Demand Forecasting for Inventory Management: The Actual Mechanics

What demand forecasting actually does in inventory management

Demand forecasting is the estimate of how many units of a SKU will sell in a future period — next week, next month, next quarter. Inventory management is everything you do with that estimate: how much to order, when to order it, and how much safety stock to carry in case the estimate is wrong. The forecast is the input. The reorder point, the purchase order quantity, and the safety stock level are the outputs.

Two formulas do the actual work:

• Reorder point = forecasted demand during lead time + safety stock
• Safety stock = a buffer sized to your desired service level and how volatile the forecast error has actually been

Get the forecast right and both numbers take care of themselves. Get it wrong and you're either sitting on cash tied up in units that won't move, or you're out of stock the week a campaign finally gets traction.

A worked example: turning a forecast into a reorder point

Say a SKU sells an average of 150 units a week, with a weekly standard deviation of 40 units — some weeks 110, some weeks 190. Supplier lead time is 5 weeks. Here's how that turns into an actual order trigger:

• Lead time demand = 150 × 5 = 750 units
• Lead time standard deviation = 40 × √5 ≈ 89 units
• At a 95% service level (z = 1.65), safety stock = 1.65 × 89 ≈ 147 units
• Reorder point = 750 + 147 = 897 units

That's the mechanical answer. The part almost nobody checks: is 150 units a week still the right number? Across the $500M+ in managed revenue we've managed across 100+ brands, the forecast error that actually costs money almost never comes from the math above — it comes from feeding last month's average into this month's order when demand already moved. A seasonal brand like Epic Gardening selling seed-starting kits has a demand curve that looks nothing like a year-round SKU from a brand like Ridge; running both through the same trailing-average forecast produces a confident, wrong number for one of them.

The main forecasting models, and when each one breaks

There's no single "demand forecasting model for inventory management" — the right one depends on how much history you have and how stable the pattern is. These are the ones that actually get used, not the exhaustive academic list.

New SKUs with zero sales history don't get a statistical forecast at all — they get a judgmental one, based on comparable launches, adjusted as real sales come in. That's the honest starting point, and it should be flagged as an assumption in your system, not treated as measured data once it's typed in.

When the forecast is wrong: what to check before you reorder

Every forecast is wrong to some degree — the question is whether it's wrong enough to change the order. Before increasing an order size or panic-buying, check these in order:

• Is the lead time assumption still true? A forecast built on a 5-week lead time is a different forecast once the supplier slips to 7.
• Did something outside the model change demand? A campaign ramping up, a competitor going out of stock, a seasonal shift arriving early — none of these show up in a trailing average until it's too late.
• Is the error a bias or just noise? Under-forecasting the same SKU three periods running is a bias — fix the model. One bad week against an otherwise accurate run is noise — leave the settings alone.

If the demand spike traces back to an ad account rather than the warehouse — a campaign that's about to outrun the stock behind it — that's a coordination problem between inventory and advertising, not a forecasting fix. That one belongs with Dr. PPC, not here.

The mistake that causes most stockouts and most overstock — including ones we've made

The most common failure isn't a bad model. It's using the right model but feeding it the wrong number, with nobody double-checking it because the spreadsheet has "always worked."

The specific version we've seen most often, including inside our own supervised accounts: the forecasting logic correctly flags rising demand off the back of a scaling campaign, but the buyer places the order off a trailing 90-day average anyway, because that's the number the reorder sheet has always used. The forecast was right. The process around it wasn't. That gap — between what the model says and what the human actually orders — is where stockouts and overstock both come from, and it's a process fix, not a math fix.

The second most common mistake: treating one bad forecast as proof the whole model is broken, and abandoning statistical forecasting for gut-feel ordering after one bad quarter. A model with a consistent 10% bias is fixable in an afternoon. A team that's stopped trusting any forecast is a much longer rebuild.

Where this fits with Dr. Stock

Demand forecasting math is the same whether you run it in a spreadsheet, an ERP, or a dedicated forecasting tool — none of that changes the formulas above. What changes is whether anyone catches the gap between the forecast and what actually got ordered. Dr. Stock, from Full Circle, is built to sit on that gap for Amazon specifically: checking whether the reorder point behind a stockout was ever right, catching FBA fee errors and storage surcharges that show up after the fact, and flagging the removal-versus-liquidation call on stock that isn't moving. Purchasing decisions stay with a human regardless of the autonomy setting. It doesn't replace an ERP or a 3PL, and it isn't a tool you buy to build your own forecasting model — it's there for when the forecast, the fee, or the shipment is already wrong and someone needs to catch it before it costs more.