Retail demand planning breaks in places that look fine on a dashboard.
The category-level forecast is accurate. The national number lands within a point or two of actual. And yet stores are out of stock on the items customers came in for, while the back room is stacked with product that will end up on a markdown rack. The aggregate was right. The distribution was wrong. That gap — between a forecast that is correct in total and an inventory position that is wrong everywhere it counts — is the central problem of retail demand planning.
This guide walks through where retail forecasting actually fails and the framework that holds up against it.
Why retail is harder than it looks
Three structural features make retail demand planning unusually hard, and most planning processes underestimate all three.
Store-SKU sparsity. The unit that matters for replenishment is the store-SKU — this item, in this store, this week. At that grain, most series are sparse: a slow item in a single store might sell zero units most weeks and three in one. Classic statistical forecasting needs a stable demand signal, and at the store-SKU level there often is not one. Forecast each store-SKU independently and you drown in noise; forecast only at the aggregate and you lose the distribution you actually need to replenish against.
Promotional volume dominates. In many retail categories, promoted weeks carry a large share of total volume, and promotional lift is several times baseline. A model that treats promo weeks as ordinary demand will over-forecast the weeks after a promotion ends and under-forecast the promo itself — a directional error that repeats every cycle. (This is forecast bias, and it is the quiet driver of retail markdown waste.)
Seasonality stacks. Retail demand carries multiple overlapping seasonal patterns — weekly day-of-week rhythm, annual holiday peaks, and category-specific seasons that do not align with the calendar. A single seasonal factor cannot capture all of them, and a model tuned to one pattern washes out the others.
A framework that holds up
The teams that forecast retail well do not solve these problems with one heroic model. They structure the problem so each piece is forecast at the level where it has signal, then reconcile.
Forecast at the level where signal lives, then disaggregate. Build the statistical forecast at a level with a stable demand signal — category-region, or product-family-cluster — where the series is dense enough to model reliably. Then disaggregate down to store-SKU using each store’s recent share of the higher-level demand. You get the modeling stability of the aggregate and the granularity of the store-SKU. This is the single highest-leverage move in retail planning, and it is the one spreadsheets handle worst, because the disaggregation has to be recomputed every cycle across thousands of stores.
Separate baseline from promotional lift. Decompose history into baseline demand and incremental promo lift, and forecast them separately. Baseline is stable and statistical. Promo lift is event-driven and tied to the promo calendar, the discount depth, and the display. Forecasting them as one blended series guarantees the post-promo over-forecast. Forecasting them separately lets you reuse the lift model the next time the same promotion runs. The same dynamic shows up in CPG demand forecasting, where promotions can be the majority of volume.
Cluster stores before you forecast them. Not every store behaves the same way. Urban and suburban, high-income and value, tourist and commuter — these clusters have different baselines, different seasonal shapes, and different promo responsiveness. Cluster stores by demand behavior, forecast at the cluster level, and the store-SKU disaggregation gets far more accurate than a single national share applied everywhere.
Reconcile top-down and bottom-up. The financial plan is top-down — the buying and merchandising team commits to a category number. The replenishment system runs bottom-up off store-SKU forecasts. When those two disagree — and they always do — you need a reconciliation step that holds the constraint that matters (usually the financial commitment) while preserving the store-level distribution. Hierarchical reconciliation does this mathematically rather than by manual override.
What good looks like
A retail demand planning process is working when three things are true at once.
The forecast is accurate at the grain you replenish against, not just in aggregate. A 5% category MAPE means nothing if the store-SKU forecast that drives the truck is off by 40%. Measure accuracy where the inventory decision is made.
Bias is near zero across segments — fast movers and the long tail, promoted and non-promoted, each cluster. Offsetting biases that net to zero in aggregate are the classic retail trap: the national number looks healthy while the top SKUs run hot and the tail runs cold.
The process runs fast enough to keep up with the merchandising calendar. Retail demand shifts weekly and promotions land on short notice. A forecasting process that takes the planning team three days to refresh is already stale when it ships. Speed is not a nice-to-have in retail; it is the difference between a forecast that reflects current demand and one that reflects last month’s.
The implementation trap
Most retail organizations know all of this. What stops them is not understanding — it is that the full-suite planning platforms that promise to handle store-SKU forecasting, promo decomposition, store clustering, and hierarchical reconciliation also take months to implement and a dedicated team to run. The transformation program becomes the project, and the actual forecasting improvement is still a year out. (It is the core tradeoff in any suite-vs-API decision.)
There is a faster path. The Geneva Forecast API handles the modeling layer — algorithm selection per series, baseline/promo decomposition, hierarchical reconciliation — through a single API call against the history you already have. You keep your existing replenishment and merchandising systems and bolt better forecasting onto them in weeks, not months. Start with one category, prove the markdown and in-stock numbers, then expand.
Running a large store network and want to see where the store-SKU forecast is leaking margin? Contact RoadMap — we will run a category against your history and show you the gap between your aggregate accuracy and your replenishment-level accuracy.
Leave a Reply