Mingwei Shen

Part 2 of “how does the (former) VP of #forecasting at Amazon, Ping Josephine Xu, think about forecasting in general?” Two sets of problems that she dealt with: top-line revenue and item-level forecasts. Jan and me call this strategic and operational forecasting problems. Those are super different. This post is about Ping’s view on operational forecasts. (i) For operational forecasts, the key to success is a fast iteration cycle from idea to experimental verification in backtests. So the backtesting and benchmark infrastructure is paramount – and more important than the modelling ideas themselves. (ii) Another important consideration is the architectural cost of having many models. The temptation is to start simple and then add models on top. E.g., one model for shoes, another one for luggage. One of cold start items, one for early life, one for end of life. However, then one may want to invent a second order model that picks the right forecasting model and then another one that mixes/ensembles them together. This leads to confusion. Ping has found that moving this complexity behind a single model has been key. For her, this is the value that deep learning has brought to her practice, independent of which specific models these are. Going from 10+ models to a single model brings architectural simplicity. (iii) Finally, it is important for the forecasters when arguing about the value of the forecast to not only be able to talk about this quantitatively (via A/B test results) but also qualitatively. For this, it is crucial to know what the forecast is used for. Forecasting is a means to an end as Jan and me always say. Therefore, it’s key to know the end goal – typically a downstream decision problem.   More in-depth discussion in Jan Gasthaus and my course on Maven “Modern Forecasting in Pratice.” Starts in 3 weeks, October 6, registrations are still open. The picture is from Jan Gasthaus Yuyang (Bernie) Wang Valentin Flunkert Christos Faloutsos and my 2020 tutorial on forecasting. This aged well! This is not related to my work at Databricks.

143

6 Comments

Think about what Amazon looked like back in 1999. Now open Amazon — notice something surprising? Amazon circa 1999. Product tiles on white background. Search bar. Basic navigation. Amazon 2026. Still, Product tiles on white background. Search bar. Basic navigation. Twenty-seven years of the most aggressive technological innovation in human history and the way we shop online looks almost identical. Every other consumer experience transformed. Stock trading went mobile-casino. Music went streaming. Travel went from agents to algorithms. Dating went from personal ads to swiping. But e-commerce? We're still looking at product photos on a white background and hoping for the best. The interface hasn't changed because the fundamental model hasn't changed: here's a picture of a thing, here are some reviews, do you want it? That works for books. It works for phone chargers. It works for anything where the product exists independently of where it ends up. But what about the category where context is everything? Where the product doesn't exist independently of the space it goes into? Furniture. Home goods. The things that define how you actually live, InHouse. We're buying couches the same way we bought books in 1999. And somehow nobody thinks that's strange. #ecommerce #spatialcommerce #AI #interiordesign #furnishing

54

20 Comments

OpenAI is killing Instant Checkout inside ChatGPT. I was one of a small number of merchants in the Shopify beta. The results were what you might expect: practically zero. The problem was never the tech. It was behavior. People used ChatGPT to research. Then they went somewhere familiar to buy. Chat intent isn’t purchase intent. The Amazon timing is worth noting. A $15B investment announced the same week OpenAI retreats from direct commerce.🤔 DTC wins for a reason. You own the customer. You own the data. You control the story, the narrative, the brand, the experience, the moment of purchase. No disintermediation. Discovery is shifting. The transaction layer isn’t. #ecommerce #DTC #ChatGPT #agenticcommerce #retail

63

23 Comments

A pattern I see across supply-chain businesses: Most problems are not caused by lack of effort. They’re caused by gaps in visibility, coordination, and decision timing. I wrote a short piece on the operating reality facing modern supply-chain organizations and how high-performing teams are responding differently. If you’re in packaging, manufacturing, or distribution, this will likely resonate. 👉 https://lnkd.in/guedw2Ty

12

Today we announced new capabilities for our Shop Direct and Buy for Me experiences on Amazon, enabling now feeds to facilitate external merchants to list reliably their product listings and reach millions of Amazon customers. Excited on this feature, that is a win/win experience for customers, merchants and brands. Excited to be part of such important project.

27