What is Demand Forecasting?

Quick answer:

Demand forecasting is the practice of estimating how much of each product customers will buy in a coming period, so that buying, staffing, and cash decisions run on an expectation instead of a reaction.

In retail it operates at the SKU level: not “will sales grow,” but how many units of this item, in this store, this month. That specificity is what makes it usable and what makes it hard.

Every purchase order is already a forecast; a store that “does not forecast” is simply forecasting with a gut feeling and last month’s invoice. The discipline is about replacing that with something the data can defend.

Here is what goes into a retail forecast, the methods from moving averages to machine learning, a worked example a spreadsheet can reproduce, and how forecasts connect to the reorder and staffing decisions they exist to serve.

What is Demand Forecasting? The Basics

A demand forecast is a statement of expected units per product per period, with a scope attached: per store, per channel, per week or month. The inputs are the store’s own history first, sales, seasonality, promotions, and stockouts, then outside signals like weather, local events, and market shifts.

One correction matters before any method: history records sales, not demand. A week the shelf sat empty is a week of demand the data never saw, so forecasting from raw sales quietly teaches the system to repeat every stockout. Good practice flags out-of-stock days and corrects for them.

Forecasts also age like produce. A number set in January degrades as the season actually unfolds, which is why the working unit of forecasting is the re-forecast cadence, not the annual plan.

The Methods, From Spreadsheet to Model

Moving average: next month equals the average of the last few. Free, stable, and blind to trend and season. Right for steady staples.
Exponential smoothing: the same idea weighted toward recent weeks, so the forecast turns when demand turns. Still spreadsheet-friendly.
Seasonal indexing: each month gets a multiplier from its history, December at 1.8x the average month, February at 0.7x, layered onto the trend. This is the workhorse for seasonal retail.
Causal models: regressions that tie demand to drivers, price, promotion, weather, so you can ask what a discount will do before running it.
Machine learning: models that read many signals at once across thousands of SKUs. This is where modern POS and planning tools earn their subscriptions, and where agentic systems increasingly act on the output directly.

Method selection is a volume question. Ten SKUs deserve a spreadsheet; ten thousand deserve software, because the arithmetic is easy and the repetition is not.

A Worked Example: One SKU, One Season

A garden store forecasts a hose nozzle for June. Sales the last three Junes: 62, 71, and 80 units, a trend of roughly +9 a year, so the trend line says about 89.

Two corrections apply. Last June the nozzle was out of stock for five selling days, roughly a sixth of the month, so true June demand ran closer to 96, lifting the trend estimate to about 105. And this June a 15% promotion is planned; the store’s history says similar promotions lift units by about a quarter, so the working forecast lands at about 130 units.

That number then does its real work: at 130 forecast units, a 10-day supplier lead time and current stock of 40, the reorder point math says order now, not at month-end. The forecast is not the deliverable. The purchase order is.

Advertisment ⓘ

Afterward, the store checks itself: actual sales against forecast, error recorded, assumptions adjusted. Forecast accuracy is a habit metric, and a consistently wrong forecast is still useful the moment you measure how it is wrong.

Why Demand Forecasting Matters for Retailers

The forecast sits upstream of the two costliest inventory mistakes. Forecast too low and the store buys stockouts: lost sales, disappointed regulars, and promotions that advertise an empty shelf. Forecast too high and it buys dead stock: cash sleeping on shelves until a markdown wakes it at a loss. Every forecasting improvement is margin recovered from those two piles.

It also sets the cash calendar. Inventory is usually a retailer’s largest working-capital line, and a forecast converts into a buying budget, which converts into what the bank balance looks like in October. Seasonal businesses live or die on this translation.

And it quietly runs the labor schedule: the same demand curve that sizes the purchase order sizes the Saturday roster, which loops back to conversion, since understaffed peak hours are where forecast misses become lost customers at the register.

The Judgment Layer the Model Cannot Replace

Every method above extrapolates history, and retail keeps scheduling events history has not seen: a competitor opening across the street, a product going viral on a Tuesday, a supplier doubling the minimum order. The forecast is a baseline for judgment, not a substitute for it.

The practical form of that judgment is the override with a note. When the buyer moves the number, the reason gets written down, and the review compares the model, the override, and reality. Half the value is finding out which of the three wins in which situations, and buyers who beat the model on new products often lose to it on replenishment.

New products deserve their own honesty: with no history, the forecast borrows from the closest comparable item’s first season, adjusted for what is different. That is a structured guess, and treating it as one, with small first orders and fast re-forecasts, is what keeps a launch from becoming a clearance event.

And promotions cut both ways. The lift during the promotion is usually real; the dip after it, as customers who stocked up stay home, is just as real and routinely missing from the plan.

Getting the Data Side Right

Clean sales history per SKU is the raw material, and it comes from the POS reports you already have. Consistent SKUs, receipts booked through purchase orders, and markdowns recorded make the history forecastable.
Flag the distortions: stockout days, one-off events, and pandemic-era months should be marked, not averaged in as if they were normal.
Let the software carry the volume: inventory platforms and better POS systems ship reorder suggestions built on exactly these methods; the software cost guide covers what that tier of tooling runs.
Feed it back into buying: a forecast that never changes a purchase order is a report, not a practice. The POS features guide covers which systems close that loop natively.

Start with the ten SKUs that matter most, forecast them monthly, and measure the error. The sophistication can grow later; the cadence is the practice.