Demand Forecasting Methods: Techniques, Models and Process

Most method guides list ten techniques and never run one. This page runs three on real weekly data and shows the errors. It sits under demand forecasting for apparel brands, which owns the accuracy metrics and why the category is hard; this page owns the methods, the process around them, and how to choose.

Demand forecasting definition, and what a forecast has to produce

Demand forecasting is the estimation of how many units customers will buy, by item and by period, early enough to act on. That is the demand forecasting definition in operational use, and every word carries weight: units not dollars, by item not in aggregate, early enough that a factory can still respond.

The definition of forecasting in supply chain work is wider, covering anything you project in order to commit resources before orders exist: demand, lead time, capacity, freight cost, returns. Asked what is supply chain forecasting, the answer is all of those, and for an importing brand the lead-time projection is as load-bearing as the demand one. The scope of demand forecasting likewise runs from a total-company season plan down to one size-colour cell in one week — different problems, different methods.

Types of forecasting in supply chain management: qualitative and quantitative

The types of forecasting in supply chain management are conventionally cut four ways: by method (qualitative or quantitative), by horizon, by level (SKU, style, class, total) and by driver (extrapolative or causal). The method split decides what you can actually run.

Qualitative methods of demand forecasting need no history at all: expert judgement, the Delphi method, sales force composites, market research and historical analogy. Quantitative methods need a series: naive and seasonal naive benchmarks, moving averages, exponential smoothing and Holt-Winters on the time-series side, regression and machine learning on the causal side.

Qualitative techniques are not a lesser option. They are the only option when the item has no history, and a meaningful share of every apparel season is new. Make judgement explicit and auditable: a Delphi round with written rationales beats a room agreeing with the loudest person.

Demand forecasting techniques, method by method

Naive and seasonal naive: the benchmark every forecast has to beat

The naive forecast is last period's actual carried forward; the seasonal naive forecast is the same period last year carried forward. Hyndman and Athanasopoulos treat both as the standard benchmarks in Forecasting: Principles and Practice, and that is their job: not to be used, but to be beaten. If your model plus three rounds of human override does not beat seasonal naive, the process is subtracting value. That comparison is called forecast value added and costs nothing to compute.

Moving average and simple exponential smoothing

A moving average forecasts the next period as the mean of the last n actuals. A four-week window is the apparel default because it spans a retail month without dragging in a stale trend.

Simple exponential smoothing weights recent observations more heavily and older ones geometrically less, controlled by a smoothing constant alpha between 0 and 1. The NIST/SEMATECH e-Handbook sets out the update form used below: the new forecast is alpha times the latest actual plus (1 − alpha) times the previous forecast. A low alpha smooths hard and reacts slowly; a high alpha tracks noise. Both methods assume no trend and no seasonality, and will lag trended data.

Holt and Holt-Winters for trend and seasonality

Holt's method adds a second smoothing equation for trend. Holt-Winters, which NIST documents as triple exponential smoothing, adds a third for seasonality, in additive and multiplicative variants. Multiplicative is usually right for apparel, because a peak week is a multiple of a base week rather than a fixed addition to it.

The constraint that rules Holt-Winters out for most brands: it needs at least two complete seasonal cycles before it can estimate seasonal indices — about two years of clean, in-stock weekly history for the same style. Most styles do not live that long. Class-level and channel-level series do.

Regression with causal drivers, and machine learning

Regression forecasts demand as a function of what you can observe or control: price, markdown depth, marketing spend, door count, temperature. It is the only family that answers "what happens if we discount 20%." The cost is data discipline: if you cannot reconstruct your price and marketing by week, the model attributes their effects to the intercept.

Gradient-boosted trees and similar machine-learning methods learn across many series at once, so a new style borrows strength from every historical style sharing its fabric, silhouette, price band and colour family. They need scale to pay off and overfit easily on short series. Among forecasting techniques in supply chain use, this is where vendor claims most outrun evidence, so apply the same out-of-sample forecast-value-added test.

Hierarchical forecasting and reconciliation across style, colour and size

This is the method that wins in apparel, for reasons of dimensionality rather than statistics. One hundred styles in three colours and six sizes is 1,800 forecast cells before any channel split, and a fashion style selling 60 units a week across 18 cells averages 3.3 units per cell per week, with tail sizes at one or two. No time-series method survives that. Forecast instead at the level where signal exists, usually style-colour or class, then disaggregate down a size curve and reconcile the parts back to the total.

Worked disaggregation of one style-colour forecast down a standard unisex size curve. The total is the 3,355 units the moving-average method produces in the next section; the ratio is the S:M:L:XL:2XL = 1:2:3:2:1 bell curve.

An inherited factory pack ratio is not your curve. Derive it from your own stockout-corrected sell-through: size curves and the SKU explosion.

SKU level demand forecasting, and when the SKU is a level you can actually forecast

SKU level demand forecasting means fitting a model to the individual size-colour cell itself rather than forecasting a parent and splitting it down. It is defensible in exactly one condition: the cell carries enough weekly volume to hold a signal.

Compare the two series on this page. The core tee colour in the worked example runs about 420 units a week across its sizes; on the 1:2:3:2:1 curve that is roughly 143 units a week in the L cell and about 46 in the smallest, both of which a moving average or exponential smoothing can fit. The fashion style in the section above sells 60 units a week across 18 cells, averaging 3.3, with tail sizes at one or two. At that density the weekly series is mostly zeros, MAPE is undefined on every zero week, and a model fitted to it is fitting noise.

So the level is a property of the item, not of the assortment. Forecast the repeating core book at SKU level, forecast fashion at style-colour or class and disaggregate down the curve, and decide which regime an item is in by counting units per cell per week rather than by what the merchandising calendar calls it.

Forecasting methods in supply chain compared by data requirement, horizon and use

The table below compares the supply chain forecasting models above on what they need and where they earn their keep. Data requirements are minimums for a usable fit, not comfortable ones.

Two notes on choosing. Demand planning forecasting models more complex than your data supports lose to seasonal naive out of sample. And retail demand forecasting methods are usually judged at the wrong horizon: a model benchmarked one week out looks excellent and says nothing about the four-month commitment you have to make.

A worked comparison: naive benchmark against a smoothed forecast

Method arguments end when someone computes the errors. Here are three demand forecasting examples run on the same series.

Basis: weekly unit sales of one core 180 gsm cotton crew tee, one colour, all sizes. Weeks 1-8 are the fit window; weeks 9-16 are held out as the test. Naive = previous week's actual. MA4 = mean of the previous four actuals. SES = exponential smoothing, alpha 0.3, seeded at 393 (the mean of weeks 1-4) and updated as 0.3 x latest actual + 0.7 x previous forecast.

Now the errors, computed on the eight test weeks.

Three readings. The naive forecast has almost perfect bias and terrible accuracy, because it chases every wobble and the wobbles cancel out; bias alone would have said it was fine. Smoothing halves the error here, because this series is noise around a flat level, the condition those methods are built for. Neither survives a trend: both lag, and Holt's trend term recovers the difference.

Run this on your own top ten styles before buying a forecasting engine. If a four-week moving average lands inside two points of the vendor's model, you are buying workflow, not accuracy.

The demand planning process, step by step

The demand planning process is the governed cycle around the maths, and it is what turns a number into a commitment. A good demand planning process enables a company to link its commercial plan to its supply commitments in one agreed set of numbers; in apparel that link is physical, because it ends in a booked dye lot and a cut ticket.

The seven demand forecasting steps in order

1. Define the item and the bucket. Style-colour by week is the usual apparel unit, and the level determines which methods are available.
2. Clean the history. Correct stockout weeks upward, strip one-off events, mark price changes. Uncorrected stockouts teach the model that demand fell when supply did.
3. Choose the forecast level in the hierarchy. Forecast where the signal is, not where the decision is; the decision level is reached by disaggregation.
4. Fit, then select against a benchmark. Two or three candidates, hold out the last eight weeks, compare each against seasonal naive.
5. Disaggregate down the size and colour curve, then reconcile the parts so they sum to the parent.
6. Apply overrides on the record. Every judgement change gets an author, a size and a reason. Unattributed overrides are how bias becomes permanent.
7. Measure and feed back. WMAPE, bias and forecast value added, at the horizon at which you commit.

That is the demand planning process flow in its simplest honest form. The demand planning methodology layer above it decides cadence, who signs, and how the demand plan is reconciled with supply: demand planning vs supply planning. Demand planning techniques such as consensus review and assumption logging live at that layer, not in the model.

The demand forecasting process also has to be timed to the sourcing calendar, a point generic guides miss. First orders run 10-16 weeks PO to FOB from India or Bangladesh, becoming 15-20 weeks PO to a US DC from India and 18-23 weeks from Bangladesh once ocean transit and clearance are added; repeat orders compress to 8-12 weeks PO to FOB. A monthly consensus cycle landing after the fabric booking date is decoration.

Demand forecasting for new products with no sales history

New styles have no series, so time-series methods do not apply. Use analogues, and make the adjustment factors explicit.

Worked analogue forecast for a new 220 gsm heavyweight crew tee at a $42 ticket. The analogue is last year's 180 gsm washed crew at $34, which sold 2,850 units in its first 12 weeks. The three multipliers are recorded planner judgement, not sourced elasticities, and each should be argued in the consensus meeting.

Then size the buy as a bet rather than a plan. Cutting 70% of the expectation gives 1,831 units, leaving the balance as a reorder option. Check it against the mill before celebrating: at 0.27 kg of fabric per tee, 1,831 units is 494 kg, right at the 300-500 kg dye-vessel minimum for a custom colour. Below that the lot cannot be run, and the buy has to be aggregated across styles sharing the colour: how fabric and dye-lot minimums set apparel MOQs.

Demand forecasting best practices for apparel brands

Eight demand forecasting best practices, ordered by what they typically return.

1. Always compute forecast value added. No benchmark, no evidence.
2. Use WMAPE, not MAPE. MAPE is undefined on zero-sales cells, and apparel is full of them.
3. Report bias separately, by size. A style can hit its total and lose money on the curve.
4. Benchmark at the commitment horizon. Each extra month adds 2-5 points of WAPE, so a one-week accuracy number is marketing.
5. Forecast the repeating book first. It has history, it is reordered, and improvements compound.
6. Correct history for stockouts before fitting anything, and log every override with author, size and reason.
7. Forecast lead time as well as demand. On a 13-week pipeline, lead-time variance usually dominates required safety stock: reorder points when your lead time is 90 days.
8. Convert the forecast into a receipt calendar and a projected on-hand position, not just a number: inventory forecasting and forward on-hand projection and open-to-buy planning for apparel brands.

The demand planning best practices that matter most are governance, not mathematics. Apparel and fashion run 35-60% MAPE at SKU level against 10-25% for FMCG staples, per Umbrex's cross-category compilation. You will not model your way out of that. You manage it by shortening the commitment horizon, stabilising lead time, and being honest about which part of the assortment is genuinely forecastable. Terms used here are defined in the Yarnstick glossary of apparel sourcing terms.

Demand forecasting for retail, supply chain and inventory: one method set, several names

The same body of method is searched for under a dozen phrasings, and most of them do not signal a different subject. A few genuinely do, and it is worth saying which.

What people call this work, what each phrasing is usually asking for, and where it is answered.

Frequently asked questions