How WGSN are building a machine for predicting design trends
Men, I’ve got some fashion news for you. Over-shirts are on their way out, and cardigans are on their way in.
If you’re a menswear retailer or label, you should start reducing your investment in over-shirts and increasing your range of cardigans over the next 2 or 3 years..
I know this not because I’m particularly fashion forward or a maven of menswear. I mean, just look at me.
No, I know this because a machine told me so.
Even if you don’t know who WGSN are, you are almost certainly wearing clothes, using make-up, driving a car or living in a home that has been designed with advice from them.
The company was founded in 1998, originally to serve the fashion world. They now help designers and retailers to tap into trends before they happen across clothing, interior design, product design and even food and beverage. Everything from the shade of your lipstick, the silhouette of your jacket, the colour of your VW Golf, or whether your latte contains ‘adaptogens’ like turmeric.
And this isn’t just fallalery, it has a materially significant impact on business. An unwise investment, over or under buying, could cost a company hundreds of thousands of pounds in lost revenue or market share.
And like many areas of businesses in recent years, the trend forecasting industry has become more data-oriented. This reflects the move of the buying industry more generally from art to science, from mood boards to spreadsheets.
It was inevitable that at some point, the industry would turn to machine learning.
Mike Burgess, the VP of Product at trend forecasters WGSN, is the person responsible for bringing trend prediction into the AI age. He was brought into the business 18 months ago on the back of a successful 20-year career in the world of digital content — which included successfully pitching Bill Gates and walking away with a $1m cheque, building some of the earliest content websites and creating some of the first conversational interfaces to use AI-driven bots and natural-language processing (NLP).
“When I arrived at WGSN, there were a lot of analysts and data scientists in the building already. Really, we’re building on what has been happening here anyway,” says Mike.
“We’re just applying machine learning into the mix.”
The data that they have access to is a data nerd’s dream:
- Posts from over 2,000 key influencers, scraped from social media over the last two years, segmented by kind of influencer
- 22m images manually tagged by fashion experts
- Image recognition that detects colours down to their ‘Coloro’ code (which is a level up from Pantone in terms of colour specificity)
- Retail data across 248 retailers and 250m items (SKU’s in the jargon)
“We also have incredible sector experts. They have an in-built data filtering capability based on their experience, what they were seeing on the catwalk, and across cultural and demographic changes. They are great at spotting trends.
“For instance, they might ask, will the success of Tiger King mean an increase in leopard print? Maybe they saw that Harry Styles had worn a colourful chunky JW Anderson cardigan and wanted to know, will cardigans go mainstream?
“Based on those insights they would speak to our data analysts who would be able to delve into the data and perhaps see that there had been a 250% increase in sell-out on cardigans or a 20% increase in their appearances in lines. Meanwhile, over-shirts, which have the same functional role as cardigans, are showing up more as on-sale. From all that information, they would be able to confirm that, yes, this is a trend.
When WGSN ran the numbers for one client, it turned out they hadn’t made the most of the cardigan opportunity at all, especially compared to their competitors. That one item alone accounted for £400,000 of value left on the table for their rivals to mop up.
“We can do that from really big picture trends, right down to very specific items like what details and trims are going to be big in womenswear in spring/summer 2021”
(Textile manipulation, self-fabric trims, and craft-led openwork and embroidery, in case you were wondering)
It’s an impressive capability, but what Mike wanted to do was to turn the relationship between expert and analyst on its head. “What if rather than a human insight being validated by data, the machine could tell us what’s interesting and humans would validate the trend?”
‘Analyst-in-the-loop’ as he calls it.
So, rather than data analysis being provided as a service, it could become a product in its own right.
Enter The WGSN Trend Curve.
Trend Curve is intended to be the next generation of trend forecasting tools. The technology that will give its users an edge over their competitors, making them earlier on trends and more specific about their implications.
Using machine learning techniques, computers can spot signals in the data that are indicators of a future surge or downturn in a trend (a ‘change-point’ in the jargon). These are nuances that humans might not be able to detect.
And since WGSN is built on actionable insights — not only what the trend is, but what a client is supposed to do about it — Trend Curve also gives a depth of information about a trend that allows clients to manage their exposure to it.
“This is part of a transformation project. Millions have been invested, but doesn’t feel like a risk. It’s just something we’ve got to do,” says Mike.“It’s not only a question of timing, when to get into a trend, or when to get out of one, but also the volume of the trend, who’s buying it and how much to invest in it”, says Mike
“We have a ton of data about everything from colours, patterns and shapes and styles of clothes. But currently there is a high bar of data literacy required to understand and interpret it,” says Mike. “The aim is to put it into the hands of someone not a data analyst. We want to democratise the data.”
The Science Bit
This section contains the technical detail of how The WGSN Trend Curve works. If that’s not for you, please feel free to skip ahead to the next section
“First of the data is normalised, so that you can properly compare like with like. And then we do regression testing,” says Mike
Regression testing, in its simplest terms, is about finding a way to fit existing data to a curve. If you had a bunch of existing data, you can reverse engineer it to work out the algorithm that would have produced that data in the first place.
Once you’ve done that, you can use that algorithm to make future predictions about data you don’t yet have, to a certain level of confidence
“So for example, you might be looking at the trajectory of animal prints over time, how it’s come in and out of fashion. Our system might look at data from, say 2016, and select a test algorithm that it thinks will correctly draw a trend curve forward. Using the data we have from 2017, it then checks to see if its prediction was right. If it doesn’t quite fit, it will then keep refining and refining it until it has an algorithm that works. Once we’re confident, we can then start to project forward from 2020.”
Colour me good
“Colour is the biggest trend that people pay attention to. It doesn’t matter whether you’re designing clothes, cars or chairs — everything has to be a colour,” Mike points out.
As a consequence, the Global Colour Forecast from WGSN is one of their most anticipated and widely used reports.
Fortunately, analysing the colour of objects is something machines are really good at.
WGSN’s own image recognition technology tracks colour trends on social media, across thousands of images, locating product elements and their colour in images to a high level of accuracy.
One of the first outings of the Trend Curve was to confirm the predictions made in the S/S 2020 Colour Forecast and, perhaps more importantly, confirm how enduring they are likely to be.
Its key findings? Neo Mint, Cantaloupe, Cassis and Purist Blue all have a fair bit of life left in them whereas Mellow Yellow may have peaked, so handle with care and only buy in small volumes.
The future of the future predictor
WGSN is making a big play on The Trend Curve, part of a wider investment in data-driven technology in general. Parent company Ascential recently purchased Chinese ecommerce analytics business Yimian and US-based Flywheel Digital, which makes analytics tools to help drive ecommerce sales.
“This is part of a transformation project. Millions have been invested, but doesn’t feel like a risk. It’s just something we’ve got to do,” says Mike.
“All the clients we’ve spoken to know this is the direction of travel for this business — but obviously need to prove it works”
Having conquered colour, WGSN are also looking to automate food trend detection. Their latest report gives the nod to the oat milk explosion, prompting one retailer to bring forward plans to launch an oat-milk based ice-cream.
And WGSN customers themselves seem to be fans:
“ We see trends emerging and growing, but without supporting quantitative data, it’s often difficult to quantify the scale of the opportunity or ensure that we’re getting the launch timing right. The Trend Curve reports are both inspirational and data led, which makes them a really useful tool for our team.”
Emily Smith — Food Trends Researcher, Marks & Spencer
The people behind the data
One interesting quirk of the data science team at WGSN is that 80% of the data scientists on the team are women. Indeed including its head, Sara Gaspar
This is in stark contrast to what is still a very male dominated field. As much as 70% male according to one report.
“It is a challenge finding people who understand domain as well as data science”, admits Mike “We are very picky about finding data scientists who know what Bayesian Inference is, and also know what puff sleeves are.”
“In this industry there is mistrust of the typical Silicon Valley data-science ‘tech bro’ whose maths says ’this is going to be a trend’, but which falls down because they simply don’t understand the market they’re playing in.”
Has there been any push-back internally from WGSN’s style mavens?
“No, and I’m not just trying to sugar-coat it. There hasn’t been a problem internally. Bear in mind we’re dealing with analysts not the editorial team. If an analyst is writing a report, they want quantitative data. This helps give them better answers”.
Cardigan Bayesian
And as for that cardigan insight?
When WGSN ran the numbers for one client, it turned out they hadn’t made the most of the cardigan opportunity at all, especially compared to their competitors. That one item alone accounted for £400,000 of value left on the table for their rivals to mop up.
After the diagnosis, what did the AI suggest as a cure? A recommended 30% increase in cardigans in their assortment of knitwear.
Our image of sentient machines is often of cold, uncaring robots. But it turns out that sometimes the machines just want to make you a little bit cosier…
If you found this interesting, please leave a comment or get in touch
And please clap or share this article if you have enjoyed it as it helps other people find it.
If you’d like to be on my (100% spam-free) mailing list to be updated when a new article is published, you can sign up here.
