Using Machine Learning for Yield Forecasting

Start with the planning decision, not the model

Imagine a hypothetical greenhouse growing cucumbers. The packing lead wants to review next week's staffing before the roster closes. The grower can describe current crop condition, but the amount reaching the agreed grade remains uncertain. The business needs an estimate early enough to inform a choice, with a clear account of what could change.

Choose a quantity that matches that decision. Total crop weight is not automatically the same as graded packing output. A weekly total also says little about which day may be busy. If the decision needs daily quantities, specify those separately rather than dividing a weekly prediction equally across seven days.

Ask what the user could do differently after seeing the forecast. If every staffing choice is already fixed before the issue time, a more accurate number may arrive too late to matter. The first improvement might be an earlier review, not a more complex model.

Kilograms alone do not state how many people or hours are needed: packing tasks, crew capability, equipment availability and the distribution of work across days would need their own local capacity assumptions. No such conversion is supplied here.

Freeze the definition before comparing estimates

In FP18-01, suppose the forecast is issued each Monday at 08:00 in the site's named time zone. Its target runs from the following Thursday at 00:00 up to, but not including, the next Thursday at 00:00. This invented schedule is a specification example, not a recommended crop horizon. Another operation may need a different lead time.

The event defining this target is grading/weighing, not the time the fruit was picked or dispatched. This local planning grade does not establish food-safety clearance, customer acceptance or available stock. Crop variety and the actual grade criteria remain local choices to document, not facts supplied by this scenario.

Keep later changes visible. If the grade rule changes, the next result may represent a different target even when the column still says yield. Do not quietly rewrite earlier forecasts to match a new definition. Preserve which rule each forecast and actual used, then decide whether they can be compared.

This article concerns estimates made before the outcome. The separate task of reconciling harvested, trimmed, held and packed quantities still matters, but a prediction cannot repair an undefined actual.

Rebuild what was known at the issue time

Kaufman and colleagues' leakage paper explains why prediction information needs legitimate timing. For FP18-01, a Sunday inspection entered after Monday's cutoff was unavailable to the original issue.

Unknown availability stays unknown. Use the weather forecast version available then, not later observed weather. Training outcomes must also have been available at the replayed issue. Fit preprocessing, feature choices and tuning on earlier development evidence, keeping later assessment outcomes separate.

Make the simple method a fair competitor

Forecasting: Principles and Practice describes evaluation using earlier training observations and later outcomes. Compare both methods at the required lead time.

Mean absolute error averages miss sizes, ignoring direction. Simple: (20 + 20 + 10 + 10) / 4 = 15 kg. Model: (5 + 15 + 20 + 30) / 4 = 17.5 kg. Here the model is worse. Four invented windows establish no crop performance. These arbitrary estimates define neither a fitted baseline nor a trained model.

For real evaluation, specify a simple candidate first, such as a locally reviewed recent-comparable-window estimate. Do not assume last week's quantity suits every crop stage.

Keep target, timing and assessed cases comparable; inspect overestimates and shortfalls separately. Repeated or overlapping windows are not independent crop trials. List every scheduled window, including missing predictions and pending actuals, before comparing.

Show a range with a meaning that can be checked

Gneiting and Raftery's scoring paper considers interval width and penalties for missed outcomes together. Simply making a range wider is not the whole assessment.

A broad range may include outcomes but distinguish no useful planning choice; a narrow one may miss too often. Check later comparable outcomes. A nominal coverage level is not achieved coverage or a guarantee for the next crop. A model's internal confidence number is not automatically a harvest probability.

Give the planner a usable brief, not a command

Return to the packing lead. A useful brief connects the estimate to a review time, the assumptions that matter and the choices still open. It should also say when it is unavailable. An old estimate carried forward silently may be more misleading than an explicit gap.

A candidate prediction model might relate previous crop records to later graded quantity. For this brief, the analyst and grower would first ask whether crop age, represented plant count and recent grading history have useful, consistently defined records. No crop-specific relationship has been demonstrated here, and a predictive pattern is not permission to change growing conditions.

AI can help summarize the inputs and explain why a forecast changed. It must not invent a missing inspection, convert an uncertain estimate into a sales promise or claim the roster has changed when only a suggestion exists. Ordinary software is better suited to versioning, arithmetic and checking required fields.

Where information is unsuitable, show that state and use the locally approved manual review or fallback method. Do not automatically label the last prediction as current.

The crop register owns group identity. The existing grading record supplies the later quantity. A read-only forecast register retains the issued brief, while the packing plan records the human choice. This separation lets the lead ask what was predicted, what was decided and what happened without treating those as the same event.

A suggestion to move staff is not an acknowledged roster change. The planner still checks the local constraints and records whether to leave the plan alone, seek another crop observation or make an authorized revision. A changed plan alone does not prove a benefit.

Keep checking whether the forecast earns its place

Start with a retained comparison alongside the existing planning method. Set review questions before looking at the results: Did the estimate arrive in time? Was its target unchanged? Did it help the planner choose among real options? Did errors or unavailable forecasts create extra work?

Revisit the comparison when crop membership, grade rules, growing methods or data collection change. A model that worked under one definition may not answer the new question.

At the review, ask the packing lead to walk through one real planning choice: what options were open, what the brief changed, and what still needed a human check. Keep this decision record even when the lead chose to leave the roster unchanged. It shows use, not proof of savings.

A pending grading record is unassessed, not zero yield. A confirmed zero is different again. Keep later corrections attached to the original actual and explain which version was used in a review. This is an outcome-status check, not a substitute for investigating crop losses.

The useful outcome is not that machine learning was installed. It is a clear, timely forecast the business can assess, challenge and use within its limits. It may remain an experiment rather than become a planning dependency.