Benchmark supported crop farm-gate price forecasting

[unes07]

Problem

Future revenue scenarios need crop-price assumptions, but national annual farm-gate series and local market prices have different meanings and must not be presented interchangeably.

Status: Blocked

Blocked by:

• Publish a farm-gate crop-price dataset for supported crops turba-data#4
• Add dataset cards, checksums, semantic versions, and automated validation turba-data#7
• Restore a reproducible training and benchmark pipeline #2
• Add model cards and artifact-integrity checks #5

Implementation must not start until the readiness gate below is satisfied.

Readiness gate

• Each supported crop has at least 10 annual observations or 36 monthly observations.
• Geographic and temporal definitions are consistent within each series.
• Missing periods do not exceed 20%.
• Wheat, barley, and maize are evaluated independently; unsupported crops remain excluded.

Proposed work

• Establish last-value and seasonal-naive baselines appropriate to each frequency.
• Use rolling-origin evaluation with no random shuffling.
• Forecast only series that pass the readiness gate.
• Keep national, regional, and local series as separate forecast products.

Acceptance criteria

• Report MAE, RMSE, sMAPE, and interval coverage against naive baselines.
• Publish point estimates, intervals, horizon, generation time, training cutoff, source dataset version, geographic level, and limitations.
• Clearly distinguish annual national baselines from local market forecasts.
• Do not present a national forecast as a guaranteed farmer sale price.
• Model cards document unsupported crops, regions, and forecast horizons.