AI demand forecasting in fashion: how it works and where it fails

Every fashion brand commits to production before it knows what will sell. Fabric is booked months ahead, factory capacity is reserved, and quantities are set on a mix of history, pre-orders and instinct. AI demand forecasting promises to make that bet more informed. It can, within limits that are worth understanding before investing in it.

How does AI demand forecasting work?

At its core, a forecasting model learns relationships between past demand and the factors that influenced it, then applies those relationships to the future. Traditional statistical methods look mainly at a product's own history and seasonality. Machine learning models can combine many more inputs at once, and they can share learning across similar products, which is useful when individual styles have short lives.

Typical inputs include:

• Historical sales by style, colour, size, channel and location
• Product attributes such as category, price point, fabric, fit and colour family
• Calendar effects: season start, holidays, promotions and markdown periods
• Wholesale pre-orders and early re-orders from retail partners
• Stock availability, so that periods of stock-out can be corrected for
• External signals where relevant, such as weather or web traffic

The output should be a distribution, for example a most likely quantity with a low and high range, rather than a single figure. Planners who see only the central number tend to treat it as certain.

Where does it work well?

The practical sweet spot is replenishment and size-level allocation for products with history. Here, models handle many items at once and react to recent sales faster than manual planning. Size curve prediction at category and store level is another area where pooled data gives useful results.

Read also

What is demand sensing and how does it work in fashion?

Where does AI forecasting fail?

Failures are predictable, which makes them manageable if they are acknowledged.

• Censored demand. When a product sells out, recorded sales show what was available, not what customers wanted. A model trained on raw sales learns to under-forecast best-sellers.
• Cold start. New styles have no history. The model relies on similar products, and the definition of 'similar' depends on how well attributes are recorded.
• Structural breaks. A new distribution partner, a pricing change or a shift in consumer behaviour makes the past less representative of the future.
• Promotions and markdowns. If discount periods are not flagged in the data, the model confuses price-driven demand with underlying demand.
• Small numbers. At store, style and size level, quantities are often so small that random variation dominates any pattern.

Why is wholesale data a strong forecasting signal?

Wholesale brands have an advantage that pure retailers lack: buyers commit to quantities before the season starts. Pre-orders are an early, real expression of demand from professionals who know their customers. Combining pre-order data with historical conversion from pre-order to final demand (re-orders, cancellations, sell-through at partner stores) gives forecasters a signal months before consumer sales begin.

The caveat is that pre-orders reflect buyer expectations, which can be collectively wrong. Re-order behaviour in the first weeks of the season is often the more reliable confirmation, which is why fast access to re-order and sell-out data matters as much as the model itself.

How should leaders evaluate a forecasting tool?

1. Ask for a back-test on your own data: the tool forecasts a past season without seeing its results, and you compare against actuals.
2. Measure accuracy separately for carry-over and new styles, because a blended figure hides where the model struggles.
3. Check that forecasts come with ranges and that planners can see the main drivers behind a number.
4. Compare against your current method, not against perfection. The question is whether the tool beats today's planning process.
5. Agree how planners can override forecasts, and track whether overrides improve or worsen results.

Accuracy also needs to be weighed by commercial impact. A forecast error on a high-volume core item costs more than the same error on a niche style, so evaluation should reflect value at risk, not just the average error across all items.

Read also

How does AI inventory optimisation reduce deadstock and stockouts in fashion?

What role remains for planners?

AI shifts the planner's job from producing numbers to managing exceptions. The model handles the bulk of stable items; planners focus on new launches, strategic bets, supply constraints and the qualitative information that never appears in data, such as a key account's change of strategy or a competitor's aggressive pricing.

The brands that get the most from AI forecasting are clear about this division of work. They let the model do what it does well, keep humans accountable for the final commitment and review forecast accuracy every season as a standing management topic.