Can AI forecast a full pre-order book from the first 20 percent of orders?

KEY TAKEAWAYS Summary by the editors

1. Early wholesale orders are one of the strongest demand signals a fashion brand has, because they reflect real buyer commitments rather than internal assumptions.
2. Research by Marshall Fisher and Ananth Raman, published in Operations Research in 1996, showed with a skiwear maker that responding to early sales could raise profits substantially compared with an informal approach.
3. AI improves on the classic method by learning, per style, colour and account type, how early orders relate to the final book, and by updating that relationship as new orders arrive.
4. A pre-order forecast is only as good as the history behind it: brands need several seasons of time-stamped order data, with cancellations and account attributes, before models become reliable.
5. The forecast should feed concrete decisions, such as fabric commitments, production splits and sales focus, and should always carry a range rather than a single number.

Yes, within limits. The first orders written in a wholesale campaign usually reveal far more about the final pre-order book than a pre-season plan does, and AI models can turn that early signal into a style-level forecast of the full book with a confidence range. The method is decades old; what AI adds is the ability to do it for thousands of SKUs, account types and markets at once, and to update the forecast every day of the campaign.

Why are early orders such a strong forecasting signal?

A pre-season plan is a set of assumptions made by the brand. Early orders are commitments made by buyers who know their own stores and customers. Even when only a small share of the eventual book has been written, those orders show which styles, colours and price points buyers are backing, and which ones they are skipping.

The idea was formalised in fashion operations research long before machine learning became common. In Making Supply Meet Demand in an Uncertain World (Harvard Business Review, 1994), Marshall Fisher, Jan Hammond, Walter Obermeyer and Ananth Raman described how a skiwear maker used early demand information to decide what to produce first and what to hold back. In a 1996 Operations Research paper, Fisher and Raman modelled this "accurate response" approach with the same company and reported profit gains of around 60 percent compared with its existing, informal way of reacting to early sales.

How does AI forecast the full pre-order book from early orders?

The core task is to learn the relationship between what had been ordered by a given point in past campaigns and what the final book eventually looked like. A machine-learning model does this at a finer level than a planner with a spreadsheet can, and it can combine several signals:

• Order curves: how quickly each category, price band or carry-over style typically fills during a campaign.
• Account mix: which retailers have ordered so far, since key accounts, independents and marketplaces order at different times and in different depth.
• Attribute similarity: for new styles without history, the model borrows from similar styles by fabric, silhouette, colour family and price.
• Campaign context: appointment calendars, trade-show dates and market-by-market timing, so that a quiet week is not mistaken for weak demand.
• Cancellation patterns: how much of a booked order historically survives to shipment.

The output should be a forecast of the final book per style and colour, with a range, plus a flag for styles where the early signal diverges sharply from plan. McKinsey has reported that AI-driven forecasting can reduce errors by between 20 and 50 percent in supply chain settings, and argues that most organisations have enough data to benefit if they handle gaps carefully. Those figures are general, not specific to fashion pre-orders, and should be treated as an indication of potential rather than a promise.

Read also

Digital showroom ROI: how to build the business case for costs and sell-in

What data does a pre-order forecasting model need?

Most brands find that the history exists but is not stored in a usable form. Order platforms often keep only the latest state of an order, so the build-up during the campaign is lost. Fixing this, by keeping daily snapshots of the order book, is usually the first and cheapest step.

How early in the campaign can the forecast be trusted?

There is no universal threshold. The point at which a forecast stabilises depends on how concentrated the customer base is and how predictable each account's timing is. A brand whose book is dominated by a few large accounts that order late may see little signal until those accounts commit. A brand with many independent retailers ordering steadily may get a usable signal much earlier.

The practical approach is to back-test: replay past campaigns, run the model as if it were week two, week three and so on, and measure how far each forecast was from the final book. That gives the team an evidence-based answer to the question "how much do we trust this number today?" for their own business.

Which decisions should the forecast drive?

1. Fabric and trim commitments: confirm materials for styles with strong early signals and hold back on those that are lagging.
2. Production splits: commit a first tranche based on the early forecast and keep a share of capacity flexible for a later decision, as in the accurate-response approach.
3. Sales focus: point sales reps towards accounts that have not yet ordered strong styles, or towards weak styles that need a push or a decision to drop.
4. Range edits: identify styles likely to miss minimum order quantities early enough to cancel or merge them.

What are the limits and risks?

Early orders can mislead. A single large account ordering heavily can make a style look stronger than it is across the market. Buyers may order early for some categories and late for others, and macroeconomic shocks or tariff changes can shift timing for a whole season. Models trained on stable years tend to underreact to these breaks.

There is also an organisational risk. If the forecast is presented as a single number, merchandisers and sales teams may either ignore it or follow it blindly. Showing a range, explaining the main drivers per style and keeping a human decision step for large commitments are simple safeguards.

Read also

How can brands use data to reduce pre-order cancellations and order cuts?

Is it worth the investment?

The return comes from better first commitments: less fabric bought for styles that never fill, fewer stock-outs on styles that do, and fewer late cancellations of production. For most brands the largest cost is not the model but the data work behind it. A sensible path is to start with one category, back-test against two or three past seasons, and only then extend the forecast to the full range.

Frequently asked questions

What is pre-order forecasting in fashion wholesale?

It is the practice of predicting the final size of a seasonal wholesale order book, per style and colour, while the order campaign is still running. The forecast uses early orders, historic order curves and account data to estimate what the book will look like when the campaign closes.

How much of the order book do you need before forecasting?

It depends on the brand's customer mix and timing. The reliable way to find out is to back-test past campaigns and measure how accurate the forecast would have been at each week. Some brands get a useful signal early, others only once their largest accounts have ordered.

Can AI forecast demand for new styles with no history?

Partly. Models can borrow from similar past styles by attributes such as category, fabric, colour family and price band, and then adjust quickly as early orders arrive. Forecasts for genuinely new categories remain uncertain and should carry wider ranges.

Why do pre-order forecasts fail?

Common causes are missing order history, distortion by one or two large accounts, shifts in order timing between seasons and cancellations that are not recorded. Forecasts also fail when teams treat a single number as certain instead of working with a range.

GuideThe complete guide to AI in fashion wholesale and B2BRead the complete guide