Three months after an early-access retailer loaded its full SKU catalog into Stockagile, the team questioned why reorder suggestions for its Barcelona store were consistently cautious. Forecasts seemed reasonable for Madrid and Valencia, yet Barcelona quantities were about 40% below expectation. The cause was a 14-month section of uploaded history covering a renovation closure. The model read 14 months of zero sales there and inferred that demand was structurally lower than at the other stores.
The tool behaved as designed. The data going in was faulty. This is the usual migration problem: buyers prioritize connecting and launching the system, while underestimating the effort needed to make its historical inputs accurate and clean.
Use the checklist below to prepare before trusting a new inventory system's output.
Step 1: Check Historical Sales Before Importing
Pull 24 months of sales history for each location and SKU from your POS or ecommerce platform. Before importing, look for these issues:
Coverage gaps. Do any locations show zero or near-zero sales across date ranges? Decide whether each is a real demand event, such as a slow period or seasonal closure, or a collection problem, such as POS downtime, a temporary store closure, or a broken integration. Keep genuine lows, but mark or exclude data gaps before they distort the baseline.
Price anomalies. Did heavily promotional pricing occur? A two-week clearance at 50% off may create a spike that looks like increased demand. Tag those dates as promotional outliers rather than baseline demand.
Inventory-constrained periods. If a SKU was out of stock for two weeks in July and sold zero units, those zeros describe supply, not demand. Treating them as demand will lower future estimates. Mark stock-constrained periods in the forecasting system instead of leaving the zeros untouched.
Step 2: Fix Your SKU Hierarchy
A messy product catalog can also damage forecasts. Before migrating, settle the key question: what is your business's forecast unit?
For apparel, that unit is often the variant: each size and color combination has its own demand pattern. A medium navy jacket will not sell at the same rate as a large navy jacket. Forecasting at parent-SKU level and allocating with a fixed ratio can leave popular sizes short and slow ones overstocked. Forecast variants when your catalog supports them.
The reverse problem is a fragmented SKU hierarchy. One physical product may have several POS IDs because of a barcode scanning issue or legacy IDs left active after a system migration. That splits demand and makes each SKU appear slower. Consolidate the IDs before migrating.
Step 3: Connect Integrations in Sequence
Connecting everything at once is tempting. Instead, work from cleaner data to poorer data, checking each integration before depending on it.
Begin with the channel holding the cleanest, most complete history, usually ecommerce. Check imported orders against your records for a sample of months. Then add POS systems store by store. Before trusting the sync, compare reported stock with a physical count at one store.
This order makes diagnosis possible. If several channels connect together and forecasts are wrong, the faulty source is unclear. Add one channel, verify it, then continue so data issues have a traceable source.
Step 4: Run the Old and New Processes Together
For at least four weeks after setup, operate the new system alongside the current process. Do not alter ordering yet. Collect weekly suggestions and compare them with the orders your previous method would have produced.
The parallel run shows three patterns. First, agreement with existing judgment, where the tool is automating sound decisions. Second, lower suggested quantities, which may expose over-buying normalized by the spreadsheet process. Third, higher suggestions, which need review because they may reflect missed demand or faulty input data.
Four weeks is the minimum. Eight weeks captures more seasonal variation and builds confidence before orders are approved directly from the system's suggestions.
Step 5: Teach Teams to Read Forecast Outputs
A data-driven forecast also carries confidence context that a spreadsheet total lacks. "Order 24 units" based on steady, well-supported demand is not the same recommendation as "order 24 units" based on three weeks of sales for a new SKU.
Buyers should know when to accept an output and when to override it. Review new SKUs with less than 8 weeks of history, items whose weekly demand varies sharply, such as a seasonal product moving between 3 and 40 units, and categories affected by unseen factors, such as a competitor closure or a promotion starting next week.
Buyers do not need to rebuild every suggestion from scratch. They need to identify which ones warrant review, which can be approved quickly, and apply that judgment consistently.
Step 6: Set a Model Accuracy Review Cycle
At month 3 and month 6 after go-live, compare system suggestions with actual sales. This checks more than tool quality. It also tests input quality and how well the model has learned your demand patterns.
For systematic over-suggestions, first inspect structural category issues, including seasonal cycles the model has not yet seen fully and promotions distorting the baseline. For systematic under-suggestions, check whether stock-constrained history was left unmarked.
This review also exposes SKUs whose demand changed after product changes, competitive shifts, or location changes. History from before a reformulation or store relocation should be down-weighted or excluded because it no longer predicts those items well.
What Migration Cannot Speed Up
The checklist addresses migration data and process, but it cannot shorten the model's learning curve. With clean, complete history covering 18 to 24 months, locations and SKUs may forecast reasonably from day one. Newer locations, recently added SKUs, and categories with little history need time to observe demand before suggestions become reliably useful.
Plan for that gap. Flag assortment areas and locations with thin or doubtful history, and keep buyer oversight higher during their first two seasons. The system learns from the clean data provided at the start, so the step one audit is essential.