Benchmark fertilizer-price forecasting with uncertainty

[unes07]

Problem

The fertilizer optimizer needs dated prices, but a forecast should only be offered for product/location series with sufficient history and a reproducible evaluation against transparent baselines.

Status: Blocked

Blocked by:

• Publish a versioned fertilizer product catalog for Morocco turba-data#1
• Define and validate fertilizer price observation records turba-data#2
• Build reproducible fertilizer-price ingestion and normalization turba-data#3
• 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 product/location series has at least 24 monthly or 52 weekly observations spanning two seasonal cycles.
• Missing periods are no more than 20% after canonical aggregation.
• Observed, quoted, imputed, stale, and expired prices remain distinguishable.
• Product IDs, currencies, units, locations, and dataset versions pass published validation.

Proposed work

• Establish last-value and seasonal-naive baselines.
• Evaluate candidate methods with rolling-origin splits and no random shuffling.
• Generate point estimates and calibrated uncertainty intervals only for qualifying series.
• Register reproducible forecast artifacts with their data cutoff and model version.

Acceptance criteria

• Report MAE, RMSE, sMAPE, and interval coverage against both baselines.
• Return series ID, point estimate, lower and upper interval, forecast horizon, generation time, training cutoff, canonical MAD unit, source dataset version, and model version.
• Publish coverage and error by product and location.
• Exclude series that fail readiness gates rather than filling them with manufactured forecasts.
• Model cards state that forecasts are uncertain scenarios, not supplier quotes.