Develop leakage-safe realized gross-margin prediction

[unes07]

Problem

A realized gross-margin model could support pre-season risk analysis, but it would be misleading without sufficient observed farm outcomes, privacy controls, temporal evaluation, and transparent deterministic baselines.

Status: Blocked

Blocked by:

• Add dataset cards, checksums, semantic versions, and automated validation turba-data#7
• Define privacy-aware farm economics and outcome records turba-data#8
• Restore a reproducible training and benchmark pipeline #2
• Reconcile model-registry metrics with published leaderboards #3
• Add spatial and grouped validation with regional error analysis #4
• Add model cards and artifact-integrity checks #5
• Add deterministic gross-margin scenario calculation turba-client#8

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

Readiness gate

• At least 500 complete farm-season observations.
• At least 100 distinct pseudonymized farms.
• At least three regions and two agricultural seasons.
• A documented missingness, consent, retention, and geographic-privacy audit.
• A stable deterministic gross-margin target definition shared with the calculator.

Proposed target and prediction point

Predict realized_gross_margin_mad_ha at a pre-season decision point. Features may include only information available when the recommendation is generated. Actual yield, realized sale price, post-season costs, and other future information must not be used as features.

Acceptance criteria

• Split evaluation by farm and time so no farm or future season leaks into training.
• Compare against mean, linear, and deterministic-scenario baselines.
• Report crop-level and regional errors plus calibrated prediction intervals.
• Publish feature-availability and leakage audits.
• Refuse publication if the model does not beat transparent baselines.
• Do not label the target net profit without complete fixed, financing, tax, and capital costs.
• Document unsupported regions, crops, farm types, and seasons in the model card.