Forecast the unit that replenishment controls.
Retail planning usually begins at store × SKU × day or week, then reconciles to category, region and network totals. A national category forecast cannot tell a store which size or colour to replenish. Conversely, an unnecessarily granular series may be too sparse to model independently.
Maintain product and store hierarchies so information can be shared and results can be reconciled. Effective assortment dates are essential: zero sales before a SKU was ranged are not evidence of zero demand.
Sales are censored when inventory is unavailable.
A stockout records fewer sales than customers wanted to buy. Training directly on those zeros can teach the model that demand falls precisely when supply fails. Capture on-hand, availability, stockout duration and lost-sales proxies where possible.
Returns also need treatment. Gross units, net units and returned units answer different questions. A future return may relate to a sale in an earlier period, so it should not blindly reduce current demand.
Promotions, price and assortment are known-future drivers.
Promotion type, discount depth, display, media and loyalty targeting can materially change demand. A model should not assume every promotion repeats the same uplift. Product cannibalisation and post-promotion dips may matter as much as the promoted SKU itself.
Store openings, closures, local events and trading calendars should be represented with the dates known at planning time. Avoid leakage from final campaign results into historical predictions.
Demand becomes an action through supply constraints.
Replenishment nets expected demand against available stock, inbound supply, lead time, case packs and service-level policy. Transfers compare shortage and excess across stores before a new purchase, but transfer cost, timing and presentation standards matter.
Evaluate availability and inventory outcomes together.
Forecast accuracy is necessary but insufficient. Measure service level, stockout duration, excess inventory, markdown, transfer acceptance and working-capital implications. Under-forecasting and over-forecasting have asymmetric costs by category.
Backtests should simulate the decision policy, not only compare point forecasts. A slightly less accurate forecast may produce better replenishment when uncertainty and case-pack constraints are handled well.
New products need attributes and analogues.
A new SKU has no direct sales history. Use category, brand, price, attributes, launch store set and comparable products, then update rapidly as actual sales arrive. Separate genuine launch demand from initial pipeline fill.
Openfluid Retail Forecasting is not yet live. These principles describe the design problem rather than a finished customer implementation.