Choose Model History by Coverage, Not Calendar Length

A longer data extract is not automatically a better model baseline. It can reveal seasonal pressure, but it can also blend equipment or process versions that no longer exist.

Choose the history around the decision and the operating conditions that could change its answer.

Make a coverage table

List the product families, pack formats, demand peaks, shift patterns and shared resources relevant to the proposed change. For each, record whether the available history contains representative runs.

Mark missing conditions separately from conditions that are rare. A calm month may adequately describe a stable process but fail to cover the mix needed for a line-consolidation decision.

If a seasonal peak is absent, obtain comparable history or model it as an explicitly uncertain scenario. Do not let a default stand in for observed performance.

Split process versions

Date material changes, equipment work, staffing changes and measurement changes. Keep obsolete and current operating states distinguishable.

An older difficult week may show a useful failure mode. It may also reflect a fault that has since been corrected. Preserve it and its explanation instead of either treating it as normal or deleting it to improve the fit.

Test the decision outside calibration

Hold back appropriate runs or periods from model fitting. Ask whether the model reproduces the outputs that matter in those conditions: good output, queues, operating time or service performance.

Set acceptable error in relation to the decision. A small uncertainty may be immaterial to one choice and reverse another. An aggregate match can hide a poor result on the line or product that governs the investment.

Document where the model has not been tested. If the proposal depends on a new operating state, calibration to current production is only part of the evidence.

The deliverable is a statement of coverage and limitations, not a badge for using twelve months. Decision confidence comes from confronting the conditions that could invalidate the recommendation.