How Can AI Improve Demand Forecasting in Retail?
AI forecasts demand per product per store, using the things that actually move it — price, promotions, weather, local events, and what similar products did. More usefully, it gives you a range rather than a single number, so your buyers can decide how much stockout risk they are willing to carry instead of guessing at it.
Traditional vs. AI-Based Demand Forecasting
| Step | Traditional Forecast | AI-Based Forecast |
|---|---|---|
| Level of detail | Category or region, then split by rule of thumb | Per product, per store, where the ordering decision is made |
| What it considers | Last year's sales, adjusted by a planner's judgement | Price, promotions, weather, calendar and local events together |
| New products | Guesswork, or a manual analogue chosen by hand | Inferred from similar products' launch curves |
| The output | A single number that hides its own uncertainty | A range, so buyers can choose their stockout risk explicitly |
| Promotions | Uplift applied from a fixed table | Learned per product, including the cannibalisation next door |
A Range Beats a Number
A single forecast tells a buyer what to expect and nothing about how wrong it might be. Two products can forecast identically and behave completely differently — one steady, one volatile — and the buyer needs to hold far more cover on the second.
A probabilistic forecast makes that explicit. You stock to a service level you have chosen rather than to a point estimate you quietly padded, and the padding stops being invisible tribal knowledge held by whoever has been buying that category longest.
It also changes the conversation about accuracy. The question becomes whether the range was honest, not whether the number was right — and an honest range is achievable in a way a right number is not.
Where to Start
One category, in stores you know well. Backtest first: run the forecast against last year and compare it to what your current process actually produced, on the same weeks and the same SKUs.
Measure the thing that costs money rather than forecast error in the abstract. Lost sales from stockouts and markdown on overstock are the numbers your finance team recognises, and a model can improve statistical accuracy while making both worse if it is tuned for the wrong objective.
Where This Fits
This is one part of our work in AI for Retail. See the full set of AI use cases for the equivalent in other industries and functions.
Frequently Asked Questions
Our sales history includes COVID. Does that break the model?
It has to be handled deliberately, or yes. Left alone, a model reads 2020 and 2021 as evidence about seasonality that simply is not true. The usual approach is to mark those periods as anomalous so the model learns the pattern around them rather than through them — and to be honest that anything genuinely structural, like a permanent shift to online, is a real change the model should learn.
How is this different from the forecasting in our ERP?
Most ERP forecasting extrapolates your own sales history and little else. It cannot tell you that a heatwave is coming, that a competitor is running a promotion, or that this SKU cannibalises the one beside it. Those external and cross-product effects are where most of the remaining error lives, which is why planners override ERP forecasts so often.
Will this replace our demand planners?
It changes what they spend time on. Planners currently spend most of their week producing numbers and very little of it on the exceptions that matter. A good system inverts that — it produces the baseline, and your planners work the products where their market knowledge genuinely beats the model, which is a real and permanent category.
How much history do we need?
Two years is comfortable, one is workable, less is difficult for anything seasonal — a model cannot learn a Christmas pattern it has only seen once. Breadth helps too: many stores with a shorter history is often better than one store with a long one, because the model can learn shared patterns across locations.
What does it do about products we have never sold before?
It borrows from products that look like them — similar price point, category and launch timing — rather than starting blind. That is usually better than a planner picking a single analogue by hand, though it is still the weakest case for any forecasting method. Expect to review new-product forecasts manually for the first few weeks.
Backtest a forecast against last year before it touches a purchase order.
Get a Free Proof of Concept within weeks.
Let’s talk
AI is here to stay
Let’s win together
Your first 45-min alignment session — and a small PoC — are free.
Or just say hello or write us an email.