In short
- 01A useful retail forecast is often simple: daily sales rate × days you are ordering for, per item, per store.
- 02Measure the rate over the days the item was actually available to sell. Dividing a new item's sales by 90 days, or by selling days only, gives the wrong answer both ways.
- 03Order the forecast minus stock on hand minus stock already on its way, rounded up to the pack you buy in. Forgetting what's on the way is how orders double.
- 04Machine-learning forecasts earn their keep with big catalogues, promotions and seasonality. Most small chains get more from clean sales data and a forecast they can check.
- 05Momentum forecasts with this arithmetic, shows its working, and a buyer approves every order. It does not use machine learning and does not reorder on its own.
On this page
To work out how much to order, estimate each item's daily sales rate at each store, multiply it by the number of days the order needs to cover, and subtract the stock you already have and the stock already on its way. That simple, checkable forecast fixes most stockouts and overstocks in small and mid-sized chains, because most ordering errors come from missing data (returns not netted, stock in transit forgotten, new items measured over the wrong period), not from a lack of artificial intelligence.
What is demand forecasting in retail?
A demand forecast estimates how many units of an item a store will sell over a coming period. Buyers use it to decide how much to order from suppliers, and chains with a warehouse use it to decide how much to send each branch.
Getting it wrong is expensive in both directions. The analyst firm IHL Group estimated that inventory distortion (out-of-stocks plus overstocks) would cost retailers worldwide $1.77 trillion in 2023, of which about $1.2 trillion was out-of-stocks and $562 billion overstocks (Food Institute summary of IHL).
Do you need AI or machine learning to forecast demand?
Not to start. Machine-learning forecasts can model promotions, prices, weather, holidays and thousands of items at once, and large retailers with the data to train them do get value from that. But a model is only as good as the sales history it learns from, and it is hard for a buyer to check.
For most chains, the bigger wins come first from getting the basics right: counting every sale and return, measuring each item over the right period, and remembering what is already on order. A transparent formula that a buyer can check and override beats a black box nobody trusts. Once that is in place, it is much easier to see where a smarter model would add something.
How do you calculate how much to order?
Five steps, per item and per store, all in the same unit (single units, not cases):
- 1Units sold over a recent window, from every channel (till, invoices, sales receipts), minus returns.
- 2Days observed: the days in that window since the item was first seen at the store (first sale or first stock), at least one. An item that arrived 13 days ago has 13 days of history, not 90.
- 3Daily rate = units sold ÷ days observed.
- 4Forecast = daily rate × the number of days the order must cover.
- 5Order = forecast − stock on hand − stock already on its way, never below zero, rounded up to the pack or case you buy in.
Why measure the rate over days observed?
New and newly stocked items break naive formulas. Take an item first stocked 13 days ago that has sold 26 units, on 8 of those 13 days:
- Daily rate
- 26 ÷ 90 = 0.29
- 30-day forecast
- 8.7
- Result
- Far too low: it runs out
- Daily rate
- 26 ÷ 8 = 3.25
- 30-day forecast
- 97.5
- Result
- Too high: quiet days ignored
- Daily rate
- 26 ÷ 13 = 2.00
- 30-day forecast
- 60
- Result
- Matches how it has actually sold
| Method | Daily rate | 30-day forecast | Result |
|---|---|---|---|
| Divide by the whole 90-day window | 26 ÷ 90 = 0.29 | 8.7 | Far too low: it runs out |
| Divide by selling days only | 26 ÷ 8 = 3.25 | 97.5 | Too high: quiet days ignored |
| Divide by days observed | 26 ÷ 13 = 2.00 | 60 | Matches how it has actually sold |
Where do simple forecasts go wrong, and how do you correct them?
- Stockouts hide demand. Days with nothing on the shelf record zero sales, which drags the rate down. If an item was out of stock for part of the window, raise the order by judgement.
- Promotions inflate it. A week of buy-one-get-one lifts the rate. Don't let a promotion set next month's normal order.
- Seasonality. A rate from summer is wrong for winter lines. For seasonal items, look at the same period last year.
- Lead time longer than the period. If the supplier takes 20 days and you order for 14, you will run out. Order for at least the lead time plus the time until your next order.
- Minimum stock. Some items must never run out. Set a floor so the order never lets them drop below it.
How does Momentum forecast demand?
Momentum by Ltiora uses exactly this arithmetic, one rule in three places: the Demand Forecast report, the purchase order prefill (the items a supplier carries, for the stores you choose) and the transfer request prefill (what a branch should ask the warehouse for). Sales from every source are counted net of returns; the rate is measured over the days the item has been at the store (reading at least the last 90 days); an order never takes stock below the item's minimum; only sellable, unexpired stock counts as on hand; stock already on its way, from a supplier or a transfer, is subtracted; and quantities are rounded up to the unit the branch usually buys in, most urgent first.
A buyer reviews and edits every line before a purchase order or transfer is raised. Momentum does not use machine learning to forecast and does not reorder automatically. Momentum AI is a separate, optional add-on, off by default, not part of the standard plans and set up per business. It helps with tasks such as reading supplier invoices and checking prices, and it suggests; it does not place orders.
Questions people ask
What is the simplest formula for how much stock to order?
Order = (daily sales rate × days to cover) − stock on hand − stock already on order, rounded up to your pack size. Measure the daily rate over the days the item was actually in the store.
How far back should sales history go?
Long enough to smooth out odd days: around 90 days is a common choice for everyday items, or the same season last year for seasonal ones. For new items, use the days they have actually been stocked.
Is AI demand forecasting worth it for a small chain?
Sometimes, if you have a large catalogue, frequent promotions and years of clean data. Most small chains get more value first from accurate sales, returns and on-order data feeding a simple forecast they can check.
Does Momentum reorder stock automatically?
No. Momentum pre-fills purchase orders and transfer requests with suggested quantities and shows the numbers behind them; a person reviews, edits and approves every order.
Good demand forecasting starts with arithmetic you can check: sold, net of returns, over the days the item was really there; times the days you are ordering for; minus what you have and what is already coming. Get that right in every store and most stockouts and overstocks disappear. Anything more sophisticated can then be judged against a forecast you trust.
Sources
See the working behind every suggested order
In a demo we'll run Momentum's Demand Forecast on a sample of your sales, pre-fill a purchase order from it, and walk through the arithmetic line by line.
