Heuritech was founded in 2013 by PhDs in Machine Learning who believed that images could tell powerful stories if one knew how to read them. They developed a proprietary image recognition technology to analyze fashion images on social media. In today’s world, where millions and millions of pictures are shared each day on Instagram and other platforms, it’s a powerful tool to analyze within seconds what’s happening out there.
Social media has changed the fashion industry: now, trends are popping up every day and it’s hard to make sense of so much information, know what’s relevant, what is not, what will last and what won’t, and brands need to be able to anticipate – but only if they have the right tools to make sense of so much data. This is the reason Heuritech harnesses the scale of data, the power of artificial intelligence, and the expertise of PhDs in machine learning to deliver the best technology to the fashion industry to anticipate and produce better.
In this article, we will deep dive into our trend forecasting technology to understand how it works.
Heuritech’s methodology to forecast fashion trends thanks to artificial intelligence
Step 1: Defining audiences panels on social media to have a holistic view
100 million pictures are shared each day on social media. One can easily get lost in the amount of data, especially fashion pictures: indeed, according to Instagram, #fashion is the 4th most used hashtag on Instagram. Furthermore, the question to ask here is what, and whom, to analyze – indeed, we want to understand the underlying influence groups on Instagram to interpret the data and analyze it with purpose.
- What to analyze: at Heuritech, we specialize in recognizing products and details within images, because brands miss 78% of insights with text-based analysis only.
- Whom to analyze: We want to be representative of the whole population to be statistically reliable.
Therefore, we have defined tailored audience panels to analyze what is relevant on Instagram and Weibo. Criteria include:
- Accounts demonstrating an interest in fashion
- A large amount of accounts that keep on posting regularly to have a holistic view
- Representative of the different style segments of the fashion market: edgy, trendy and mainstream accounts
Since the goal is to be representative of what is happening on social media, we work only on aggregated data, i.e. anonymized data, so any personal information is discarded.
Let’s deep dive a little more into how we built these panels.
We automatically collect a random sample from social media to analyze. Random sampling allows us to prevent selection bias and thus to be representative of the social media population and what is happening in the market, i.e. how consumers are adopting products.
It gives us the following panels:
- Edgy accounts: People with bold and distinctive style. They represent the smallest segment. Still, their content is very niche and they tend to post often. They cover all the main fields of the fashion industry, whether they are professionals (stylists, journalists, influencers) or industry authorities (luxury, sports, fast fashion, high street, beauty).
- Trendy accounts: Fashionable people looking for the latest styles and who tend to help spread trends across the market. They are more substantial in number than edgy accounts but remain much more qualified and rare than mainstream accounts. They also tend to post often.
- Mainstream accounts: People looking for safe clothing choices and who will follow a trend rather than shape it. They represent the largest panel. Their content is more casual and occasional.
Furthermore, we create specific geographic panels to deep-dive into certain zones – Europe as well as specific countries, like the USA or China (particularly on Weibo and WeChat) to understand the regional differences and the common dynamics.
These panels allow us to capture the synergies of the market – whether they are coming from edgy, trendy or mainstream accounts. As we will see later, early signals will help us predict fashion trends.
Step 2: Applying our computer vision technology to millions of social media pictures
In any given picture, Heuritech’s proprietary computer vision technology tailored to fashion can detect and categorize more than 2,000 components, from shapes and attributes to fabrics, prints, and colors.
Combining these components allows us to precisely identify trends in details adopted by our panels. For example, we can recognize boots that are beige, pointed and have a kitten heel. Thanks to the millions of images analyzed each day, we are able to assess the volume of the beige, pointed, kitten heel boot trend. To ensure we are tracking a reliable trend signal and not noise from the wild, we make sure that we have enough social media images on any trend.
This way, we monitor relevant trends through time in order to help a merchandiser, buyer or designer understand how the market has been responding and help back their product choices with data. We can quantify at scale how a trend has evolved, qualify behavior and compare it with other trends in its category. Past data is an essential condition to get a 360° view on the market.
Step 3: Use our machine learning algorithms to predict fashion trends up to one year in advance
It goes without saying that the main interest when focusing on such a gold mine of data would be to know how these trends will behave. Therefore, we need to predict the trend’s behavior and indicate how it will evolve in the year ahead.
As we all know, predicting the future is notoriously hard, and a tricky endeavour. At Heuritech, we tackle this by using the best tools available in the forecaster toolkit: a series of forecasting methods and a master algorithm, whose purpose is to arbitrate between these methods to automatically obtain the best consensus.
Scientific literature is teeming with expert forecast methods. Most of them stem from a classical statistics approach*, relying only on the past dynamics of a trend to predict its future. While this approach provides good overall results and is state-of-the-art in many applications, it ignores the general context of a trend and may fail to predict what is most interesting: sudden, emerging, niche trends which will become the new, mainstream fashion trends.
To be able to forecast these seemingly random evolutions, we developed an in-house deep-learning approach** that feeds off of what we call early signals: those blips of activity among edgy influencers or sub-cultures segments of the trend which have the potential to tilt the way we perceive and wear fashion items. Using these early signals, our algorithms are able to detect the onset of new emerging trends before they realize their full market potential.
Finally, we trained a master algorithm*** to unify the different forecasts we obtain for each fashion item. Based on general information about the trend (product category, geography, past dynamic, etc.), the master algorithm is able to select the best combination of forecast methods.
The more we go back in time and the more we acquire data and learn about a trend, the better we can predict it. Therefore, even though our forecast model uses past data to predict the future, our forecast is thus constantly adjusted and accurate as we approach the projected date of interest.
Furthermore, as our technology can really go into detail, we can deep dive into different metrics: precise growth, compare the upcoming season with the previous one, as well as back assortment intuitions in the product development and buying stages.
* State-of-the-art statistical toolkit – outperforms all pure machine-learning approaches: Makridakis, S. et al. (2018) Statistical and Machine Learning forecasting methods
** Derived from Number 1 forecasting method of M4 competition: Smyl, S. (2019) A hybrid method of exponential smoothing and recurrent neural networks for time series forecasting
*** Derived from Number 1 ensembling algorithm of M4 competition: Montero-Manso, P. et al. (2019). FFORMA: Feature-based forecast model averaging
Curious about the top trends, brands, and cities from SS21 Fashion Weeks?
Step 4: Insert Heuritech’s data into our market intelligence platform
Heuritech’s exclusive data is integrated every day into The Suite, a dashboard used by brands to back their intuitions and make data-driven decisions on products, marketing, merchandising, stock, etc. The breadth and depth of Heuritech’s data science analysis allow leading brands worldwide to forecast more accurately throughout the product’s lifecycle, from trends, to demand and stock planning. Heuritech impacts brand sales, stock, and sell-through.