of Shopify, recently told his employees in an internal memo: “Before asking for more headcount and resources, teams must exhibit why they can not get what they need done using AI”.
Having worked in startups for the past 6 years, asking for more headcount or more resources is frequently not an option anyhow. Constraints are tight and also you generally have to carefully put money into projects you might be confident can be impactful. So in these situations, Tobi would probably rephrase: “Suck it up and just use AI when you can”.
As a Data Scientist, I need to know how our work is evolving with AI. Tech Executives are clearly expecting every team to be more efficient and more creative. But can a multi-billion parameter model, even though it has read all the Web, be systematically helpful at solving your personal problems? To tackle this query, I’m proposing the next framework: let me undergo all of the projects I even have worked on for the reason that starting of my profession and assess how much AI would have helped.
Today, we return to 2020. I’m a junior Data Scientist at an organization that has been hit pretty bad by the pandemic: Rent the Runway.
What the Project was about
Rent the Runway was launched in 2009. The corporate experienced rapid growth from 2016 to 2020, after introducing their hottest product: a monthly “unlimited” subscription to fashion, aka “Closet within the Cloud”, allowing you to rent an enormous variety of high end clothes at an unbeatable price. The product was a success for each woman wanting to frequently wear something latest at work, night outs, parties, special events etc. So obviously, when Covid began in March 2020, and everybody stopped going out for weeks… well, it kinda killed the vibe.
The “Netflix of fashion” (yes, some people really used that nickname) ended up with an insane amount of unused inventory, a whole season of things that may just must “sit” in a warehouse, and naturally an enormous revenue decrease. It was urgent to seek out a brand new revenue stream to survive financially. Not the proper time to ask for more resources or headcount, as a third of the workforce was furloughed.
Here got here a superb idea: what if we were trying to return back to the retail business? That’s, selling items as second hand as an alternative of renting them. But here was the large query: because the lockout goes to finish sooner or later and individuals are going to return to renting, what items should we keep for now vs. sell for a reduction? And how much should this discount be?
The 2020 Solution
The goal of the project is to get for every product the optimal price, that can be the right balance between renting and selling. You’ll be able to get the optimal price p as the worth that may maximize the next:
Which is simple to seek out… assuming you realize the future rental revenue (the “RentalRev” on this equation) and the price elasticity (the chances on this equation).
In early 2020, I used to be already working on RTR unit economics and revenue forecasting. I used to be constructing a model to predict, based on an item rental history, what number of more times it could possibly be rented and what additional revenue it might generate.
The missing piece was having an idea of pricing elasticity, i.e answering the query: given a price for an item, what could be the probability of selling it? To know more about this model, I’d redirect you to this very detailed and well-written blog article by my teammate Meghan Solari.
It is vital to notice that some business constraints needed to be applied to be certain that we’d not unload a whole style and keep some units for rentals.
How AI could have helped
This project is near a classic demand and provide problem, with the twist of the rental vs retail revenue that makes it a bit more interesting. But finding the equation that offers the optimal price is just not the principal challenge. The principal challenge is tips on how to estimate each parameter given insufficient data.
Indeed, predicting future demand is tough: you only have a couple of months of history (at best) for every style, and you must predict a big horizon (mainly as much as end of life). Rapid changes in fashion trends require a deep understanding of the industry to be predicted, if predictable in any respect. And the uncertainty created by the early Covid period made any time series models very hard to construct.
Estimating pricing elasticity is just not any easier. As Rent the Runway was not a retail business, sales data was by design limited.
And that’s exactly where the challenge would come for any AI-driven solution as well. An AI can only be nearly as good as the info it’s being provided.
Solving for the sparse style-level data
Although each style has limited history, there’s a wealth of data in similar items. That is a primary use case for transfer learning and shared embeddings that would have been made easier by the access to pre-trained LLMs. Shared style-level embeddings could have allowed us to make strong assumptions on latest styles based on metadata: color, brand, price, fabric, silhouette… We could have more effectively built models that learn tips on how to predict demand curves from a couple of data points, drawing from patterns in historically similar items. An organization like Stitch Fix has been pioneering this space by using item metadata to create deep embeddings that generalize across latest inventory.
Maintaining with Fast fashion cycle
LLMs could have made it easier to follow and understand ever-changing fashion trends and work on external signals to predict potential shifts in all the industry. That was not something that was easy in 2020, since it requires scrapping massive amounts of information, checking out what’s relevant and interpreting weak signals. Today, that is precisely what LLMs are good at. Corporations like Trendalytics do exactly that, scanning TikTok, Google Trends, and social media to surface emerging patterns in silhouettes, colours, or influencers’ posts. That data would have been extremely helpful to make an accurate demand forecast.
Constructing a dynamic pricing Agent
One last item that would have been fun to explore, given today’s technology, is to construct an agent that will have modified the costs in real time and learnt, through reinforcement learning, the optimal pricing strategies by interacting with the environment. That would have allowed us to be certain the prices rely on the style’s historical and future demand but in addition on the customer features, i.e personal rental and buy history, engagement, taste, etc. That would have brought us closer to what top RL teams at Airbnb or Uber do, constantly adjusting prices based on real time demand and booking probability.
These are a number of the ideas that I selfishly would have been super excited to work on, but note two essential things:
1. From a product perspective, it is actually hard to estimate (especially now that I don’t have access to the info anymore) what the impact on overall revenue would have been.
2. These ideas could have also been built in-house back in 2020, given the nice team of ML engineers we had at Rent the Runway. Nevertheless it would have represented months — if not years — of research and development with high risks, which we couldn’t afford at the moment.
And that’s probably my principal takeaway to this point on LLMs: they don’t trivialize the issues we used to bang our heads at 5 years ago (or not yet) but they make it easier to check ideas that will have taken an unrealistically very long time to develop back in the times. This changes the paradigm through which Data teams typically operate and opens latest opportunities of partnership with Product teams.
