Demand Sensing in Retail

Demand sensing in retail is the in-season practice of estimating how fast demand is running now from recent signals, and using that estimate to change the decisions that can still move inside their lead times. The signals are point-of-sale sell-through, e-commerce orders and add-to-cart, returns, the on-hand and in-transit position and, for wholesale brands, retailer or dealer sell-through; the decisions are allocation, reallocation, replenishment triggers, chase orders, receipt flow and markdown timing. It does not replace the pre-season demand forecast that sized the buy months earlier. It works on a horizon of weeks, and its value is bounded by the decisions still open when the signal arrives.

The term covers two settings. Where the same item sits on the shelf for years, sensing is a short-horizon correction to a long-run item forecast, with a replenishment order always waiting to act on it — the standing-position setting core and NOOS replenishment programs covers. In seasonal and fashion-led categories the item has weeks of history and the buy is committed, so the question is narrower: given the first weeks of selling, which remaining levers should move, and how far. This guide is written for the second setting, across all ten verticals.

What demand sensing is

Three things separate sensing from reading a sales report. It produces a rate, not a total — units per week at a stated grain, adjusted for demand at positions that had no stock. It compares that rate with something already committed — the plan, the position, the on-order — so the output is a gap. And it attaches the gap to a decision with a deadline: the next allocation run, the replenishment trigger, the last week a chase can land and sell. Sensing is finished only when a decision has moved, or has been deliberately held with a date to look again.

Demand sensing vs demand forecasting vs demand planning

All three produce a number about demand; they differ in question, horizon, grain and what happens next. The one-paragraph form of this distinction sits in the demand sensing glossary entry. The demand vs merchandise vs supply planning guide covers where demand planning meets merchandise planning, and demand forecasting by vertical covers the pre-season half across the same ten categories.

The practical boundary is recency. A forecast weights a long history to stay stable; sensing weights the last few weeks to respond, and that responsiveness is what exposes it to noise. A sensed rate is a short-horizon estimate, and projecting it to season end turns it into a forecast it was never built to be. A sensed shift that persists belongs in the reforecast — the reforecast cadence guide covers when.

The signals: what each measures, how late it arrives, and at what grain

Each signal arrives with a delay, stops being trustworthy below some grain, and has a way of recording something other than demand.

Sell-through is the backbone, and its limit is fundamental: sales history records what sold where stock existed, not what customers wanted. It also thins out fast. A style spread across a fleet of doors and sizes can sell well under one unit per door-size position per week, so the readable rate sits at style-color across doors, pushed down through the size curve. Interest signals matter more for censoring than for volume: a click on an out-of-stock size is evidence of demand at a position that recorded no sale. Returns turn gross demand into net on a lag, so a style with a high return rate is read net of expected returns before it drives a chase.

For wholesale brands the order book is sell-in, and an account's reorder is a delayed, filtered version of its sell-through; shared retailer POS is the direct view of the consumer in that channel. Weather and event calendars are context, not volume: weather explains when demand arrives more reliably than how much, so an early cold week raises the question of whether the season got bigger or only earlier.

The sensing horizon: which decisions a signal can still change

A signal is worth reading only against a decision whose lead time has not passed. The in-season chase guide calls this each lever's last responsible week.

The pre-season buy is outside the sensing horizon. Once production is committed, a week-three signal changes how the buy is distributed, whether the remaining open-to-buy funds a chase, and when markdowns start — not how many units were bought. Sensing redistributes and times inventory that exists or is committed, and adds a bounded quantity where a chase is still open; the quantity lesson goes to the season hindsight. How much budget stays uncommitted for that chase is covered in open-to-buy by vertical.

Each lever on the same style has its own horizon, so the sensing cadence follows the fastest decision still open, and latency is judged against the lever: a weekly retailer report can inform a chase that closes in a month but is too slow to direct this week's holdback release.

Data prerequisites: accurate position, the sellable grain, and censored demand

Sensing amplifies input errors, because it weights recent weeks where errors have had no time to average out. Data readiness for merchandise planning has the wider checklist.

On-hand accuracy. The position is the denominator of forward weeks of supply, the replenishment minimum and every transfer. An overstated count hides a stockout and understates the sensed rate at once, because the position looks stocked while selling nothing — see inventory accuracy. Sellable excludes floor samples, testers, damages and goods on memo.

The sellable grain. Breakage happens at the grain the customer buys — the size, the size run, the shade, the finish, the piece — and has to be measured there. The rate is estimated at the coarsest grain the volume supports and pushed down. Confusing the two produces either invisible breakage or a noisy rate.

Censored demand. Sales history is demand filtered by availability. An empty position records zero, a broken size run records slow sales on the sizes left, an unallocated door records nothing, and a late receipt records weak weeks that were an empty floor. Every one of those reads as low demand unless the rate is adjusted for availability. In-stock rate shows where demand was censored and lost sales estimates how much. Without them, a winner that sells out early looks like a style that slowed down, exactly when it should be chased. Promotional weeks distort the other way and are read apart from full-price weeks.

Illustrative worked example: one style-color at week three

The figures below are illustrative, chosen to divide cleanly. They are not benchmarks or targets, and they are not drawn from any brand.

A women's linen shirt in one color is bought at 2,500 units for a 12-week full-price window: 2,000 allocated across 50 doors in 5 sizes, 500 held back in the DC. The plan is back-loaded toward the summer peak and calls for 100 units a week in weeks one to three. Fabric is held for a chase with a five-week lead time.

Step one, sense at a grain the evidence supports. Week three sold 200 units across 250 door-size positions, 0.8 per position. No single position carries a readable rate; the style-color across the fleet does.

Step two, adjust for censoring. The 200 in-stock positions sold one unit each. If the 50 empty positions had demand at the same rate, week-three demand was 250 units. That is a first approximation, and more likely to understate than overstate, because the positions that stocked out were the fastest ones.

Step three, compare with the position. Doors hold 2,000 − 500 = 1,500 units; with the holdback, 2,000. That is 8.0 weeks of forward supply at the adjusted rate against 9 weeks remaining, or 10.0 weeks at the observed rate. The two rates bracket the decision.

Step four, attach decisions. Release the holdback now, against the empty sizes in doors where the rest of the size run is selling: both rates say the 50 empty positions are costing sales, and the lever takes days. Hold the chase one week, with the rule written in advance. A chase placed at the end of week three lands for week nine; placed a week later it lands for week ten, when the position at the adjusted rate would be 2,000 − (6 × 250) = 500 units against 750 units of demand for weeks ten to twelve — so a 250-unit chase still closes the gap. The extra week buys the first read with the holdback on the floor. The rule: chase nine weeks of the week-four adjusted rate less the 2,000-unit position at the start of week four, and nothing if that is zero or below — 250 units at 250 a week, none at 200. Markdown stays shut: at the adjusted rate the style is 250 units short, and even the censored observed rate leaves only 200 units of excess cover — one week at 200 a week — too little to open price.

Sensing did not change the 2,500-unit buy. It redistributed 500 units, pre-committed a chase of up to 250, and left price alone. The output of sensing is a decision with a threshold, not a number on a dashboard.

Failure modes

Reacting to noise

A promotion, a marketing send, a local event or a receipt landing mid-week moves one week's number without moving demand. Three tests separate a shift from noise: it persists into a second read, it shows in more than one signal, and it points the same way across clusters. Match the evidence bar to how reversible the lever is — a transfer on one read, a chase or a markdown on more.

Sensing at a grain finer than the evidence

The example's 0.8 units per door-size position is the problem in one number: a rate at that grain is zeros and ones, and corrections chase each other. Estimate at the coarsest grain that still changes the decision and push down through the size curve, the store cluster or a door index — the rule in allocation and replenishment by vertical.

Sensing without a decision attached

Sell-through against plan that reaches a meeting with no owner, lever or deadline looks like diligence and changes nothing. The symptom is the same style flagged three weeks running until the only lever left is price. Every sensed exception names the lever, the decider and the last week the decision can land; the weekly trade meeting is where that rhythm lives.

Projecting the launch, ignoring the on-order

Launch and floor-set weeks carry novelty and placement that later weeks do not, so a launch rate projected to season end sizes the chase to a peak; scenario pairs keep the decision inside the evidence. And a rate compared only with on-hand double-orders — the position that matters is on-hand plus confirmed on-order, less commitments, the logic of available to sell.

How to evaluate a demand sensing solution

These questions are vendor-neutral and apply to a dedicated tool, a planning platform's sensing layer or an internal build. Questions to ask planning software vendors covers the wider evaluation.

1. →Which decisions does the sensed rate write back to? Allocation, replenishment parameters, chase quantities, markdown timing — or only a report. Ask to see the decision record.
2. →At what grain is the rate estimated, and how is it pushed down? An unpooled rate at door by size by week is noise wherever volume per position is low.
3. →How are stockouts and broken size runs treated? Is demand adjusted for availability, and is the adjustment visible?
4. →How is in-stock defined? Planned positions or all positions; snapshot or time in stock; sellable units only.
5. →Which signals, at what latency, and what happens when one is late? The system has to tell a missing retailer report from a week of zero sales.
6. →Does the rate sit beside on-order and remaining open-to-buy? A chase recommendation without the budget that funds it cannot be approved.
7. →How are promotions and returns separated from baseline demand?
8. →Can a planner see why a recommendation was made, and override it with the override recorded?
9. →How is it backtested on your history? Use data as it existed on each date, not restated history, or the result looks better than it could have been live.
10. →What does it do with no history, and where does it stop? Ask about launches, new doors and new shades, and how it exchanges numbers with any statistical forecasting system you run.

Demand sensing by vertical

The mechanics hold in every category; three inputs change with it — the trustworthy signal, the lever the lead time leaves open, and what hides demand. The planning calendar by vertical sets the exit dates each lever runs against.

Apparel

The trustworthy signal is full-price sell-through by style-color, pooled across doors and DTC and read against plan by delivery window, with DTC add-to-cart as the earliest read. The open levers are holdback release, reallocation between doors and between the DTC and wholesale pools, a chase against held fabric or greige, and the markdown cadence. A size curve that broke in week two is recorded as a style that slowed in week three, because the core sizes sold out while the fringe sizes kept trickling. Late deliveries, unallocated doors and floor sets that move a style off the front table hide demand the same way. Carryover styles give sensing a history; newness has only its first weeks, which is where scenario pairs matter.

Footwear

Footwear senses at the model-color and reads availability at the size run: pairs sold on the ends of a broken run describe the run, not the demand. A door counts as in stock only while the run's core sizes are present. The open levers are run repair from DC stock, at-once fill for wholesale accounts against their prebooks, and rationing a second width to doors whose fit history supports it; factory reorders on lasts and tooling land inside the season only where capacity was held. Broken size runs, missing widths and absent half sizes all record as weak demand for the model. For wholesale brands the prebook is a commitment, and account sell-through is the demand view. See assortment planning for footwear brands.

Accessories & bags

The evergreen core runs on replenishment, so sensing updates the rate its minimum is set against on leather and hardware lead times. Seasonal colorways are allocated once, so sensing reallocates them between doors and sets next collection's depth by color, reading the hero color apart from the secondary colors. Attached items follow their host — the belt follows the denim — so attach rate directs sensing to the host. An attached item whose host stocked out looks weak on its own sales, and cutting its depth on that evidence compounds the host's stockout. Colorways missing from a door hide demand, and leather goods MOQs limit how finely a sensed change can be ordered.

Home & furniture

Ocean lead times define the reach. A sensed change informs the next container's mix, or a cancel or push on a container not yet loaded — not the receipt already on the water. The trustworthy signals are written orders by model and finish, which record demand at commitment rather than delivery, and special orders, which record demand the stocked assortment missed. Container quantities round every decision, priced at landed cost. A finish that is not on the floor cannot be seen, and a stocked-out finish quoted at a long delivery date converts fewer written orders, so both understate demand. Sense at model by finish at the DC, not at door by SKU. See merchandise planning for home and furniture brands.

Outdoor

Dealer prebooks commit the model year's production before the season, so sensing works on the at-once pool, DTC, and the split between them. Trustworthy signals are DTC sell-through and dealer sell-through where dealers report; an at-once reorder is a delayed version of the same. In snow and other weather-exposed categories, a late first snow moves when demand arrives before it moves how much, so the first question is smaller season or later season; the counter-seasonal summer line draws on the same OTB and is sensed on its own calendar. MAP pricing constrains price as an in-season rebalancing lever, so reallocation between dealers and channels does the work until closeout. Dealer sell-in read as sell-through hides demand — stockroom units are not sales — and the signal lands mostly in next model year's prebook guidance. See merchandise planning for outdoor brands.

Sporting goods

Team and roster orders are committed demand with dates, so sensing concerns at-once product, the retail channel and consumables — balls, grips, tape — which replenish on min/max, with the DC-to-door lead time setting each door's reorder point and the supplier lead time setting the DC's. Season-defined categories anchor the window to the sport's season start, and a late start shifts timing. Dealer prebooks and model years constrain the rest of the line as in outdoor, with MAP constraining price so reallocation does the in-season work. A closed team order window hides demand completely: the team that missed it orders nothing, and the history shows no demand. Closeout pricing on the prior model year inflates its rate and suppresses the new model's, so the two are sensed apart. See merchandise planning for sporting goods brands.

Beauty & wellness

The sellable grain is shade by door, and the ends of the shade ladder break first. Retailer POS by door and shade is the direct consumer view where a retailer shares it; DTC orders by shade arrive fastest. Launches are read over a defined launch period before moving to core replenishment. A missing or empty tester suppresses a shade's sales as surely as an empty shelf, so testers belong in the in-stock definition. Shelf life and PAO dating cap how far a sensed increase can be acted on — a batch justifies only the cover that sells before its dating ceiling, as dating rules and weeks of supply computes. Where the retailer controls replenishment, the brand's levers are its fill to the retailer's DC and the next batch.

Toys & games

Q4 concentration puts the loudest signal after the factory lever has closed: a factory reorder placed on a peak read lands after the peak. Sensing inside Q4 redistributes domestic DC stock between retailers and DTC and shapes next year's retailer commitments; the useful in-season read is retailer POS in the pre-peak weeks, which directs the domestic holdback. Retailer commitments are sell-in, not demand. Licensed windows stop every lever at the window end minus the lead time, and safety standards mean a reworked item needs testing before it can be chased. A retailer shelf running empty while stock sits in the retailer's own DC hides demand. See planning a licensed product window.

Baby & juvenile

Registry demand records intent weeks or months before purchase, which makes registry adds a forward read for long-tail hard goods replenishment. The signal is censored too: a registrant who sees an item shown as unavailable swaps it, so availability suppresses the registry signal itself. Safety standards and recalls constrain the levers — a certified configuration cannot be chased with a substitute, and a recall is a lot-level event, so records carry the production lot. Model-year changeovers shift registrants to the new model once listed, so the outgoing model's adds are sensed apart. The open levers are supplier replenishment, allocation between the registry channel and doors, and the pace of the run-down. See planning with registry demand.

Jewelry & watches

Inventory is piece-level, with a unit or two of depth per style in a door, and a piece sells once, so a weekly rate at the piece grain is zero or one. Sense at a coarser grain — collection, metal, stone and price band, by door or region — and place pieces within it. Try-on requests, appointments and ring-size enquiries record demand a missing piece cannot. Memo and consignment make rotation between doors the main lever, with bench or made-to-order refill behind it. Metal cost moves reprice pieces at refill and can shift them into a different price band, so sensing by band keeps the read stable. Single-piece depth hides demand: one sale takes the style out of the door. See merchandise planning for jewelry and watch brands.

Where RetailNorthstar fits

RetailNorthstar is a merchandise planning platform — OTB, assortment and line planning, buy planning, allocation, sizing, PO and WIP tracking and analytics on a shared data model. Its AI-assisted planning uses in-season demand signals, read as sell-through against plan, to inform buy-depth and allocation recommendations: where committed inventory goes and how deep a chase should be. Because plan, buy, allocation and the PO and WIP record share one data model, a sensed gap is read against position, on-order and remaining open-to-buy in one place.

It is not statistical supply-chain forecasting at the scale of ToolsGroup or o9; where a brand runs deep statistical forecasting, it connects through integrations, and the merchandise plan supplies the financial envelope. The ToolsGroup comparison sets out that boundary, and allocation planning covers the decisions the sensed rate informs.