Machine Learning for Fashion and Apparel Brands

Every season, the history resets

A supermarket sells the same tin of beans for years. A fashion brand sells a style for twelve weeks and then it is gone, replaced by something similar but not identical. You are always forecasting products that have never sold, which is why so many apparel businesses still buy on instinct and pay for it in markdowns.

Imagine a direct-to-consumer womenswear brand with 900 styles a year across three sizes of fit, selling online and through a few concession spaces. Returns run high, because they do in online fashion, and a fifth of each season's buy ends up discounted. Both problems are prediction problems, and machine learning for fashion brands is about both.

The fashion machine learning use cases worth considering

• New style demand forecasting. Predicting sales of an unseen style from attributes such as category, silhouette, fabric, colour family, price point and launch timing.
• Size curve prediction. Deciding how many of each size to buy per style, which is where a lot of broken stock and markdown starts.
• Fit and size recommendation. Suggesting the right size to a shopper based on their past purchases and returns, to cut bracketing where customers order two sizes and return one.
• Returns prediction. Flagging styles with fit problems early, from return reasons and review text, before reordering.
• Markdown optimisation. When and how deeply to discount each style to clear stock at the best overall margin.
• Attribute tagging from images. Using image models to label neckline, pattern and colour consistently, which feeds every other model here.
• Recommendations. Complete-the-look and similar-item suggestions online.

If recommendations are your priority, our piece on recommendation engines for ecommerce covers them in depth. Here we focus on buying, sizing and returns, where the margin sits.

Attributes are the whole game

Because every style is new, the model can only learn from what the style is like. If one merchandiser records a colour as "navy", another as "dark blue" and a third as "midnight", the model sees three colours. If sleeve length is recorded on half the range, it is nearly useless.

Image tagging with current vision models is now good enough to back-fill several seasons of attributes cheaply, with a merchandiser checking a sample. That one step often makes a forecasting project possible where it was not a year earlier.

Returns: prediction helps, but it is not only a model problem

Size-related returns are the largest single category for most online fashion brands. A fit recommendation model helps, particularly for repeat customers whose past orders and returns are known. For first-time shoppers it has much less to go on, and a clear, accurate size guide measured from the actual garments may do as much.

1. Record return reasons properly, with a short fixed list, before building anything
2. Find styles whose size-related return rate is well above their category, and check the pattern measurements
3. Add fit guidance on those product pages, such as "runs small, consider sizing up"
4. Then build a personalised size recommendation for returning customers

The order matters. Plenty of returns problems are really a grading or pattern problem in a handful of styles, which a model will politely work around rather than fix.

What it costs

Indicative ranges: an attribute clean-up and image tagging exercise, four to six weeks; a new style and size curve forecasting model tested against a past season, eight to twelve weeks; a fit recommendation feature on the website, ten to fourteen weeks including testing; a markdown model, eight to ten weeks once forecasting exists.

For a brand turning over a few million pounds, a small improvement in buy accuracy often outweighs these costs quickly. For a brand selling a few hundred styles a year through a Shopify store, good analytics and a disciplined review of last season's sell-through will likely do more than a model. Our guide to ecommerce for clothing and apparel covers the platform side.

Where fashion models go wrong

• Forecasting from sales that were capped by stockouts, which teaches the model that bestsellers were average
• Ignoring marketing: a style pushed in the email and on the homepage sells more for reasons unrelated to its attributes
• Trend shifts the history cannot see, where the model is confidently wrong about the new thing
• Buyers ignoring outputs they cannot interrogate

A model will tell you how a style like this sold before. It cannot tell you that everyone suddenly wants wide-leg trousers. That is still the buyer's job.

A sensible starting point

SpiderHunts would begin with a backtest: clean the attributes for two past seasons, train on one, forecast the other at the point the buy was placed, and compare against what the buyers ordered. If the model would have reduced both stockouts and end-of-season surplus, you have a case. If not, you have a clean attribute dataset, which is worth having regardless. Our machine learning services page explains how that backtest becomes a production tool.

Frequently asked questions