When AI stock forecasts go wrong

A bad forecast usually comes from the data, not the maths. Most of the causes below can be spotted by anyone who knows the shop.

One-off sales

A single big order, a clearance or a school fundraiser can lift one week far above the rest. The tool sees a high number and may treat it as normal. In statistics, a value that stands far from the others is called an outlier.¹ The cure is simple. Mark those weeks, or take them out, before you ask for a forecast.

How to notice: the suggested order for a product is far above what you've ever sold in a normal week.

Stockouts that hide demand

When an item runs out, sales stop. The history then shows low demand when the real problem was no stock. Wikipedia notes that substitution disguises true demand, since shoppers buy something else when their first choice is gone.² The tool may then suggest ordering less of the item you ran short of.

How to notice: look for weeks where on-hand stock was zero or near it, and where sales fell. Add a note column for those weeks. Inventory data can help find out-of-stocks, but only depending on its accuracy.² If the system counts are shaky, you may not even know which weeks were empty.

New products and ended ones

A product with only a few weeks of sales gives the tool little to work from. A product you've stopped stocking may still show up in the list. Either way the suggestion is guesswork. Leave new items out of automatic suggestions and decide those orders yourself.

How to notice: any suggestion for an item with a short history, or for something you've dropped.

Seasons and trends

Averages assume demand stays steady. A standard safety stock formula assumes steady demand. Wikipedia says that with a summer peak and a winter dip, it leads to repeated shortages in summer and surplus in winter.³ The same goes for products that are growing or fading. If you sell seasonal goods, give the tool the right weeks, and tell it which season is coming.

How to notice: the suggestion looks right for last month and wrong for next month.

Wrong lead times

A suggestion depends on how long the supplier takes. Wikipedia says lead time is extremely hard to quantify in complex purchase environments.³ A tool using a lead time you never checked will get the cover wrong. A supplier who has slowed down means you need to order earlier than the tool says.

How to notice: ask the tool what lead time it assumed. Compare it with your last few deliveries.

Bad counts and messy codes

If the on-hand figure is wrong, so is the answer. The same goes for duplicate rows or two codes for one product. See Getting your stock data ready for an AI tool.

How to notice: a suggestion to order stock you can see on the shelf, or not to order something that's empty.

Made-up details from the tool

Chat tools can invent numbers or change them on the way through. See What AI can and cannot do with stock and reorders.

How to notice: redo the sum for one product by hand, using the check list for order quantity requests.

A routine that catches most of these

Compare forecast and result on a regular schedule. Forecast error needs to be calculated using actual sales as a base.⁴ You don't need a formula. Put last period's suggestion beside what you really sold, and note the gap. Microsoft's guidance makes a related point: starting a forecast before the last historical point lets you compare the forecasted series to the actual data.⁵

If the gaps are big for one product, look at the causes above before you blame the tool. The cause is often in the records. And whatever you find, a person approves the order. Stock counts and checks a person must still do explains what to check on the shelf.