| --- |
| license: other |
| license_name: mixed-public-domain |
| language: |
| - en |
| tags: |
| - agriculture |
| - time-series |
| - commodities |
| - crop-production |
| - weather |
| - climate |
| - USDA |
| - FAOSTAT |
| - foundation-model |
| - pretraining |
| - fertilizer |
| - crop-insurance |
| pretty_name: 'TFSMAG: Time-series Foundation Model for Agriculture' |
| size_categories: |
| - 10M<n<100M |
| configs: |
| - config_name: calendar_spreads |
| data_files: |
| - split: train |
| path: calendar_spreads/train-* |
| - split: validation |
| path: calendar_spreads/validation-* |
| - split: test |
| path: calendar_spreads/test-* |
| - config_name: calendar_spreads_deep |
| data_files: |
| - split: train |
| path: calendar_spreads_deep/train-* |
| - config_name: cftc_cot |
| data_files: |
| - split: train |
| path: cftc_cot/train-* |
| - split: validation |
| path: cftc_cot/validation-* |
| - split: test |
| path: cftc_cot/test-* |
| - config_name: climate_indices |
| data_files: |
| - split: train |
| path: climate_indices/train-* |
| - split: validation |
| path: climate_indices/validation-* |
| - split: test |
| path: climate_indices/test-* |
| - config_name: commodity_futures |
| data_files: |
| - split: train |
| path: commodity_futures/train-* |
| - split: validation |
| path: commodity_futures/validation-* |
| - split: test |
| path: commodity_futures/test-* |
| - config_name: faostat |
| data_files: |
| - split: train |
| path: faostat/train-* |
| - split: validation |
| path: faostat/validation-* |
| - split: test |
| path: faostat/test-* |
| - config_name: india_mandi |
| data_files: |
| - split: train |
| path: india_mandi/train-* |
| - split: validation |
| path: india_mandi/validation-* |
| - split: test |
| path: india_mandi/test-* |
| - config_name: nass_prices |
| data_files: |
| - split: train |
| path: nass_prices/train-* |
| - split: validation |
| path: nass_prices/validation-* |
| - split: test |
| path: nass_prices/test-* |
| - config_name: usda_nass_production |
| data_files: |
| - split: train |
| path: usda_nass_production/train-* |
| - split: validation |
| path: usda_nass_production/validation-* |
| - split: test |
| path: usda_nass_production/test-* |
| - config_name: usda_nass_progress |
| data_files: |
| - split: train |
| path: usda_nass_progress/train-* |
| - split: validation |
| path: usda_nass_progress/validation-* |
| - split: test |
| path: usda_nass_progress/test-* |
| - config_name: usda_psd |
| data_files: |
| - split: train |
| path: usda_psd/train-* |
| - split: validation |
| path: usda_psd/validation-* |
| - split: test |
| path: usda_psd/test-* |
| - config_name: usda_rma_insurance |
| data_files: |
| - split: train |
| path: usda_rma_insurance/train-* |
| - split: validation |
| path: usda_rma_insurance/validation-* |
| - config_name: wb_commodities |
| data_files: |
| - split: train |
| path: wb_commodities/train-* |
| - split: validation |
| path: wb_commodities/validation-* |
| - config_name: weather_power |
| data_files: |
| - split: train |
| path: weather_power/train-* |
| - split: validation |
| path: weather_power/validation-* |
| - split: test |
| path: weather_power/test-* |
| --- |
| |
| # TFSMAG: Time-Series Foundation Model for Agriculture |
|
|
| The largest open agricultural time-series corpus. ~547,000 series totaling ~38 million observations across 14 configs spanning commodity markets, global supply/demand, crop phenology, agroclimatology, soil moisture, climate teleconnections, wholesale mandi prices, fertilizer prices, and crop insurance. |
|
|
| No existing TSFM pretraining corpus (LOTSA, Chronos KernelSynth, TimesFM) contains agricultural data. TFSMAG fills that gap. |
|
|
| ## Split protocol (v2) |
|
|
| All configs share **one global temporal cutoff** so that no training observation postdates any evaluation observation, in any series, in any config: |
|
|
| - **train**: observations through **2023-12-31** |
| - **validation**: calendar year **2024** |
| - **test**: observations from **2025-01-01** onward |
|
|
| A series appears in a split only if it has at least one observation in that window; `n_obs`, `start_date`, and `end_date` are recomputed per split. Configs whose history ends before a window simply have no rows there (e.g. `calendar_spreads_deep` ends in 2004 and is train-only; `wb_commodities` and `usda_rma_insurance` end in 2024 and have no test rows). At evaluation time, condition on everything before the window you are scoring (train for validation; train + validation for test). For short annual series, filter `n_obs` in eval to taste. |
|
|
| v1 of this dataset used a per-series chronological 80/20 split, which put every series' boundary at a different date (median train cutoff ranged from 1999 to 2026 across configs), leaked each series' holdout period into other series' training windows, and withheld the most recent regimes from training. The v1 files remain available under the `v1-per-series-split` tag: |
|
|
| ```python |
| load_dataset("Viharikvs/TFSMAG", "faostat", revision="v1-per-series-split") |
| ``` |
|
|
| ## Dataset summary |
|
|
| | Config | Train (series / obs) | Validation (series / obs) | Test (series / obs) | Freq | Date Range | |
| |---|---|---|---|---|---| |
| | `faostat` | 488,456 / 21,035,524 | 406,169 / 406,169 | 14,206 / 14,206 | A | 1961-2024 | |
| | `weather_power` | 528 / 8,244,042 | 528 / 193,248 | 528 / 192,720 | D | 1981-2025 | |
| | `calendar_spreads_deep` | 550 / 3,030,408 | -- | -- | B | 1959-2004 | |
| | `usda_psd` | 22,683 / 1,012,314 | 16,636 / 16,636 | 16,636 / 17,553 | A | 1960-2025 | |
| | `usda_nass_progress` | 2,756 / 798,078 | 2,440 / 37,621 | 2,437 / 34,987 | W | 1979-2026 | |
| | `usda_nass_production` | 8,079 / 533,531 | 3,322 / 3,322 | 3,255 / 3,661 | A | 1866-2025 | |
| | `cftc_cot` | 430 / 391,575 | 420 / 21,611 | 399 / 25,746 | W | 1995-2026 | |
| | `commodity_futures` | 51 / 323,853 | 51 / 12,892 | 51 / 16,566 | B | 1971-2026 | |
| | `nass_prices` | 924 / 273,102 | 334 / 3,435 | 328 / 3,967 | M | 1970-2026 | |
| | `usda_rma_insurance` | 19,816 / 254,933 | 8,497 / 8,497 | -- | A | 1989-2025 | |
| | `india_mandi` | 2,244 / 239,538 | 1,954 / 282,759 | 1,552 / 142,085 | D | 2023-2025 | |
| | `wb_commodities` | 51 / 34,517 | 48 / 576 | -- | M | 1960-2024 | |
| | `climate_indices` | 17 / 16,861 | 16 / 192 | 16 / 233 | M | 1854-2026 | |
| | `calendar_spreads` | 88 / 5,721 | 284 / 34,441 | 370 / 108,115 | B | 2021-2026 | |
| | **Total** | **546,673 / 36,193,997** | **440,699 / 1,021,399** | **39,778 / 559,839** | | | |
|
|
| **Grand total: 37,775,235 observations across train + validation + test.** |
|
|
| ## Schema |
|
|
| Every config uses a uniform 14-column schema. Each row is one time series stored as parallel arrays. `len(timestamps) == len(values) == n_obs` is guaranteed. All timestamps are in YYYY-MM-DD format, strictly increasing within a series. Zero NaN, zero Inf across all values. |
|
|
| `series_id` (unique key per config), `ticker`, `name`, `field`, `domain`, `category`, `exchange`, `unit`, `freq`, `n_obs`, `start_date`, `end_date`, `timestamps` (list[str]), `values` (list[float64]). |
|
|
| ## Usage |
|
|
| ```python |
| from datasets import load_dataset |
| |
| train = load_dataset("Viharikvs/TFSMAG", "commodity_futures", split="train") |
| val = load_dataset("Viharikvs/TFSMAG", "commodity_futures", split="validation") |
| test = load_dataset("Viharikvs/TFSMAG", "commodity_futures", split="test") |
| |
| series = train[0] |
| print(series["name"], len(series["values"]), "observations") |
| ``` |
|
|
| ## Configs |
|
|
|
|
| ### `commodity_futures` (51 series, daily) -- 9/10 |
| 28 instruments: 13 ag futures (corn, soybeans, wheat HRW/SRW, soybean meal/oil, cotton, coffee, sugar, cocoa, cattle, hogs, feeder cattle), 4 energy, 5 FX, 6 ag equities. Source: Yahoo Finance. As-of: 2026-04-17. |
| |
| ### `usda_psd` (22,683 series, annual) -- 10/10 |
| Complete USDA FAS PSD. All 5 commodity groups, 44 commodities, 209 countries, 12 S&D attributes. Source: apps.fas.usda.gov/psdonline. As-of: 2025-04-09. |
|
|
| ### `usda_nass_progress` (2,761 series, weekly) -- 9/10 |
| Crop progress + condition. 9 crops, 49 states, 28 phenological stages + condition ratings. Source: NASS Quick Stats bulk. As-of: 2026-04-18. |
|
|
| ### `usda_nass_production` (8,079 series, annual) -- 8/10 |
| Production, yield, area, prices. 16 major crops, 52 states. Goes back to 1866. Source: NASS Quick Stats bulk. As-of: 2026-04-18. |
|
|
| ### `faostat` (488,456 series, annual) -- 10/10 |
| COMPLETE UNFILTERED FAO data. 301 production items, 616 trade items, 235 price items. 283 countries. Production (74,873), Trade (391,262), Prices (52,751). Source: fenixservices.fao.org bulk downloads. As-of: 2024-12-23. |
|
|
| ### `climate_indices` (17 series, monthly) -- 9/10 |
| ONI, MEI, SOI, PDO, NAO, AMO, IOD, PNA, AO, and 4 Nino SST regions (absolute + anomaly). Source: NOAA CPC/PSL/NCEI. As-of: 2026-03. |
| |
| ### `india_mandi` (2,680 series, daily) -- 4/10 |
| AGMARKNET modal prices. 5 commodities (Potato, Onion, Wheat, Tomato, Rice), 1,598 mandis, 30 states. Source: Kaggle scrape. As-of: 2025-06-11. |
|
|
| ### `weather_power` (528 series, daily) -- 8/10 |
| NASA POWER agroclimatology. 12 parameters (temp mean/max/min, precip, solar, humidity, wind, root-zone soil moisture, profile soil moisture, dewpoint, surface pressure, evapotranspiration). 44 global ag regions spanning US, Brazil, Argentina, India, Ukraine, Russia, China, Australia, France, South Africa, Thailand, Indonesia, Canada, Pakistan, Bangladesh, Myanmar, Philippines, Kenya, Ethiopia. Source: power.larc.nasa.gov. As-of: 2025-12-31. |
| |
| ### `wb_commodities` (51 series, monthly) -- 9/10 |
| World Bank Pink Sheet. 5 fertilizers (DAP, TSP, Phosphate rock, Potassium chloride, Urea), grains, oilseeds, edible oils, softs, livestock, energy, fruits. 1960-2024. Source: worldbank.org CMO Historical Data. As-of: 2024-12. |
|
|
| ### `usda_rma_insurance` (19,816 series, annual) -- 8/10 |
| USDA RMA Summary of Business. Premium, indemnity, liability, loss ratio, insured acres, policies. 246 crops, all states, 1989-2024. Source: pubfs-rma.fpac.usda.gov. As-of: 2024. |
|
|
| ### `cftc_cot` (430 series, weekly) -- 9/10 |
| CFTC Commitments of Traders disaggregated + legacy. 21 agricultural markets (grains, oilseeds, softs, livestock, dairy, OJ, lumber). 21 position fields per market: commercial/non-commercial long/short/spread (legacy, 1995+), plus PMPU / Swap dealer / Managed money / Other reportable long/short/spread (disaggregated, 2010+). Source: CFTC bulk COT zips. As-of: 2026-04-14. |
| |
| ### `nass_prices` (925 series, monthly) -- 8/10 |
| USDA NASS Prices Received. Farm-gate monthly prices for 19 commodities (corn, soybeans, wheat, cotton, sorghum, oats, barley, rice, sunflower, peanuts, canola, hay, rye, flaxseed, milk, cattle, hogs, eggs, turkeys) at national and state level. 1970-2026. Source: NASS Quick Stats API. As-of: 2026-02-27. |
|
|
| ### `calendar_spreads` (370 series, daily) -- 7/10 |
| Calendar spreads reconstructed from Yahoo individual contract months. 10 commodities (corn, soybeans, wheat SRW, soy meal, soy oil, cotton, coffee, sugar, live cattle, lean hogs). All valid front-vs-deferred pairs within rolling 15-month window per commodity, not just adjacent. History limited to 2021-2026 by Yahoo contract retention. Source: Yahoo Finance via yfinance. As-of: 2026-04-17. |
| |
| ### `calendar_spreads_deep` (550 series, daily) -- 9/10 |
| Deep historical calendar spreads from TurtleTrader free individual contract archive. 16 commodities (corn, soybeans, wheat SRW, cotton, coffee, sugar, oats, OJ, feeder cattle, lean hogs, pork bellies, crude oil, gold, silver, copper, heating oil). All valid spread pairs, 1959-2004. 3M+ observations. Complements `calendar_spreads` (2021-2026) with 45 years of deep history. Source: turtletrader.com/hpd. As-of: 2002-2004 (static archive, no longer updated). |
|
|
| ## Data quality |
|
|
| - **Zero NaN, zero Inf** across all values |
| - **Zero duplicate series_ids** within any config/split (26 colliding spread-pair ids in v1 `calendar_spreads` were disambiguated with `_2` suffixes) |
| - All timestamps normalized to YYYY-MM-DD, strictly increasing within each series |
| - **Global temporal split**: train <= 2023-12-31 < validation <= 2024-12-31 < test (see Split protocol) |
| - ~40k duplicated (series, timestamp) pairs at the v1 faostat train/test seam were deduplicated (first occurrence kept) |
| - All-zero and constant series dropped |
| - SOI climate index: timestamps sorted and deduplicated |
| - Indian mandi: series with 60+ constant-price days dropped |
| |
| ## Licensing |
| |
| - **USDA (NASS, FAS PSD, RMA)**: US Government, public domain |
| - **NOAA**: US Government, public domain |
| - **NASA POWER**: US Government, public domain |
| - **FAOSTAT**: CC BY-NC-SA 3.0 IGO |
| - **World Bank**: CC BY 4.0 |
| - **AGMARKNET**: Indian Government open data (NDSAP) |
| - **Yahoo Finance**: Verify terms for commercial redistribution |
| - **TurtleTrader**: Free public archive, verify terms for redistribution |
| |
| ## Citation |
| |
| ```bibtex |
| @misc{tfsmag2026, |
| title={TFSMAG: Time-Series Foundation Model for Agriculture}, |
| author={Vihari Kanukollu}, |
| year={2026}, |
| url={https://huggingface.co/datasets/Viharikvs/TFSMAG}, |
| } |
| ``` |
| |
