How does AI inventory optimisation reduce deadstock and stockouts in fashion?
Fashion stock is perishable: too much ends in markdowns, too little in lost sales. Here is how AI models balance the two, what evidence exists and what they need to work.
KEY TAKEAWAYS Summary by the editors
- AI inventory optimisation combines demand forecasts at style, colour, size and location level with optimisation rules that decide how much stock to buy, hold, move or discount.
- Fashion is harder than most categories because many products are new each season, selling windows are short and demand is split across sizes and locations.
- In a peer-reviewed Interfaces paper, Zara's model for store shipments was reported to have increased sales by 3 to 4 percent, an early example of optimisation linking stock levels to demand.
- Zalando's published forecasting work treats pricing as a major lever for reducing the risk of end-of-season overstock, linking inventory and markdown decisions.
- Stockouts hide true demand: research on footwear found that nearly 25 percent of demand lost to stockouts spilled over to adjacent sizes, so models must treat sold-out periods explicitly.
AI inventory optimisation reduces deadstock and stockouts by forecasting demand at a fine level (style, colour, size, location, week) and then using optimisation rules to decide how much to buy, where to hold it and when to move or reprice it. The aim is not zero leftover stock, which would mean many lost sales, but the most profitable balance between the cost of excess and the cost of shortage.
Why is inventory so hard to manage in fashion?
Fashion combines several difficult features. A large share of each range is new, so there is no sales history for the exact item. Selling seasons are short, and production lead times are often longer than the selling window, so most quantities are committed before the first sale. Demand is split across many sizes and colours and across stores, web and wholesale partners, which multiplies the number of small, noisy forecasts.
The two failure modes have different costs. Excess stock is visible: it ends in markdowns, off-price channels or write-offs. Shortages are largely invisible, because a customer who cannot find their size simply leaves and no system records the lost sale. A good inventory model has to estimate both.
How does AI inventory optimisation work?
Most systems follow the same chain of steps, whatever the vendor or in-house approach:
- Clean the history: align product codes, remove one-off events and flag periods when items were out of stock.
- Forecast demand: estimate unconstrained demand by product, size and location, including price and promotion effects.
- Model uncertainty: produce a range of likely outcomes rather than a single number, because buying decisions depend on risk.
- Optimise decisions: choose quantities that maximise expected margin given costs of excess and shortage, budgets and constraints.
- Re-plan in season: update forecasts with actual sales and adjust replenishment, transfers and prices.
The step that is most often skipped is the uncertainty step. A buy quantity for a basic T-shirt that sells every week should be treated very differently from a trend piece with the same average forecast but much wider spread.
What evidence exists that it works?
Public, independently reviewed evidence is limited, because most results are reported by vendors. Two well-documented cases stand out. Zara worked with academic researchers on a model that determines inventory shipments from its central warehouses to its stores. According to the paper published in Interfaces in 2010, the system, which explicitly links stock levels and demand to choose replenishment quantities, increased sales by 3 to 4 percent, with estimated additional revenues of about 233 million US dollars for 2007 and 353 million for 2008.
Zalando has published how it forecasts demand for pricing decisions: a transformer-based model forecasting weekly demand 26 weeks ahead, in use in different versions since 2019. The authors describe pricing as a major lever to reduce the risk of being overstocked at the end of the season, which shows how closely inventory and markdown decisions are linked in practice.
Other brands have described their aims more generally. H&M told Retail Dive in 2020 that its AI team, set up in 2018, was working on quantifying how much of each item to buy, with the goal of getting the right product to the right place at the right time. Levi Strauss said in 2021 that AI-driven demand forecasting had improved accuracy and was expected to support more precise inventory investment and fewer markdowns.
Why do stockouts distort the data?
Sales are not demand. When a size sells out, recorded sales stop even though customers keep asking. A model trained on raw sales will therefore under-forecast exactly the sizes and stores that performed best. Research by Akchen and Caro, forthcoming in Manufacturing & Service Operations Management, studied a footwear retailer and found that nearly 25 percent of demand left unmet by stockouts spilled over to the next size up or down; the remainder did not turn into a purchase of a neighbouring size. Correcting for these effects, by flagging stockout periods and estimating the missing demand, is one of the most valuable things an inventory model does.
| Lever | Typical decision | What AI adds | Key data |
|---|---|---|---|
| Initial buy | Quantity per style and size before the season | Demand ranges for new items based on similar products | Product attributes, history of analogous items |
| Initial allocation | Stock per store or channel at launch | Location-level demand estimates | Sales by location, store profiles |
| Replenishment | Top-ups from warehouse | Frequent re-forecasting with recent sales | Daily sales and stock by location |
| Transfers | Moving stock between locations | Identifies where stock sells faster | Stock and sales by location, transfer costs |
| Markdowns | Timing and depth of discounts | Price response models | Price and promotion history |
What data and organisation does it need?
The data list is predictable: clean product master data, sales and stock history at size and location level, price and markdown history, and supply constraints such as lead times and minimum order quantities. Wholesale brands also need sell-out data from retail partners, otherwise they only see what was shipped, not what was sold.
Organisation matters as much. Buying, allocation and pricing are often run by different teams with different targets. An optimisation model that suggests a smaller initial buy and more in-season replenishment only creates value if supply contracts and warehouse processes allow that flexibility.
What are the limits?
- Forecast error for genuinely new items remains high; AI narrows it but does not remove it.
- Optimisation assumes cost inputs (markdown cost, lost-sale cost) that the business has to estimate and agree on.
- Small stores and slow sellers produce very sparse data, so store-level forecasts need pooling across similar stores.
- Vendor-reported improvement figures are rarely independently verified and depend heavily on the starting point.
Treated as a disciplined decision process rather than a tool purchase, AI inventory optimisation is one of the most practical applications of AI in fashion, precisely because its results show up in stock and margin figures that finance already tracks.
Frequently asked questions
Can AI eliminate deadstock in fashion?
No. Some leftover stock is the price of not losing sales. AI aims for the most profitable balance between excess and shortage, which usually reduces both markdowns and stockouts compared with manual planning.
What is the difference between sales and demand in inventory planning?
Sales are what was sold when stock was available. Demand includes customers who wanted an item that was sold out. Models must estimate this hidden demand, otherwise they repeatedly under-buy the best sellers.
How long does it take to see results from AI inventory optimisation?
Replenishment and markdown use cases can be tested within one season by running the model alongside the current process. Initial buy improvements take longer because results are only visible after the full season.
Do small fashion brands benefit from AI inventory tools?
They can, but sparse data limits accuracy at size and store level. Smaller brands often gain most from basic disciplines first: clean product data, stockout tracking and regular in-season re-planning.
One edition every weekday morning. Read in five minutes. Free for industry professionals.
SOURCES
- London Business School: Zara uses operations research to reengineer its global distribution system
- arXiv: Deep Learning based Forecasting: a case study from the online fashion industry (Zalando)
- UCL Discovery: Akchen and Caro, On Size Substitution and Its Role in Assortment and Inventory Planning
- Retail Dive: H&M's AI operation helps make its supply chain more sustainable
- Consumer Goods Technology: Levi's Doubling Down On AI to Power Demand Forecasting and CX
