How do you forecast demand for a fashion product with no sales history?
Most of a fashion range is new each season, so classic time-series forecasting has nothing to work with. How AI borrows history from similar products, images, attributes and early sales, and where it still falls short.
KEY TAKEAWAYS Summary by the editors
- AI forecasts a new fashion product without its own sales history by learning from comparable past products, described through attributes, images, prices and channels, rather than by extrapolating a time series.
- Attribute-based demand models estimate how much customers value features such as colour, fit or material, which lets a retailer forecast items it has never stocked.
- In a published study by Fisher and Vaidyanathan (Management Science, 2014), an attribute-based model forecast sales shares of new products with mean absolute percentage errors between 16.2% and 28.7% across three categories.
- Research on the public VISUELLE dataset of 5,577 new fast-fashion products found that adding Google Trends signals to image and metadata models improved forecast error modestly, not dramatically.
- Because pre-season forecasts for newness remain uncertain, the largest gains usually come from combining them with fast re-forecasting on early sales and flexible stock allocation.
A fashion product with no sales history is forecast by borrowing history from similar products. AI models describe each new item through its attributes, images, price and planned distribution, find past products that resemble it, and learn how those features translated into sales. The forecast is then refined quickly once the first days of real selling data arrive.
Why is forecasting new fashion products so hard?
Classic forecasting methods extend a product's own past into the future. That works for a continuous white T-shirt, but much of a fashion range is new each season: new silhouettes, prints and colourways with no sales record at all. Researchers who built a public benchmark for this problem describe new product forecasting as a challenging task that classical methods cannot solve, precisely because there are no past sales to model.
The difficulty is compounded by the fashion calendar. Buy decisions for many products are taken months before launch, demand depends on trends that shift during that time, and sales are often capped by the stock that was bought. A forecast that is wrong in either direction becomes markdowns or missed sales.
How does AI forecast a product that has never been sold?
Most approaches replace the missing time series with a description of the product and then learn from analogues. In practice that description is built from several layers of data.
- Structured attributes: category, fabric, fit, length, colour family, price band and season.
- Images: computer vision turns product photos or sketches into numerical features that capture shape and style.
- Commercial plan: planned price, number of doors or channels, launch date and marketing support.
- External signals: search interest or social data that indicate whether a style is rising or fading.
- Early sales: the first days or weeks of real demand, once the product is live.
Fisher and Vaidyanathan, writing in Management Science in 2014, showed the attribute logic clearly. Their model treats each product as a bundle of attribute levels, estimates demand for each level and the probability that customers substitute when their first choice is missing, and uses this to forecast products the retailer does not yet carry. The categories studied were not fashion (snack cakes, tyres and car care products), but the method transfers directly to apparel attributes.
How accurate are AI forecasts for new products?
Published results are a useful reality check. In the Fisher and Vaidyanathan study, forecasts of sales shares for new products had mean absolute percentage errors of 16.2%, 19.1% and 28.7% in the three categories, which the authors compared favourably with an earlier benchmark of 30.7%. Errors of that size are an improvement, not certainty.
In fashion specifically, Skenderi and colleagues released VISUELLE, a dataset of 5,577 new products sold by an Italian fast-fashion company between 2016 and 2019, with images, metadata, sales and related Google Trends series. Their multimodal model outperformed the baselines they tested, and adding Google Trends improved weighted absolute percentage error by 1.5%. That is a meaningful but modest gain, a reminder that external data refines a forecast rather than transforming it.
| Approach | How it works | Data needed | Main limitation |
|---|---|---|---|
| Like-for-like analogue | Planner picks one or more past items and adjusts their sales | Clean history of past styles | Subjective choice of analogue |
| Attribute-based model | Learns the demand contribution of each product attribute | Consistent attribute tagging across seasons | Misses effects not captured by attributes |
| Image and multimodal model | Combines image features, metadata and external signals | Product images, metadata, sales history | Needs large volumes of comparable products |
| Early-sales re-forecast | Updates the forecast from the first days of real demand | Fast, reliable sell-through data | Only useful if stock can still be moved or re-ordered |
What data does a brand need before it starts?
The quality of a new-product forecast depends less on the algorithm than on the archive behind it. Attributes must be tagged consistently across several seasons, otherwise the model cannot recognise that two products are similar. Sales history must be cleaned for stockouts, because a product that sold out early looks less popular than it was. Researchers working with the online retailer Rue La La addressed exactly this, using machine learning to estimate the sales lost on sold-out items before forecasting demand for products the retailer had never sold, according to Harvard Business School's Working Knowledge.
Images need to be comparable (same angle, background and crop) if they are used as features. And the commercial plan, including price and distribution breadth, must be recorded, because a product shown in fifty doors is not comparable with one shown in five.
Where does human judgement still matter?
Models learn from the past, so they are weakest exactly where fashion is most interesting: genuinely new categories, a collaboration, a trend with no precedent in the archive. Merchandisers and buyers add context the data does not hold, such as a marketing push, a celebrity moment or a supplier constraint. The sensible pattern is a model that proposes and a planner who adjusts, with the adjustments logged so that their accuracy can also be measured.
How do early sales change the forecast?
The most valuable data point for a new product is its own first week of selling. Zara's work with academics, published in Operations Research in 2015, built a system that updates forecasts from early sales and decides how much stock to send to each store initially versus hold back in the warehouse. In a controlled field experiment with 34 articles in the 2012 season, the system increased average season sales by about 2% and reduced unsold units at the end of the regular season by about 4%.
Those results illustrate the realistic shape of AI benefits in new-product planning: incremental percentage gains that compound across a large range, achieved by pairing a forecast with an operational decision that can still change.
How should a brand get started?
- Audit two to three seasons of product data for consistent attributes, images and price records.
- Correct historical sales for stockouts and for the number of doors each product was offered in.
- Start with one category where newness is high and the archive is large, and benchmark the model against the current planner method.
- Measure accuracy at the level decisions are taken (style, colour, size or store), not only in aggregate.
- Connect the forecast to a decision that can flex, such as initial allocation or a reserved re-order, so better forecasts actually change outcomes.
Frequently asked questions
Can AI predict sales for a new clothing product?
Yes, within limits. AI predicts sales for new products by learning from similar past products, using attributes, images, price and distribution plans. Published studies show measurable improvements over simpler methods, but errors remain substantial, so forecasts should be treated as ranges.
What is an analogue product in forecasting?
An analogue is a past product judged similar enough to a new one that its sales can guide the new forecast. Traditionally a planner chooses analogues manually; AI models select and weight many analogues automatically based on attributes and images.
What data do you need to forecast new fashion products?
You need several seasons of consistently tagged product attributes, comparable product images, sales history corrected for stockouts, and records of price and distribution breadth. External signals such as search trends can help but are a refinement rather than a foundation.
How much do Google Trends improve fashion forecasts?
In the VISUELLE study of 5,577 new fast-fashion products, adding Google Trends to an image and metadata model improved weighted absolute percentage error by 1.5%. The gain was real but modest, which suggests external signals complement rather than replace product and sales data.
One edition every weekday morning. Read in five minutes. Free for industry professionals.
SOURCES
- arXiv: Well Googled is Half Done, multimodal forecasting of new fashion product sales with image-based Google Trends (Skenderi et al.)
- IDEAS/RePEc: A Demand Estimation Procedure for Retail Assortment Optimization with Results from Implementations (Fisher and Vaidyanathan, Management Science 2014)
- Harvard Business School Working Knowledge: How do you predict demand and set prices for products never sold before?
- IDEAS/RePEc: Initial Shipment Decisions for New Products at Zara (Gallien et al., Operations Research 2015)
