Why Store-Level Forecasting Changes What You Order
Ten units of average demand across five stores can hide one location selling out and four locations holding dead stock. The average is mathematically correct, but it tells you almost nothing about where the next ten units belong.
This is the central problem with forecasting only at chain level. An aggregate forecast answers, “How many units will the business sell?” A useful replenishment decision also needs to answer, “Which location will sell them, when, and with what risk of running out?” Those are different questions, supported by different data.
How aggregation masks the real demand pattern
Consider five stores that sell the same lamp. Over the past four weeks, their combined sales were 50 units, or 10 units per store on average.
| Store | Units sold | Opening stock | Closing stock |
|---|---|---|---|
| Madrid | 18 | 18 | 0 |
| Valencia | 14 | 20 | 6 |
| Seville | 10 | 18 | 8 |
| Bilbao | 6 | 20 | 14 |
| Zaragoza | 2 | 18 | 16 |
A chain level forecast might recommend another 50 units. That recommendation is not useless, but it leaves the important allocation problem unresolved. Madrid has already sold through its opening stock. Valencia is moving quickly. Zaragoza has 16 units remaining after selling only two.
If the next delivery is divided equally, each store receives 10 units. Madrid may sell out again before the next delivery, while Zaragoza reaches 26 units of stock for a product with very low recent movement. The business has purchased the right total quantity and made the wrong operational decision.
Aggregation also hides changes in demand. Suppose a new residential development increases demand around one location, while a nearby competitor reduces traffic at another. The chain total may remain stable even though the store level distribution has changed substantially. A single average cannot distinguish those situations.
The minimum data needed for a location forecast
A per location forecast starts with a clean history for each SKU at each location. As a practical minimum, you need at least 8 to 12 weeks of sales history per location per SKU. More history is preferable when the product has seasonal behavior, irregular replenishment, or a long sales cycle.
The basic data set should include:
- Units sold by date, SKU, and location.
- Stock on hand, ideally captured at a consistent daily or intraday point.
- Receipts and replenishment dates, so a zero sales day is not mistaken for zero demand.
- Stockout periods, because observed sales during a stockout understate what customers wanted to buy.
- Returns, cancellations, and transfers, separated from ordinary customer sales.
- Price changes, promotions, and relevant product status changes.
- Opening and closing dates for locations, including temporary closures.
Sales history alone is not always demand history. If a store had zero units available for five days, recorded sales of zero do not mean demand was zero. They mean the store could not fulfill demand during that period. A forecast that treats every stockout as weak demand will systematically recommend too little stock for fast sellers.
The same principle applies to transfers. If one store repeatedly receives emergency stock from another, its sales record may look ordinary while the transfer record reveals unmet demand and an unstable allocation process.
What the model is estimating
For each SKU and location, the forecast estimates future unit demand over a defined horizon, such as the next seven or fourteen days. The forecast can then be compared with available stock, expected receipts, supplier lead time, and a target service level.
Several measures help evaluate whether the forecast is useful. Mean absolute error, or MAE, is the average absolute difference between predicted and actual units. It is easy to interpret because it remains in units. Bias measures whether forecasts tend to be too high or too low. A model with low average error but consistent underforecasting can still create frequent stockouts.
Evaluation should happen at the level where decisions are made. A chain level MAE can look acceptable while individual stores have large errors that cancel one another out. Measure accuracy by SKU and location, then review aggregate performance as a separate view.
A five location example
Imagine a home decor retailer with five locations in Madrid, Valencia, Seville, Bilbao, and Zaragoza. Each store carries the same 4,000 SKU catalog. The catalog is shared, but the customer mix is not.
In Madrid, small furniture and premium lighting sell quickly in relatively consistent quantities. The Valencia store has stronger demand for outdoor products during warm months. Seville sells more tableware and decorative items suited to larger household gatherings. Bilbao has steadier demand for practical home goods. Zaragoza has lower footfall and a higher proportion of planned purchases, so many products move slowly but in occasional bursts.
Take one set of ceramic serving bowls. Over 12 weeks, the five stores record the following weekly average sales:
| Location | Average weekly sales | Recent weekly sales | Interpretation |
|---|---|---|---|
| Madrid | 8.5 | 11 | Increasing velocity |
| Valencia | 6.0 | 7 | Stable demand |
| Seville | 5.5 | 4 | Moderating demand |
| Bilbao | 3.0 | 3 | Low but regular demand |
| Zaragoza | 1.5 | 0 | Sparse demand |
The chain average is 4.9 units per store per week. A replenishment rule based on that average would overstate Zaragoza’s near term requirement and understate Madrid’s. It could also miss the difference between a regular three unit weekly pattern in Bilbao and intermittent demand in Zaragoza.
A location level forecast can produce separate expectations, such as 10 units for Madrid, 7 for Valencia, 4 or 5 for Seville, 3 for Bilbao, and 1 or 2 for Zaragoza, before safety stock and delivery timing are considered. The value is not that every number will be exact. The value is that the allocation reflects observed behavior instead of distributing an average by default.
Location forecasts also improve exception handling. A buyer can review a forecast increase in Madrid alongside recent sales and stockouts, then decide whether it reflects genuine demand or a one time event. Without the location detail, that review starts with a blended number that has already removed the evidence.
The sparsity problem
Location level forecasting is not automatically better for every SKU. At the bottom of the sales curve, each location may have too little history to support a reliable individual estimate. A product that sells twice in 12 weeks at one store does not provide enough observations to identify trend, seasonality, or a meaningful distribution of demand.
This is a statistical problem, not a software setting. A forecast with excessive precision can create false confidence. If the observed data is sparse, the system may need to use a broader reference pattern, such as the product category, similar locations, or the chain level history. That approach sacrifices some local detail to reduce random variation.
One practical method is to segment SKUs by sales frequency and importance:
- High volume, regular sellers can receive a direct SKU and location forecast.
- Moderate volume products can use local history combined with category or chain patterns.
- Very sparse products may be managed with reorder rules, minimum presentation stock, or manual review.
- New products require an initial assumption based on comparable items until their own history develops.
The right level of detail can also differ by decision. A buyer may forecast demand by location, but purchase from a supplier using the chain total. The important point is to preserve the location signal until the allocation and replenishment decisions have been made.
When location level forecasting is overkill
There are cases where a separate forecast for every location adds complexity without improving the decision. Consider a retailer with two stores in the same shopping district, a shared stockroom, and same day transfers between them. If staff can move inventory quickly and customers can buy through either location, the effective selling unit may be the combined operation.
Location detail may also be excessive for a product with extremely sparse demand, especially when the item is non seasonal, low value, and easy to reorder. Maintaining a detailed forecast for a SKU that sells once every several months at each store can produce more maintenance work than useful signal.
The test is operational. Ask whether a location specific forecast would change the order quantity, the allocation, the transfer plan, or the timing of a decision. If it would not, aggregate planning may be sufficient. If it would prevent repeated stockouts in one store while reducing excess in another, the additional detail has a clear purpose.
Putting the forecast into the replenishment process
Location level forecasting works best when it is connected to the decisions that follow it. A forecast should be visible alongside current stock, inbound stock, lead time, recent sales, and known stockouts. It should also be refreshed often enough to respond to new evidence without reacting excessively to one unusual day.
Stockagile runs forecasts per SKU per location nightly. Its Growth tier supports up to 10 locations and 25,000 active SKUs. The Scale tier supports unlimited locations and 100,000 or more SKUs. For a retailer, the relevant question is not simply how many forecasts a system can calculate. It is whether those forecasts arrive at the level where the replenishment team makes its decisions.
A sensible review process begins with exceptions rather than asking a buyer to inspect every number. Highlight locations where projected demand exceeds available stock, where forecast bias is persistent, where a stockout may have distorted sales, and where local demand diverges materially from the chain pattern. This keeps statistical output connected to business context.
Location level forecasting does not remove judgment from inventory planning. It gives that judgment a more accurate starting point. The remaining decisions still include whether to transfer stock between locations, whether a local spike is temporary, and how to forecast a new store ramp up when it has no sales history of its own.
