Demand Forecasting With AI: What It Needs to Work

Google published a forecasting model this week that tops three leaderboards at 330 million parameters, which is small enough to run almost anywhere and cheap enough that nobody needs to think about the cost. Forecasting has quietly stopped being a capability problem. What decides whether it works for your business is something much less interesting: whether your own history is worth learning from.

It is not only about stock

Demand forecasting gets discussed as an inventory topic, which leaves out most small businesses. If you sell services, the same technique answers questions you probably guess at every month.

How many bookings should we expect in March. How many crews will we need in the third week of July. How many enquiries come in after the long weekend. When does cash actually arrive, given how our customers really pay rather than when the invoice says.

The technique does not care what the numbers represent. It cares whether there is a pattern to learn.

History is the whole constraint

That bottom row matters more in Canada than the general advice suggests. Most businesses here have a strong annual cycle, and a model that has only watched nine months has no concept of what February does to you. It will still produce a number, and the number will look exactly as authoritative as a good one.

Check what you already own

Before buying anything, look at your accounting package, your scheduling system, and your point of sale. Forecasting is turning up inside tools businesses already have, because the models got small enough to include as a feature rather than sell as a product.

There is also a practical argument for the built-in version beyond price. A forecast that lives where your data already sits gets looked at. A separate system that needs a monthly export gets used twice and then quietly abandoned, which is the same fate as most standalone reporting. If your operation runs on a large ERP the picture is different, and forecasting inside Oracle SCM covers that end.

Backtest before you trust it

This takes an afternoon and it is worth more than every vendor accuracy figure combined, because it is measured on your business rather than someone else's.

Give the model your data up to a point six months ago. Ask what it predicts for the six months after that. Compare against what actually happened.

You will learn two things immediately: whether it is roughly right, and where it goes wrong. The second is more useful. A model that is reliable except in December is a model you can use, as long as you know that about it.

It cannot know about the new competitor

A forecasting model learns patterns from the past. It handles seasonality, trend, and day-of-week effects well. It handles a competitor opening across the street, a supplier collapsing, or a regulatory change not at all.

And it will produce a confident number anyway, because producing a number is what it does. This is why a forecast should inform a decision rather than make one, and why the person reviewing it needs to know what is actually happening in your market. The pattern is the same as everywhere else AI touches operations: excellent at the regularities, blind to the thing you read about in the local paper last week.

Keep score

Once it is running, record the forecast next to what actually happened. Every month, one line.

After a few months you will know whether to plan around it, and you will know the size of the error rather than having a vague feeling about it. That is what turns a forecast from an interesting output into something you can build a staffing decision on. It also makes the inventory version worth doing, which we covered separately in AI inventory management, and none of it works if the underlying records are inconsistent, which is the argument in getting your data ready for AI.

Frequently Asked Questions