By Laas NuryFor a wholesaler or distributor, the forecast sits behind almost every number that matters: how much stock you hold, how often you run out, how much you write off, how much cash is tied up in the warehouse. Most companies still forecast with a spreadsheet, a moving average in the ERP, and the buyer's experience. That works, until the range grows, the promotions multiply and the experienced buyer retires.
AI demand forecasting can do better, but only on the right items and with the right data. This article explains how it works, what data it needs, how to measure whether it actually beats what you do today, and how a better forecast turns into better purchasing.
How it works
A classic forecast looks at one item's history and extends the pattern: last year's sales, a moving average, a seasonal factor. An AI forecast learns across all items at once, and adds the factors that drive demand:
- History: sales or orders per item, per location, per customer group.
- Calendar: seasons, holidays, school holidays, paydays, the day of the week.
- Your own actions: promotions, price changes, new listings at customers.
- Outside factors where they matter: weather for seasonal products, known events, market indicators.
Because it learns across items, it can forecast items with a short history by borrowing patterns from similar ones. In the large M5 forecasting competition on retail sales data, machine-learning methods that learned across many items at once outperformed the classic statistical benchmarks. That matches what is seen in practice: the gain is largest where there are many items and many factors.
Which data you need
| Data | Why it matters | Minimum |
|---|---|---|
| Sales or order history | The basis of every forecast | Two years, per item and ideally per location |
| Stock-outs | Without them, the model learns that no stock means no demand | A record of when items were not available |
| Promotions and price changes | They explain the peaks the model would otherwise treat as noise | Dates and depth of each promotion |
| Lead times and order quantities | Needed to turn a forecast into a purchase proposal | Per supplier and item |
| Item attributes | Let the model learn across similar items | Group, brand, pack size, season |
You do not need perfect data to start. You do need to know where it is wrong, especially the stock-outs: they are the most common reason a forecast systematically underestimates demand.
Measuring whether it is better
A new forecast is only worth it if it beats the current one on your data. Measure both the same way, on a period the model has not seen:
- Accuracy per group of items, for example with the weighted absolute percentage error (WAPE): how far off the forecast was, weighted by volume.
- Bias: does it consistently forecast too high or too low? A small but steady bias costs more than a larger random error.
- Forecast value added: how much does each step improve the forecast, including the buyer's manual adjustments? Sometimes the adjustments make it worse.
- The business result: availability, stock levels and write-offs in a test period, not only the error figures.
Split the results by type of item. Fast movers with stable demand will improve most; slow movers and new items stay hard for any method. That is normal, and it tells you where the buyer's attention should go.
From forecast to purchasing
A forecast on its own changes nothing. The value comes from what it drives:
- The forecast per item and location, updated daily or weekly.
- Safety stock based on how uncertain that forecast is, rather than a fixed number of days for everything.
- Purchase proposals that take lead times, minimum order quantities and supplier agreements into account.
- The buyer decides: approves, adjusts or rejects, with the reasons for each proposal visible.
- Exceptions first: the items where the forecast is uncertain, demand is changing or a supplier is late come to the top of the list.
In most companies the buyer's job shifts from calculating to judging: less time on the items that behave normally, more on the ones that do not.
How to start
- Pick one product group or one warehouse with enough volume and history.
- Collect the data in the table above, and mark the stock-outs.
- Build the forecast next to the current process and compare it on a period the model has not seen.
- Run it in parallel for a few weeks: the buyers work as usual and see the proposals alongside.
- Switch the product group over when the numbers are better, and expand from there.
Check first whether your ERP's built-in forecasting already does enough for your range; many systems include it. If you run SAP or Dynamics 365, AI on top of SAP and Microsoft Dynamics explains how a forecast next to the ERP connects to it. And the AI readiness assessment helps you check whether the data is there before you begin.
