How can AI help decide buy depth and production quantities in fashion?
Buying too deep creates markdowns and waste; buying too shallow loses sales. How AI combines demand forecasts, uncertainty and lead times to set production quantities, and why flexibility matters as much as accuracy.
KEY TAKEAWAYS Summary by the editors
- AI helps set fashion buy depth by turning a demand forecast into a probability range and choosing the quantity that best balances the cost of unsold stock against the cost of lost sales.
- The right quantity is rarely the average forecast: when margins are high relative to the cost of leftovers it pays to buy above the median, and when leftovers are costly it pays to buy below.
- Splitting the buy between an initial commitment and a later re-order or held-back reserve lets early sales data correct the forecast before all stock is committed.
- In Zara's published field experiment, a system that updated forecasts and allocated held-back stock increased season sales by about 2% and cut end-of-season unsold units by about 4%.
- The EU's ecodesign regulation (EU) 2024/1781 prohibits the destruction of unsold apparel, clothing accessories and footwear for large companies, which raises the cost of over-buying.
AI helps decide buy depth by converting a demand forecast into a range of likely outcomes and choosing the quantity that best balances the cost of leftover stock against the cost of lost sales. It does this per style, colour and size, taking lead times, minimum order quantities and re-order options into account. The best results come when AI sizes not just one buy but a sequence: an initial commitment followed by flexible top-ups as real demand appears.
Why is buy depth so difficult in fashion?
Production quantities for seasonal products are often committed months before the first sale, when uncertainty is highest. Too much stock leads to markdowns, storage costs and, increasingly, regulatory and reputational pressure over unsold goods. Too little stock means stockouts on products customers wanted, which also hide true demand from next season's data. Minimum order quantities, fabric commitments and long lead times limit how much these decisions can be corrected later.
How does AI decide how much to make?
The core logic is a classic inventory trade-off. For each product, the planner asks: if I buy one more unit, is the expected profit from selling it greater than the expected loss if it does not sell at full price? AI improves each input to that calculation.
- Demand distribution: instead of one number, the model forecasts a range, for example the quantity it expects to sell with 50%, 70% or 90% probability.
- Economics: full-price margin, markdown recovery, the cost of holding or disposing of leftovers, and the cost of a stockout including lost customer goodwill.
- Constraints: minimum order quantities, fabric availability, factory capacity and lead times.
- Options: whether a re-order, a held-back reserve or a later production run is possible.
The choice of forecast matters. Forecasting: Principles and Practice notes that minimising mean absolute error produces median forecasts while minimising squared error produces mean forecasts. For buy depth neither is automatically right: the target quantity depends on the economics, so high-margin core items may justify buying above the median, while low-margin trend items with poor clearance value may justify buying below it.
| Product situation | Cost of leftovers | Cost of stockout | Typical buy position |
|---|---|---|---|
| High-margin continuing line | Low (can carry over) | High | Above median forecast |
| Seasonal fashion item | High (markdown required) | Medium | Near or below median |
| Trend item, short life | Very high | Low to medium | Below median, with re-order option |
| Hero product with marketing support | Medium | Very high | Above median, reserve fabric |
| Fringe sizes | High relative to volume | Customer frustration | Tested by size curve, often shallow |
Why should the buy be split into stages?
The single largest improvement often comes from not committing everything at once. Early sales are the most informative data a new product generates, so keeping part of the quantity open, as unallocated warehouse stock, a reserved fabric position or a later production slot, lets the forecast be corrected before the full investment is made.
Zara's work with academics, published in Operations Research in 2015, illustrates the principle at the allocation stage. The decision support system balanced larger initial shipments to stores, which reduce early lost sales, against keeping stock in the warehouse to replenish once early sales were observed. In a controlled field experiment with 34 articles during the 2012 season, it increased total average season sales by about 2% and reduced unsold units at the end of the regular season by about 4%.
Which data improves buy depth decisions?
- Several seasons of sales history, corrected for stockouts, so that demand is not understated where products sold out.
- Consistent product attributes for matching new items to comparable past items.
- Size curves by channel, region or store cluster.
- Early demand signals before production, such as wholesale pre-orders, showroom feedback or online pre-launch interest.
- Supplier data: minimum quantities, lead times, capacity and the cost of reserving fabric or production slots.
Correcting for stockouts is essential. Researchers working with the online retailer Rue La La first used machine learning to estimate sales lost on sold-out items before forecasting demand for new products, as reported by Harvard Business School's Working Knowledge. Without this step, a model learns that successful products were less popular than they were, and recommends buying them too shallow again.
For brands that sell through wholesale, pre-orders from retail partners are a valuable early signal, because they arrive before production is finalised. Their usefulness depends on how well sell-in has historically predicted sell-out for comparable products, which is worth measuring before relying on it.
How does regulation change the cost of over-buying?
In the EU, Regulation (EU) 2024/1781 on ecodesign for sustainable products creates a framework to prevent unsold consumer products from being destroyed. Its recitals explain that destruction of unsold consumer apparel, clothing accessories and footwear is prohibited, that micro and small enterprises are exempted, and that the prohibition applies to medium-sized enterprises six years after the regulation entered into force. Companies other than micro and small enterprises must also disclose the number and weight of unsold consumer products they discard each year. Over-buying therefore no longer ends with disposal: leftover stock must be sold, donated, reused or carried, which raises its cost in the buy calculation.
What are the limits of AI for production quantities?
- Forecast uncertainty for new products remains high, so AI narrows rather than removes the risk.
- Supplier constraints often dominate: a minimum order quantity can override an optimal quantity.
- Models depend on accurate cost inputs, especially the true cost of leftovers and stockouts, which many businesses do not track.
- Recommendations need buyer review for strategic products, launches and collaborations where history is a weak guide.
Used well, AI changes the buy conversation from a single number to a set of scenarios with explicit risks, which makes trade-offs between margin, availability and waste visible before money is committed.
Frequently asked questions
How do fashion brands decide how much to produce?
Brands combine a demand forecast with product economics, supplier constraints and lead times. AI tools forecast a range of likely demand and recommend the quantity that best balances the cost of unsold stock against lost sales, often splitting the buy into an initial order and later re-orders.
What is buy depth in fashion?
Buy depth is the quantity bought per style or style-colour, as opposed to buy breadth, which is the number of different styles. Deep buys suit proven sellers with low leftover risk; shallow buys suit uncertain or trend-driven items, ideally with a re-order option.
Can AI reduce overproduction in fashion?
AI can reduce overproduction by improving demand forecasts, quantifying uncertainty and supporting staged buying based on early sales. Published results show incremental gains, such as about 4% fewer unsold units in Zara's 2012 field experiment, rather than eliminating surplus altogether.
Is it illegal to destroy unsold clothes in the EU?
Regulation (EU) 2024/1781 prohibits the destruction of unsold consumer apparel, clothing accessories and footwear, with micro and small enterprises exempted and medium-sized enterprises covered six years after entry into force. Companies other than micro and small enterprises must also disclose how many unsold products they discard each year.
One edition every weekday morning. Read in five minutes. Free for industry professionals.
SOURCES
- IDEAS/RePEc: Initial Shipment Decisions for New Products at Zara (Gallien et al., Operations Research 2015)
- Hyndman and Athanasopoulos, Forecasting: Principles and Practice (3rd ed.): Evaluating point forecast accuracy
- Harvard Business School Working Knowledge: How do you predict demand and set prices for products never sold before?
- EUR-Lex: Regulation (EU) 2024/1781 establishing a framework for ecodesign requirements for sustainable products
