Demand forecasting for brands: why integrated data has to come first

Brand teams often want a demand forecast early: how much to order from the OEM, how much to send to FBA, how much cash to tie up in inventory. The tempting move is to jump straight to a model—spreadsheet smoothing, a forecasting tool, a consultant’s spreadsheet. That skips the real prerequisite. A forecast is only as good as the history you feed it, and that history has to be one coherent story per SKU and channel. If sales, inventory, and cost live in disconnected places with different definitions, you are not forecasting demand—you are extrapolating confusion.

What “integrated data” means here (it is not only “one big database”)

For a brand that buys from OEMs and sells through marketplaces or distributors, integrated does not have to mean a single enterprise server on day one. It means you can answer, for each product and channel: how many units sold, when, net of what, at what true unit cost, with how much stock on hand or in transit—without reconciling three spreadsheets that use the same word for different things.

Practically, that is often a small integrated layer: one table or warehouse subject that joins marketplace exports, your order system, receiving records, and landed-cost assumptions—keyed by a stable SKU (or SKU × channel) ID. The integration is logical first (same keys, same definitions), even when the physical data still lives in a few systems.

Why forecasting fails when the base data is fragmented

Demand forecasting needs a time series of comparable observations: units or revenue per period, at a consistent grain (weekly or monthly), with returns and promos handled the same way over time. If Amazon net sales sit in one export, wholesale in another, and “our” revenue in QuickBooks with a different cut-off, you cannot stitch a clean history without decisions—those decisions are exactly what an integrated view makes explicit.

Inventory makes the problem worse. A forecast feeds replenishment. If available stock is split between your warehouse, 3PL, and FBA—and each system uses slightly different units or timing—you will either over-order or under-order while the model looks mathematically fine. The model did not fail; the inputs did.

Cost data has the same effect on how much risk you can take on inventory. Landed cost spread across freight spreadsheets and OEM invoices without a joined view turns “target weeks of cover” into guesswork. You forecast demand; you decide order quantity with margin—margin needs the same integrated backbone.

What to put in place before you trust the forecast

1. One identifier map. Brand SKU ↔ OEM part ↔ channel listing ↔ barcode, written down and owned. Without it, every export joins wrong eventually.

2. One definition of “sale.” Net of returns or not, which fees deducted, which orders count (cancelled, pending). The forecast inherits your definition; mixing definitions across months creates fake seasonality.

3. One timeline for inventory. Where units sit and when they became sellable, close enough for planning—not perfect audit precision, but not three incompatible “on hand” numbers.

4. Then build or buy the forecast—rule of thumb, statistical model, or tool—against that integrated slice. The sophistication of the model is secondary to the integrity of the series.

What to take away

For brand-led companies, integrated data is the precondition for useful demand forecasting, not a luxury you add after the model. You can start small—a single subject-area table, a disciplined weekly job to refresh it—but you cannot skip joining sales, inventory, and cost at a stable grain. Get that layer right, and even a simple forecast beats a sophisticated one built on fragmented truth. Skip it, and you optimize the wrong story.