What data does a forecasting pilot need?
Start with the records that describe the decision you want to improve. If you replenish by item and warehouse each week, a monthly company revenue total will not support that test. Define a stable item identifier, location identifier, date, quantity, and unit of measure before selecting a model.
There is no single history length that guarantees a successful pilot. Seasonal patterns need enough history to observe relevant cycles; new products may need a different approach. Document what is missing instead of silently filling gaps with invented demand.
The minimum data inventory
| Dataset | Useful fields | Validation question |
|---|---|---|
| Demand history | Item, location, date, quantity, unit, returns | Does a zero mean no demand or no record? |
| Item master | Item ID, category, status, unit conversion | Are replacements and discontinued items identified? |
| Location master | Warehouse ID, region, timezone | Are transfers distinct from customer demand? |
| Inventory | Snapshot time, on-hand, reserved, blocked stock | Which quantities are actually available? |
| Supply | Order line, expected receipt, quantity, status | Are cancelled or late receipts still counted? |
| Policy | Lead time, order multiple, minimum quantity | Who maintains these assumptions? |
| Business context | Promotions, closures, lost sales, launches | Was this information known at forecast time? |
Use a small, representative sample first. Retain the original export and document transformations. An unexplained conversion from cases to individual units can overwhelm any modelling improvement.
How to handle stockouts and missing weeks
Observed sales can understate demand when an item is unavailable. Flag known stockouts and discuss how the pilot will treat them. Do not automatically assume every low-sales week represents weak demand. Equally, do not replace all low values with an optimistic estimate.
Create a calendar of expected records. Investigate missing dates, duplicate item-location-date rows, negative quantities, and abrupt changes in item codes. Decide whether returns belong in net demand or a separate flow and use that definition consistently.
Define success before sharing the extract
Choose the forecast horizon, review frequency, included segments, and comparison baseline. Reserve later periods for evaluation. If the team plans four weeks ahead, testing only one period ahead answers a different question. The time-series cross-validation reference explains why training data must precede test observations.
An illustrative acceptance sheet might ask for: error versus the current baseline; bias for priority items; time required to review exceptions; and successful reconciliation of one approved plan. These are suggested measures, not promised GrepEye outcomes.
Who needs to attend the pilot review?
Include a planner who understands exceptions, a data owner who can explain the extract, a receiving-system owner, and the person who approves operational changes. Assign a named owner to every unresolved data issue. Keep a log of excluded records so that a strong result cannot hide a weak segment.
For a reusable starting point, download the planning pilot checklist. Explore GrepEye Supply Chain and the software evaluation guide to turn the data pack into a practical demo agenda.