Access note: verified organisation email

5daab46 verified 5 days ago

9.42 kB

---
license: cc-by-nc-sa-4.0
base_model: amazon/chronos-2
pipeline_tag: time-series-forecasting
library_name: chronos-forecasting
language:
- en
tags:
- time-series
- forecasting
- chronos
- agriculture
- india
- mandi
- prices
datasets:
- judicialmind/MandiBench
extra_gated_prompt: 'This model 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
---
# Mandi-Chronos

<!-- 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 -->

Mandi-Chronos is [Chronos-2](https://huggingface.co/amazon/chronos-2) fine-tuned to forecast daily wholesale (mandi)
prices in India. Given a market's daily modal price history for one commodity variety (Rs per quintal, with gaps on
days without trade), it returns probabilistic forecasts (quantiles) for the next days. It was fine-tuned on the
[MandiBench](https://huggingface.co/datasets/judicialmind/MandiBench) series before 2025-06-01, leaving out the nine
benchmark held-out states and union territories, and on MandiBench it beats zero-shot Chronos-2 at every horizon,
both overall and in the held-out states. Architecture and API are those of Chronos-2 (about 120M parameters).

## Quickstart

```bash
pip install -r requirements.txt   # chronos-forecasting, torch, transformers, pandas, ...
```

```python
import pandas as pd
import torch
from chronos import Chronos2Pipeline
from huggingface_hub import hf_hub_download
repo = "judicialmind/Mandi-Chronos"
pipe = Chronos2Pipeline.from_pretrained(repo, device_map="cpu")  # device_map="cuda" on a GPU
# daily modal prices (Rs per quintal) of three Agmarknet series over the 512 days before 2025-06-01
df = pd.read_csv(hf_hub_download(repo, "example.csv"), parse_dates=["ts"])
origin = pd.Timestamp("2025-06-01")
days = pd.date_range(origin - pd.Timedelta(days=512), periods=512, freq="D")
ids, contexts = [], []
for sid, g in df.groupby("series_id"):
    ids.append(sid)
    contexts.append(torch.tensor(g.set_index("ts")["target"].reindex(days).to_numpy(), dtype=torch.float32))
quantiles, mean = pipe.predict_quantiles(contexts, prediction_length=30, quantile_levels=[0.1, 0.5, 0.9])
for sid, q in zip(ids, quantiles):  # q: (1 variate, 30 days, 3 quantiles)
    print(" | ".join(sid.split("|")[1:5]))
    print("  day 1 :", [round(v) for v in q[0, 0].tolist()], " day 30:", [round(v) for v in q[0, -1].tolist()])
```

**Forecast quantiles, Rs per quintal (CPU, float32)**

| state | district | market | commodity | day 1 q0.1 | day 1 median | day 1 q0.9 | day 30 q0.1 | day 30 median | day 30 q0.9 |
|---|---|---|---|---:|---:|---:|---:|---:|---:|
| Odisha | Dhenkanal | Hindol | Tomato | 2,496 | 2,798 | 3,116 | 2,440 | 3,491 | 5,174 |
| Uttar Pradesh | Prayagraj | Ajuha | Onion | 1,163 | 1,208 | 1,262 | 1,496 | 1,805 | 2,121 |
| Uttar Pradesh | Raebarelli | Raibareilly | Wheat | 2,425 | 2,436 | 2,445 | 2,392 | 2,458 | 2,516 |

**Input:** one 1-D float tensor per series, the daily price on a calendar-day grid ending the day before the first
forecast day, with `NaN` for days without trade. The MandiBench protocol uses the last 512 days. **Output:**
`predict_quantiles` returns, per series, a tensor of shape (1, prediction_length, len(quantile_levels)) in the
input's unit, and the mean forecast. `pipe.predict_df(df, prediction_length=30, id_column=..., timestamp_column=...,
target=...)` takes a long pandas table instead.

### Command line

`mandi_chronos.py` builds the daily contexts from a long table (`series_id, ts, target`, csv or parquet) and writes
forecasts in the MandiBench submission format (`series_id, origin, h, ts, q0.1 ... q0.9`):

```bash
python mandi_chronos.py example.csv --origin 2025-06-01 --out forecasts.csv
# 3 series x 30 days -> forecasts.csv
```

In Python: `from mandi_chronos import load, forecast; forecast(load(), df, "2025-06-01")`. On a GPU it loads in
bfloat16 (the precision used for the benchmark forecasts); float32 on CPU agrees with them to about 0.1% (median
relative difference of the quantiles).

## Results

[MandiBench](https://huggingface.co/datasets/judicialmind/MandiBench) test windows: 11 monthly origins from
2025-06-01 to 2026-04-01, horizons of 7, 14 and 30 days. WQL (weighted quantile loss): lower is better; dir_acc_7d:
share of windows where the median forecast moves in the same direction as the price within the first week. Mandi-Chronos is `chronos2_ft_a` in the
MandiBench files.

**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 | dir_acc_7d |
|---|---|---|---|---|
| **Mandi-Chronos** | **0.0665** | **0.0778** | **0.0990** | 0.6271 |
| Chronos-2 (zero-shot) | 0.0677 | 0.0799 | 0.1025 | 0.6156 |
| TimesFM 2.5 (zero-shot) | 0.0672 | 0.0793 | 0.1012 | 0.6109 |
| TiRex (zero-shot) | 0.0709 | 0.0827 | 0.1038 | 0.6204 |
| LightGBM (global) | 0.0700 | 0.0831 | 0.1087 | 0.6025 |
| ETS (damped trend) | 0.0708 | 0.0862 | 0.1153 | 0.6109 |
| Seasonal naive (7 days) | 0.0845 | 0.0957 | 0.1193 | 0.1037 |

**Held-out states** (Chandigarh, Himachal Pradesh, Madhya Pradesh, Manipur, Meghalaya, Odisha, Rajasthan, Tamil Nadu,
West Bengal; not seen in fine-tuning; windows: 13,176 / 13,875 / 14,479)

| model | WQL h7 | WQL h14 | WQL h30 | dir_acc_7d |
|---|---|---|---|---|
| **Mandi-Chronos** | **0.0532** | **0.0641** | **0.0847** | 0.6150 |
| Chronos-2 (zero-shot) | 0.0541 | 0.0652 | 0.0858 | 0.6008 |
| TimesFM 2.5 (zero-shot) | 0.0541 | 0.0653 | 0.0869 | 0.5955 |
| TiRex (zero-shot) | 0.0559 | 0.0665 | 0.0863 | 0.6137 |
| LightGBM (global) | 0.0551 | 0.0666 | 0.0894 | 0.6002 |
| ETS (damped trend) | 0.0569 | 0.0703 | 0.0947 | 0.6014 |
| Seasonal naive (7 days) | 0.0682 | 0.0783 | 0.0990 | 0.1106 |

Bold: lowest WQL in the table. Mandi-Chronos also has the highest dir_acc_7d in both tables. It needs only the price
history. The full leaderboard (MASE and interval coverage included), all forecasts and the scoring script are in the
MandiBench dataset.

## Files

| file | content |
|---|---|
| `model.safetensors`, `config.json` | Chronos-2 weights and configuration (float32) |
| `mandi_chronos.py` | `load`, `daily_contexts`, `predict`, `forecast` and the command line above |
| `example.csv` | three MandiBench series (`series_id, ts, target`), 2024-01-06 to 2025-05-31 |
| `requirements.txt` | pinned runtime dependencies |

## Coverage

Fine-tuned and evaluated on 12,863 MandiBench series: daily modal prices of 176 food commodities in 2,126 Agmarknet
markets in 24 states and union territories. The fine-tuning data ends before 2025-06-01; the test windows run from
2025-06-01 to 2026-04-30.

## Source data

Training and example data: 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.

<!-- 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{judicialmind2026mandichronos,
  title        = {Mandi-Chronos: Chronos-2 fine-tuned for daily mandi price forecasting in India},
  author       = {JudicialMind},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/judicialmind/Mandi-Chronos}}
}
```

<!-- 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.

Base model(s): `amazon/chronos-2` (apache-2.0). Their licence terms and notices continue to apply to the upstream components; see `THIRD_PARTY_NOTICES.md`.
<!-- /jm:license -->

	---
	license: cc-by-nc-sa-4.0
	base_model: amazon/chronos-2
	pipeline_tag: time-series-forecasting
	library_name: chronos-forecasting
	language:
	- en
	tags:
	- time-series
	- forecasting
	- chronos
	- agriculture
	- india
	- mandi
	- prices
	datasets:
	- judicialmind/MandiBench
	extra_gated_prompt: 'This model 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
	---

	# Mandi-Chronos

	<!-- 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 -->

	Mandi-Chronos is [Chronos-2](https://huggingface.co/amazon/chronos-2) fine-tuned to forecast daily wholesale (mandi)
	prices in India. Given a market's daily modal price history for one commodity variety (Rs per quintal, with gaps on
	days without trade), it returns probabilistic forecasts (quantiles) for the next days. It was fine-tuned on the
	[MandiBench](https://huggingface.co/datasets/judicialmind/MandiBench) series before 2025-06-01, leaving out the nine
	benchmark held-out states and union territories, and on MandiBench it beats zero-shot Chronos-2 at every horizon,
	both overall and in the held-out states. Architecture and API are those of Chronos-2 (about 120M parameters).

	## Quickstart

	```bash
	pip install -r requirements.txt # chronos-forecasting, torch, transformers, pandas, ...
	```

	```python
	import pandas as pd
	import torch
	from chronos import Chronos2Pipeline
	from huggingface_hub import hf_hub_download

	repo = "judicialmind/Mandi-Chronos"
	pipe = Chronos2Pipeline.from_pretrained(repo, device_map="cpu") # device_map="cuda" on a GPU

	# daily modal prices (Rs per quintal) of three Agmarknet series over the 512 days before 2025-06-01
	df = pd.read_csv(hf_hub_download(repo, "example.csv"), parse_dates=["ts"])
	origin = pd.Timestamp("2025-06-01")
	days = pd.date_range(origin - pd.Timedelta(days=512), periods=512, freq="D")
	ids, contexts = [], []
	for sid, g in df.groupby("series_id"):
	ids.append(sid)
	contexts.append(torch.tensor(g.set_index("ts")["target"].reindex(days).to_numpy(), dtype=torch.float32))

	quantiles, mean = pipe.predict_quantiles(contexts, prediction_length=30, quantile_levels=[0.1, 0.5, 0.9])
	for sid, q in zip(ids, quantiles): # q: (1 variate, 30 days, 3 quantiles)
	print(" \| ".join(sid.split("\|")[1:5]))
	print(" day 1 :", [round(v) for v in q[0, 0].tolist()], " day 30:", [round(v) for v in q[0, -1].tolist()])
	```

	Forecast quantiles, Rs per quintal (CPU, float32)

	\| state \| district \| market \| commodity \| day 1 q0.1 \| day 1 median \| day 1 q0.9 \| day 30 q0.1 \| day 30 median \| day 30 q0.9 \|
	\|---\|---\|---\|---\|---:\|---:\|---:\|---:\|---:\|---:\|
	\| Odisha \| Dhenkanal \| Hindol \| Tomato \| 2,496 \| 2,798 \| 3,116 \| 2,440 \| 3,491 \| 5,174 \|
	\| Uttar Pradesh \| Prayagraj \| Ajuha \| Onion \| 1,163 \| 1,208 \| 1,262 \| 1,496 \| 1,805 \| 2,121 \|
	\| Uttar Pradesh \| Raebarelli \| Raibareilly \| Wheat \| 2,425 \| 2,436 \| 2,445 \| 2,392 \| 2,458 \| 2,516 \|

	Input: one 1-D float tensor per series, the daily price on a calendar-day grid ending the day before the first
	forecast day, with `NaN` for days without trade. The MandiBench protocol uses the last 512 days. Output:
	`predict_quantiles` returns, per series, a tensor of shape (1, prediction_length, len(quantile_levels)) in the
	input's unit, and the mean forecast. `pipe.predict_df(df, prediction_length=30, id_column=..., timestamp_column=...,
	target=...)` takes a long pandas table instead.

	### Command line

	`mandi_chronos.py` builds the daily contexts from a long table (`series_id, ts, target`, csv or parquet) and writes
	forecasts in the MandiBench submission format (`series_id, origin, h, ts, q0.1 ... q0.9`):

	```bash
	python mandi_chronos.py example.csv --origin 2025-06-01 --out forecasts.csv
	# 3 series x 30 days -> forecasts.csv
	```

	In Python: `from mandi_chronos import load, forecast; forecast(load(), df, "2025-06-01")`. On a GPU it loads in
	bfloat16 (the precision used for the benchmark forecasts); float32 on CPU agrees with them to about 0.1% (median
	relative difference of the quantiles).

	## Results

	[MandiBench](https://huggingface.co/datasets/judicialmind/MandiBench) test windows: 11 monthly origins from
	2025-06-01 to 2026-04-01, horizons of 7, 14 and 30 days. WQL (weighted quantile loss): lower is better; dir_acc_7d:
	share of windows where the median forecast moves in the same direction as the price within the first week. Mandi-Chronos is `chronos2_ft_a` in the
	MandiBench files.

	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 \| dir_acc_7d \|
	\|---\|---\|---\|---\|---\|
	\| Mandi-Chronos \| 0.0665 \| 0.0778 \| 0.0990 \| 0.6271 \|
	\| Chronos-2 (zero-shot) \| 0.0677 \| 0.0799 \| 0.1025 \| 0.6156 \|
	\| TimesFM 2.5 (zero-shot) \| 0.0672 \| 0.0793 \| 0.1012 \| 0.6109 \|
	\| TiRex (zero-shot) \| 0.0709 \| 0.0827 \| 0.1038 \| 0.6204 \|
	\| LightGBM (global) \| 0.0700 \| 0.0831 \| 0.1087 \| 0.6025 \|
	\| ETS (damped trend) \| 0.0708 \| 0.0862 \| 0.1153 \| 0.6109 \|
	\| Seasonal naive (7 days) \| 0.0845 \| 0.0957 \| 0.1193 \| 0.1037 \|

	Held-out states (Chandigarh, Himachal Pradesh, Madhya Pradesh, Manipur, Meghalaya, Odisha, Rajasthan, Tamil Nadu,
	West Bengal; not seen in fine-tuning; windows: 13,176 / 13,875 / 14,479)

	\| model \| WQL h7 \| WQL h14 \| WQL h30 \| dir_acc_7d \|
	\|---\|---\|---\|---\|---\|
	\| Mandi-Chronos \| 0.0532 \| 0.0641 \| 0.0847 \| 0.6150 \|
	\| Chronos-2 (zero-shot) \| 0.0541 \| 0.0652 \| 0.0858 \| 0.6008 \|
	\| TimesFM 2.5 (zero-shot) \| 0.0541 \| 0.0653 \| 0.0869 \| 0.5955 \|
	\| TiRex (zero-shot) \| 0.0559 \| 0.0665 \| 0.0863 \| 0.6137 \|
	\| LightGBM (global) \| 0.0551 \| 0.0666 \| 0.0894 \| 0.6002 \|
	\| ETS (damped trend) \| 0.0569 \| 0.0703 \| 0.0947 \| 0.6014 \|
	\| Seasonal naive (7 days) \| 0.0682 \| 0.0783 \| 0.0990 \| 0.1106 \|

	Bold: lowest WQL in the table. Mandi-Chronos also has the highest dir_acc_7d in both tables. It needs only the price
	history. The full leaderboard (MASE and interval coverage included), all forecasts and the scoring script are in the
	MandiBench dataset.

	## Files

	\| file \| content \|
	\|---\|---\|
	\| `model.safetensors`, `config.json` \| Chronos-2 weights and configuration (float32) \|
	\| `mandi_chronos.py` \| `load`, `daily_contexts`, `predict`, `forecast` and the command line above \|
	\| `example.csv` \| three MandiBench series (`series_id, ts, target`), 2024-01-06 to 2025-05-31 \|
	\| `requirements.txt` \| pinned runtime dependencies \|

	## Coverage

	Fine-tuned and evaluated on 12,863 MandiBench series: daily modal prices of 176 food commodities in 2,126 Agmarknet
	markets in 24 states and union territories. The fine-tuning data ends before 2025-06-01; the test windows run from
	2025-06-01 to 2026-04-30.

	## Source data

	Training and example data: 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.

	<!-- 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{judicialmind2026mandichronos,
	title = {Mandi-Chronos: Chronos-2 fine-tuned for daily mandi price forecasting in India},
	author = {JudicialMind},
	year = {2026},
	publisher = {Hugging Face},
	howpublished = {\url{https://huggingface.co/judicialmind/Mandi-Chronos}}
	}
	```

	<!-- 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.

	Base model(s): `amazon/chronos-2` (apache-2.0). Their licence terms and notices continue to apply to the upstream components; see `THIRD_PARTY_NOTICES.md`.
	<!-- /jm:license -->