Demand forecasting in retail and CPG, explained

What demand forecasting is

Demand forecasting is the practice of predicting how many units of a product will sell in a future period, store by store and week by week, so the people downstream can order, ship, and stock against a number instead of a guess. I spent eight years owning this number for a natural snack brand, pulling sell-through every Monday from Walmart Retail Link and KeHE Connect, and the forecast was the line everyone in the building either trusted or quietly worked around.

When the forecast was good, the S&OP meeting was short. When it was off by 20%, we either ran out of a hero SKU at Sprouts during a promotion or sat on six weeks of slow-moving inventory in the UNFI Atlanta DC. Both cost real money, and both traced back to the same forecast.

Demand forecasting vs demand planning

The two get used interchangeably in job titles and rarely mean the same thing in practice. Demand forecasting produces the number: how many units of this item will sell, in this store, in this week. Demand planning is what the business decides to do about that number once it exists, which includes the parts a forecast cannot settle on its own.

The distinction is not academic, because the two fail differently. A forecasting problem shows up as a miss against actuals, and you fix it by improving the model or the input data. A demand planning problem shows up as a forecast everyone agreed was right and nobody acted on, and no amount of model work touches it.

If a hero SKU is forecast correctly and still stocks out because the co-packer could not run it, that is a planning constraint, not forecast error, and scoring the forecast harder will never surface it. The replenishment planning side of the same loop is where that constraint becomes visible.

Baseline plus seasonality plus promo lift

A working retail demand forecast is three pieces added together, not one magic regression. Most demand forecasting methods differ only in how they estimate those pieces, so build each one separately: each breaks in a different way, and a blended model hides which one failed.

Take a single SKU: a 10 oz bag of roasted almonds at 140 Sprouts stores. Baseline runs 9 units per store per week. December indexes at 1.3x on seasonality. A planned 25%-off TPR has historically pulled a 2.1x lift on top of that. So the December promo-week forecast per store is 9 x 1.3 x 2.1 = roughly 25 units, and across 140 stores that is about 3,500 units for the week. Miss the seasonality multiplier and you under-order by a third going into your biggest month.

The reason brands separate lift from baseline is the same reason it shows up in inventory management work: you cannot plan replenishment off a blended average that silently buries a promotion inside it. The promotion ends, the average stays high, and you over-ship for a month.

Why POS-driven beats shipment-driven

Here is the distinction that took me two years to fully respect. You can forecast off what you shipped to the distributor, or off what actually sold through to a shopper. They are not the same series, and the gap between them is the difference between consumption and shipment data.

Shipment data is lumpy. A distributor places a forward buy ahead of a price increase and your shipment line spikes 40% in a week when true consumer demand did not move at all. Forecast off that spike and you chase a phantom. POS data (Retail Link, SPINS, Circana) shows the demand signal itself, smoothed of the distributor's ordering behavior. I learned to treat shipments as a financial number and POS as the demand number, and to never let the two get confused in the same model.

Forecast accuracy: MAPE and bias

A forecast you do not score is a wish. The two numbers I lived by were MAPE (mean absolute percentage error, how far off you are on average) and bias (whether you consistently over- or under-forecast). MAPE tells you the size of the miss; bias tells you the direction, which matters more for inventory than people expect.

Here is a four-week forecast-versus-actual on that almond SKU, in units per store per week:

MAPE is the average of the percentage errors: (11.1 + 14.3 + 13.6 + 8.3) / 4 = 11.8%. Bias is total forecast minus total actual: 58 - 57 = +1 unit, or about +1.8%, so this forecast runs very slightly hot but is close to neutral. A MAPE near 12% on a promoted natural SKU is respectable; the promo week (W3) is where most of the error lives, which is exactly where you would expect it. If bias had been +15% week after week, no amount of good MAPE would save you. You would be quietly building a stockpile.

Demand sensing, and the volatility it is meant to catch

Demand sensing is short-horizon forecasting: using last week's actual sell-through to correct the next one to four weeks, rather than waiting for the monthly cycle to re-plan. The statistical forecast above sets the shape of the quarter. Demand sensing adjusts the front of it with signal that arrived after the plan was locked.

The reason it exists is demand volatility, which is not the same thing as seasonality. Seasonality is the part that repeats and is already in the model. Volatility is the week-to-week variation left over once you remove it, and its size decides how much a short-horizon correction can buy you.

On the almond SKU above, the promo week (W3) carried most of the forecast error, and that is the general pattern: the volatile weeks are where the miss lives, so they are the only weeks a weekly correction pays for. A staple at ±10% does not need sensing, it needs to be left alone.

Lead time is what caps the whole idea. If your co-packer needs six weeks and the distributor holds four, a signal that arrives on Monday cannot change what ships this month. Demand sensing helps where you still have a decision left to make: allocation between retailers, a promo cut-in, an expedite. Below that horizon it is just a more current number to be wrong with.

Where Scout fits

Scout is a demand-side analytics layer. It connects your SPINS, Circana, or retailer POS exports so the baseline, seasonality, and promo-lift signal you forecast against come from real sell-through, and so you can score prior forecasts against actuals (MAPE and bias) by retailer and SKU without rebuilding the spreadsheet every Monday. It is not an S&OP system or a supply-planning engine. It does not place orders or generate the replenishment plan. It measures the demand signal and the accuracy of your calls so the planning system downstream has a cleaner number to work from.

The short version

• Demand forecasting predicts future unit sales by period and location so ordering and stocking run off a number, not a guess.
• A working forecast is baseline plus seasonality plus promo lift, each built and checked separately.
• POS-driven forecasts beat shipment-driven ones because shipments carry the distributor's ordering noise (forward buys), not pure demand.
• Score every forecast with MAPE (size of miss) and bias (direction). A small MAPE with a steady positive bias still builds dead inventory.