_hmis_s1` | 0.1784 | 0.9177 | 0.8161 | 0.5048 | 587,740 |
| 3 | `naive` | 0.1821 | 1.0405 | 0.7854 | 0.4570 | 587,740 |
| 3 | `ets_lite` | 0.1897 | 1.0110 | 0.7333 | 0.4346 | 587,740 |
| 3 | `seasonal_naive` | 0.2997 | 1.2552 | 0.5786 | 0.3425 | 587,740 |

Held-out states:

| h | model | WQL | MASE | cov_80 | win rate vs chronos2 | windows |
|---|---|---|---|---|---|---|
| 1 | `tirex` | 0.1662 | 0.8465 | 0.8343 | 0.5269 | 107,543 |
| 1 | `timesfm25` | 0.1669 | 0.8513 | 0.7692 | 0.5220 | 107,543 |
| 1 | `chronos2` | 0.1698 | 0.8530 | 0.8335 | – | 107,543 |
| 1 | `chronos2_ft_hmis_s0` | 0.1740 | 0.8543 | 0.8112 | 0.4782 | 107,543 |
| 1 | `lightgbm` | 0.1772 | 0.8560 | 0.8052 | 0.4986 | 107,543 |
| 2 | `timesfm25` | 0.1805 | 0.9203 | 0.7603 | 0.5284 | 107,873 |
| 2 | `tirex` | 0.1826 | 0.9173 | 0.8438 | 0.5317 | 107,873 |
| 2 | `chronos2` | 0.1862 | 0.9252 | 0.8395 | – | 107,873 |
| 2 | `chronos2_ft_hmis_s0` | 0.1888 | 0.9248 | 0.8069 | 0.4816 | 107,873 |
| 2 | `lightgbm` | 0.1888 | 0.9228 | 0.7983 | 0.5041 | 107,873 |
| 3 | `timesfm25` | 0.1929 | 0.9747 | 0.7516 | 0.5352 | 108,054 |
| 3 | `tirex` | 0.1968 | 0.9731 | 0.8485 | 0.5402 | 108,054 |
| 3 | `lightgbm` | 0.1984 | 0.9737 | 0.7911 | 0.5147 | 108,054 |
| 3 | `chronos2` | 0.2006 | 0.9838 | 0.8410 | – | 108,054 |
| 3 | `chronos2_ft_hmis_s0` | 0.2048 | 0.9848 | 0.8045 | 0.4804 | 108,054 |

Per-indicator slices (`headline=<name>`) are in `leaderboard.json` and the
[leaderboard Space](https://huggingface.co/spaces/judicialmind/HMIS-Bench-Leaderboard).

Models:
- `chronos2`: [amazon/chronos-2](https://huggingface.co/amazon/chronos-2), zero-shot, each series alone.
- `chronos2_ft_hmis_s0`: [HMIS-Chronos](https://huggingface.co/judicialmind/HMIS-Chronos), Chronos-2 fine-tuned on
  HMIS-Bench months before August 2021 (held-out states excluded), forecasting all indicators of a sub-district
  together; `chronos2_ft_hmis_s1` is the same model trained with another seed.
- `timesfm25`: [google/timesfm-2.5-200m-pytorch](https://huggingface.co/google/timesfm-2.5-200m-pytorch), zero-shot.
- `tirex`: [NX-AI/TiRex](https://huggingface.co/NX-AI/TiRex), zero-shot.
- `lightgbm`: one global LightGBM model on lagged values, trained on months before August 2021 without the
  held-out states.
- `naive` (last value), `seasonal_naive` (same month a year earlier) and `ets_lite` (damped trend exponential
  smoothing), with quantiles from each series' own past errors.

## Coverage

5,622 sub-districts in 704 districts of 33 States/UTs; 24 months, April 2020 to March 2022 (FY2020-21 and
FY2021-22, the COVID-19 years); 48 headline indicators out of the 320 HMIS indicators in the source file. The
district extension adds 38 indicators for FY2010-11 to FY2012-13 where the older district sheets name the same
item.

## Source data

Derived from the HMIS sub-district monthly reports (indicator-wise, FY2020-21 and FY2021-22) of the Ministry of
Health and Family Welfare, published on the Open Government Data Platform India
([data.gov.in](https://data.gov.in)), snapshot May 2026, under the
[Government Open Data License - India](https://data.gov.in/government-open-data-license-india). The district
extension uses the HMIS district-wise (DICR) sheets for FY2010-11 to FY2012-13 from the same Ministry.

Providers: Ministry of Health and Family Welfare (HMIS).

<!-- jm:source -->
Original data: published by Ministry of Health and Family Welfare (Health Management Information System, HMIS) on the Open Government Data Platform India ([data.gov.in](https://data.gov.in)) and, where marked, NITI Aayog's National Data and Analytics Platform ([NDAP](https://ndap.niti.gov.in)), under the [Government Open Data License - India](https://data.gov.in/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.
<!-- /jm:source -->

## Citation

```bibtex
@misc{judicialmind2026hmisbench,
  title={HMIS-Bench: Probabilistic Forecasting of Indian Sub-district Health Service Series},
  author={JudicialMind},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/judicialmind/HMIS-Bench}}
}
```

<!-- jm:license -->
## License

[CC BY-NC-SA 4.0](https://creativecommons.org/licenses/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.
<!-- /jm:license -->

	---
	license: cc-by-nc-sa-4.0
	pretty_name: HMIS-Bench
	language:
	- en
	task_categories:
	- time-series-forecasting
	tags:
	- health
	- india
	- hmis
	- time-series
	- forecasting
	- benchmark
	- probabilistic-forecasting
	size_categories:
	- 1M<n<10M
	configs:
	- config_name: bench
	data_files: series/bench.parquet
	default: true
	- config_name: all
	data_files: series/all.parquet
	- config_name: district_extended
	data_files: series/district_extended.parquet
	extra_gated_prompt: '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.'
	extra_gated_fields:
	Name: text
	Email: text
	Organization: text
	I accept the CC BY-NC-SA licence terms (research and non-commercial use only): checkbox
	extra_gated_button_content: Accept and request access
	---

	# HMIS-Bench

	<!-- jm:top -->
	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.
	<!-- /jm:top -->

	A probabilistic forecasting benchmark built from India's Health Management Information System (HMIS): monthly
	counts that public and private health facilities report for each sub-district, April 2020 to March 2022, for 48
	headline indicators of maternal health, deliveries, immunisation, family planning, disease and outpatient
	services. The task is to forecast each series 1, 2 and 3 months ahead from six rolling origins (August 2021 to
	January 2022) with nine quantiles. Four states are held out entirely, so models can be scored on sub-districts
	they never saw. The release ships the series, the task definitions, the forecasts of nine models and an
	`evaluate.py` that recomputes the leaderboard from those forecasts exactly.

	## Quickstart

	```python
	import pandas as pd

	# after `huggingface-cli download judicialmind/HMIS-Bench --repo-type dataset --local-dir HMIS-Bench`
	bench = pd.read_parquet("HMIS-Bench/series/bench.parquet") # or "hf://datasets/judicialmind/HMIS-Bench/series/bench.parquet"
	print(bench.series_id.nunique(), "series,", len(bench), "rows")
	print(bench.loc[bench.series_id == "hmis\|10\|188\|1112\|1.1\|total", ["ts", "target"]].head(4).to_string(index=False))
	```

	```
	100000 series, 2364616 rows
	ts target
	2020-04-01 0.0
	2020-05-01 1721.0
	2020-06-01 1207.0
	2020-07-01 1255.0
	```

	With the `datasets` library: `load_dataset("judicialmind/HMIS-Bench", "bench", split="train")` (configs `bench`,
	`all`, `district_extended`).

	### Score the shipped forecasts

	```bash
	cd HMIS-Bench
	pip install -r requirements.txt # numpy, pandas, pyarrow
	python evaluate.py --check # recompute every leaderboard row from forecasts/ (about 2 minutes, 5 GB RAM)
	```

	```
	h model wql mase cov_80 win_rate n_windows
	1 tirex 0.1435 0.8348 0.8278 0.5175 585323
	1 chronos2 0.1458 0.8380 0.8267 NaN 585323
	...
	checked 1377 of 1377 leaderboard rows against leaderboard.json: 0 mismatching values
	```

	### Evaluate your own model

	1. For each origin in `splits.json`, get the benchmark windows: `make_windows(root, origin)` in `evaluate.py`
	returns `series_id`, a 36-month `context` (oldest first, NaN = month not reported) and the 3 observed
	`future` values. A window exists when the series has at least 8 reported months in the context and at least
	one reported month in the 3 months from the origin.
	2. Forecast 3 months with the quantiles 0.1, 0.2, ..., 0.9 and write `forecasts/<your_model>/<origin>.parquet`
	with columns `series_id, origin, h, ts, q0.1 ... q0.9, mean, target` (one row per series and step `h` = 1..3;
	`target` = the observed value from `future`).
	3. Run `python evaluate.py --models chronos2,<your_model>` (add the other models to compare on the same windows).

	A simple example (last value ± a widening normal band):

	```python
	import os, sys
	import numpy as np, pandas as pd
	sys.path.insert(0, "HMIS-Bench")
	from evaluate import make_windows, QCOLS

	origin = "2021-08-01" # repeat for each origin in splits.json
	w = make_windows("HMIS-Bench", origin)
	last = np.array([c[np.isfinite(c)][-1] for c in w.context]) # last reported value
	spread = np.array([np.nanstd(c) for c in w.context])
	z = np.array([-1.2816, -0.8416, -0.5244, -0.2533, 0, 0.2533, 0.5244, 0.8416, 1.2816])
	rows = []
	for h in (1, 2, 3):
	q = np.maximum(last[:, None] + np.sqrt(h) * spread[:, None] * z[None, :], 0)
	df = pd.DataFrame(q, columns=QCOLS)
	df.insert(0, "series_id", w.series_id); df.insert(1, "origin", pd.Timestamp(origin).date())
	df.insert(2, "h", h); df.insert(3, "ts", (pd.Timestamp(origin) + pd.DateOffset(months=h - 1)).date())
	df["mean"] = df["q0.5"]; df["target"] = [f[h - 1] for f in w.future]
	rows.append(df)
	os.makedirs("HMIS-Bench/forecasts/my_naive", exist_ok=True)
	pd.concat(rows).to_parquet(f"HMIS-Bench/forecasts/my_naive/{origin}.parquet")
	```

	```bash
	cd HMIS-Bench && python evaluate.py --models chronos2,my_naive # scored on the windows both models cover
	```

	```
	h model wql mase cov_80 win_rate n_windows
	1 chronos2 0.1624 0.9946 0.7843 NaN 96211
	1 my_naive 0.1673 1.0557 0.7489 0.5065 96211
	...
	```

	## Files

	\| file \| rows \| content \|
	\|---\|---\|---\|
	\| `series/bench.parquet` \| 2,364,616 \| the benchmark: 100,000 sub-district series, 24 months each at most \|
	\| `series/all.parquet` \| 5,840,115 \| every headline series: 264,835 series, 5,622 sub-districts, 704 districts, 33 States/UTs \|
	\| `series/district_extended.parquet` \| 809,458 \| 24,260 district series of 38 indicators, April 2010 to March 2022 (district sheets FY2010-11 to FY2012-13 plus sub-district sums FY2020-21 to FY2021-22) \|
	\| `headline_indicators.parquet` \| 48 \| headline name, HMIS indicator code and text \|
	\| `indicators.parquet` \| 320 \| every HMIS indicator of the sub-district file with coverage statistics \|
	\| `legacy_alignment.parquet` \| 255 \| how each older district-sheet item maps to a current HMIS indicator \|
	\| `splits.json`, `tasks.json` \| \| origins, horizons, held-out states; one task per origin × horizon (fev-style) \|
	\| `forecasts/<model>/<origin>.parquet` \| \| forecasts of the nine leaderboard models, 6 origins each \|
	\| `leaderboard.json`, `leaderboard.parquet` \| 1,377 \| every model × horizon × slice score \|
	\| `evaluate.py`, `requirements.txt` \| \| the scoring code (numpy + pandas) and its pinned dependencies \|

	## Schema

	`series/bench.parquet` and `series/all.parquet` (one row per series and month):

	\| column \| type \| meaning \|
	\|---\|---\|---\|
	\| `series_id` \| string \| `hmis\|<state_lgd>\|<district_lgd>\|<subdistrict>\|<indicator>\|total` \|
	\| `ts` \| date \| month (first day) \|
	\| `target` \| float \| reported count, public + private facilities \|
	\| `state_lgd`, `district_lgd`, `subdistrict` \| int \| LGD codes of the State/UT, district and sub-district \|
	\| `state_name`, `district_name`, `subdistrict_name` \| string \| names \|
	\| `indicator`, `indicator_name` \| string \| HMIS indicator code (2017 format) and text \|
	\| `headline` \| string \| short name of the headline indicator, e.g. `anc_registered`, `penta3`, `opd_allopathic` \|
	\| `type` \| string \| `total` \|
	\| `cov_public`, `cov_private`, `cov_rural`, `cov_urban` \| float \| the same count split by facility type and area (covariates) \|

	`series/district_extended.parquet`: `series_id` (`hmisd\|<state_lgd>\|<district_lgd>\|<indicator>\|total`), `ts`,
	`target`, `state_lgd`, `district_lgd`, `indicator`, `type`, `origin_src` (`legacy` = FY2010-13 district sheets,
	`subdistrict_sum` = FY2020-22 sub-district values summed to the district).

	`indicators.parquet`: `indicator`, `indicator_name`, `section`, `n_rows`, `n_subdistricts`, `n_months`, `total`,
	`zero_share` (share of reported zeros), `identity_fail_share` (share of cells where public + private differs from
	rural + urban).

	`legacy_alignment.parquet`: `legacy_name` (item text in the older sheets), `indicator`, `indicator_name`, `score`
	(text similarity), `method` (`exact`, `jaccard` or `none` = not aligned).

	Forecast files: `series_id`, `origin`, `h`, `ts`, `q0.1` ... `q0.9`, `mean`, `target` (see above).

	## Benchmark

	- Series: 100,000 sub-district × headline-indicator series, drawn in equal numbers from every
	(headline indicator, State/UT) stratum (smaller strata are taken whole) among the series with at least 18
	reported months and a non-zero total.
	- Origins: 2021-08-01, 2021-09-01, 2021-10-01, 2021-11-01, 2021-12-01, 2022-01-01. Each forecast uses only
	months before the origin.
	- Horizons: 1, 2 and 3 months; quantiles 0.1 to 0.9.
	- Held-out states: Uttar Pradesh (9), Bihar (10), Odisha (21) and Andhra Pradesh (28): 18,555 of the 100,000
	series. Slice `heldout` = these states, `indist` = the rest.
	- Metrics: WQL (weighted quantile loss, pooled over a task), MASE (seasonal period 12, falling back to lag 1
	when no lag-12 pair exists in the context), `cov_80` (empirical coverage of the 0.1-0.9 interval) and win rate
	against zero-shot Chronos-2 (share of windows with lower MASE). Scores are averaged over the 6 origins, on the
	599,714 windows that every model forecast.

	## Results

	All series (lower WQL and MASE are better; a calibrated 80% interval covers 0.80):

	\| h \| model \| WQL \| MASE \| cov_80 \| win rate vs chronos2 \| windows \|
	\|---\|---\|---\|---\|---\|---\|---\|
	\| 1 \| `tirex` \| 0.1435 \| 0.8348 \| 0.8278 \| 0.5175 \| 585,323 \|
	\| 1 \| `chronos2` \| 0.1458 \| 0.8380 \| 0.8267 \| – \| 585,323 \|
	\| 1 \| `timesfm25` \| 0.1459 \| 0.8378 \| 0.7568 \| 0.5193 \| 585,323 \|
	\| 1 \| `lightgbm` \| 0.1473 \| 0.8349 \| 0.7989 \| 0.5070 \| 585,323 \|
	\| 1 \| `naive` \| 0.1540 \| 0.9081 \| 0.7441 \| 0.4833 \| 585,323 \|
	\| 1 \| `chronos2_ft_hmis_s0` \| 0.1553 \| 0.8224 \| 0.8019 \| 0.4985 \| 585,323 \|
	\| 1 \| `chronos2_ft_hmis_s1` \| 0.1565 \| 0.8249 \| 0.8191 \| 0.4954 \| 585,323 \|
	\| 1 \| `ets_lite` \| 0.1593 \| 0.8854 \| 0.6756 \| 0.4603 \| 585,323 \|
	\| 1 \| `seasonal_naive` \| 0.3161 \| 1.2229 \| 0.5726 \| 0.3744 \| 585,323 \|
	\| 2 \| `lightgbm` \| 0.1574 \| 0.8917 \| 0.7959 \| 0.5050 \| 587,203 \|
	\| 2 \| `tirex` \| 0.1575 \| 0.8893 \| 0.8404 \| 0.5222 \| 587,203 \|
	\| 2 \| `timesfm25` \| 0.1582 \| 0.8944 \| 0.7521 \| 0.5217 \| 587,203 \|
	\| 2 \| `chronos2` \| 0.1594 \| 0.8949 \| 0.8363 \| – \| 587,203 \|
	\| 2 \| `chronos2_ft_hmis_s0` \| 0.1671 \| 0.8738 \| 0.8013 \| 0.5073 \| 587,203 \|
	\| 2 \| `chronos2_ft_hmis_s1` \| 0.1682 \| 0.8760 \| 0.8184 \| 0.5028 \| 587,203 \|
	\| 2 \| `naive` \| 0.1709 \| 0.9823 \| 0.7683 \| 0.4674 \| 587,203 \|
	\| 2 \| `ets_lite` \| 0.1775 \| 0.9563 \| 0.7102 \| 0.4439 \| 587,203 \|
	\| 2 \| `seasonal_naive` \| 0.3073 \| 1.2408 \| 0.5766 \| 0.3490 \| 587,203 \|
	\| 3 \| `lightgbm` \| 0.1634 \| 0.9326 \| 0.7923 \| 0.5099 \| 587,740 \|
	\| 3 \| `timesfm25` \| 0.1662 \| 0.9359 \| 0.7472 \| 0.5247 \| 587,740 \|
	\| 3 \| `tirex` \| 0.1665 \| 0.9291 \| 0.8476 \| 0.5305 \| 587,740 \|
	\| 3 \| `chronos2` \| 0.1683 \| 0.9376 \| 0.8409 \| – \| 587,740 \|
	\| 3 \| `chronos2_ft_hmis_s0` \| 0.1770 \| 0.9153 \| 0.8022 \| 0.5094 \| 587,740 \|
	\| 3 \| `chronos2_ft_hmis_s1` \| 0.1784 \| 0.9177 \| 0.8161 \| 0.5048 \| 587,740 \|
	\| 3 \| `naive` \| 0.1821 \| 1.0405 \| 0.7854 \| 0.4570 \| 587,740 \|
	\| 3 \| `ets_lite` \| 0.1897 \| 1.0110 \| 0.7333 \| 0.4346 \| 587,740 \|
	\| 3 \| `seasonal_naive` \| 0.2997 \| 1.2552 \| 0.5786 \| 0.3425 \| 587,740 \|

	Held-out states:

	\| h \| model \| WQL \| MASE \| cov_80 \| win rate vs chronos2 \| windows \|
	\|---\|---\|---\|---\|---\|---\|---\|
	\| 1 \| `tirex` \| 0.1662 \| 0.8465 \| 0.8343 \| 0.5269 \| 107,543 \|
	\| 1 \| `timesfm25` \| 0.1669 \| 0.8513 \| 0.7692 \| 0.5220 \| 107,543 \|
	\| 1 \| `chronos2` \| 0.1698 \| 0.8530 \| 0.8335 \| – \| 107,543 \|
	\| 1 \| `chronos2_ft_hmis_s0` \| 0.1740 \| 0.8543 \| 0.8112 \| 0.4782 \| 107,543 \|
	\| 1 \| `lightgbm` \| 0.1772 \| 0.8560 \| 0.8052 \| 0.4986 \| 107,543 \|
	\| 2 \| `timesfm25` \| 0.1805 \| 0.9203 \| 0.7603 \| 0.5284 \| 107,873 \|
	\| 2 \| `tirex` \| 0.1826 \| 0.9173 \| 0.8438 \| 0.5317 \| 107,873 \|
	\| 2 \| `chronos2` \| 0.1862 \| 0.9252 \| 0.8395 \| – \| 107,873 \|
	\| 2 \| `chronos2_ft_hmis_s0` \| 0.1888 \| 0.9248 \| 0.8069 \| 0.4816 \| 107,873 \|
	\| 2 \| `lightgbm` \| 0.1888 \| 0.9228 \| 0.7983 \| 0.5041 \| 107,873 \|
	\| 3 \| `timesfm25` \| 0.1929 \| 0.9747 \| 0.7516 \| 0.5352 \| 108,054 \|
	\| 3 \| `tirex` \| 0.1968 \| 0.9731 \| 0.8485 \| 0.5402 \| 108,054 \|
	\| 3 \| `lightgbm` \| 0.1984 \| 0.9737 \| 0.7911 \| 0.5147 \| 108,054 \|
	\| 3 \| `chronos2` \| 0.2006 \| 0.9838 \| 0.8410 \| – \| 108,054 \|
	\| 3 \| `chronos2_ft_hmis_s0` \| 0.2048 \| 0.9848 \| 0.8045 \| 0.4804 \| 108,054 \|

	Per-indicator slices (`headline=<name>`) are in `leaderboard.json` and the
	[leaderboard Space](https://huggingface.co/spaces/judicialmind/HMIS-Bench-Leaderboard).

	Models:
	- `chronos2`: [amazon/chronos-2](https://huggingface.co/amazon/chronos-2), zero-shot, each series alone.
	- `chronos2_ft_hmis_s0`: [HMIS-Chronos](https://huggingface.co/judicialmind/HMIS-Chronos), Chronos-2 fine-tuned on
	HMIS-Bench months before August 2021 (held-out states excluded), forecasting all indicators of a sub-district
	together; `chronos2_ft_hmis_s1` is the same model trained with another seed.
	- `timesfm25`: [google/timesfm-2.5-200m-pytorch](https://huggingface.co/google/timesfm-2.5-200m-pytorch), zero-shot.
	- `tirex`: [NX-AI/TiRex](https://huggingface.co/NX-AI/TiRex), zero-shot.
	- `lightgbm`: one global LightGBM model on lagged values, trained on months before August 2021 without the
	held-out states.
	- `naive` (last value), `seasonal_naive` (same month a year earlier) and `ets_lite` (damped trend exponential
	smoothing), with quantiles from each series' own past errors.

	## Coverage

	5,622 sub-districts in 704 districts of 33 States/UTs; 24 months, April 2020 to March 2022 (FY2020-21 and
	FY2021-22, the COVID-19 years); 48 headline indicators out of the 320 HMIS indicators in the source file. The
	district extension adds 38 indicators for FY2010-11 to FY2012-13 where the older district sheets name the same
	item.

	## Source data

	Derived from the HMIS sub-district monthly reports (indicator-wise, FY2020-21 and FY2021-22) of the Ministry of
	Health and Family Welfare, published on the Open Government Data Platform India
	([data.gov.in](https://data.gov.in)), snapshot May 2026, under the
	[Government Open Data License - India](https://data.gov.in/government-open-data-license-india). The district
	extension uses the HMIS district-wise (DICR) sheets for FY2010-11 to FY2012-13 from the same Ministry.

	Providers: Ministry of Health and Family Welfare (HMIS).

	<!-- jm:source -->
	Original data: published by Ministry of Health and Family Welfare (Health Management Information System, HMIS) on the Open Government Data Platform India ([data.gov.in](https://data.gov.in)) and, where marked, NITI Aayog's National Data and Analytics Platform ([NDAP](https://ndap.niti.gov.in)), under the [Government Open Data License - India](https://data.gov.in/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.
	<!-- /jm:source -->

	## Citation

	```bibtex
	@misc{judicialmind2026hmisbench,
	title={HMIS-Bench: Probabilistic Forecasting of Indian Sub-district Health Service Series},
	author={JudicialMind},
	year={2026},
	publisher={Hugging Face},
	howpublished={\url{https://huggingface.co/datasets/judicialmind/HMIS-Bench}}
	}
	```

	<!-- jm:license -->
	## License

	[CC BY-NC-SA 4.0](https://creativecommons.org/licenses/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.
	<!-- /jm:license -->