Stitch Fix is a fashion retailer in the sense that they take inventory, do merchandising, buy at wholesale price and sell at retail price. But that is about where the similarity with a traditional retailer ends.

Stitch Fix’s business model involves users submitting detailed body measurement data and style preferences which is then used to feed into the Stitch Fix machine learning model. Stitch Fix will combine algorithm recommendation and the wisdom of a human stylist to pick out 5 pieces of clothing and sent them to users in a box which is called a Fix. Users don’t know what clothes they are going to get until they receive it. So there is a surprise and delight element. Users will pick what they like and return what they don’t like.

Stitch Fix’s Data Flywheel

Stitch Fix is very exciting to me because it benefits from the Virtuous Cycle of Data. Google search is the classic example here. As more users use Google search, Google acquires more data which makes its search results more accurate which attracts more users. More users generate more data which makes the search result even more accurate. And the cycle continues.

Virtuous Cycle of Data

Stitch Fix’s business clearly benefits from Virtuous Cycle of Data as well. As more users use Stitch Fix, they collect more data which helps to improve the algorithm’s recommendation accuracy. This should attract more users and the cycle continues.

Many investors seem to view Stitch Fix as a consumer subscription business. Stitch Fix’s founder Katrina Lake has been very clear from Day 1 that Stitch Fix is a personalization company. And at the core of personalization, it is all about using data and machine learning to help customers find the clothes that they personally like and fits them well.

Similar to web searches, fashion has a very long tail i.e. it has a large number of SKUs and the longevity of each SKU is limited and hence the search cost is very high. Machine learning is very well suited to solve this kind of problem.

Okay, it is all well and good that Stich Fix benefits from the Virtuous Cycle of Data but its competitors, such as Amazon, should also benefit from the same dynamics. So is there really any competitive edge to Stitch Fix?

Stitch Fix vs TikTok

I think Stitch Fix still has unique advantages over its competitors because its data quality structurally better and this insight comes from understanding TikTok’s success.

As explained by Eugene Wei’s blog post – Seeing Like an Algorithm, he wrote:

“TikTok fascinates me because it is an example of a modern app whose design, whether by accident or, uhh, design, is optimized to feed its algorithm as much useful signal as possible. It is an exemplar of what I call algorithm-friendly design”

He meant that, unlike Facebook / Twitter where the vertical flow has a lot of content going on the screen at any point in time, TikTok’s screen only has one video playing at any one point in time. This led to an incredible clean data signal of the user’s preference for the video. For example, with three posts on a screen at any point in time, Facebook gets strong signal about which post the user like if the user clicks on one of the three posts but doesn’t know with a high degree of certainty if the user doesn’t like the other two posts. Since there is only one video on display each time, TikTok knows very precisely how the user feels about each video – both positive and negative sentiments. In this way, TikTok is able to collect very high-quality data from users which are then used in a machine learning model to serve more targeted videos to its users. Given TikTok’s better precision, users find it acceptable to watch one video at a time. Eugene argued that Tiktok’s innovation comes from its algorithm-friendly as opposed to user-friendly product design of one video per screen to maximize the power of the machine learning algorithm.

This is a very useful insight because one can easily draw parallels to see if there are other companies that design their product to optimize for high-quality data collection.

I believe that Stitch Fix’s business is designed to be algorithm-friendly. First, Stitch Fix took away users’ ability to choose and try on clothes means that users are willingly sharing very high-quality data with Stitch Fix because they want clothes that match their style and actually fit them. No other fashion retailers are able to persuade users to share this data.

Similar to TikTok, Stitch Fix knows very precisely how customers feel about each piece of clothing, both positive and negative feelings, that was sent to them. Users are then asked to provide detailed feedback on each piece of clothing. Overtime, Stitch Fix’s algorithm should have a distinct advantage in its algorithm’s predictive power derived from its data quality.

Again no other fashion retailers have such high-quality data. Traditional fashion retailers such as Zara and Uniqlo only know what customers on aggregate likes and dislike. Online fashion retailers, such as Asos, know what individual customer likes. But only Stitch Fix knows what individual customers like AND dislike.

However, unlike TikTok, Stitch Fix is about moving atoms rather than bits and hence it takes a much longer time to collect data for Stitch Fix. While TikTok collects a new data point every time a user swipes the video and users spent hours on TikTok every day, Stitch Fix users on average order a fix on monthly basis and that means only 12 data points collected each year. So Stitch Fix’s data accumulation process takes much longer.

As Stitch Fix continues to gain scale, its algorithm is going to get better. Said in another way, Stitch Fix’s moat strengthens with scale. And fashion retailing is a big market. So there is a very long growth runway for Stitch Fix. This combination of a big market and strengthening moat with scale is a recipe for huge value creation in the long run.

Direct Buy and Consignment Model

And Stitch Fix is doing very interesting things to open up a larger addressable market with Direct Buy and injecting third-party inventory model.

Stitch Fix started with a “5-piece-in-a-box” model and utilized a surprise and delight model. The advantage of starting with this model is that users are willing to share their detailed body measurement data with Stitch Fix. But the disadvantage is that only a small fraction of the population likes the surprise and delight model and the unit economics is not far from satisfactory (return logistics is very costly).

There are three types of consumer demand:

1) full certainty demand – consumers know exactly what they want to buy. For example a pair of sneakers of Nike

2) semi-certain demand – consumers know they want to buy a pair of jeans but they don’t know which cut/fit /color and brand

3) zero certainty demand – very weak purchase intent and just want to explore what is available; and if they see something they like, they might buy it. And they don’t care if it is going to be a jean or a shirt.

Clearly, semi-certain demand has the biggest market size while the market sizes for full certainty and zero certainty demand are smaller. Stitch Fix mostly serves zero certainty demand while retailers such as Zara and Uniqlo mostly serve semi certain demand. It is clear that Stitch Fix is hitting the ceiling of the zero-certainty market size as evidenced by less than stellar revenue growth in recent years.

Stitch Fix Revenue 2015 – 2020

The next stage of development for Stitch Fix is to move into the semi-certain demand to continue to grow its business. Given what Stitch Fix already knows about a customer, can Stitch Fix recommend the right jeans for a customer if he/she is specifically on the lookout for jeans? With Direct Buy, Stitch Fix is exactly trying to serve this market.

And Stitch Fix is opening up its inventory model to increase supply. Historically, Stitch Fix would buy its own inventory. Now, it is exploring how to do a consignment model where fashion brands can place inventory with Stitch Fix to sell for them. Stitch Fix only charges a fee when the inventory is sold.

I believe with Direct Buy and consignment model, Stitch Fix’s flywheel should accelerate.

Stitch Fix’s AI-first organizational structure

In this video below, Andrew Ng (legendary AI researcher) explained that a shopping mall with a website doesn’t make it an Internet company. In the same vein, a company that uses machine learning doesn’t make it a first-rate AI company. He believes that first-rate AI companies need to elevate the importance of the AI team within the business.

Lecture 1.5 — What makes an AI company? — [AI For Everyone | Andrew Ng]

Stitch Fix was lucky to have Eric Colson who clearly understood Andrew’s point and designed Stitch Fix to be a first-rate AI company from the ground up.

We decided from the beginning to elevate the data science team to a top-level department. This isn’t just symbolic. To bring transformative change to the business, the data science team needs the level of accountability and influence that officer representation provides.” – Eric Colson

And Eric explained Stitch Fix’s unique org structure here – https://cultivating-algos.stitchfix.com/. Incidentally, Stitch Fix created this webpage as a recruitment tool to attract the best data scientist.

Data Science is a top-level department within Stitch Fix with its own budget

The data scientists team is organized around business capabilities. For example, the merchandising team will partner with the data scientist team that is responsible for the demand forecasting algorithm.

Data scientists are organized by business capabilities and able to take end to end responsibilities for each algorithm

We decided early on to primarily organize our data scientists by business capability (though still centralized for better knowledge-sharing and career pathing). Groups of related capabilities form teams and collections of teams make-up the department. Roles are verticalized, orienting them towards business impact and making them more agnostic to technical function. This enables a tenacious data scientist to work on any part of a solution: from conception, modeling, and ETL to implementation and measurement.

This creates a very unique organizational structure that integrates AI capabilities into the very fabric of Stitch Fix. This will translate into excellent execution competence relative to its competitors who are still stuck in the Internet company era. While it is hard to assess the quantum of this advantage, I believe it will help Stitch Fix to expand its advantage over competitors.

Valuation & investment thoughts

Powered by its data flywheel and unique product design format, there is a reasonable chance that Stitch Fix could become a bigger business with a stronger moat in the next 5-10 years. But it is far from a guarantee. There are quite a few risks. Stitch Fix’s effort to expand the market could fail and it would be confined to the zero-certainty market segment. Fashion changes very quickly and maybe its machine learning model needs to be constantly retrained with new data which might diminish its data-driven competitive moat.

That said, I think the valuation is reasonable in the sense that I still get a large enough upside to compensate for the risks.

If Stitch Fix is able to expand its addressable market through Direct Buy and consignment model, then its revenue growth should accelerate from its current rate of mid-teens growth rate. Assuming that it can grow revenue at 40% CAGR for the next 5 years, its revenue could be USD 9bn. Assuming that it can achieve a 5% net profit margin (benchmarking online and offline fashion retailers), then its net profit in 2025 is USD 450m. Putting a 20x multiple on USD 450m gets to USD 9bn of market value. This implies ~100% upside from the current market cap of USD 4bn.

Consider the above as a base case scenario. In reality, I think if my investment thesis for Stitch Fix indeed works out, it should be able to maintain a high revenue growth rate for many years beyond 2025 which means the terminal multiple of 20x is too conservative.

If Stitch Fix fails to expand its market size and accelerate growth, then it could probably still manage 10% revenue growth. In which I calculate the downside to be ~50% from the current market value.

So on balance, this feels like a pretty good risk-reward to me. But only time will tell.