What Is Demand Sensing Software?
Demand sensing software estimates near-term demand, days to a few weeks out, from high-frequency signals such as daily POS, web traffic, open orders, weather and promotions, and feeds the adjustment into replenishment, allocation and reforecast decisions.
Demand sensing software is software that estimates near-term demand, days to a few weeks out, from high-frequency signals such as daily point-of-sale sales, web traffic, open orders, weather and promotions, and feeds the adjustment into replenishment, allocation and reforecast decisions. It works at the grain where those signals arrive, store by SKU by day, and on the horizon where they can still change something: the next replenishment order, the next allocation run and the next reforecast rather than the season's buy. It is a different system from the demand planning software that builds a statistical baseline months out, and from the merchandise planning software that sets the financial and assortment plan the adjustment lives inside.
This guide defines the category: what the software is and is not, what it ingests and at what grain, how its horizon and cadence compare with demand planning and merchandise planning software, why it grew up in consumer packaged goods and grocery and why seasonal apparel is a harder fit, what it outputs and which decision each output can still move, the data it needs, how it sits next to the merchandise plan, a worked example, evaluation criteria, where it pays and cannot help in ten verticals, and the failure modes. It is the software companion to demand sensing in retail, which covers the concept and the practice in full; the one-paragraph definition is in the glossary.
What demand sensing software is, and what it is not
The practice came first. Demand sensing is the in-season discipline of estimating the rate demand is running at now and using it to change the decisions still open; the pillar guide sets it out, and a planner with a weekly sell-through report and a trade meeting does it by hand. Demand sensing software is what that practice becomes when it runs every day, at store-SKU grain, across a fleet, by a system rather than a person: a pipeline that lands high-frequency signals, a short-horizon model that turns them into a revised rate for the next days and weeks, and a write-back that pushes the revision into replenishment, allocation and the reforecast.
Three parts, and the third decides whether the category is worth paying for. A sensed rate that stops at a dashboard is a report, and a planning team already has reports. The software earns its place when the revised rate reaches an order quantity, an allocation split or a plan line without a person re-keying it, and when the owner of that decision can see why the number moved and override it with the override recorded.
What it is not is as useful to say. It is not a forecasting suite: it corrects a baseline rather than building one, and expects the baseline to come from a demand planning system, a merchandise plan or a simple model of its own. It is not a replenishment engine, although a suite can bundle the two; the engine decides the order against a rate, and sensing supplies the rate. It is not business intelligence: BI reads what happened, sensing estimates what is about to. And it is not merchandise planning software, which decides how much money to commit, which options to carry and how deep, before most of the signals sensing reads exist. Sensing adjusts within a plan; it does not make one.
What it ingests, and at what grain
The category is defined by its inputs as much as its outputs, because the inputs set the grain and the grain sets what the software can see. The native grain is store by SKU by day: the unit a point-of-sale system writes, the unit a perpetual inventory holds, and the unit a replenishment order is placed in. Anything coarser can be rolled up from it; nothing finer can be recovered once it has been aggregated away. Web signals arrive finer still, by session and by minute, and are rolled up to the day to join the rest.
| Feed | Native grain | Typical arrival | What the model uses it for | What breaks it |
|---|---|---|---|---|
| Store point-of-sale | Transaction line by SKU by store, aggregated to the day | Daily batch; intra-day where the POS streams | The sales signal itself, by location | Late batches, returns netted inconsistently, promotions not tagged to the sale |
| E-commerce orders | Order line by SKU, timestamped | Near real time | Fleet-wide demand with no shelf in the way; the earliest read on newness | Cancellations and returns not netted; a marketing send read as a trend |
| Sessions, product views, add-to-cart | Product page by SKU or style, by session | Near real time | Interest, including interest in sizes and colors that could not convert | Traffic source mix, placement, bots |
| Open orders | Order line by SKU by requested ship date | Daily | Committed demand ahead of shipment: wholesale bookings, pre-orders, backorders | Orders the account will cut or push; the same order held in two systems |
| On-hand and in-transit | Store or DC by SKU by day; shipment by SKU by expected date | Daily | Turning a rate into weeks of supply; flagging positions where demand was censored | Count errors, unsellable units counted as sellable, ship dates that slip without an update |
| Promotions, price and marketing calendar | Event by SKU or category by store by date, with depth | Ahead of the event | Separating lift from baseline; anticipating the next event | Events not recorded; depth or dates changed after the record was made |
| Weather | Region by day, observed and forecast | Daily, days ahead | Timing of weather-sensitive categories | A timing shift read as a volume change |
| Returns | Return line by SKU by date, with reason | Days to weeks after the sale | Netting gross demand to kept demand | Lag; bracketing; reason codes nobody fills in |
| Retailer and dealer sell-through | Account by door by SKU by week, or coarser | Weekly or irregular | The consumer view for a wholesale brand | Partial coverage; a missing report read as a week of zero |
Two properties of the table matter more than any row. The first is the join. Every feed has to land against the same product master, the same location master and the same calendar, or the model is correlating a SKU in one system with a different code for the same item in another. A brand that sells through its own stores, its own site, marketplaces and wholesale accounts holds product and location identifiers in four systems, and reconciling them is a large part of the implementation work in this category; the data readiness guide covers what has to be true first.
The second is what the signals mean. POS and web orders are demand that converted. Add-to-cart and product views are demand that may not have, including demand at positions with no stock. Open orders are demand committed but not yet shipped. The on-hand position is not demand at all; it is the denominator that turns a rate into weeks of supply and the record of where sales were censored by an empty shelf. A model that treats them all as the same kind of number produces a confident rate that describes availability rather than demand, the distortion the pillar guide calls censored demand and the glossary sizes as lost sales.
Horizon and cadence: sensing against demand planning and merchandise planning software
The three systems a buyer meets in a search for demand sensing software are confused with one another because each produces a number about future demand. They differ in the question asked, the horizon, the cadence, the grain and, most of all, in what happens to the number afterwards. Demand vs merchandise vs supply planning separates the disciplines; what is merchandise planning software? and supply chain software vs merchandise planning software draw the system boundaries on the planning side.
| Demand sensing software | Demand planning and forecasting software | Merchandise planning software | |
|---|---|---|---|
| Question it answers | How fast is demand running now, and what should change this week? | How much of each existing item will sell, by location, over the coming months? | How much to invest, which options to carry, how deep by size and pack, and where to place them |
| Horizon | Days to a few weeks | Months to a year or more | A season to a year or more, re-planned in season |
| Cadence | Daily, sometimes intra-day | Weekly or monthly forecast refresh | Pre-season build; weekly or monthly reforecast |
| Native grain | Store or node by SKU by day | Item by location by week or month | Department and class by month, down to style-color and size by delivery |
| Main inputs | Daily POS, web traffic and orders, open orders, on-hand and in-transit, weather, promotions and price | Sales history, seasonality, causal factors, attributes for items with no history | Financial targets, the hierarchy and calendar, history and analogs, the line plan, vendor costs and minimums |
| Model | A short-horizon correction to a baseline, weighted toward the most recent signals | A statistical baseline from history, with promotional and causal effects layered on | A plan: top-down targets reconciled with a bottom-up build, with judgement on newness |
| Output | A revised near-term rate and the gap against baseline and position | A forecast by item, location and period | The financial plan, open-to-buy, assortment, buy, size and allocation plans |
| Decisions it feeds | Replenishment quantity, allocation and transfers, holdback release, chase and markdown timing, the near-term reforecast | Supply commitments, replenishment parameters, inventory targets, purchasing and production | The buy itself: how much money, which options, how deep, when and where |
| Heritage | Consumer packaged goods and grocery continuous replenishment | Supply chain planning and S&OP | Retail merchandising and buying |
| Weakest on | Items with no history, one-shot buys, sparse grain | Newness, short lifecycles, assortment decisions | High-frequency item-store replenishment at grocery scale |
The three are layered rather than rival. The plan is the envelope, the baseline is the expectation, and sensing is the correction: merchandise planning decides what will be owned and how much money stands behind it; demand planning or forecasting software, where a brand runs one, builds the item-location expectation inside that envelope; sensing adjusts the expectation for the next few weeks from what the floor and the site are doing today. A brand can run all three, two or one. What it cannot do is substitute one for another: a sensed rate projected to season end is not a forecast, a forecast refreshed daily is not sensing if nothing downstream moves on it, and neither decides which options to carry; that decision has its own category, covered in what is assortment planning software?.
The difference a planner feels is in the horizon and the cadence. Forecasting software is built to be stable, so that a buy sized on it in March still describes the item in July; it weights a long history and damps the last week. Sensing software is built to be responsive, so that a change on the floor reaches this week's replenishment order; it weights the last few days and discounts the history. Stability and responsiveness are a trade, not a scale, and a product that claims both for the same number is describing two numbers.
Where it comes from: CPG and grocery, and why fashion is a harder fit
Demand sensing as a software category was built for consumer packaged goods manufacturers and grocery retailers, and the reason is structural. Four conditions hold in those businesses, and each is what a short-horizon model at store-SKU-day grain needs.
The item lives for years. A detergent or a yoghurt sits in the same store at the same shelf position through many seasons, so there is a long item-store history to learn a baseline from, and the baseline is worth learning because it will still be the right item next year. The signal is dense. The item sells units every day in every store, so a day's POS at one store is a meaningful observation and a week across a region is a very good one. The demand shock is scheduled and repeated. The main thing that moves demand off baseline is a promotion, and promotions are planned ahead, recorded with dates and depth, and run again, so their lift can be measured and anticipated. And there is always an order waiting. Replenishment is continuous and lead times are days, so a revised rate on Monday changes a truck on Wednesday; the adjustment has a decision to land in every single week.
Fashion and seasonal apparel reverse all four, and the reversal is mechanical rather than a matter of which product is chosen. Short lifecycles mean a seasonal style sells for a few weeks and is gone; by the time it has enough history to estimate a baseline, the window to act on it has mostly closed. Newness without history means the styles a season introduces have nothing of their own to learn from, so the model falls back on an analog or an attribute cohort, which is a pre-season forecasting technique rather than sensing, as demand forecasting for fashion covers. Size-level sparsity means that at the grain the customer buys, one size of one color in one store on one day, the observation is almost always zero or one, so the rate has to be estimated at style-color across the fleet and pushed down through the size curve, a coarser grain than the category was built for. One-shot buys mean there is often no order waiting: the quantity was committed to a factory months earlier, so the sensed rate can change how the committed units are distributed and timed, and whether a held chase is released, but not how many were bought.
Two further frictions follow. Apparel is returned in a way groceries are not, and the return lands weeks after the sale, so gross daily demand overstates kept demand until the returns come back. And a good part of in-season apparel demand is created by price, through the markdown cadence, so the rate the model senses after a markdown is partly its own output.
None of this makes the category useless to an apparel brand. It moves where the category pays. The parts of an apparel, footwear, accessories or beauty business that look like grocery, a replenished core, NOOS basics, a core shade range, a core size run on a carryover model, have the long life, the dense signal and the waiting order the model needs, and for them a daily sensed rate feeding the replenishment trigger is the right tool. The seasonal part of the same business is better served by the practice the pillar guide describes, run at the grain the evidence supports and pointed at the levers still open. The vertical section below draws that line for ten categories.
What it outputs, and which decision each output can still move
The output of demand sensing software is a set of revised numbers about the next few weeks, and each is only worth producing if a decision can still move on it. The pillar guide sets out each lever's lead time and the week it closes; the table maps the software's outputs onto those levers.
| Output | The decision it can move | Lead time the lever consumes | The output arrives too late when |
|---|---|---|---|
| A revised near-term rate by SKU and location | Replenishment quantity and trigger on continuing items | Supplier or DC lead time | The item's exit date sits inside the lead time |
| An exception: the gap between sensed rate, baseline and position | Reallocation between doors and channels, transfers, holdback release | Days | The holdback is empty, or the remaining weeks will not repay the move |
| A projected stock-out date, or forward weeks of supply | Chase against held material or capacity; cancel or push on open orders | Production plus transit | The chase's last responsible week has passed, or material is cut |
| A promotion lift estimate | The buy and allocation for the next event, and the return to baseline after it | The event's own lead time | The event is already allocated |
| A sustained deviation flag | Markdown timing and depth | One price-change cycle | The season exit date, the last lever to close |
| A near-term reforecast adjustment | The weekly or monthly reforecast of the plan | The reforecast cadence | Never, as long as it actually reaches the plan |
Three things about the table decide whether a product is doing its job. First, the write-back is the product. An exception that appears on a screen and waits for a planner to re-key it into an allocation tool has the latency of the planner's week, not the signal's day, and the category's whole case rests on closing that gap. Ask which rows a product writes to directly, and where the decision record lives.
Second, the rate is only half of each row. The replenishment quantity needs the rate, the position and the lead time; the chase needs the rate, the on-order and the remaining open-to-buy; the markdown needs the rate, the exit date and the remaining units. A sensed rate presented without the committed position next to it is a number nobody can act on, which is why forward weeks of supply, on-hand plus on-order against the sensed rate, is the output a planner can act on directly.
Third, every row has a last week. Sensing a signal that cannot change a decision in time is the characteristic waste of the category, and it is easy to commit because the signal stays interesting after the lever has closed. A perfect read of a seasonal style's demand in its final fortnight changes the markdown date and nothing else; a perfect read of a toy's demand in the second week of December changes next year's commitment and, this year, only where the last domestic cartons go. The in-season chase guide covers the last responsible week for each lever.
The data a brand needs before it is useful, and what breaks it
Demand sensing software amplifies the quality of its inputs, because it weights the most recent days, where errors have had no time to average out. The pillar guide covers the planning prerequisites; this is what the software specifically needs in order to run.
- Daily POS at store-SKU grain, with the sale tagged to its price and any promotion. Weekly totals are a forecasting input, not a sensing input, and a sale recorded without its promotional context teaches the model that the promotion was baseline.
- A perpetual inventory that is right at SKU-store level, including in-transit. The position is the denominator of every weeks-of-supply calculation and the only way the model can tell an empty shelf from a slow seller. The inventory accuracy formula is the measure; a position wrong by a carton is a rate wrong by a week.
- One product master and one location master across every channel, so that the SKU sold on the site, the SKU rung in the store and the SKU in the wholesale account's report are the same code. In apparel that is the size and color code as well as the style; in footwear, size within width; in beauty, the shade.
- A promotion and price calendar recorded ahead of time, with dates and depth, including the retailer's events for a wholesale brand, so that lift is separated from baseline before the event rather than explained after it.
- A defined in-stock rule. Sellable units only, at the positions planned to carry the item, measured over the day rather than at a snapshot. Without it in-stock rate cannot be computed and censored demand is invisible.
- Returns netted on a lag the model knows about, so that a style with a high return rate is sensed on kept demand rather than gross.
What breaks it is the mirror image, and most of it is organisational rather than technical. POS that arrives in a weekly batch turns a daily model into a weekly one with a daily cost. A SKU re-coded mid-season splits one item's history into two short ones. A promotion run without a record teaches a permanent lift. A stockout nobody flagged is read as a slow week. A retailer report that fails to land is read as a week of zero unless the pipeline distinguishes missing from nothing. A hierarchy restructured in season breaks every aggregation the model pools on. And a sensed rate that lands where no decision is taken, because allocation still lives in a spreadsheet, is a pipeline with no outlet. Budget the data work as the larger part of the project; the model is the smaller part.
How it sits next to the merchandise plan: the envelope and the adjustment
The relationship between demand sensing software and the merchandise plan is easy to misdescribe in a sales conversation. The plan is the envelope and sensing adjusts within it. The merchandise financial plan sets the sales, margin and inventory the business intends to carry by month; the open-to-buy converts it into receipts the business can afford; the assortment and the buy decide which options and how deep; allocation decides where they go. All of that is decided before most of the signals sensing reads exist, and none of it is changed by a sensed rate directly.
What sensing changes is inside the envelope. On continuing items, it revises the rate the replenishment trigger is set against, and the resulting order spends open-to-buy the plan set aside for replenishment. On seasonal items, it redistributes and re-times inventory already committed, releases a holdback, moves units between doors and channels, and decides whether a held chase is placed and how deep, spending open-to-buy the plan held back for exactly that. It informs when markdowns start and how deep they go, against the planned reductions line the plan already carries. The pillar guide's worked example makes the point in units: sensing redistributed, pre-committed a bounded chase and left price alone, and did not change the buy.
The two systems exchange numbers in both directions. Downward, the plan supplies the baseline the model corrects, the on-order it reads position against and the remaining open-to-buy that bounds any chase; a sensed increase with no open-to-buy behind it is a recommendation nobody can approve. Upward, a sensed shift that persists, confirmed on a second read and visible in more than one signal, goes into the reforecast on the plan's cadence, where it changes the month's sales and receipt lines and resets open-to-buy for the periods still open. How often to reforecast covers that cadence, and the distinction matters: the sensed rate moves the week's decisions; the reforecast moves the plan of record. A product that writes a daily sensed rate straight into the financial plan has removed the step where a planner decides whether the shift is real.
An illustrative worked example: a sensed rate reaches the reforecast
The figures below are illustrative and chosen to divide cleanly. They are not benchmarks or targets, and they are not drawn from any brand; the method is the point.
A replenished core item, a men's black crew-neck tee, sells through 40 doors, read as one pool, with sizes pooled across the fleet because no single door-size position carries a readable daily rate. The plan rate is 300 units a week, flat. The supplier lead time is six weeks, and the standing rule is to order up to eight weeks of cover: six weeks of lead time, one week between order reviews and one week of safety stock, all at the rate in force. At the plan rate the rule would not trigger an order until the week-7 review. At the start of week 5 the position is 2,100 units on hand, 1,400 in doors and 700 in the DC, with 600 units on order landing at the start of week 7, so the inventory position including on-order is 2,700.
Four weeks of signals have landed:
| Week | Plan (units) | Actual sales | Doors in stock on core sizes, of 40 | In-stock rate | Availability-adjusted demand |
|---|---|---|---|---|---|
| 1 | 300 | 300 | 40 | 100% | 300 |
| 2 | 300 | 310 | 40 | 100% | 310 |
| 3 | 300 | 350 | 40 | 100% | 350 |
| 4 | 300 | 380 | 38 | 95% | 400 |
Step one, adjust for availability. In week 4, two doors were out of stock on core sizes all week, so the 380 units sold came from 38 doors: 380 ÷ 0.95 = 400 units of demand, on the first approximation that the two empty doors would have sold at the fleet rate. Weeks 1 to 3 were fully in stock and need no adjustment.
Step two, sense the rate. The rule, written before the season, is the mean of the two most recent availability-adjusted weeks: (350 + 400) ÷ 2 = 375 units a week. Two numbers bracket it: the four-week average, (300 + 310 + 350 + 400) ÷ 4 = 1,360 ÷ 4 = 340, and the plan, 300. The sensed rate is 25% above plan and the four-week average a little over 13% above it; the shift has grown for three weeks running, which passes the pillar guide's persistence test for a shift rather than noise; the second read in week 5 confirms or reverses it.
Step three, revise the weeks of supply. On hand only: 2,100 ÷ 300 = 7.0 weeks at plan, and 2,100 ÷ 375 = 5.6 weeks at the sensed rate. Including the 600 on order: 2,700 ÷ 300 = 9.0 weeks at plan, and 2,700 ÷ 375 = 7.2 weeks sensed.
| Line | At the plan rate (300 a week) | At the sensed rate (375 a week) |
|---|---|---|
| Weeks of supply, on hand (2,100) | 7.0 | 5.6 |
| Forward weeks of supply, on hand plus on order (2,700) | 9.0 | 7.2 |
| Order-up-to level, eight weeks of cover | 8 × 300 = 2,400 | 8 × 375 = 3,000 |
| Order needed now (order-up-to less the 2,700 position) | None; 300 over | 300 units |
| Position when a week-5 order lands, start of week 11 | 2,700 − 1,800 = 900 | 2,700 − 2,250 = 450 |
| Position if the order waits for the week-7 review, landing week 13 | 2,700 − 2,400 = 300 | 2,700 − 3,000 = −300 |
Step four, attach the decision. At the plan rate there is nothing to do: the position covers 9.0 weeks against a six-week lead time, the order-up-to level of 2,400 is already exceeded by 300 units, and the week-7 review would land an order at the start of week 13 with 300 units still on hand. At the sensed rate the same schedule fails: 2,700 units last 7.2 weeks, so the item runs out early in week 12, a week before the week-7 order can land, and the position at the start of week 13 would be 300 units short. The decision is to bring the replenishment order forward from week 7 to week 5 and size it to the sensed rate: an order-up-to level of 8 × 375 = 3,000, less the 2,700 position, is a 300-unit order this week, landing at the start of week 11 when the position would be 2,700 − (6 × 375) = 450 units, or 1.2 weeks of cover at the sensed rate, tight but not out. The second read in week 5 either confirms the rate, and the week-7 review sizes its own order at 375 too, or it does not, and 300 units of a core black tee is the small cost of being early on a replenished basic.
Step five, carry it to the reforecast. The sensed rate has moved the order; it has not yet moved the plan. The reforecast rule is that the forward rate for the cover window, weeks 5 to 12, is revised to 375 and reviewed on the monthly cadence, while the season total is held at the plan rate until the second read confirms. If it confirms, the month's sales line rises, the replenishment receipts line rises with it, and the open-to-buy for the periods still open is reset; the remaining question for the plan is whether the extra receipts are funded from replenishment open-to-buy or from elsewhere in the class. The sensed rate moved the week's decision; the reforecast moves the plan of record, and the software's job is to hand one to the other without re-keying.
Evaluation criteria and questions to ask
The pillar guide gives ten vendor-neutral questions about the sensing method itself: which decisions the rate writes back to, the grain, censoring, the in-stock definition, latency, on-order visibility, promotions and returns, override, backtesting and newness. They apply to every product in this category and are not repeated here. The questions below are about the software as a purchase, and sit alongside questions to ask planning software vendors, the selection guide and how to evaluate the implementation model.
- Which architecture is this? A standalone sensing engine that needs a baseline and a replenishment system on either side; a module inside a demand planning or supply chain suite, sharing its data model; or a sensing capability inside a planning platform, reading the plan as its baseline. Each is right for some brands, and each puts the integration work in a different place.
- Where does the baseline come from, and where does the correction go? If the answer to either is a file, the write-back the category depends on does not exist yet.
- What does it do with the parts of the assortment that look like fashion? Ask to see a new style in its second week, a size run with two sizes missing, and a style three weeks from exit. The honest answer for a grocery-shaped product is that it pools, falls back on an analog, or stops; a confident daily rate for all three is not reading the evidence.
- How much of the project is data engineering, and who owns it after go-live? The feeds in the ingestion table are the project. Ask who builds the joins to the product and location masters, who maintains them when a channel changes its export, and what it costs when a feed breaks on a Monday.
- Does it read the plan's constraints? On-order, remaining open-to-buy, the exit date, the planned markdown. A recommendation produced without them has to be checked against them by hand, which puts the latency back.
- Is the horizon honest? Ask what the product shows at week 12 of a 12-week style, and whether it projects a sensed rate to season end. The right answer is a short horizon and a handoff to the reforecast.
- How is it backtested on history as it existed on each date? Restated history, with stockouts cleaned and late feeds filled, makes any model look better than it would have been live.
- What is the cost of being wrong on the lever it writes to? A sensed rate that moves a transfer is cheap to get wrong; one that moves a chase or a markdown is not. Ask how the evidence bar scales with the reversibility of the lever, and whether a planner sets it.
Start the evaluation with the second question. A product that cannot name the system its correction lands in, and the decision record it leaves there, is a dashboard, however good the model behind it. The build vs. buy guide applies here with unusual force, because a brand with a clean POS feed and a replenishment rule can build a serviceable sensed rate in a spreadsheet, and the question is whether it wants to own the pipeline for a decade.
Where sensing pays and where it cannot help, by vertical
The mechanics are the same in every category. What changes is whether the four CPG conditions hold for the part of the assortment in question: a long-lived item, a dense signal, a scheduled shock and an order waiting to act. Where they hold, demand sensing software pays; where they do not, the practice still applies but the software's native grain and cadence buy little. The demand sensing in retail guide covers the trustworthy signal, the open lever and what hides demand for each vertical; this section is only about where the software earns its keep.
Apparel
It pays on the replenished core: NOOS basics, carryover styles and the essentials programme, where a style-color lives across seasons, sells daily across the fleet and has a replenishment order waiting at the supplier's lead time. Sensed daily at style-color across doors and pushed down through the size curve, it keeps the core sizes in stock without a planner reading forty door reports. It cannot help a fashion drop: the style was bought once against held fabric or not at all, sells for a few weeks, and its decisions are the holdback release, the reallocation between doors and DTC, the held chase and the markdown date, which the weekly trade meeting reads at style-color against plan. Sensing there is a practice, not a product.
Footwear
It pays on core runs by size on carryover models: the model-colors that continue across model years, where pairs by size within width sell every week and a run repair from DC stock or a factory reorder on held capacity is always possible. The software's grain is the size run, and its in-stock rule has to count a door as stocked only while the core sizes are present. It cannot help the model-year changeover or the dealer prebook: the production run was sized from the prebook months earlier on lasts and tooling that will not change, so a sensed rate on a new model moves the at-once pool and next year's prebook guidance, and nothing at the factory this season unless capacity was held. Width fits and half sizes make the grain sparser than the category expects, so pooling across doors is the default.
Accessories and bags
It pays on the evergreen core, where the leather article, the hardware finish and the colorway continue for years and replenish on a rule against the tannery's lead time; sensed daily, it resets the minimum the rule is held against. It cannot help the seasonal colorway, allocated once and exited, where the decision is reallocation between doors and next collection's depth by color. The attach relationship is the trap: an attached item whose host stocked out senses weak on its own sales, so the software has to read the host's position, and a product built for grocery has no reason to.
Home and furniture
It pays on replenishable SKUs: core bedding and textiles, tabletop, lighting and decor in continuing finishes, and stocked upholstery in core fabrics, where the signal is written orders and web orders by model and finish and the order waiting is the next container. The grain is model by finish at the DC, not door by SKU, and the lead time is the ocean. It cannot help the container already on the water, nor special-order and made-to-order pieces, where demand is recorded at commitment and nothing is stocked to adjust. Container minimums round every recommendation, so a sensed increase becomes a mix change on the next container rather than a quantity change this week, and dealer prebooks net against the buy before the software sees them. The floor set, which decides which finishes are seen at all, stays a planning decision.
Outdoor
It pays on the at-once pool and the DTC channel for carryover models, and on consumables and accessories that replenish year round, read against dealer sell-through where dealers report. It cannot help the model-year production run, committed from the dealer prebook before the season, so a sensed rate on a new model moves the split between at-once and DTC and next year's prebook guidance. Weather moves when demand arrives before it moves how much, so a product that reads a warm November as a smaller season rather than a later one cuts the at-once pool exactly when it is about to be needed; MAP pricing constrains price as an in-season lever until closeout, so the write-back is to reallocation, not to markdown.
Sporting goods
It pays in season, by sport: consumables such as balls, grips and tape on min/max replenishment, and the at-once retail product in a sport's window, where the DC-to-door lead time is short and the season start anchors demand. The software has to carry seasonality by sport, because a store runs several seasons at once and a single calendar reads one sport's peak as another's noise. It cannot help the team and institutional prebook, which is committed demand with dates, or the closed team order window, which hides demand completely. Closeout pricing on the prior model year inflates its rate and suppresses the new model's, so the two are sensed apart or not at all.
Beauty and wellness
It pays on the replenishment of core shades. A foundation or lipstick range is bought and refilled as a set, each shade sells daily, the retailer's replenishment runs on a short lead time, and the ends of the shade ladder break first; sensed at shade by door, with testers inside the in-stock rule, the software keeps the range whole, which is the whole job in beauty. Shelf-life dating caps how far a sensed increase can be acted on: a batch justifies only the cover that can sell inside its unopened shelf life and the retailer's date-code requirement at delivery, so the write-back has to read batch dating from the batch system; PAO is the customer's period after opening and does not set the ceiling. It cannot help a launch in its launch period, which is read as an event, or a gondola reset, which decides when the range can change on shelf and is set by the retailer's calendar rather than by demand.
Toys and games
It pays on the evergreen catalogue, the games and classic lines that sell year round and replenish on a domestic lead time, and on the domestic holdback in the pre-peak weeks, where retailer POS directs which account gets the last cartons. It cannot help the gifting peak that was bought once: the factory order for the fourth-quarter peak closed months before the loudest signal arrives, so a sensed rate in December moves nothing at the factory this year, only the last cartons in the domestic DC and next year's commitment. Licensed windows close every lever at the window end minus the lead time, and retailer commitments are sell-in, which the software must not read as demand.
Baby and juvenile
It pays on the long-tail hard goods and consumables that replenish on a steady rate as the child grows, and it has a forward signal of its own: registry adds, which record intent weeks ahead of purchase and can feed replenishment before the sale. The software has to carry the age or stage band as the planning grain, and it has to sense the outgoing model apart from the new one at a model-year changeover. It cannot help a recall, which removes a lot from sale in a day and is an inventory event rather than a demand one, and a registrant who sees an item as unavailable swaps it, so the registry signal is censored by availability exactly as sales are.
Jewelry and watches
It pays least here, and honestly so. Depth is one or two pieces per style in a door, a piece sells once, and a daily rate at the piece grain is zero or one; the useful read is at collection, metal, stone and price band, by door or region, which is a planning roll-up rather than a store-SKU-day signal. The collection calendar and the precious-metal cost base, not the daily rate, decide the economics: metal moves reprice pieces at refill and can shift them into a different price band, and memo and consignment make rotation between doors the main lever, with bench or made-to-order refill behind it. Try-on requests and ring-size enquiries record demand a missing piece cannot, and they are the signal worth sensing, by hand.
The pattern across the ten is the one the CPG conditions predict: the software pays on whatever part of the assortment is long-lived, dense, replenished and still open to an order, and buys little on whatever was bought once, sells once or closed before the signal arrived.
Failure modes
The pillar guide's four failure modes of the practice, reacting to noise, sensing at a grain finer than the evidence, sensing with no decision attached and projecting the launch while ignoring the on-order, all apply to the software, which makes two of them easier to commit because it produces a confident number every day. Three more belong to the software specifically.
Sensing a signal that cannot change a decision in time. A daily pipeline does not know which levers are still open, so it will produce a precise read on a seasonal style in its last two weeks, on a container that sailed last month and on a toy in the second week of December, and each costs attention that belongs on a decision that can still move. The fix is in the configuration: every item carries its exit date and the lead time of each lever, and the software stops producing recommendations for levers that have closed and says so.
Chasing noise. A model weighted toward the last few days is built to respond, and responsiveness is indistinguishable from twitchiness without an evidence bar. A marketing send, a local event, a receipt that landed mid-week or a competitor's stockout moves a day's number without moving demand, and a product that acts on single reads shuttles units between doors and raises and lowers replenishment orders with every one. The fix is the pillar guide's three tests, persistence into a second read, agreement across more than one signal and the same direction across clusters, scaled to the reversibility of the lever: one read for a transfer, more for a chase or a markdown.
Over-fitting one promotion. The CPG heritage shows here. A model that learned promotional lift from items that run the same event every few weeks will fit an apparel brand's single friends-and-family weekend as a repeating pattern, carry the lift forward as baseline, and over-order against demand that was pulled forward rather than created. The post-event dip is then read as a slowdown and replenishment is cut, exactly when the baseline is returning. The fix is to record the event with its depth and dates ahead of time, to read the weeks after it apart from the weeks before, and to require a second event before any lift is treated as learned.
Underneath all three is the failure the category cannot fix for a brand: a pipeline without an outlet. A sensed rate that lands where allocation still lives in a spreadsheet and the reforecast is rebuilt by hand has the latency of the spreadsheet, and the brand has bought a faster number for a process that cannot use it. The replace spreadsheets page covers that process problem; it comes first.
Where RetailNorthstar fits
RetailNorthstar is not demand sensing software, and this guide is not a case for buying it as one. RetailNorthstar is a merchandise planning platform: open-to-buy, assortment and line planning, buy planning, allocation, sizing, PO and WIP tracking and analytics on a shared data model. As the demand sensing in retail guide states, 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, which is the envelope-and-adjustment relationship this guide describes.
It is not statistical supply-chain forecasting or store-SKU-day demand sensing at the scale of a dedicated engine; where a brand runs one, it connects through integrations, and the merchandise plan supplies the financial envelope. Allocation planning covers the decisions a sensed rate informs, and the evaluation questions above apply to RetailNorthstar as much as to any product a brand compares it with.
Related resources
- What Is Markdown Optimization Software? — the software category: the decisions it owns, how the price-response model and the optimizer work, and a worked example of two markdown paths
- Demand Sensing in Retail — the concept and the practice in full, with the signals, the horizon, a worked example and ten verticals
- Demand Sensing, Demand Forecasting, Demand Planning and In-Stock Rate — the terms
- Demand vs Merchandise vs Supply Planning, What Is Merchandise Planning Software?, What Is Assortment Planning Software? and Supply Chain Software vs Merchandise Planning Software — the neighbouring categories
- Demand Forecasting by Vertical and Demand Forecasting for Fashion — the pre-season half of the split
- The In-Season Chase, How Often to Reforecast a Merchandise Plan and Core and NOOS Replenishment Programs — the decisions a sensed rate feeds
- Weeks of Supply and Forward Weeks of Supply — the arithmetic in the worked example
- Questions to Ask Planning Software Vendors — the wider evaluation
See how RetailNorthstar reads in-season sell-through against plan, on-order and open-to-buy to inform buy depth and allocation.
Start Free Trial →No credit card. No commitments.
Common questions
What is demand sensing software?
Demand sensing software is software that estimates near-term demand, days to a few weeks out, from high-frequency signals such as daily point-of-sale sales, web traffic, open orders, weather and promotions, and feeds the adjustment into replenishment, allocation and reforecast decisions. It works at store-SKU-day grain, corrects a baseline rather than building one, and earns its place only when the revised rate reaches an order quantity, an allocation or a plan line without being re-keyed. It is a different system from demand planning software, which builds the statistical baseline months out, and from merchandise planning software, which sets the financial and assortment plan the adjustment lives inside.
How is demand sensing software different from demand planning or forecasting software?
They differ in horizon, cadence and purpose. Demand planning and forecasting software estimates how much of each existing item will sell by location over the coming months, from a long sales history with seasonality and causal effects, and is built to be stable so that a buy sized on it still describes the item months later. Demand sensing software estimates how fast demand is running over the next days and weeks, weights the most recent signals and is built to be responsive, so that a change on the floor reaches this week's replenishment order. Stability and responsiveness are a trade, not a scale: a sensed rate projected to season end is not a forecast, and a forecast refreshed daily is not sensing if nothing downstream moves on it.
How is demand sensing software different from merchandise planning software?
Merchandise planning software decides what the business will own: how much money to commit by month, which options to carry, how deep by size and pack, and where to place them, and it does so before most of the signals sensing reads exist. Demand sensing software adjusts within that plan. On continuing items it revises the rate the replenishment trigger is set against; on seasonal items it redistributes and re-times inventory already committed, releases holdbacks, sizes a held chase and informs markdown timing. The plan is the envelope and sensing adjusts within it; a sensed increase with no open-to-buy behind it is a recommendation nobody can approve, and a sensed shift that persists reaches the plan through the reforecast rather than by writing straight into it.
Why does demand sensing software come from CPG and grocery, and does it work for fashion?
Because four conditions hold there: the item lives for years at the same store, so there is a long item-store history to learn a baseline from; it sells units every day in every store, so a day of point-of-sale data is a meaningful observation; the main demand shock is a scheduled, repeated promotion whose lift can be measured; and replenishment is continuous, so every revised rate has an order waiting to act on it. Fashion reverses each: short lifecycles, newness without history, size-level sparsity and one-shot buys remove the history to learn from and the order waiting to act. The category still pays in apparel, footwear, accessories and beauty wherever the assortment looks like grocery: NOOS basics, carryover styles, core size runs and core shades on replenishment. For fashion drops the practice applies, run at style-color across the fleet and pointed at the levers still open, but the software's store-SKU-day grain buys little.
What data does a brand need before demand sensing software is useful?
Daily point-of-sale sales at store-SKU grain with each sale tagged to its price and any promotion; a perpetual inventory that is right at SKU-store level including in-transit; one product master and one location master across every channel, so the same item carries the same code in store, online and in wholesale reports; a promotion and price calendar recorded ahead of time with dates and depth; a defined in-stock rule covering sellable units at planned positions; and returns netted on a known lag. What breaks it is mostly organisational: weekly POS batches, SKUs re-coded mid-season, promotions run without a record, stockouts nobody flagged, retailer reports that fail to land and are read as zero, and a hierarchy restructured in season. Budget the data work as the larger part of the project.
Which decisions can demand sensing software still change once the season has started?
Only the ones whose lead time has not passed. A revised near-term rate moves the replenishment quantity and trigger on continuing items; an exception against baseline and position moves reallocation, transfers and holdback release within days; a projected stock-out date or forward weeks of supply moves a chase against held material, or a cancel or push on open orders, until the last responsible week; a sustained deviation moves markdown timing and depth until the season exit; and a near-term reforecast adjustment moves the plan of record on the reforecast cadence. The pre-season buy, once production is committed, is outside the horizon. Sensing a signal that cannot change a decision in time is the characteristic waste of the category.
Is RetailNorthstar demand sensing software?
No. RetailNorthstar is a merchandise planning platform: open-to-buy, 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, and 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 or store-SKU-day demand sensing at the scale of a dedicated engine; where a brand runs one, it connects through integrations, and the merchandise plan supplies the financial envelope.
Share this guide with your team
Copy a link or a pre-written message for Slack, Teams, or email.
// Know where your operation stands
Apply this to your planning operation.
The free Apparel Planning Maturity Assessment benchmarks your operation and tells you exactly which gaps to fix first.