What is demand forecasting?
Demand forecasting means estimating how many units of each product you will sell in the coming weeks. You use those estimates to decide what to reorder, and when.
The goal is modest and valuable: fewer sold-out best-sellers and less cash sitting in products nobody is buying.
What data do you need?
- Sales by product and by date, ideally 12 to 24 months.
- Stock levels, so you can tell low sales from no stock.
- Promotion and discount dates.
- Supplier lead times and minimum order quantities.
- Launch dates for new products.
One quiet problem is that sales drop to zero on days an item was out of stock. That is a supply problem, not a demand problem, and your forecast should not learn from it.
How do you start with a simple baseline?
Take the same period last year, adjust for how much your business has grown, and add the effect of any planned promotion. For steadier products, a moving average of recent weeks works well. This takes an afternoon in a spreadsheet, and it often performs better than people expect.
How do you turn a forecast into a reorder?
A forecast on its own does not place an order. You need two more things: how long the supplier takes, and how much safety stock you want for surprises. A simple rule of thumb ties them together.
| Piece | Meaning |
|---|---|
| Daily demand | Forecast units per day |
| Lead time | Days from order to delivery |
| Safety stock | Extra units for surprises |
| Reorder point | Daily demand × lead time + safety stock |
When stock on hand falls to the reorder point, it is time to order.
How do you handle promotions and new products?
Promotions lift sales and then often dip them afterwards, so mark them in your data. For new products, there is no history. Use a similar product as a guide, start small and review weekly. Be honest that these are judgment calls.
When does machine learning help?
Once you have a baseline, a machine learning model can learn patterns such as seasonality, promotion effects and product groups. Test it on past data you held back. If it does not beat the baseline by a clear margin, keep the simple method. It is easier to explain and to trust.
What mistakes should you avoid?
- Forecasting every product the same way.
- Ignoring stockout days in the data.
- Trusting a single number instead of a range.
- Letting the system order without a person approving.
- Never checking forecast accuracy against what actually sold.
A worked example of a reorder point
Say a product sells about 10 units a day. The supplier takes 14 days to deliver, and you want a cushion of 40 units for surprises. Daily demand times lead time is 140, plus 40 safety stock, which gives a reorder point of 180.
When stock falls to 180 units, place the order. If you expect a promotion to double sales for a week, raise the daily demand for that period and the reorder point rises with it. The numbers are made up, but the method is real, and it works in a spreadsheet.
Which tools can you use?
- A spreadsheet: enough for a baseline and reorder points.
- A dashboard tool: useful once data comes from several sources and you want live views.
- A small script or application: when you want nightly updates and alerts.
- Machine learning libraries: worth adding after the simple methods are working and measured.
How often should you review forecasts?
Weekly is a good rhythm for fast-moving products and monthly for the rest. Each review should compare what you forecast with what sold, and note why they differed.
When is forecasting not worth the effort?
If you sell a handful of products with stable demand and short lead times, a simple reorder rule is enough. Forecasting earns its keep when you have many products, long supplier lead times or strong seasonality.
Related: AI and workflow automation · D2C and Commerce · All articles