Prepare the data¶
A statistical model cannot repair broken keys, incompatible units, or a partially loaded month. Establish the data contract before configuring the forecast.
1. Register materials¶
Every historical sale must reference an existing material. At minimum, define a stable identifier, description, active status, default and sales units, and lifecycle dates when relevant.
Discontinued materials may remain in history but should not automatically enter the future horizon. New materials with no history require an explicit succession, analogy, or collaboration rule.
2. Register locations¶
A location is the point where demand is observed. It may represent a customer, channel, store, region, distribution center, or another planning unit.
Keep its meaning stable throughout the time series. Changing the business meaning of an ID without rebuilding history breaks comparability.
3. Select the historical source¶
| Source | What it records | Best used when | Interpretation risk |
|---|---|---|---|
| Sell-out | Sale observed at the selling location when identifying the final destination is not relevant or possible | Consumer sale at a store without identifying the individual consumer | Partial channel coverage or delayed capture |
| Sell-in | Sale with an identified internal origin and customer or destination location | Direct channels, billing, or customer-level shipments | Inventory at the customer can separate sales from consumption |
Both sources may coexist in different scopes. Review the sales data overview before selecting a source for each channel.
Sell-in and sell-out are not interchangeable versions of one series.
In collaborative planning with a distributor, retain sell-in from the plant or distribution center to the distributor and, when shared, the distributor's sell-out and inventory. Use projected sell-out and inventory evolution to estimate future replenishment sell-in. Do not add the two series; define the point in the flow represented by each one.
4. Set the time bucket¶
Use a bucket size consistent with the decision:
- daily for high-frequency, short-response demand;
- weekly for recurring operations and day-of-week effects;
- monthly for tactical cycles, S&OP, and monthly MPS.
In a monthly process, load only closed months. If history is complete through June, the plan should start in July. A partial July should not be treated as closed history.
5. Reconcile before running¶
Check that:
- every material and location ID exists;
- each unit exists and can be converted;
- document IDs are unique;
- historical rows contain no future dates;
- the final bucket is complete;
- quantities reconcile by period and source;
- returns and negative quantities follow a known rule;
- active and lifecycle status matches the horizon.
Received does not mean usable
A technically accepted load may still be incomplete for planning. Record the effective window, accepted and rejected rows, and quantity reconciliation.
Next step¶
Define forecast levels and clusters, then review the data relationship map.