How Data Driven Demand Forecasting Reduces Stockouts in Retail Operations

A shopper walks in for oat milk. The shelf is bare. She doesn't wait for next week's delivery, she grabs another brand or walks to the Tesco across the road. Online it takes even less time: one grey "out of stock" button and the tab gets closed. And most forecasts still come from last year's Excel file, with a 5% uplift added on top. That worked when demand was boring. It isn't boring anymore.

Why the Old Spreadsheet Can't See What's Happening Today

Start with the money. IHL Group puts the global cost of empty shelves and excess stock at around $1.7 trillion a year in its 2026 study. That's 6.2% of all retail sales in the world. And the bigger slice of that pie? Empty shelves. Not dusty pallets in some warehouse, as most people would guess.

Look at what happens at the big global chains once they switch on modern consumer goods technology solutions and wire store sales data straight into their distribution centres: suddenly the whole flow of goods is in view, end to end and in real time, from the till back to the loading bay. A planner can open one screen and see what sold in Croydon at 11 a.m., what's still on a truck outside Daventry and how many units the website promised for tomorrow's delivery.

A spreadsheet shows none of that. It works with monthly totals. It looks at the whole category, not the single SKU. And it has no idea that a hot weekend in Manchester will clear bottled water from 40 stores by Saturday lunch.

Familiar picture? Ask any planner. They'll tell the same story, usually with a sigh.

What a Good Forecast Is Actually Built From

The formula matters less than people think. What matters is the data going in, and how small the slices are. SKU by store by day. For some fresh categories, by the hour.

POS Data Integration Comes First

Start with the till. Every beep at checkout is a tiny vote for a product. Feed those votes to the planning system every night, not every Friday, and the forecast gets sharper within a few weeks. SAP Integrated Business Planning has a demand sensing module built around exactly this idea. It takes the latest sales and adjusts the next few days instead of waiting for the monthly planning meeting.

Now, a trap. POS data shows what sold, not what people wanted. Say a yoghurt was out of stock from Monday to Wednesday. The history says zero sales. Real demand was probably the usual 25 pots a day. A decent system marks those days as "censored" and fills the gap. A weak one learns that nobody buys that yoghurt on Mondays. Next order gets smaller. The shelf goes empty again. Funny how that works.

Outside Signals That Earn Their Place

The till is only part of it. Other data that usually pays off:

• Weather by postcode. Charcoal in May. Umbrellas on a drizzly Thursday in Glasgow.

• Promo calendars with real detail. Discount depth, end-cap display or not, TV support. A simple "promo: yes" flag is close to useless.

• Local events. Derby matches, school holidays, a big concert in the area.

• Competitor prices. A rival cutting price on nappies by 20% changes demand next door.

• Online browsing data. Interest on the website often climbs two or three days before store sales do.

• Real supplier lead times. Not the 7 days in the contract. The 11 days the supplier actually takes.

Does every retailer need all six? No. A DIY chain gets more out of weather and housing data. Pharmacies? They watch flu numbers like hawks, and half the rest is noise with a pretty dashboard on top.

Machine Learning or Plain Statistics?

Vendors love to say "AI-powered". Okay. But toilet paper doesn't need a neural network, good old exponential smoothing copes just fine. ML starts paying its way when three things land at once, say a two-for-one on Pimm's, a 29-degree Saturday and a bank holiday Monday. A moving average just shrugs at that kind of week.

The sensible setups run both. Simple models for the steady 60% of the range, ML for the volatile rest. Cheaper to run, easier to explain to a category manager, and nobody has to defend a black box in the Monday meeting. Sounds reasonable, right?

Where the Stockout Actually Gets Stopped

Here's the uncomfortable part. A brilliant forecast sitting in a dashboard fills zero shelves. It has to push replenishment orders automatically, or it's just a nice chart.

Safety Stock That Recalculates Itself

Most retailers still set safety stock once. "Two weeks of cover" across the board, signed off, and nobody touches it again until next January, when someone finally notices the backroom is full of last summer's paddling pools. The result is predictable. Slow lines pile up in the backroom. And the twitchy, fast stuff runs dry by Wednesday.

Safety Stock Optimization flips that. Every night, each SKU gets its buffer recounted. How far off was the forecast last week? How late did the lorry turn up? Is the target 95% service or 98%? Take a grocer with roughly 150 shops. It switches to daily automatic recalculation. Within two quarters, out-of-stock incidents on its top 2,000 products fall by roughly 38%. Total stock barely changes, because units shift from tired long-tail items to the fast sellers. No new warehouse. No extra cash tied up.

So is that a software win? Partly. The other half is people. Planners had to stop overriding the numbers every Monday morning out of habit. That part is harder than the maths, honestly.

One trick that tends to work: a monthly override scorecard. Who beat the system, who made it worse, by how many units. Overrides often halve within a quarter. Nobody wants to be the red bar on that chart twice in a row.

Omnichannel Logistics and One Shared Pool of Stock

Omnichannel makes all of this messier. Take one tin of paint in a Leeds store. At 2 p.m. a walk-in customer picks it up and puts it back. At 2:10 a click-and-collect order claims it. By 3 the website has promised it to somebody in Harrogate too. Whose tin is it? Somebody's getting an apology email.

Tools built for this mess:

• Manhattan Active Omni. Manhattan Associates' answer to "where is that tin right now?", across shops, DCs and fulfilment sites.

• Blue Yonder Luminate Planning. ML forecasts hooked to store replenishment. Grocers love it.

• Dynamics 365 Supply Chain Management. Planning, warehouses, stock visibility, and it plugs into Dynamics 365 Commerce. Handy for anyone already deep in Microsoft.

• Oracle Fusion Cloud SCM. Demand and supply planning, a natural fit for shops already running Oracle Retail.

Who Does What in the Market

Nobody does everything well. Three or four tools glued together is normal for a big chain, and that's fine.

Rough cheat sheet. SAP IBP if S/4HANA is already in the building. Blue Yonder when store replenishment is the sore spot. Manhattan when orders leave the DC too slowly. Dynamics 365 SCM for mid-sized chains that don't fancy a three-year rollout.

Brands come at it from the other end. With Salesforce Consumer Goods Cloud, a field rep walks the aisle with a phone, snaps the gap, logs the visit, moves on. Those photos catch stuff the till never will.

And somebody has to wire it all together. DXC Technology, other integrators and the vendors' own consultants do the plumbing: data pipelines, ERP upgrades, master data clean-up, keeping things running after go-live. Dull? Sure. Skip it, though, and the shiny engine runs on rubbish.

Numbers That Prove the Forecast Works

No KPIs? Then by month six it's just people arguing in meetings. Track these:

1. OSA, or on-shelf availability. How often the product is actually there to buy. Strong grocers sit at 97–98%.

2. WMAPE. Weighted forecast error at SKU-store-week level. Category-level accuracy looks great and hides almost everything.

3. Forecast bias. Always 4% too high, week after week? That quietly stuffs the backroom. Random misses hurt less than a steady lean one way.

4. Inventory turnover. How many times stock sells through in a period. Turnover and OSA going up together is the best sign there is.

5. DC-to-store fill rate. If the warehouse ships only 92% of store orders, no store forecast will save the shelf.

6. Estimated lost sales. Modelled demand during stockout days, priced at retail. The CFO will read this one first.

Look at them side by side. Availability bought with 30% more inventory isn't progress. It's just an expensive way to look good.

Why So Many Projects Stall

Funny thing: the maths almost never breaks. So what breaks?

• Messy master data. Duplicate SKUs, wrong case sizes, lead times last updated in 2021.

• Phantom stock. System: 12 units. Shelf: empty. Reorder: never sent, because on paper all is well. RFID and regular cycle counts help a lot.

• Untracked manual overrides. Planners tweak numbers "on gut feel". Measure those tweaks. Many make the forecast worse.

• Rolling out everywhere at once. Switching on 500 stores on a Monday is asking for trouble. Start with chilled desserts or pet food, prove it, then grow.

• No owner. Merchandising blames supply chain, supply chain blames stores. Someone has to sign their name next to the OSA target.

• Leaving suppliers out. A supplier who only sees the purchase order is always a step behind. CPFR, basically shared forecasts, fixes a lot of that.

Worth the Effort?

Yes. Not magic, mind. A supplier will still be late now and then, and some TikToker will still wipe out a shelf of Stanley cups in one evening. What's new is that someone owns the availability number and checks it every morning, not in a Friday panic. Granular POS data, self-adjusting safety stock, one view of stock across channels. Get those three right and the empty shelf becomes the exception.