Fit, Facts, Flair, and Figures: A Foray Into StichFix’s Data-Driven Styling Services
Stitchfix, a personal styling company that relies on data and algorithms to make recommendations to people, has yet to properly balance its use of human capital and compute power.
The same people who can’t decide whether or not Domino’s is a tech or pizza company will be similarly troubled when they think about StitchFix. StitchFix is an e-commerce company that offers highly-customized styling to men, women, and children. At a high level, the business model works by pairing customers with a (human) stylist who ships them (on a cadence of the customer’s choosing) boxes of hand-picked clothing, footwear, and jewelry that flatter customer’s bodies, meet their needs (clothing for work, parties, exercise, etc), and are affordable. Customers return what they hate and buy what they love. Data is collected and used extensively and in-real-time throughout the customer journey.
StitchFix gathers and uses data implicitly and explicitly, on both the front-end and back-end to efficiently and personally select and send clothing based on customers’ preferences. To cite a small number of examples of how these processes work, from StitchFix’s technical blog: (1) Specific data about customers’ bodies, taste in clothing, and price range are acquired through a 15-20 minute survey customer fills out upon making an account. (2) StitchFix algorithms select inventory based on customer’s responses. Stitchfix stylists (who filter the inventory once the algorithm picks it) are paired with customers using a complex algorithm whose inputs include, among other things, the history of relationship between stylist and client and the overlap between stylist and client tastes. (3) StitchFix closely monitors clients’ feedback on stylist selections to iterate for future shipments. (4) Data is constantly collected to optimize shipping routes for clothing.
Example of StitchFix’s data gathering method during new user onboarding:
The result of all of this? A not-so-lean, very mean, personal styling machine! Data is plentiful and cheap, but its utility is not. Deriving value from data depends on both the quantity and quality of data scientists working for Stitchfix. Such experts (who are generally expected to have at least an MS if not a PhD in data science) are in short supply and very expensive. According to Indeed, the average salary for a data scientist at StitchFix is a whopping (but well-deserved!) $240,000 per year. The reason might be found in Anaconda’s recent research that 63% of companies surveyed are worried that they won’t be able to find data scientists they need in the future. Of the almost 6,000 employees at StitchFix, over 100 are data scientists. And a lot more than fluency in statistics and Python is demanded of them. According to StitchFix’s blog, data scientists are hired to be lightly-supervised self-starters who are constantly asking and answering questions about how to use data to improve the user experience. StitchFix needs to think about how it will overcome the challenge of recruiting these highly-skilled, motivated individuals in the future.
While people might be the priciest, StitchFix has had to make other, substantial investments to create a “robust data platform.” Culture and digital infrastructure are critical to the success of harnessing value from data. Eric Colson, the former Chief Algorithms OFficer (yes this is a thing and shows you how seriously Stitchfix takes its data!) argues that the entire organization must embrace “things like learning by doing, being comfortable with ambiguity, and balancing long- and short-term returns.” It’s unclear how capital is deployed to create a culture like this, but one can imagine significant amounts of time are demanded from leaders across the organization to create it. It serves to note that at StitchFix, data science is an independent department whose head reports to the CEO, a rare division of labor, even in tech companies. Digital infrastructure is more quantifiable. Data scientists as StitchFix have been given access to compute power and cloud systems to support their work.
It would be irresponsible to conclude this blog post without acknowledging the fact that StitchFix’s performance, at least according to its stock price, is in the gutter. When the company went public in 2017, it was trading at ~$15 per share, and now it is down to ~$3. This disappointment can be attributed to several factors-including the company’s over-optimistic growth projections, TAM, a declining macro-environment, the rollout and then reversal of a cannibalizing line of business. But data-or its misuse-is worthy of blame as well. The Motley Fool’s analysis blames the company’s over-reliance on data and algorithms, instead of stylists. (These algorithms apparently recommended sweaters to people in July.) Stitchfix hasn’t quite struck the right balance (yet) between maximizing human and technical capabilities.
Nothing kills (or makes) dreams like data.
Sources:
https://algorithms-tour.stitchfix.com/#recommendation-systems
https://multithreaded.stitchfix.com/career/data-scientist/
https://www.indeed.com/cmp/Stitch-Fix/salaries/Data-Scienti
https://know.anaconda.com/rs/387-XNW-688/images/ANA_2022SODSReport.pdf
This is cool stuff, Aliza! I have two thoughts. The first is your point about the company’s over-reliance on data and algorithms, instead of stylists. It seems like the company’s structure is centered around data – which in turn establishes its resources, processes and priorities, rather than data being added into its processes. I wonder how StitchFix can restructure its RPPs to incorporate a more design-focused process. This brings me to my second point, which you noted when you began your post. It seems like many companies embrace the “Tech company” brand to attract funding opportunities. However, this branding then also seeps into RPPs and company culture. If StitchFix wants to be a tech-enabled clothing company, do they need to separate their external brand from internal processes, organizational culture and reporting structure? Has the tech company brand stopped them from implementing processes better suited to a clothing company?
Poor stock performance aside, I think one core issue with StitchFix has been their desire to use data to create what appears to be a human-to-human relationship, but ultimately feels empty once issues arise. Having a stylist is both a practical and an emotional relationship, and professional stylists often share an intimate connection with their customers in terms of understanding and choosing their “look”. StitchFix is trying to use data to replicate this at scale, but once something goes wrong, there’s no human there to physically apologize or learn, and so the customer becomes very (quickly) aware that at the end of the day, they are talking to an algorithm. The irony is that digital and human stylists may make the same mistakes over and over again, but the emotional ties of working with a person means customers may be more forgiving of the latter. I sense a parallel to ChatGPT here: when it’s working great, everything is awesome; but when it falls short, it can feel acutely dehumanizing.
Thanks for this, Aliza! I’m curious to understand whether you think this was a good idea to begin with? It may seem odd given that it’s a big business (though obviously performing poorly), but I wonder whether the use of big data brings anything genuinely unique to the equation. If one gave top-tier stylists from different backgrounds, the same individual’s measurements, could they not more cheaply and perhaps more creatively suggest outfits and styles that were genuinely stand out? Put differently, is big data primarily useful to capture the mean preference and if so, to what degree does that defeat the purpose in arenas like fashion where individual expression is typically intended not to conform to the mean?