AI for Merchandising Part 1: The science of similarity

AI has already transformed fashion retail’s ability to sell to their customers

E-Commerce has driven a technology revolution in retail, capitalising on the rich data living in every online store to help connect customers with the things they actually want to buy.

Platforms like Shopify have also democratised AI, allowing any brand to “switch on” new capabilities with a few clicks in their app store - whether it be smarter emails or better on-site product recommendations.

But as sophisticated as the customer-facing part of the business is getting, we aren’t quite seeing the same love from the tech world for the back of house where all the most important decisions are being made.

Where merchandising is today

Access to information is key, and fast fashion retailers with short lead times can take advantage of a test-and-learn strategy allowing brands to back their winners and limit the impact of bad buying.

For example, DTC giants like Revolve and Shein employ this strategy, testing each new style with a minimal number of units and only producing or reordering at a mass level if the demand is there.  Through this intelligence, brands are able to better control stock levels and initiate fewer markdowns.

When will we see an AI revolution for the fashion buyer?

‍

One way that AI can boost merchandising is by preventing mistakes from happening in the first place and identifying the opportunities you’re not capitalising on.

‍

Do you really need to “test and learn” a fluoro lace crop top that looks very similar to something that absolutely tanked last year? And how far through the process will you get before somebody more experienced notices?

‍

‍

You also can’t “test and learn” every silhouette, every fabric, every occasion and every price point all at once - and AI can help you focus your efforts.

‍

And while these both seem like quite different problems, they are actually different sides of the same coin - both being similarity problems.

Know Your Algorithm - introducing “kNN”

Ever used Google image search? Or uploaded a photo of something you like to a fashion store to find things that are similar?

Chances are, that experience was powered by a variant of the K-Nearest Neighbors algorithm, commonly known as kNN.

kNN is a supervised machine learning algorithm, which in this case means it’s pre-trained on products you’ve ranged in the past, and analyses which items are in closest proximity to each other, i.e. each product’s “nearest neighbours”.

The “k” in “kNN” refers to a parameter of the algorithm - how many nearest neighbours to analyse (more is better, but also more expensive to compute).

It can simply be fed product images, or can be further enhanced by rich product attribution, and will turn your range into a multi-dimensional “map” of all your styles and where they sit relative to each other. Furthermore, it can predict where on the map an entirely new style will sit.

‍

Stop Repeating Your Mistakes

When you combine this virtual map with data about how each product has performed, then you can exploit this to predict how new products will perform - and to dodge outright dogs.

kNN will be able to show you the most similar products you’ve ranged in the past, including products that are similar in ways you might not expect, breaking down the silos and allowing you to learn from the mistakes (and winners!) of adjacent categories or occasions.

‍

Finding the White Space

As a general rule of thumb, if you have substantial demand for something, you’re going to have some demand for something that’s adjacent.  This insight can help buyers and merchandisers forge into new territory - whether it be extending their size range, or launching entirely new categories or even expanding into new markets.

...if you have substantial demand for something, you’re going to have some demand for something that’s adjacent.

But where is the opportunity? High price, or low price? Black or green? Rompers or dresses? Casual or cocktail? It’s probably easy to make each of these decisions independently, but there are so many different dimensions to a well-crafted range that it can be hard to hit the “bullseye” of all the right combinations that your customers want - but you don’t yet have.

Once again our hero kNN comes to the rescue. By slightly adapting the kNN algorithm, it’s possible to instead identify the “holes” in the range where similar products could exist, but don’t.  And then based on how the products around that gap perform, you’ll know where the safest bets can be made.

kNN can also be used in this way to identify high-performing products that hit their own unique sweet spot - but stand like “lone mountains” in your range (with plenty of space for new friends!).

‍

Cohesion vs Cannibalism

In its final form, kNN can give you a new health indicator for your range - a cohesion score, which measures how similar your range “feels” at any point in time.

This score can assist with your top-down planning: how many different colours or styles should we have? How many white dresses?

A range that’s too similar will cannibalise itself, but at the same time, a range that’s completely erratic has no identity or “brand cohesion”. Striking the right balance can be a challenge and artificial intelligence could help you understand what composition made your best seasons fly.

Want to know what happens next?  Sign up in the footer for Part 2 of Style Arcade’s AI for Merchandising series.

‍

Style Arcade is an AI-powered tool that can automate Trade Reports, calculate reorder and sizing quantities, collaborate on Range Plans in real-time, and predict opportunities and risks in your stock mix.  It’s an indispensable tool for your day-to-day merchandising and planning tasks.