Demand Sensing in CPG: Why It Misses Stockouts & Shelf Gaps

Table of Contents

• 1. What Demand Sensing Is, and Why CPG Teams Have Adopted It
• 2. The Flat Line Problem: One Signal, Three Causes
• 3. Why POS Data Isn't Ground Truth
• 4. The Confounds a Forecasting Model Can't See
• 5. Why This Isn't a Modeling Problem
• 6. The Fix: Feeding Demand Sensing a Shelf-Truth Signal
• 7. Closing the Blind Spot: What This Means for a Demand Sensing Stack
• 8. Frequently Asked Questions

Key Takeaways

• Demand sensing fine-tunes a short-term forecast using near-real-time data – POS, orders, shipments – layered on top of standard demand planning, not a replacement for it.
• A demand-sensing model can't tell a real demand drop from a stockout or a lost shelf position; all three produce the same flat or falling line in POS data.
• POS data shows what happened at the register, not at the shelf, so confounds like phantom inventory or a lost display go undetected.
• The fix is better input, not a better model: real-time shelf-truth data that confirms availability before the forecast reacts.

Demand sensing is one of the most widely adopted answers to a persistent supply chain problem: traditional forecasting is too slow to catch short-term shifts in real demand. By pulling in near-real-time signals instead of relying on historical averages, it promises CPG brands a sharper read on what is about to sell. For most programmes, though, the gap between demand sensing as a concept and demand sensing as a reliable operational tool comes down to one problem: the data feeding the model was never designed to tell a real demand shift apart from a shelf-execution failure.

Demand sensing, the practice of using high-frequency data such as POS transactions, orders, and shipments to adjust short-term forecasts, is well understood as a discipline. The challenge is not knowing what demand sensing is or how to build one. The challenge is that its inputs record what happened at the register, not what happened at the shelf, and a sales dip looks identical in that data whether it was caused by falling consumer demand, a stockout, or a lost shelf position.

Industry researchers put the annual global cost of these blind spots, out-of-stocks and overstocks combined, at $1.73 trillion (IHL Group, 2025), a scale of loss that no amount of forecasting sophistication has resolved on its own.

This blog looks at why that blind spot exists, why it survives even the most advanced forecasting models, and what closes it: a shelf-truth signal that most demand-sensing programmes are still missing. If you're looking at how this fits into a broader demand-planning process, read our guide on predictive demand planning for CPG brands.

1. What Demand Sensing Is, and Why CPG Teams Have Adopted It

Demand sensing is a short-term forecasting method. Instead of relying only on historical sales patterns, it uses high-frequency signals – POS scans, order data, shipments, sometimes weather or promotional calendars – to adjust a forecast days or weeks out rather than months out. The goal is to catch a demand shift while there is still time to act on it: reorder faster, adjust production, or reallocate inventory before a stockout or an overstock compounds. –

Demand sensing in CPG usually sits downstream of a broader demand-planning process, refining a monthly or quarterly forecast with faster-moving data rather than replacing it. That is the difference between demand sensing and demand forecasting: forecasting sets the baseline, sensing corrects it in near real time. On paper, this works well. In practice, its accuracy depends entirely on how much the near-real-time data it uses can be trusted, and that is where the model runs into a problem it was never built to solve.

2. The Flat Line Problem: One Signal, Three Causes

A demand-sensing model reads a sales dip as a single signal. That signal, though, can mean three very different things, and nothing in the data tells the model which one it is looking at.

Consider a hypothetical example: a SKU loses a quarter of its weekly unit sales in a region. There are at least three plausible explanations:

• A real drop in consumer demand: shoppers are genuinely buying less of this product, for reasons ranging from a shifting trend to a stronger competitor promotion.
• A stockout: the product sold out mid-week and simply was not there to be bought for the rest of it.
• A lost shelf position: the SKU is still in stock in the back room, but it was pulled from its planogram slot during a reset, pushed to a lower shelf, or crowded out by a competitor's secondary display.

To a demand-sensing model reading POS or order data, these three scenarios are indistinguishable. All of them produce the same flat or falling line. Yet the correct response to each is completely different: cut the forecast and reallocate inventory elsewhere, expedite replenishment, or fix a merchandising failure that has nothing to do with consumer demand at all. A model that cannot tell these apart will, at best, average out the error over time. At worst, it will act on the wrong one, undersupplying a SKU that never lost real demand or overcorrecting for a shelf problem no amount of extra inventory will fix.

3. Why POS Data Isn't Ground Truth

POS data is not a measurement of shelf conditions. It is a byproduct of a completed transaction, which means it only tells you something when a purchase happens. When a purchase doesn't happen, POS data goes quiet, and quiet looks the same for every reason a purchase might not happen.

This is easy to miss because POS data feels authoritative. It is granular, timestamped, and tied to real transactions rather than a survey or an estimate. Compared to older inputs like distributor sell-in or shipment volumes, it is a genuine improvement, since it reflects what consumers actually bought rather than what a warehouse shipped out. That improvement, though, has a ceiling. Sell-in and shipment data tell you what left the distribution centre. POS data tells you what left the shelf. Neither tells you what was actually available on that shelf for a customer to pick up in the first place, correctly faced, at the right price, in the right spot.

That gap, between what a demand-sensing model can see and what actually happened at the point of sale, is where the blind spot lives. Closing it does not mean replacing POS data. It means adding a layer underneath it that can confirm or rule out what happened at the shelf before the model has to guess.

4. The Confounds a Forecasting Model Can't See

Several shelf-level events routinely produce the same signature as a genuine demand drop, and none of them show up in POS, order, or shipment data on their own.

Phantom Inventory

A SKU can show as in stock in the inventory system while being physically absent from the shelf, a mismatch commonly caused by scanning errors, theft, or misplaced stock in the backroom. This is not a rare glitch: research from ECR Retail Loss, based on more than 1.3 million stock audits across six grocery retailers (Rekik, Syntetos, and Glock, 2026), found that roughly 60% of stock records contained inaccuracies when checked against physical shelf counts. To a demand-sensing model, this looks exactly like waning demand: the product is technically available, so a sales dip gets read as a real signal rather than as inventory that was never actually sellable.

Planogram Drift and Non-Compliance

Shelves rarely stay exactly as planned. A reset, a rushed restock, or a store manager's local decision can leave a SKU facing reduced, moved to a worse position, or dropped from the set entirely. None of this is visible in transactional data. It simply shows up later as fewer units sold, with no indication that the cause was execution rather than demand.

Lost Secondary Displays

Promotional lift is often built around end-caps, off-shelf displays, or secondary placements. When that display comes down, whether the promotion ended or it was reclaimed by another brand, sales usually fall. A demand-sensing model has no way to attribute that drop to the missing display rather than to a genuine cooling of demand for the product itself.

Competitor Shelf-Share Gains

A shelf reset that favours a competitor, whether through a bigger facing count, a better position, or an added display, can quietly erode a brand's sales without any change in category demand at all. The category is doing fine. The brand's share of the shelf is not, and that distinction never reaches a forecasting model built on sales data alone.

5. Why This Isn't a Modeling Problem

None of the confounds above are solved by a better algorithm. Stacking a more sophisticated model on top of the same POS and order data does not give that model new information about what happened at the shelf; it just processes the same blind spot more efficiently. Machine learning can find patterns in the data it is given. It cannot infer the presence of a stockout, a phantom-inventory mismatch, or a lost display from a sales number alone, because that information was never captured in the first place.

This is worth stating plainly: ParallelDots does not build demand-sensing or forecasting models, and this isn't an argument for switching to a different one. It is an argument that the demand-sensing problem sits upstream of the model, in the data feeding it, and that no forecasting improvement can substitute for knowing what actually happened at the shelf.

6. The Fix: Feeding Demand Sensing a Shelf-Truth Signal

Closing this gap means adding a signal that demand-sensing models don't currently have access to: real-time, shelf-level data on whether a product was actually available, correctly placed, and priced as planned. Call it shelf-truth data. Where POS data answers whether someone bought a product, shelf-truth data answers the question that comes before it: whether the product was actually there to be bought, the way it was supposed to be.

In practice, this comes from on-shelf availability and stockout-detection capabilities that capture shelf conditions directly, rather than inferring them from what did or didn't sell. The range here is wide: on-shelf availability scores across CPG brands commonly fall anywhere between 40% and 85%, which is too broad a spread for a model reading sales data alone to know where any given store or SKU sits on it at a given moment. When a sales dip shows up, a shelf-level signal can confirm whether the product was in stock and correctly displayed at the time, ruling stockouts and shelf-execution failures in or out before the forecast has to interpret the dip as a demand signal.

That doesn't replace a brand's existing demand-sensing or forecasting system. It gives that system an input it was missing, so it can spend its modeling effort on genuine demand shifts instead of shelf-execution noise. For more on how stockouts affect retail performance more broadly, see our breakdown of how out-of-stocks impact CPG sales.

This is the role ParallelDots plays in a brand's demand-sensing stack: supplying the shelf-truth signal through on-shelf availability and stockout detection, not the forecasting layer that consumes it. ShelfWatch captures this data directly from shelf images across a brand's retail network, turning shelf conditions into a structured signal that a demand-sensing or forecasting system, whichever one a brand already runs, can use to tell a real demand change from a shelf-execution failure.

7. Closing the Blind Spot: What This Means for a Demand Sensing Stack

A demand-sensing programme built entirely on POS, order, and shipment data will carry this blind spot regardless of how advanced its forecasting model becomes. Closing it isn't a matter of switching algorithms or planning platforms. It's a matter of giving the existing one a signal it doesn't currently have: confirmation of what actually happened at the shelf.

For CPG teams evaluating their own demand-sensing accuracy, the more useful diagnostic question is simpler than “is the model good enough”: for any given sales dip, can the team currently tell whether it reflects a real demand change, a stockout, or a shelf-execution failure? Where the answer is no, that's a data gap rather than a modeling one.

That is the gap ShelfWatch's on-shelf availability and stockout-detection data is built to close, feeding whatever demand-sensing or forecasting system a brand already runs with the shelf-truth signal it was missing.

‍