Demand Forecasting by Vertical

Demand forecasting is the practice of estimating how much will sell over a defined selling period — a season, a model year, a licensed window — at a stated grain, before the quantity is committed, using comparable sales history, an analog for anything with no history of its own, the planned price and the known changes to assortment, doors and channel. It differs from demand sensing, which works inside the season on a horizon of weeks and changes only the decisions still open, and from demand planning, which turns an agreed view of demand into committed item-by-location supply. Forecasting sizes the buy; sensing redistributes and times what the buy already committed. This guide covers the pre-season half of that split, cut by vertical.

It belongs to the by-vertical series alongside the merchandise hierarchy, which defines the level each category plans at; the planning calendar, which dates the commitment this forecast has to beat; open-to-buy by vertical, which holds the budget; and assortment planning by vertical, which spends it. Demand forecasting for fashion is the single-vertical treatment this guide generalizes.

What demand forecasting is

Three things separate a forecast from a sales projection pasted into a plan. It states a grain and a period, so it can be scored later against something specific. It carries its assumptions in the open — the analog used, the price assumed, the door count and channel mix planned, the delivery dates taken as given — because a forecast that misses is only useful if the wrong assumption can be identified. And it comes with a range, not a point: a base case with an upside and a downside, sized so the buy can be split between a committed quantity and a quantity held for a later decision. Scenario planning covers how those pairs are built.

The estimate is judged at one date: the date the quantity was locked. Everything learned after that belongs to sensing and to the season hindsight, not to the forecast's score.

Forecasting, sensing and planning: where this guide starts

Demand vs merchandise vs supply planning sets out the three disciplines in full, and the demand sensing guide carries the question-by-question comparison, including the inputs each one takes and the decisions each one feeds. Three rows of that comparison scope this page.

Everything below sits inside the first and third rows: a horizon that opens on the commitment date, and a score taken on the vintage that bought the goods.

The three inputs that change per vertical

1. What history is comparable

Comparability comes in three strengths, and the strongest available one should be used. The same item with its own history — a carryover model, an evergreen core colorway, a replenished consumable — supports a velocity forecast and needs only an adjustment for price, doors and channel. The same item under a different name — a new colorway on a carryover body, a new finish on an existing frame, a new shade inside an established franchise — inherits the parent's history, adjusted for the colorway or finish itself. An attribute neighbourhood — the same silhouette and fabrication in the same price band, the same format at the same price point under a different licensed property — is the weakest, and it is what newness falls back on when nothing closer exists.

Two tests decide whether an analog earns its place. It has to be comparable in price band, channel and door count, because a style that sold at one price through one channel is not evidence about another, and a forecast built by averaging analogs across price bands quietly forecasts a product that was never sold. And it has to be comparable in availability: an analog that stocked out in week four is an analog of a censored season, and carrying its recorded sales forward builds that stockout into the new plan as if it were demand. The hindsight analysis is where analogs earn or lose their standing for next season.

2. What the lead time leaves open

The forecast's real horizon runs from the date the quantity locks to the end of the selling window, not from the season start. That distance decides how much of the plan is evidence and how much is a bet. Where a chase against reserved material can still land and sell, a forecast can be smaller and the rest held as a decision to make later — the shape the in-season chase describes. Where the quantity closes at a dealer prebook, a factory cut-off or a licensed window, the whole quantity is a bet placed before any consumer read exists, and the only flexibility left is one bought before the season: an at-once layer, a domestic holdback, reserved fabric or capacity.

This is why the same forecast error costs differently by category. A forecast horizon that ends before the first consumer read is a forecast that cannot be corrected, so the useful work moves to the range rather than the point: what the downside costs in markdown, what the upside costs in lost sales, and which of the two the business can carry.

3. What censors the signal, and how the history is cleaned

An analog can arrive already damaged, because sales history records what could be bought rather than what was wanted. An empty position contributes a zero, a run missing its core sizes contributes a low number taken from the sizes left behind, a door that never received the colorway contributes nothing, and a delivery that landed late contributes weeks of empty floor. The pre-season consequence is cumulative: each of those reads as low demand, and a quantity built on them is committed short before the season opens. In-stock rate shows where history was censored and lost sales sizes it. The sensing guide covers how censoring distorts a live in-season read; the pre-season question is different and narrower. History is cleaned before it is used as an analog, or last season's stockout becomes this season's plan.

Cleaning is an operation with steps, not a caveat attached to a number. Mark the weeks and positions where the item was not fully available, using the in-stock definition the category actually sells on rather than a generic one — core sizes present for a run, a working tester beside a shade, the finish standing on the floor. Take the rate from the clean weeks alone instead of averaging the censored ones in, then rebuild the period total from that rate on the season's own phasing. Where a position never existed at all — a door that never received the colorway, a shade never shipped, a finish never displayed — the gap is filled from the sibling positions that did sell, because the item's own zero is a record of the assortment decision rather than of demand.

Three disciplines keep the result honest. Every adjustment carries its evidence: which weeks were dropped, what stood in for them, and how far the total moved. The adjusted history and the recorded history are both kept, so the next season can see how much of an analog was reconstructed. And the size of the adjustment is read as a signal in its own right, since an analog that needed heavy reconstruction deserves a wider range around it, not a confident point. The hindsight is where each of those adjustments is finally tested against what the season did.

Apparel: style-color for the season, pushed down through the size curve

Apparel forecasts the style-color for a delivery window inside SS or FW, then distributes it through the size curve and a door or cluster index. Carryover and core programs carry their own history and adjust for price, doors and channel. Seasonal newness has none, so the forecast comes from an attribute analog — silhouette, fabrication, fit and price band — and the range around it is wider by design. Newness and carryover are forecast on different evidence and should not share one accuracy expectation.

The size curve is a second forecast, not a rounding step. A style total that is right with a curve that is wrong sells the middle of the run out early and marks the ends down. An apparel analog is therefore cleaned twice: the style total is rebuilt from the weeks the full run stood on the floor, and the curve is refitted on the doors that held every size, because a curve measured across broken doors describes what was left rather than what was wanted. Promotional weeks and the doors a style never reached are excluded from both before either is copied forward. Fabric booked at a mill minimum locks the quantity before the style-color exists, which is the constraint planning against fabric minimums sets out; where colorways are left unassigned, part of the forecast can be deferred to a chase. Extended sizes and kidswear age bands each carry their own curve and are cleaned separately, because a curve borrowed across bands is not evidence about either.

Footwear: the last and the model, distributed across the size run

Footwear forecasts the model-color and distributes it across the size run and, where the brand offers them, widths. The analog unit is the last and the model rather than the colorway: a carryover model on an existing last brings a real history, and a new colorway on that last inherits it with a colorway adjustment, while a new last is newness whatever the upper looks like. Tooling and lasts commit early and at model level, so the quantity conversation starts before colorway decisions are finished.

For wholesale, the prebook converts the brand's forecast into an account commitment, which means the brand is forecasting the account's forecast plus its own at-once and DTC demand. Run completeness is the filter a footwear analog passes through before it becomes evidence. A door-week counts only while the run's core sizes were on the shelf; the model's rate comes from those weeks, and the run curve is refitted on the doors that carried the full run, half sizes and second width included. Weeks that fail the test are dropped rather than averaged in, and the count of dropped weeks travels with the forecast, so the number of pairs that came from reconstruction rather than from sales is visible at the prebook conversation. Half sizes and secondary widths thin the volume per position quickly, which is why the model-color stays the forecast grain and the run stays the distribution. See assortment planning for footwear brands and size and pack optimization.

Accessories & bags: long-lived carryover and demand that follows a host

Accessories carry fewer sizes and longer lives, so history per style-colorway is denser than in apparel and the evergreen core — black leather, house hardware, stable velocity — supports a velocity forecast rather than a season forecast. The work concentrates in two places. Seasonal colorways are bought once against a collection date, with the hero color forecast apart from the secondary colors grouped beside it. And attached items are forecast from their host: a belt that sells with a denim program is forecast through the attach rate on the host's plan, not from its own thin history.

An attached item is cleaned through its host rather than on its own sales. The attach rate is computed only across the weeks the host was in stock, then carried onto the host's new plan; the weeks the host stood empty say nothing about the belt and only drag the rate down when they are averaged in. Where the same style also sells standalone, the two streams are separated before either is used, since a blended rate moves every time the host's plan does. Tannery and hardware minimums round the answer, so a forecast that lands between two minimums is a decision about which one to take rather than an arithmetic result — the point planning accessories lines develops. Small leather goods and cold-weather accessories run on a gifting calendar that concentrates demand into weeks, so the annual total is the wrong unit to forecast in.

Home & furniture: model by finish, on container lead times

Home and furniture forecasts the model by finish or option, grouped by supplier, for a landing window rather than a month. The container and factory slot lock the quantity a long way ahead, so the forecast horizon runs from the booking date through the ocean crossing to the end of the window, and the useful flexibility is the mix of the next container rather than a change to the one on the water. Quantities round to container cube, which makes the forecast a mix decision as much as a volume one, valued at landed cost.

Cleaning has an unusual advantage in this category: the demand the floor could not serve is written down. Written orders date demand at commitment rather than at delivery, so a model's history can be rebuilt on order dates, and a long quoted delivery date then shows up as the conversion penalty it was instead of as weak demand. Special orders name the finish, size or configuration a customer wanted and could not take from the floor, and adding them back at model level is the adjustment that turns a floor-limited record into a demand record. A model that was never on the floor has no history to clean at all, so it is forecast from the collection it belongs to rather than from its own zero. Weekly SKU-level volume in the long tail is too thin to read, so collections and finishes are the forecast grain and the SKU is the distribution. See merchandise planning for home and furniture brands.

Outdoor: model year changeover, prebooks and weather that moves timing

Outdoor forecasts the model inside a model year, and the analog is the same model in the prior year adjusted for spec changes, with carryover models carrying more evidence than refreshed ones. The dealer prebook closes the quantity before the season, so the brand is forecasting what dealers will commit, plus its own at-once pool and DTC demand — three different questions that collapse easily into one number. Planning a model year changeover covers the transition, and the prior year's closeout inflates its own recorded rate while suppressing the new model's, so the two are read apart.

Weather-exposed categories need timing and volume kept separate in the forecast. A season that opens warm asks one question ahead of every other — smaller season or later season — and the same weekly numbers answer it differently depending on which is assumed. The cleaning step follows from that: a prior year is re-phased onto its own weather anchor, the week that season actually started, before its shape is copied onto a calendar, or a late start in the analog year becomes a permanent dip in the plan. Forecast the season total and phase it, rather than reading the phasing as the total. The dealer side is cleaned differently again, because prebook units are a commitment rather than a sale: the model's demand history is rebuilt from the dealers who report sell-through and scaled on the share of the fleet those dealers represent, with the scale factor written down rather than assumed. Counter-seasonal lines run on their own calendar against the same budget. See merchandise planning for outdoor brands.

Health & beauty: the shade ladder, dating and the regimen

Beauty forecasts the franchise — a foundation, a lipstick line, a serum — and distributes it across the shade ladder, because a single shade at a single door falls below a readable rate while the franchise does not. The ends of the ladder are thin, and component and batch minimums round them up, so the shade distribution is where forecast error turns into either write-off or a permanently out-of-stock shade. The in-stock definition a beauty analog is cleaned against has to include the tester, not only the stock behind it: shade-weeks that ran without a working tester come out of the record before the ladder is refitted, and the ladder itself is refitted on the doors that carried every shade. Ends that were never shipped are filled from the shades beside them rather than from their own zeros, which is the difference between a shade that is thin and a shade that was never given a chance to sell.

Shelf life caps the horizon. A batch can only be forecast as far forward as unopened shelf life and the retailer's dating requirement at delivery allow it to be sold, so a forecast that justifies more cover than the dating ceiling supports is a forecast of a write-off, the arithmetic dating rules and weeks of supply sets out. Gift-with-purchase and gratis consume units without generating a demand signal and need their own line rather than being buried in the franchise total. Regimen attach — cleanser to serum to moisturizer — means a step's demand is partly a function of the step before it, so a franchise launch forecast carries the attached steps with it. Where the retailer controls replenishment, the brand forecasts its fill to the retailer's DC and the retailer's reset calendar sets the dates. See merchandise planning for health and beauty brands.

Sporting goods: season-anchored windows, team cycles and consumables

Sporting goods splits into three forecasts that behave differently. Season-defined hardgoods anchor to the sport's season start and a model year, with dealer prebooks closing the quantity as in outdoor; the analog is the prior model year adjusted for spec and rule changes, since an equipment specification change can invalidate a model's history outright. Team and league business is committed demand with dates rather than a forecast — a roster order is an order — but the order windows themselves are forecastable, and a window that closed early leaves nothing behind to clean: there is no partial sale to adjust upward, only an absence. Team volume is forecast from the roster and league calendar the window served, not from the order total it recorded, and a season whose window closed early is flagged in the analog set rather than quietly restated.

Consumables — balls, grips, tape, shuttles — replenish against min and max, carry dense history and support a genuine velocity forecast, which makes them the part of the line where statistical method pays. Prior-model-year closeout pricing distorts both sides of the changeover and is excluded from the analog. League calendars and event dates concentrate demand into weeks, so phasing carries as much weight as the annual total. See merchandise planning for sporting goods brands.

Toys & games: a Q4 peak behind a factory cut-off

Toys and games forecast a property in a format at a price point, and the structural fact is that the factory cut-off lands well before the Q4 peak the forecast is aimed at. The quantity decision is made before any consumer read on the peak exists, so the range matters more than the point, and the flexibility that remains is a domestic DC holdback bought earlier rather than a reorder placed later. Quantities round to case pack, which is the same mix-versus-volume constraint containers impose in furniture — see case packs and planned depth.

Licensed windows bound the forecast on both ends: a property's demand is tied to a release calendar and the window closes on a date, so the forecast covers the window rather than a fiscal year, and any quantity that cannot sell inside it is planned as an exit. Planning a licensed product window covers the mechanics and the minimum guarantee behind it. A property's analog is built on consumer sales, not on what retailers committed. Where a retailer shares POS, the prior property's season is rebuilt from it; where none is shared, the weeks a retailer's own DC held stock back are marked unknown rather than counted as demand, because a shipment that never reached a shelf is evidence about logistics. Case-pack rounding is undone before any comparison, since a property bought in packs records the quantity the pack chose rather than the one the forecast asked for. See merchandise planning for toy and game brands.

Baby & juvenile: registry intent and age-band graduation

Baby and juvenile has a forward signal the category's own channel produces. Registry adds record purchase intent weeks or months before the sale, which makes them an input to the forecast rather than only an in-season read, and planning with registry demand covers how the two connect. The signal censors itself, and the cleaning is specific: registry adds count only for the weeks an item was listed as available, and an add placed on a substitute while the first choice showed unavailable belongs to both items — to one as intent it never recorded, to the other as a sale it won by default. Reading the substitute's adds as its own demand is the adjustment that has to happen before either item becomes an analog.

Demand here is a cohort flow rather than a trend. A child ages out of a band, so an item's demand follows the population entering its age band rather than the growth of the item itself, and a forecast built on the item's own trend line misreads a cohort change as a product change. Apparel-side product plans on age bands where sizes sit, as planning a kids and baby apparel line sets out. Hard goods are long-lived certified configurations, and the certification is what makes their analogs stable. A platform that carries forward unchanged brings its whole multi-year history with it, while a re-certified configuration starts a new record even where the product looks the same, so the certification date is the cut point for the analog rather than the season boundary. Certification and the production run lock the quantity, so the clean history has to be ready before that date, not after it. A withdrawn lot comes out of the history entirely rather than being read as a collapse in demand, which is one more reason the production lot travels with the record.

Jewelry & watches: bands and collections, not pieces

Depth in jewelry and watches runs to a piece or two per style in a door, and each piece sells once, which puts a piece-level weekly history permanently at zero or one — too thin to carry a quantity decision. The collection, metal, stone and price band is the forecast grain; the piece is the distribution. The band is also the stable unit through a cost change: metal price movement reprices pieces at refill and can move the same design into a different band, so a forecast held at the design level drifts for reasons that have nothing to do with demand, while a band-level forecast stays comparable and planning margin on a moving cost base handles the margin side.

Demand concentrates into gifting peaks — the Q4 and spring occasions the gifting calendar dates — so the annual total is close to meaningless as a planning unit and the phasing carries the decision. Single-piece depth makes cleaning unavoidable rather than optional. One sale empties the door, so a band's history is assembled from the door-weeks in which that band had a piece present: a door that sold its only piece in week one contributes one week of evidence, not a season of zeros. Appointments, try-on requests and ring-size enquiries are added back at band level as demand the empty weeks could not record, and pieces out on memo are taken out of the selling record of the door that owns the stock, because they were displayed somewhere else. See merchandise planning for jewelry and watch brands.

What a forecast should be held to

A forecast is scored to improve the next one, which only works if the score measures the thing the forecaster actually decided. Four rules make that true.

Score the vintage that bought the goods. A forecast revised in week six is not the forecast that committed the quantity in the prior season. Keep the version live at the commitment date and score that; revisions are a separate, useful record of how fast the view changed.

Score against demand, not recorded sales. A style that sold out in week four is recorded as having met its plan when it beat it, and scoring the censored number teaches the next forecast to under-buy exactly the styles that ran hot. Adjust the actual for availability before the comparison, using the same in-stock rate and lost sales work the season already required.

Keep bias and error apart. Signed bias asks whether the forecast runs high or low across a portfolio; unsigned error asks how far off it is on each line. They fail differently and are fixed differently. A portfolio with near-zero bias and large per-style error is describing newness behaving like newness. A portfolio with small per-style error and persistent one-way bias is describing a process problem — a target pushed down into a forecast, or an analog set that no longer resembles the assortment.

Score at the grain the decision was made at. This is the one that surprises people, because errors offset on the way up.

Illustrative example: the same season, scored two ways

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

Four styles in one class, each planned at 1,000 units, 4,000 for the class. The actuals below are demand, already adjusted for availability as the second rule requires, and class demand lands at exactly 4,000.

At class level the forecast is perfect: zero variance, zero bias, and the open-to-buy envelope it sized was exactly right. At style level the absolute error is 1,800 units on a 4,000-unit plan. Both numbers are correct, and they answer different questions. The class number validates the money: the receipt plan, the stock target and the budget were the right size. The style number describes what the season actually felt like — A and C sold out early, with the demand their empty weeks could not record added back, while B and D went to markdown — and it is where the margin went.

That is why the two are reported together and never substituted for one another. A forecast reviewed only at class level looks accurate for as long as its errors keep cancelling, and the cancelling is luck rather than method. A forecast reviewed only at style level condemns newness for being newness. Decomposing a plan miss separates the causes, and the plan vs actual variance formula carries the arithmetic. The grain to report at is the grain the decision was made at, which means class or department for the money and style-color, model, franchise or band for the mix.

The handoff into in-season sensing

The forecast's last act is to hand the season something sensing can test. That handoff has three parts. The plan phased by week or by delivery window, so a variance has a date attached rather than being a season-to-date total — how to phase a sales plan covers the phasing itself. The assumption list, written down: the analog used, the price assumed, the door count, the channel mix, the delivery dates. And the scenario pair, so the upside and downside are already priced when the first read arrives and the chase or the markdown decision is a threshold rather than an argument.

Sensing then does the job the forecast cannot: it estimates the rate demand is actually running at, adjusted for availability, and moves the decisions still open inside their lead times. It does not replace the forecast: the demand sensing guide sets out why a sensed rate cannot be stretched to season end, and how often to reforecast covers the cadence on which a shift that persists is taken back into the plan. At season end, the hindsight closes the loop: which analog held, which assumption broke, and which styles were scored against a censored actual. That record is the input to next season's analog set, which is the only part of a forecast that compounds.

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 a style's sell-through history, or the history of analogous styles for newness, to inform buy quantities, and a plan can be tested against upside and downside scenarios before the buy is committed. In season, sell-through against plan informs buy-depth and allocation recommendations. Because the plan, the buy, the allocation and the PO record share one data model, a forecast change reaches the buy quantity rather than sitting in a separate file — the disconnection demand forecasting for fashion describes.

It is not a statistical supply-chain forecasting engine at the scale of a dedicated demand planning system. 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. OTB planning holds the budget the forecast sizes, and assortment planning is where the forecast becomes options and depth.

Related resources

• →Demand Sensing in Retail — the in-season half of this split, across the same ten verticals
• →Demand Forecasting for Fashion — the apparel-only treatment, including attribute-based forecasting for newness
• →Demand vs Merchandise vs Supply Planning — where each discipline stops
• →The Planning Calendar by Vertical — the commitment dates this forecast has to beat
• →Open-to-Buy by Vertical — the budget the forecast sizes
• →How to Hindsight a Season — how the score feeds the next analog set