Shelf-out has three causes. Only one is a forecasting problem.

We arrived at a regional grocery chain in the interior of Brazil on a Tuesday in February. The operations director had booked an hour. We stayed the whole afternoon.

His question was blunt. “Shelf-out is at 6%, I want a better forecasting model.” He’d already gotten quotes from two forecast vendors and had a budget number in his head. He just wanted to know which one to buy.

We asked for one thing before answering. Let us see where the shelf-out actually comes from, SKU by SKU, store by store, over the last six weeks. He said yes. It was the best hour he spent that month.

Shelf-out on the shelf has three causes

Anyone who has worked the operation knows that “shelf-out” on the dashboard is a single number. On the shelf, it’s three different problems with the same symptom: the customer shows up, and the product isn’t there.

Cause one: a forecasting miss. You bought too little. Demand came in above what the system expected, stock hit zero before the next order. This is the cause a better model fixes.

Cause two: a delivery miss. The product was ordered, it’s invoiced, but it didn’t reach the store on time. The truck was late, the DC prioritized another route, the receiving window closed at 11 and the load showed at 11:40. The stock exists. It’s just not where the customer looks.

Cause three: phantom shelf-out. The system says there’s stock in the store. Sometimes there is, in the back room, in a case nobody moved to the shelf. Sometimes there isn’t, and the count has been wrong since the last inventory. The shelf is empty and the system is full.

All three show up as the same 6%. Each one has a different fix, a different owner, a different timeline.

What we found that the director hadn’t said

We split the six weeks by cause. The result changed the conversation in the room.

Of total shelf-out, about a third was forecasting. The rest split between delivery and phantom, with phantom pulling hardest on perishables and high-turn items, exactly the ones that hurt the sale most.

One of the chain’s buyers already knew this, the way an operator knows things before the data proves them. “Shelf-out here isn’t a lack of product, it’s a lack of arriving on time.” He told me that in the aisle, over coffee. He’d just never seen the number split out to show the leadership.

So the director was one purchase order away from spending a full year of forecast budget to attack a third of the problem. The other two thirds have nothing to do with forecasting. No model, however good, would touch them.

What we changed wasn’t the model

We didn’t swap the forecast. We changed two cheaper, more boring things.

First, the store’s replenishment trigger. The signal that told someone to refill the shelf looked at system stock, not the shelf. We changed it so a physical count of the top-200 turn items, done by the stocker on the morning round, fired the refill before the system “noticed” the gap.

Second, the delivery triage. Every receipt outside the window started generating an alert to the buyer the same day, not in Monday’s report. The delay became visible while there was still time to call the DC and fight for the load.

Neither one is a model. One is a trigger. The other is an alert. The system comes after; the decision about what to measure comes first.

That criterion comes before any model purchase, the way we wrote about what to measure in week one. And it’s why we work from inside the operation: splitting the three causes doesn’t come out of a report, it comes out of walking the floor with the stocker. That’s how we work.

The number that moved

In nine weeks shelf-out dropped from 6% to 3.4%.

Now the honest part, because without it this turns into an ad. That number didn’t fall only from what we did. Midway through, the chain opened a new DC that shortened the route for two regions, and that alone improved the “delivery” cause independent of us. My estimate: half the drop was our work, half was the DC. I won’t pretend we earned the full 2.6 points.

Here’s what matters. The forecasting piece, the one the director was going to buy first, moved the least. It’s still sitting there, waiting. The difference is he now knows what it’s worth: real, but the third problem in line, not the first.

What we got wrong

We were slow to distrust the store count.

Early on we instrumented phantom shelf-out assuming the back-room stock was correct and the problem was just not getting it to the shelf. Wrong. A good chunk of phantom was a bad count in the back room itself. The system said 12, there were 3.

We lost a week chasing a replenishment problem that was, underneath, an inventory problem. When the count lies, everything downstream lies with it. We should have tested how much to trust the stock number before building on top of it. Noted for next time.

What we left running

We left two things with the internal team, and walked out.

A shelf-out triage card: for every gap on the shelf, the stocker marks which of the three causes. Forecast, delivery, or phantom. The buyer reads it, and at month end the chain knows the proportion, not just the total. That’s what was missing the whole time: the 6% never said where it came from.

And a daily shelf audit on the top-200, run by the store itself, without needing us there for it to happen.

Now think for a second about your operation. Open your shelf-out dashboard. It gives you a number. Ask it which of the three causes it’s measuring. It can’t answer. The store team can, but nobody asks them the right way.

That’s the gap: between the number the head office reads on Monday and what the stocker sees walking the shelf at ten in the morning. The head office doesn’t come down to the shelf. The stocker doesn’t write reports. The number sits in the middle, alone.

If that shape of gap sounds familiar, describe in one paragraph where your shelf-out hurts most. If the shape confirms a useful observation, we go in two weeks embedded and come out with a one-page document your team keeps: what we saw, what your shelf-out number missed, and the three changes that separate forecast from delivery from phantom. The better forecast can wait until you know whether it’s actually the one costing you the sale.