Demand Forecasting for Apparel Brands

In this cluster

This is the cluster pillar for demand planning inside the Yarnstick knowledge base, for a founder or ops lead who buys imported apparel and is tired of choosing between a stockout and a markdown. It covers the planning vocabulary you will be held to, the two accuracy metrics that matter, and why apparel forecasting is harder than the textbooks admit. The pages beneath it carry the forecast into a purchase order, a budget and a size run, and go deeper on method, tooling and inventory projection than a pillar should.

What an apparel demand forecast has to produce, and when

A demand forecast for an apparel brand is not a revenue number. It is a unit number, by SKU, by week, produced early enough that a factory can still act on it. A revenue forecast can be revised on Monday; a production commitment cannot, because once the yarn is allocated, the dye vessel loaded and the marker cut, it is physical.

Forecasting apparel is hard not because the mathematics is hard. It is hard because you must commit before you know, and the length of that commitment window is set by your supply chain, not by your model. A brand shipping from Tirupur to a New Jersey DC forecasts at a four-to-six month horizon whether its planner wants to or not. Improving the model shaves points off the error at that horizon; shortening the horizon changes what the error costs.

Demand planning and forecasting: one activity, several names

Job adverts, software categories and search results use half a dozen names for this work, and the differences between them are real but small. Knowing which name signals which scope saves an argument in every vendor call and every interview.

What each term usually signals about scope, and where you will meet it.

Two distinctions are worth defending. Demand planning is the process; forecasting is the statistical step inside it — how demand planning and supply planning divide the work draws that line formally. And a forecast in units is not a revenue plan: you can hit revenue and still buy the wrong sizes. The mathematics of each approach, from naive benchmarks to attribute models, sits in demand forecasting methods, method by method.

Retail demand planning metrics an apparel forecast is judged against

Every planning conversation with a buyer, a lender or a board runs on the same formulas. The table below is the standard retail-math set as published in Toolio's retail math reference.

Three are routinely misread. Sell-through is meaningless without a window — "62% sell-through" says nothing until you state 62% at four weeks or season-to-date. GMROI uses inventory at cost, which is what makes it the cleanest comparison between a slow high-margin category and a fast low-margin one. Open-to-buy is a control, not a plan: it is what remains after committed POs, and for an importer on a 15-23 week PO-to-DC lead time it is near zero by the time the season starts. Mechanics, the cost-versus-retail distinction and a worked table live on open-to-buy planning for apparel brands; this page stops at the definition on purpose.

A worked read on one style: sell-through, weeks of supply and GMROI

Assumptions: a core cotton crew tee, one colour, 3,000 units received, retail ticket $34, landed cost $6.08 a unit at non-peak ocean freight — a build worked line by line in landed cost for apparel imports.

• Four-week sales 1,140 units. ST% (4-week) = 1,140 ÷ 3,000 = 38.0%.
• On hand 1,860 units, average weekly sales 285. WOS = 1,860 ÷ 285 = 6.5 weeks.
• Forecast average weekly sales 240. FWOS = 1,860 ÷ 240 = 7.75 weeks.
• Season close: 2,760 units at an average realised $29.00 = $80,040 revenue; COGS $16,781. Gross margin $ = $63,259. Average inventory at cost $9,850. GMROI = 6.4.

That 6.4 is flattering: a single-receipt style measured over one window carries almost no inventory in the denominator. GMROI is built for category comparison across a full year; at style level, use sell-through and forward weeks of supply.

Measuring apparel forecast accuracy: MAPE, WMAPE and why WMAPE wins

Two metrics dominate published guidance, and one is actively wrong for apparel.

• MAPE (mean absolute percentage error) is the mean of |actual − forecast| ÷ actual across cells. It breaks down at low volumes and is undefined when the actual is zero.
• WMAPE / WAPE (weighted absolute percentage error) is Σ|actual − forecast| ÷ Σ actual. It weights by volume and stays defined when individual cells are zero.

At style-colour-size level, apparel data is full of cells that sold 0-3 units in a week. That makes MAPE structurally misleading, not merely noisy: cells it cannot compute get dropped, so the metric quietly excludes your worst-behaved SKUs.

The table below shows both metrics computed on the same eight-week read of one style-colour across six sizes. Forecast and actual are units.

• MAPE = (81.8 + 18.9 + 12.2 + 25.0 + 37.9) ÷ 5 = 35.2%, and it had to discard the 3XL cell to compute at all.
• WMAPE = 118 ÷ 500 = 23.6%.

The gap is not cosmetic. MAPE is driven by the S size, where an 18-unit miss on a 22-unit base reads as 81.8% error and carries the same weight as the L size, where a 25-unit miss on 205 units reads as 12.2%. WMAPE puts both misses on the same scale — the one your inventory lives on.

Two more metrics belong in any planning review. Tracking signal (cumulative forecast error ÷ mean absolute deviation) detects a model that has drifted. Forecast value added (FVA) compares your process against a naive benchmark; if your model plus three rounds of human override does not beat naive, the process is destroying value. FVA is the most under-used metric in apparel planning and is free to compute.

Why a forecast bias check matters more than a forecast accuracy score

WMAPE is magnitude-only. It tells you how wrong you were and says nothing about which direction.

Bias (mean forecast error) = Σ(forecast − actual) ÷ n, or as a percentage of demand, Σ(forecast − actual) ÷ Σ actual. On the table above: (512 − 500) ÷ 500 = +2.4%. That looks benign. It is not.

Read the same table by size. The forecast ran over on S by 18 units, XL by 24, 2XL by 11 and 3XL by 12 — 65 units on the tails — and under on M by 28 and L by 25, 53 units in the middle. The net is a tidy +2.4%. The signal is that the style total was nearly perfect and the size curve was wrong, a different problem with a different fix: see size curves and the SKU explosion.

A forecast can post an excellent WMAPE and a ruinous bias. Persistent positive bias is systematic over-buying, and it surfaces not as an error metric but as markdown dollars one to two quarters later. Report WMAPE and bias together, at style and size level, every month.

What apparel forecast accuracy actually looks like

Typical SKU-level forecast error by category, compiled by Umbrex. Orientation, not targets.

On WAPE: under 20% is good, 10-15% strong, and under 10% is exceptional and realistic only for stable, high-volume items. Critically, each additional month of forecast horizon typically adds 2-5 percentage points of WAPE. Benchmark at 1, 3, 6 and 12-month horizons rather than reporting one number, because the only horizon that matters commercially is the one at which you must commit.

Why apparel demand forecasting is structurally hard

Eight structural reasons, roughly ordered by damage done. None is a modelling failure. They are properties of the category.

The aggregate cost is worth stating plainly: stockouts plus overstock are estimated at $1.73 trillion a year globally, around 6.5% of global retail sales.

Supply chain demand planning: your lead time sets your forecast horizon

Here is the join almost no forecasting article makes: your forecast horizon is not a planning preference, it is an output of your sourcing decision.

Indicative apparel lead-time stages, August 2026. The first six rows sum to the PO-to-FOB time; add the ocean transit row plus clearance, drayage and DC receiving to get PO to DC.

First orders run 12-20 weeks PO to FOB globally and 10-16 weeks PO to FOB from India or Bangladesh, which becomes 15-20 weeks PO to a US DC from India and 18-23 weeks from Bangladesh once transit and clearance are added; repeat orders compress to 8-12 weeks PO to FOB by eliminating sampling. Bangladesh runs 75-120 days PO-to-FOB on knits and adds 15-25% in the Q3 peak. Holiday shutdowns are additive: Chinese New Year (two to four effective weeks), Tet, two Eid periods, Diwali.

Now apply the horizon-degradation rule. A process running 15% WAPE at a one-month horizon, committing at four months, adds three months at 2-5 points each: best case 21% WAPE, worst case 30%.

Nothing about the model changed; the supply chain moved the measurement point. That is why the highest-return planning investment for most importing brands is not a better algorithm but anything that shortens or stabilises the pipeline — a held fabric platform, a nearshore replenishment tail, or knowing your own lead-time distribution well enough to stop treating it as a constant.

The forecastable book versus fashion and seasonal drops

Not all of your assortment is equally forecastable, and treating it as one book is the most common structural error in apparel planning.

The repeating book is where forecasting pays first. It has real history at size and colour level, so the statistics work. It is reordered, so an improvement compounds across every future cycle instead of being spent once. And because the spec is stable, the fabric can be platformed — one greige quality, one dye recipe held at the mill — the biggest single lever on lead time and therefore on safety stock. Most brands do the opposite, spending planning attention on new styles while running the core book on gut and a reorder email.

Improving supply chain forecasting: what to fix first

Improving supply chain forecasting is usually attempted in the wrong order — tool first, method second, data last. Reverse it. Ranked by what moves money for a brand importing on a 15-23 week PO-to-DC pipeline:

1. Fix the demand history before the model. Weeks in which a size was out of stock record demand as zero. Any model trained on that record will under-forecast your best sellers and re-buy last season's shortage. Correct the censored weeks first: you cannot forecast demand you never let yourself observe.
2. Shorten or stabilise the pipeline. Every month removed from the commitment horizon is worth 2-5 points of WAPE that no algorithm has to earn.
3. Choose the method to match the data you have. A new style has no series to fit; a core tee has three years of it. Demand forecasting methods, method by method sets out which technique the data supports.
4. Forecast the inventory position, not only the demand. A demand number becomes actionable only once you project on hand plus on order forward week by week: projecting an inventory position forward week by week.
5. Split the channel. Wholesale front-loads, DTC builds with paid media, and marketplace stock limits bite separately: forecasting for a DTC and marketplace channel.
6. Buy tooling last, and against a decision. Vendors sell supply chain predictions; what changes your margin is whether a prediction reaches a purchase order early enough to alter it. Compare what the demand planning software categories actually do and where machine learning genuinely beats a planner against your own naive benchmark before signing anything.

Where the forecast meets the sourcing constraint

A forecast that stops at a number is half a decision. Three constraints turn it into a buy.

Lead time turns a forecast into an order date. A reorder point is average demand across the lead time plus safety stock, and at a 13-week PO-to-DC lead time rather than one week, lead-time variability dominates that calculation: reorder points when your lead time is 90 days.

Money turns a forecast into a receipt calendar. With a 90-day PO-to-DC pipeline the buy decision lands a quarter before the receipt month: open-to-buy planning for apparel brands.

The size curve turns a style number into a cut ticket. Size ratios are cut into the marker and packed into the ratio carton before the goods ship, so a wrong ratio is baked in a quarter before it reaches your sales data: size curves and the SKU explosion.

One constraint sits underneath all three: your order quantity has a floor you do not control. Dye vessels have minimum loads, typically 300-500 kg of fabric per custom colour — roughly 1,000-2,000 tees' worth — and tier-1 Vietnamese mills commonly quote 800-1,200 kg. A forecast that says "buy 640 units of sage" collides with a mill that cannot run a lot that small: how fabric and dye-lot minimums set apparel MOQs and sourcing apparel direct from factories.

Duty closes the loop, because forecast error is priced at landed cost, not FOB. Under the Section 301 forced-labor tariffs effective 24 July 2026, apparel from India and Bangladesh carries an additional 10% and from Vietnam and China 12.5%, on top of MFN rates in the U.S. International Trade Commission's Harmonized Tariff Schedule — verify against the current HTSUS. See landed cost for apparel imports and the Yarnstick glossary of apparel sourcing terms.

Frequently asked questions