Datasets:

judicialmind
/

MandiBench

---
license: cc-by-nc-sa-4.0
pretty_name: MandiBench
language:
- en
task_categories:
- time-series-forecasting
tags:
- agriculture
- india
- mandi
- agmarknet
- prices
- forecasting
- benchmark
size_categories:
- 10M<n<100M
configs:
- config_name: series
  data_files: series/all.parquet
- config_name: catalog
  data_files: series/catalog.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
---
# MandiBench

<!-- 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 -->

MandiBench is a probabilistic forecasting benchmark for daily wholesale (mandi) prices in India. Each of its 12,863
series is the daily modal price of one commodity variety in one Agmarknet market (Rs per quintal), with history
from 2001 to May 2026. Forecasters predict the next 7, 14 and 30 days from 11 monthly origins (2025-06-01 to
2026-04-01), as nine quantiles (q0.1 ... q0.9). Nine state/UT groups are held out entirely, to test transfer to
markets a trained model has never seen. The repository ships the series, the forecasts of nine models, the
leaderboard and `evaluate.py`, which recomputes the leaderboard from the shipped forecasts and scores new models.

**Headline result:** our Chronos-2 fine-tunes top the leaderboard at every horizon, on all series and in the held-out
states: the group-context fine-tune reaches WQL 0.0930 at 30 days (0.0796 held-out) and the released
[Mandi-Chronos](https://huggingface.co/judicialmind/Mandi-Chronos) 0.0990 (0.0847), against 0.1012 for TimesFM-2.5
and 0.1025 for zero-shot Chronos-2.

## Quickstart

```python
from huggingface_hub import snapshot_download
root = snapshot_download("judicialmind/MandiBench", repo_type="dataset")
import pandas as pd
series = pd.read_parquet(f"{root}/series/all.parquet")
print(series.shape)
print(series[["series_id", "ts", "target", "cclass"]].head(3).to_string())
```

Output: `(27437227, 14)` (27,437,227 rows, 14 columns), then:

**First three rows of `series/all.parquet`**

| series id | date | target | commodity class |
|---|---|---:|---|
| mandi\|Andhra Pradesh\|Anantapur\|Anantapur\|Mousambi(Sweet Lime)\|Mousambi | 2014-05-27 | 1,600 | fruit |
| mandi\|Andhra Pradesh\|Anantapur\|Anantapur\|Mousambi(Sweet Lime)\|Mousambi | 2014-06-07 | 1,570 | fruit |
| mandi\|Andhra Pradesh\|Anantapur\|Anantapur\|Mousambi(Sweet Lime)\|Mousambi | 2014-06-20 | 1,400 | fruit |

Other loaders: `datasets.load_dataset("judicialmind/MandiBench", "series", split="train")`, or DuckDB
(`SELECT commodity, avg(target) FROM 'series/all.parquet' GROUP BY 1`).

### Evaluate a model

`evaluate.py` (numpy, pandas, pyarrow; `pip install -r requirements.txt`) exposes the benchmark windows and their
contexts as arrays. Row `i` of every array is window `bench.windows.iloc[i]` (`series_id, origin, state, cclass,
heldout`). Input: `bench.context` (windows x 512) holds the daily prices of the 512 days before the origin (NaN = no
trade that day). Output: an array of shape (windows, 30, 9), the forecast for days origin ... origin + 29 at the nine
quantile levels in `QUANTILES`.
```python
import sys
import numpy as np
from statistics import NormalDist
sys.path.insert(0, root)
from evaluate import QUANTILES, Benchmark
bench = Benchmark(root)
print(len(bench.windows), "windows")
# toy forecaster: last observed price, normal spread of 5% x sqrt(h)
z = np.array([NormalDist().inv_cdf(q) for q in QUANTILES])
h = np.arange(1, 31)[:, None]
Q = bench.last[:, None, None] * (1 + 0.05 * np.sqrt(h) * z)   # (windows, 30 days, 9 quantiles)

res = bench.score(Q, "last_value")
print(res[res["slice"].isin(["all", "heldout"])][["h", "slice", "wql", "mase", "cov_80"]].round(4).to_string(index=False))
```
Output: `96648 windows` (96,648 benchmark windows), then:
**Toy forecaster `last_value`, by horizon and slice**
| horizon (days) | slice | WQL | MASE | 80% coverage |
|---:|---|---:|---:|---:|
| 7 | all | 0.0791 | 1.1059 | 0.7254 |
| 7 | heldout | 0.0649 | 1.0358 | 0.7948 |
| 14 | all | 0.0936 | 1.4293 | 0.7511 |
| 14 | heldout | 0.0787 | 1.2900 | 0.8131 |
| 30 | all | 0.1204 | 1.8335 | 0.7807 |
| 30 | heldout | 0.1036 | 1.6976 | 0.8343 |
From the command line (run inside the downloaded folder):
```bash
python evaluate.py                                    # recompute the leaderboard from forecasts/ -> "REPRODUCED"
python evaluate.py --write-windows windows.csv        # the 96,648 (series_id, origin) windows to forecast
python evaluate.py --submission my_forecasts.parquet --name my_model --out my_model.csv
```
A submission file (or a folder of parquet/csv files) has one row per window and day: `series_id, origin, h, q0.1,
q0.2, ..., q0.9`, with `origin` the first forecast day (ISO date) and `h` = 1 ... 30 (day `origin + h - 1`). Use only
observations dated before `origin`. The output adds `win_rate`, the share of windows where the model's MASE beats
`chronos2`. To appear on the [MandiBench-Leaderboard](https://huggingface.co/spaces/judicialmind/MandiBench-Leaderboard),
open a discussion on this repository with the `--out` csv and a short model description.

## Tasks and metrics

- **Origins:** the first day of each month from 2025-06-01 to 2026-04-01 (11). The context is every observation
  before the origin; a model may use the full history (the shipped models read at most the last 512 days).
- **Horizons:** 7, 14 and 30 days. One call per (series, origin) forecasts 30 days; a horizon-h task scores days
  1 ... h. Only days with an observed price count.
- **Held-out states:** Chandigarh, Himachal Pradesh, Madhya Pradesh, Manipur, Meghalaya, Odisha, Rajasthan,
  Tamil Nadu, West Bengal (`splits.json`). The trained models (LightGBM and the fine-tuned Chronos-2 variants) used
  only data before 2025-06-01 from the other states. Slices: `all`, `heldout`, `indist`, and one per commodity class
  (`cclass=cereal`, ...).
- **Metrics** (computed per origin x horizon, then averaged over origins):
  - `wql`: weighted quantile loss, the mean over the 9 quantiles of 2 x summed pinball loss / summed |y|, pooled
    over the windows of the task. Lower is better.
  - `mase`: mean absolute error of q0.5 divided by the mean absolute 7-day change of the 512-day context (lag 1
    when no 7-day pair is observed), averaged over windows. Lower is better.
  - `cov_80`: share of observed days inside [q0.1, q0.9]; 0.80 is ideal.
  - `dir_acc_7d`: share of windows where q0.5 moves in the same direction as the price from the last observed
    value, at the last observed day within the first week (windows with no change are skipped).
  - `n_windows`: windows with a defined MASE.
  The leaderboard scores the 96,648 (series, origin) windows that all nine models forecast.
## Results

Leaderboard, sorted by WQL at 30 days (`leaderboard.parquet`; per-class slices and `win_rate` vs `chronos2` are in
`leaderboard.json`).

**All series** (windows: 88,899 at h=7, 92,776 at h=14, 96,341 at h=30)

| model | WQL h7 | WQL h14 | WQL h30 | MASE h7 | MASE h14 | MASE h30 | cov_80 h30 | dir_acc_7d |
|---|---|---|---|---|---|---|---|---|
| chronos2_ft_b | 0.0633 | 0.0736 | 0.0930 | 1.0660 | 1.3494 | 1.6990 | 0.7289 | 0.6729 |
| chronos2_ft_c | 0.0656 | 0.0765 | 0.0969 | 1.1077 | 1.4087 | 1.7858 | 0.7678 | 0.6370 |
| chronos2_ft_a ([Mandi-Chronos](https://huggingface.co/judicialmind/Mandi-Chronos)) | 0.0665 | 0.0778 | 0.0990 | 1.1126 | 1.4177 | 1.8018 | 0.7215 | 0.6271 |
| timesfm25 | 0.0672 | 0.0793 | 0.1012 | 1.0833 | 1.4031 | 1.8046 | 0.7410 | 0.6109 |
| chronos2 | 0.0677 | 0.0799 | 0.1025 | 1.1487 | 1.4675 | 1.8759 | 0.7490 | 0.6156 |
| tirex | 0.0709 | 0.0827 | 0.1038 | 1.2017 | 1.5219 | 1.9185 | 0.7605 | 0.6204 |
| lightgbm | 0.0700 | 0.0831 | 0.1087 | 1.2280 | 1.6134 | 2.1358 | 0.7704 | 0.6025 |
| ets_lite | 0.0708 | 0.0862 | 0.1153 | 1.0865 | 1.4099 | 1.8143 | 0.8532 | 0.6109 |
| seasonal_naive | 0.0845 | 0.0957 | 0.1193 | 1.2682 | 1.5662 | 1.9484 | 0.8003 | 0.1037 |
**Held-out states** (windows: 13,176 at h=7, 13,875 at h=14, 14,479 at h=30)
| model | WQL h7 | WQL h14 | WQL h30 | MASE h7 | MASE h14 | MASE h30 | cov_80 h30 | dir_acc_7d |
|---|---|---|---|---|---|---|---|---|
| chronos2_ft_b | 0.0511 | 0.0609 | 0.0796 | 1.0379 | 1.2741 | 1.6773 | 0.7126 | 0.6531 |
| chronos2_ft_c | 0.0523 | 0.0626 | 0.0819 | 1.0734 | 1.3154 | 1.7252 | 0.7608 | 0.6264 |
| chronos2_ft_a ([Mandi-Chronos](https://huggingface.co/judicialmind/Mandi-Chronos)) | 0.0532 | 0.0641 | 0.0847 | 1.0734 | 1.3243 | 1.7485 | 0.7221 | 0.6150 |
| chronos2 | 0.0541 | 0.0652 | 0.0858 | 1.1387 | 1.3944 | 1.8477 | 0.7466 | 0.6008 |
| tirex | 0.0559 | 0.0665 | 0.0863 | 1.1729 | 1.4408 | 1.8897 | 0.7618 | 0.6137 |
| timesfm25 | 0.0541 | 0.0653 | 0.0869 | 1.0336 | 1.2881 | 1.7270 | 0.7269 | 0.5955 |
| lightgbm | 0.0551 | 0.0666 | 0.0894 | 1.0864 | 1.3739 | 1.9001 | 0.7963 | 0.6002 |
| ets_lite | 0.0569 | 0.0703 | 0.0947 | 1.0352 | 1.2910 | 1.7046 | 0.8391 | 0.6014 |
| seasonal_naive | 0.0682 | 0.0783 | 0.0990 | 1.1941 | 1.4292 | 1.8256 | 0.7936 | 0.1106 |

Models:

| model | description |
|---|---|
| `seasonal_naive` | price on the same weekday of the last week; spread from past 7-day changes (its median often equals the last price, which scores as a wrong direction in `dir_acc_7d`) |
| `ets_lite` | damped additive-trend exponential smoothing per series; spread from one-step residuals |
| `lightgbm` | global LightGBM quantile regression on scaled lags, rolling statistics and calendar features |
| `chronos2` | `amazon/chronos-2`, zero-shot |
| `timesfm25` | `google/timesfm-2.5-200m-pytorch`, zero-shot |
| `tirex` | `NX-AI/TiRex`, zero-shot |
| `chronos2_ft_a` | Chronos-2 fine-tuned on MandiBench series, univariate: released as [Mandi-Chronos](https://huggingface.co/judicialmind/Mandi-Chronos) |
| `chronos2_ft_b` | Chronos-2 fine-tuned and run with group context: the series of one state and commodity are forecast jointly |
| `chronos2_ft_c` | Chronos-2 fine-tuned with daily soil-moisture and reservoir-storage covariates (NRSC and CWC data from data.gov.in, not included here) |

## Files

| file | content |
|---|---|
| `series/all.parquet` | the benchmark series, one row per series and trading day (27,437,227 rows) |
| `series/catalog.parquet` | all 178,352 candidate market x commodity x variety series with coverage counts; `kept` marks the 12,863 benchmark series |
| `splits.json` | origins, horizons, held-out states, season length (7) |
| `tasks.json` | the 33 origin x horizon tasks in fev-style task format (id/timestamp/target columns, horizon, quantile levels) |
| `forecasts/<model>/<origin>.parquet` | the forecasts of the nine models, one file per origin |
| `leaderboard.parquet`, `leaderboard.json` | the leaderboard (324 rows: model x horizon x slice); the json adds `win_rate` |
| `evaluate.py`, `requirements.txt` | evaluation harness |

## Schema

`series/all.parquet`

| column | type | description |
|---|---|---|
| `series_id` | string | `mandi|state|district|market|commodity|variety` |
| `ts` | date | trading day |
| `target` | float64 | modal price, Rs per quintal (median over the grades reported that day) |
| `state`, `district`, `market` | string | Agmarknet market location |
| `commodity`, `variety` | string | Agmarknet commodity and variety names |
| `cclass` | string | commodity class: cereal, pulse, oilseed, vegetable, fruit, spice, sugar, dairy_meat_fish, top (tomato, onion, potato) |
| `grade` | string | always `NA` (grades are merged) |
| `district_lgd` | int32 | empty in this release |
| `cov_min_price`, `cov_max_price` | float64 | minimum and maximum price that day (median over grades), Rs per quintal |
| `n_grades` | int64 | number of grade rows merged into the day |

`series/catalog.parquet`: `state, district, market, commodity, variety, cclass`, `first_ts`/`last_ts` (first and last
trading day), `n_obs_pre` / `first_pre` (days and first day before 2025-06-01), `n_obs_post` (days from 2025-06-01),
`n_mkt_days` (trading days of the market from `first_pre` to 2025-06-01), `n_obs_recent` / `mkt_days_recent` (days
of the series and of its market in the two years before 2025-06-01), `obs_ratio` (= `n_obs_recent /
mkt_days_recent`), `series_id`, `kept`.
`forecasts/<model>/<origin>.parquet`: `series_id, origin, h, ts, q0.1 ... q0.9, mean, target` (one row per window and
forecast day; `target` = observed price, null when the market did not trade).

## Coverage

12,863 series from 2,126 markets in 457 districts of 24 states and union territories, 176 food commodities in
nine classes (vegetables 4,622 series, tomato/onion/potato 2,030, cereals 1,742, fruit 1,510, spices 1,285,
oilseeds 798, pulses 725, sugar 129, dairy/meat/fish 22). Daily prices from 2001-01-10 to 2026-05-15; 25,729,325
rows before the first origin and 1,707,902 from it. A series is included when it has at least two years of history
before 2025-06-01 and traded on at least half of its market's trading days in the two years before that date.

The prices come from 83.2 million Agmarknet rows. 14.0 million exact paging duplicates were removed, as were
non-food commodities (4.6 million rows) and 90,130 flagged rows (unit slips, outliers and non-positive prices).

## Source data

Derived from 1 resource published on the Open Government Data Platform India (data.gov.in), snapshot May 2026,
under the [Government Open Data License - India](https://data.gov.in/government-open-data-license-india).

Providers: Directorate of Marketing and Inspection (Agmarknet), Ministry of Agriculture and Farmers Welfare.

- Directorate of Marketing and Inspection (Agmarknet), Ministry of Agriculture and Farmers Welfare. "Variety-wise Daily
  Market Prices Data of Commodity". data.gov.in resource
  35985678-0d79-46b4-9ed6-6f13308a1d24, https://api.data.gov.in/resource/35985678-0d79-46b4-9ed6-6f13308a1d24.
  Snapshot: May 2026.

GODL-India does not cover personal information; this dataset holds market prices only.

<!-- jm:source -->
Original data: published by Directorate of Marketing and Inspection (Agmarknet), Ministry of Agriculture and Farmers Welfare 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{judicialmind2026mandibench,
  title        = {MandiBench: a probabilistic forecasting benchmark for daily mandi prices in India},
  author       = {JudicialMind},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/judicialmind/MandiBench}}
}
```

<!-- 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: MandiBench
	language:
	- en
	task_categories:
	- time-series-forecasting
	tags:
	- agriculture
	- india
	- mandi
	- agmarknet
	- prices
	- forecasting
	- benchmark
	size_categories:
	- 10M<n<100M
	configs:
	- config_name: series
	data_files: series/all.parquet
	- config_name: catalog
	data_files: series/catalog.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
	---

	# MandiBench

	<!-- 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 -->

	MandiBench is a probabilistic forecasting benchmark for daily wholesale (mandi) prices in India. Each of its 12,863
	series is the daily modal price of one commodity variety in one Agmarknet market (Rs per quintal), with history
	from 2001 to May 2026. Forecasters predict the next 7, 14 and 30 days from 11 monthly origins (2025-06-01 to
	2026-04-01), as nine quantiles (q0.1 ... q0.9). Nine state/UT groups are held out entirely, to test transfer to
	markets a trained model has never seen. The repository ships the series, the forecasts of nine models, the
	leaderboard and `evaluate.py`, which recomputes the leaderboard from the shipped forecasts and scores new models.

	Headline result: our Chronos-2 fine-tunes top the leaderboard at every horizon, on all series and in the held-out
	states: the group-context fine-tune reaches WQL 0.0930 at 30 days (0.0796 held-out) and the released
	[Mandi-Chronos](https://huggingface.co/judicialmind/Mandi-Chronos) 0.0990 (0.0847), against 0.1012 for TimesFM-2.5
	and 0.1025 for zero-shot Chronos-2.

	## Quickstart

	```python
	from huggingface_hub import snapshot_download

	root = snapshot_download("judicialmind/MandiBench", repo_type="dataset")

	import pandas as pd

	series = pd.read_parquet(f"{root}/series/all.parquet")
	print(series.shape)
	print(series[["series_id", "ts", "target", "cclass"]].head(3).to_string())
	```

	Output: `(27437227, 14)` (27,437,227 rows, 14 columns), then:

	First three rows of `series/all.parquet`

	\| series id \| date \| target \| commodity class \|
	\|---\|---\|---:\|---\|
	\| mandi\\|Andhra Pradesh\\|Anantapur\\|Anantapur\\|Mousambi(Sweet Lime)\\|Mousambi \| 2014-05-27 \| 1,600 \| fruit \|
	\| mandi\\|Andhra Pradesh\\|Anantapur\\|Anantapur\\|Mousambi(Sweet Lime)\\|Mousambi \| 2014-06-07 \| 1,570 \| fruit \|
	\| mandi\\|Andhra Pradesh\\|Anantapur\\|Anantapur\\|Mousambi(Sweet Lime)\\|Mousambi \| 2014-06-20 \| 1,400 \| fruit \|

	Other loaders: `datasets.load_dataset("judicialmind/MandiBench", "series", split="train")`, or DuckDB
	(`SELECT commodity, avg(target) FROM 'series/all.parquet' GROUP BY 1`).

	### Evaluate a model

	`evaluate.py` (numpy, pandas, pyarrow; `pip install -r requirements.txt`) exposes the benchmark windows and their
	contexts as arrays. Row `i` of every array is window `bench.windows.iloc[i]` (`series_id, origin, state, cclass,
	heldout`). Input: `bench.context` (windows x 512) holds the daily prices of the 512 days before the origin (NaN = no
	trade that day). Output: an array of shape (windows, 30, 9), the forecast for days origin ... origin + 29 at the nine
	quantile levels in `QUANTILES`.

	```python
	import sys

	import numpy as np
	from statistics import NormalDist

	sys.path.insert(0, root)
	from evaluate import QUANTILES, Benchmark

	bench = Benchmark(root)
	print(len(bench.windows), "windows")

	# toy forecaster: last observed price, normal spread of 5% x sqrt(h)
	z = np.array([NormalDist().inv_cdf(q) for q in QUANTILES])
	h = np.arange(1, 31)[:, None]
	Q = bench.last[:, None, None] * (1 + 0.05 * np.sqrt(h) * z) # (windows, 30 days, 9 quantiles)

	res = bench.score(Q, "last_value")
	print(res[res["slice"].isin(["all", "heldout"])][["h", "slice", "wql", "mase", "cov_80"]].round(4).to_string(index=False))
	```

	Output: `96648 windows` (96,648 benchmark windows), then:

	Toy forecaster `last_value`, by horizon and slice

	\| horizon (days) \| slice \| WQL \| MASE \| 80% coverage \|
	\|---:\|---\|---:\|---:\|---:\|
	\| 7 \| all \| 0.0791 \| 1.1059 \| 0.7254 \|
	\| 7 \| heldout \| 0.0649 \| 1.0358 \| 0.7948 \|
	\| 14 \| all \| 0.0936 \| 1.4293 \| 0.7511 \|
	\| 14 \| heldout \| 0.0787 \| 1.2900 \| 0.8131 \|
	\| 30 \| all \| 0.1204 \| 1.8335 \| 0.7807 \|
	\| 30 \| heldout \| 0.1036 \| 1.6976 \| 0.8343 \|

	From the command line (run inside the downloaded folder):

	```bash
	python evaluate.py # recompute the leaderboard from forecasts/ -> "REPRODUCED"
	python evaluate.py --write-windows windows.csv # the 96,648 (series_id, origin) windows to forecast
	python evaluate.py --submission my_forecasts.parquet --name my_model --out my_model.csv
	```

	A submission file (or a folder of parquet/csv files) has one row per window and day: `series_id, origin, h, q0.1,
	q0.2, ..., q0.9`, with `origin` the first forecast day (ISO date) and `h` = 1 ... 30 (day `origin + h - 1`). Use only
	observations dated before `origin`. The output adds `win_rate`, the share of windows where the model's MASE beats
	`chronos2`. To appear on the [MandiBench-Leaderboard](https://huggingface.co/spaces/judicialmind/MandiBench-Leaderboard),
	open a discussion on this repository with the `--out` csv and a short model description.

	## Tasks and metrics

	- Origins: the first day of each month from 2025-06-01 to 2026-04-01 (11). The context is every observation
	before the origin; a model may use the full history (the shipped models read at most the last 512 days).
	- Horizons: 7, 14 and 30 days. One call per (series, origin) forecasts 30 days; a horizon-h task scores days
	1 ... h. Only days with an observed price count.
	- Held-out states: Chandigarh, Himachal Pradesh, Madhya Pradesh, Manipur, Meghalaya, Odisha, Rajasthan,
	Tamil Nadu, West Bengal (`splits.json`). The trained models (LightGBM and the fine-tuned Chronos-2 variants) used
	only data before 2025-06-01 from the other states. Slices: `all`, `heldout`, `indist`, and one per commodity class
	(`cclass=cereal`, ...).
	- Metrics (computed per origin x horizon, then averaged over origins):
	- `wql`: weighted quantile loss, the mean over the 9 quantiles of 2 x summed pinball loss / summed \|y\|, pooled
	over the windows of the task. Lower is better.
	- `mase`: mean absolute error of q0.5 divided by the mean absolute 7-day change of the 512-day context (lag 1
	when no 7-day pair is observed), averaged over windows. Lower is better.
	- `cov_80`: share of observed days inside [q0.1, q0.9]; 0.80 is ideal.
	- `dir_acc_7d`: share of windows where q0.5 moves in the same direction as the price from the last observed
	value, at the last observed day within the first week (windows with no change are skipped).
	- `n_windows`: windows with a defined MASE.
	The leaderboard scores the 96,648 (series, origin) windows that all nine models forecast.

	## Results

	Leaderboard, sorted by WQL at 30 days (`leaderboard.parquet`; per-class slices and `win_rate` vs `chronos2` are in
	`leaderboard.json`).

	All series (windows: 88,899 at h=7, 92,776 at h=14, 96,341 at h=30)

	\| model \| WQL h7 \| WQL h14 \| WQL h30 \| MASE h7 \| MASE h14 \| MASE h30 \| cov_80 h30 \| dir_acc_7d \|
	\|---\|---\|---\|---\|---\|---\|---\|---\|---\|
	\| chronos2_ft_b \| 0.0633 \| 0.0736 \| 0.0930 \| 1.0660 \| 1.3494 \| 1.6990 \| 0.7289 \| 0.6729 \|
	\| chronos2_ft_c \| 0.0656 \| 0.0765 \| 0.0969 \| 1.1077 \| 1.4087 \| 1.7858 \| 0.7678 \| 0.6370 \|
	\| chronos2_ft_a ([Mandi-Chronos](https://huggingface.co/judicialmind/Mandi-Chronos)) \| 0.0665 \| 0.0778 \| 0.0990 \| 1.1126 \| 1.4177 \| 1.8018 \| 0.7215 \| 0.6271 \|
	\| timesfm25 \| 0.0672 \| 0.0793 \| 0.1012 \| 1.0833 \| 1.4031 \| 1.8046 \| 0.7410 \| 0.6109 \|
	\| chronos2 \| 0.0677 \| 0.0799 \| 0.1025 \| 1.1487 \| 1.4675 \| 1.8759 \| 0.7490 \| 0.6156 \|
	\| tirex \| 0.0709 \| 0.0827 \| 0.1038 \| 1.2017 \| 1.5219 \| 1.9185 \| 0.7605 \| 0.6204 \|
	\| lightgbm \| 0.0700 \| 0.0831 \| 0.1087 \| 1.2280 \| 1.6134 \| 2.1358 \| 0.7704 \| 0.6025 \|
	\| ets_lite \| 0.0708 \| 0.0862 \| 0.1153 \| 1.0865 \| 1.4099 \| 1.8143 \| 0.8532 \| 0.6109 \|
	\| seasonal_naive \| 0.0845 \| 0.0957 \| 0.1193 \| 1.2682 \| 1.5662 \| 1.9484 \| 0.8003 \| 0.1037 \|

	Held-out states (windows: 13,176 at h=7, 13,875 at h=14, 14,479 at h=30)

	\| model \| WQL h7 \| WQL h14 \| WQL h30 \| MASE h7 \| MASE h14 \| MASE h30 \| cov_80 h30 \| dir_acc_7d \|
	\|---\|---\|---\|---\|---\|---\|---\|---\|---\|
	\| chronos2_ft_b \| 0.0511 \| 0.0609 \| 0.0796 \| 1.0379 \| 1.2741 \| 1.6773 \| 0.7126 \| 0.6531 \|
	\| chronos2_ft_c \| 0.0523 \| 0.0626 \| 0.0819 \| 1.0734 \| 1.3154 \| 1.7252 \| 0.7608 \| 0.6264 \|
	\| chronos2_ft_a ([Mandi-Chronos](https://huggingface.co/judicialmind/Mandi-Chronos)) \| 0.0532 \| 0.0641 \| 0.0847 \| 1.0734 \| 1.3243 \| 1.7485 \| 0.7221 \| 0.6150 \|
	\| chronos2 \| 0.0541 \| 0.0652 \| 0.0858 \| 1.1387 \| 1.3944 \| 1.8477 \| 0.7466 \| 0.6008 \|
	\| tirex \| 0.0559 \| 0.0665 \| 0.0863 \| 1.1729 \| 1.4408 \| 1.8897 \| 0.7618 \| 0.6137 \|
	\| timesfm25 \| 0.0541 \| 0.0653 \| 0.0869 \| 1.0336 \| 1.2881 \| 1.7270 \| 0.7269 \| 0.5955 \|
	\| lightgbm \| 0.0551 \| 0.0666 \| 0.0894 \| 1.0864 \| 1.3739 \| 1.9001 \| 0.7963 \| 0.6002 \|
	\| ets_lite \| 0.0569 \| 0.0703 \| 0.0947 \| 1.0352 \| 1.2910 \| 1.7046 \| 0.8391 \| 0.6014 \|
	\| seasonal_naive \| 0.0682 \| 0.0783 \| 0.0990 \| 1.1941 \| 1.4292 \| 1.8256 \| 0.7936 \| 0.1106 \|

	Models:

	\| model \| description \|
	\|---\|---\|
	\| `seasonal_naive` \| price on the same weekday of the last week; spread from past 7-day changes (its median often equals the last price, which scores as a wrong direction in `dir_acc_7d`) \|
	\| `ets_lite` \| damped additive-trend exponential smoothing per series; spread from one-step residuals \|
	\| `lightgbm` \| global LightGBM quantile regression on scaled lags, rolling statistics and calendar features \|
	\| `chronos2` \| `amazon/chronos-2`, zero-shot \|
	\| `timesfm25` \| `google/timesfm-2.5-200m-pytorch`, zero-shot \|
	\| `tirex` \| `NX-AI/TiRex`, zero-shot \|
	\| `chronos2_ft_a` \| Chronos-2 fine-tuned on MandiBench series, univariate: released as [Mandi-Chronos](https://huggingface.co/judicialmind/Mandi-Chronos) \|
	\| `chronos2_ft_b` \| Chronos-2 fine-tuned and run with group context: the series of one state and commodity are forecast jointly \|
	\| `chronos2_ft_c` \| Chronos-2 fine-tuned with daily soil-moisture and reservoir-storage covariates (NRSC and CWC data from data.gov.in, not included here) \|

	## Files

	\| file \| content \|
	\|---\|---\|
	\| `series/all.parquet` \| the benchmark series, one row per series and trading day (27,437,227 rows) \|
	\| `series/catalog.parquet` \| all 178,352 candidate market x commodity x variety series with coverage counts; `kept` marks the 12,863 benchmark series \|
	\| `splits.json` \| origins, horizons, held-out states, season length (7) \|
	\| `tasks.json` \| the 33 origin x horizon tasks in fev-style task format (id/timestamp/target columns, horizon, quantile levels) \|
	\| `forecasts/<model>/<origin>.parquet` \| the forecasts of the nine models, one file per origin \|
	\| `leaderboard.parquet`, `leaderboard.json` \| the leaderboard (324 rows: model x horizon x slice); the json adds `win_rate` \|
	\| `evaluate.py`, `requirements.txt` \| evaluation harness \|

	## Schema

	`series/all.parquet`

	\| column \| type \| description \|
	\|---\|---\|---\|
	\| `series_id` \| string \| `mandi\|state\|district\|market\|commodity\|variety` \|
	\| `ts` \| date \| trading day \|
	\| `target` \| float64 \| modal price, Rs per quintal (median over the grades reported that day) \|
	\| `state`, `district`, `market` \| string \| Agmarknet market location \|
	\| `commodity`, `variety` \| string \| Agmarknet commodity and variety names \|
	\| `cclass` \| string \| commodity class: cereal, pulse, oilseed, vegetable, fruit, spice, sugar, dairy_meat_fish, top (tomato, onion, potato) \|
	\| `grade` \| string \| always `NA` (grades are merged) \|
	\| `district_lgd` \| int32 \| empty in this release \|
	\| `cov_min_price`, `cov_max_price` \| float64 \| minimum and maximum price that day (median over grades), Rs per quintal \|
	\| `n_grades` \| int64 \| number of grade rows merged into the day \|

	`series/catalog.parquet`: `state, district, market, commodity, variety, cclass`, `first_ts`/`last_ts` (first and last
	trading day), `n_obs_pre` / `first_pre` (days and first day before 2025-06-01), `n_obs_post` (days from 2025-06-01),
	`n_mkt_days` (trading days of the market from `first_pre` to 2025-06-01), `n_obs_recent` / `mkt_days_recent` (days
	of the series and of its market in the two years before 2025-06-01), `obs_ratio` (= `n_obs_recent /
	mkt_days_recent`), `series_id`, `kept`.

	`forecasts/<model>/<origin>.parquet`: `series_id, origin, h, ts, q0.1 ... q0.9, mean, target` (one row per window and
	forecast day; `target` = observed price, null when the market did not trade).

	## Coverage

	12,863 series from 2,126 markets in 457 districts of 24 states and union territories, 176 food commodities in
	nine classes (vegetables 4,622 series, tomato/onion/potato 2,030, cereals 1,742, fruit 1,510, spices 1,285,
	oilseeds 798, pulses 725, sugar 129, dairy/meat/fish 22). Daily prices from 2001-01-10 to 2026-05-15; 25,729,325
	rows before the first origin and 1,707,902 from it. A series is included when it has at least two years of history
	before 2025-06-01 and traded on at least half of its market's trading days in the two years before that date.

	The prices come from 83.2 million Agmarknet rows. 14.0 million exact paging duplicates were removed, as were
	non-food commodities (4.6 million rows) and 90,130 flagged rows (unit slips, outliers and non-positive prices).

	## Source data

	Derived from 1 resource published on the Open Government Data Platform India (data.gov.in), snapshot May 2026,
	under the [Government Open Data License - India](https://data.gov.in/government-open-data-license-india).

	Providers: Directorate of Marketing and Inspection (Agmarknet), Ministry of Agriculture and Farmers Welfare.

	- Directorate of Marketing and Inspection (Agmarknet), Ministry of Agriculture and Farmers Welfare. "Variety-wise Daily
	Market Prices Data of Commodity". data.gov.in resource
	35985678-0d79-46b4-9ed6-6f13308a1d24, https://api.data.gov.in/resource/35985678-0d79-46b4-9ed6-6f13308a1d24.
	Snapshot: May 2026.

	GODL-India does not cover personal information; this dataset holds market prices only.

	<!-- jm:source -->
	Original data: published by Directorate of Marketing and Inspection (Agmarknet), Ministry of Agriculture and Farmers Welfare 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{judicialmind2026mandibench,
	title = {MandiBench: a probabilistic forecasting benchmark for daily mandi prices in India},
	author = {JudicialMind},
	year = {2026},
	publisher = {Hugging Face},
	howpublished = {\url{https://huggingface.co/datasets/judicialmind/MandiBench}}
	}
	```

	<!-- 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 -->