5 NOVEMBER 2026 · 17:00 GST · 5 MIN READ · DUBAI CAREER NOTES

How to forecast demand for your online store with AI: the Baseline, Bump, Bounce rule

By Prudhvi Polavarapu, supply chain and e-commerce operations leader in Dubai

If you want to forecast demand for your online store with AI, do not start by buying software. Start with a spreadsheet of last year's weekly sales and a language model that can explain what that spreadsheet is telling you.

The pallet that arrived one week late

A scene I have lived more than once. A promotion is planned, the creative is approved, the campaign goes live on a Friday, and the stock for the hero product is still on a ship. Or the opposite: the stock arrived early, 4,000 units sit in the racks, and the sale underperforms because nobody checked what happened the last time the same offer ran. Both mistakes come from the same place. Someone guessed the number, and nobody wrote the guess down so it could be checked later.

I have planned for peaks above 5,000 orders a day, and the lesson never changed. A forecast is not a prophecy. It is a written number with a reason attached, so that you can learn from the gap between what you expected and what happened.

What can AI actually do for demand forecasting?

There are two different jobs hiding under the word AI. The first is statistical forecasting: software that reads your order history and projects it forward, accounting for trend and seasonality. Many inventory planning tools do this, and some marketplace and store platforms include basic versions. The second is what a language model does: you give it your weekly sales table and your promotion calendar, and it helps you explain the pattern, spot odd weeks, and draft the questions you should ask the buying team.

For a small or mid-sized UAE store, the second job is the one you can start tomorrow at no cost. It will not replace a planner, but it makes a planner faster. It reads the table, flags the weeks that look strange, and writes the summary you would otherwise skip because it takes an hour.

The Baseline, Bump, Bounce rule

I split every forecast into three parts, and I call it the Baseline, Bump, Bounce rule.

Baseline is what the product sells in a normal week with no promotion and no event. Take the quiet weeks only. This is the number your reorders should protect.

Bump is the extra demand from something you can name: a sale, Ramadan, back to school, a national day offer, a marketplace campaign, a new ad. Each bump needs a label and a size, even if the size is a rough estimate.

Bounce is what happens after. Customers who bought early do not buy again, so the weeks after a big sale often run below baseline. Teams that forget the bounce over-order for the following month and then wonder why the racks are full.

A forecast is a written number with a reason attached.

This is where AI earns its place. Paste in two years of weekly sales for one product, mark which weeks had a promotion, and ask the model to separate the three parts and show its reasoning. Then check it against what you remember. If it says March was a bump and you know March was a stock-out, you have just found a data problem, which is worth more than any forecast.

What data do I need for AI demand forecasting?

Less than you fear, but cleaner than you have. You need weekly units sold per product for at least a year, ideally two so that a season repeats. You need a simple calendar of promotions, price changes and stock-out weeks. A stock-out week shows low sales that were really lost sales, and if you feed it in unmarked, the forecast will learn that the product is slow when it was simply missing.

That last point connects directly to inventory accuracy. If your system says 12 and the shelf says 9, your sales history and your forecast are built on a false picture. I wrote about fixing that in how to fix inventory accuracy with cycle counts. In my own warehouses, 99.6% inventory accuracy mattered as much for planning as for picking, because every forecast starts from what you truly hold.

How do I run a first forecasting test?

Pick ten products: five best sellers and five slow ones. Hide the last eight weeks of sales, give the model the earlier history, and ask for a weekly forecast for those eight weeks with the Baseline, Bump and Bounce shown separately. Then reveal the real numbers and compare. Count how many weeks it landed within a range you could live with, and note where it missed and why.

Do this before you trust any tool, paid or free. If a simple method matches your planner, you have saved money. If it misses the same weeks your planner misses, you have found the weeks that need a human conversation, such as a supplier delay or a campaign nobody logged. Keep the habit from my earlier note on starting with the boring AI: let the tool do the reading and the arithmetic, and keep the deciding for a person who knows the market.

What to keep humans responsible for

The order itself. A purchase order commits money, shelf space and supplier relationships, so a named person approves it. AI also cannot know what is not in your table: a competitor's launch, a supplier's lead time changing, a new marketplace listing going live. Add those facts to the prompt yourself. And take care with data: remove customer names, phones and addresses before pasting anything into a tool, and follow your company's policy on approved tools. The UAE has made artificial intelligence a national priority, with a dedicated UAE Artificial Intelligence Office, so operators who learn these habits early are working with the direction of the country, not against it.

You do not need a data science team to plan better. You need honest history, a named reason for every spike, and the discipline to compare your guess with reality every week. Write the number down, and let the gap teach you.