Fundamentals
What is demand forecasting?
Demand forecasting estimates how much of a product or service people will want in a coming period, so you can buy, staff, and price before the rush — not after it has already walked in.
Updated Aug 20, 2026·9 min read
The plain answer
Demand forecasting is an estimate of how much people will want in a coming window — units of a SKU, covers in a restaurant, beds in a ward — broken into the slices you actually plan on. It is not a guarantee, and it is not last year’s spreadsheet with a nicer font.
Sales and demand are cousins, not twins. Sales is what you rang up. Demand is what would have walked out if the shelf had not been empty. Stockouts make history look calmer than the customers were. If you train only on sales, you teach the model to repeat your shortages.
Weekly demand — 4 seasons + a promo
What you actually forecast
The noun matters. “Demand” is vague until you pick a unit, a place, and a clock. A grocer cares about cases of milk by store by day. A hospital cares about arrivals by department by hour. Same idea, different pain.
- Unit: pieces, cases, kWh, visits, seats.
- Place: SKU, store, region, channel — the grain a PO is written at.
- Clock: hour, day, week. Finer looks smarter and is usually noisier.
Most teams need more than one of those. That is fine. Forecast the grain the decision lives on, then roll up. Splitting a monthly number into days after the fact is how you smuggle a made-up shape into operations.
How a demand forecast is built
Strip the marketing off and you are left with a loop. You take a history of the target. You add anything that regularly moves it — price, promo, weather, holidays. You ask a model to fill the next ticks. You hide some past ticks and score the miss. You ship the one that lost least, then you do it again when the world moves.
Classical methods (smoothing, ARIMA) are still useful on a well-behaved series. They drown when you have thousands of SKUs and a messy promo calendar. Machine learning eats more drivers. Foundation models can draw a first path with little local history. None of that replaces the loop. A pretty architecture with no backtest is a mood board.
The data that moves the number
Start with a clean target at the right grain. Then a calendar. Then one driver you actually understand — usually promotions or price. Weather helps categories that melt or freeze. Marketing spend helps when you can timestamp it. “All of our tables” does not help. It overfits last quarter’s accident.
Hierarchy is underrated. A new flavor has no spine of its own. The parent brand does. Borrow the season, scale by distribution, and get off the analog as soon as this SKU can beat it. That is cold start, not a different method family.
Mistakes that quietly wreck plans
The common ones are boring, which is why they survive. Forecasting sales and calling it demand. Scoring MAPE on a catalog full of zeros. Using one heroic 12-month path for both overnight replenishment and ocean freight. Fitting once in January and wondering in August why the number smells.
Another: treating the center line as a fact. Warehouses live on bands. p90 is “don’t stock out.” p10 is “don’t drown in yogurt.” A single line in a slide is how finance and ops end up with two unofficial numbers.
Using the number, not just reporting it
A forecast that nobody buys against is a report. Wire it into the calendar that already exists: the buy, the roster, the production slot. Alert when actuals leave the band, not when they cross a threshold someone typed in 2023. When a promo is on the table, run the path with it on and with it off. That is the whole product.
If you want a deeper cut on the clock itself, read time series forecasting. If the fight in the room is “is 12% good?”, start with WAPE.
FAQ
- What is demand forecasting in simple terms?
- It is an estimate of how much customers will want in a future window — next Tuesday, next week, next season — broken down by the slices you actually plan on, such as SKU, store, or channel.
- Is demand forecasting the same as sales forecasting?
- Not quite. Sales is what you rang up. Demand is what people would have taken if the item had been on the shelf. Stockouts make sales look lower than true demand, which is why many teams forecast demand and treat sales as a censored version of it.
- How far ahead should a demand forecast look?
- As far as the decision that consumes it. Replenishment might need 1–14 days. Buying from overseas might need 8–16 weeks. A horizon longer than your slowest lead time is usually wasted precision.
- What data do you need to start?
- A clean history of the thing you sell or serve, at the grain you plan on, plus a calendar. Promotions, price, and weather help once the baseline is honest. You can start thinner than people think; you cannot start with a messy target.
Keep going
Guide
Time series forecasting, explained
Time series forecasting predicts the next values of something that was recorded in order — sales by day, load by hour, tickets by week — by learning from its own past, and from anything that regularly moves with it.
Guide
Forecasting with covariates
A covariate is an extra series the model is allowed to see — a promo flag, a temperature, a price. Used well, it explains swings the target's past cannot. Used as a dump of every column you have, it just overfits last quarter.
Guide
WAPE, explained
WAPE (weighted absolute percentage error) is total absolute error divided by total actuals. It behaves when some periods are zero, and it does not let tiny SKUs dominate the score the way MAPE does.
Use case
Retail demand forecasting
Anticipate demand by product, store, channel, and region before buying or allocation decisions are locked.
Use case
Food demand and waste forecasting
Predict demand for perishable products while accounting for shelf life, substitutions, events, and weather.