Here's the link to the code pipeline: github.com/deep-logit-d...

We can also capture hard-to-quantify characteristics such as aesthetic similarity captured by images and functional benefits mentioned in reviews. Finally, our method can be scaled very easily across categories.

👉 What are advantages of our approach? We side-step the need define which product characteristics are relevant in a given category and instead extract information from product descriptions and reviews which likely mention the characteristics most important to consumers.

👉 Does this work well? In an online choice experiment we show that our approach predicts second choices better than characteristics-based models and it predicts substitution pattern well in real-world data across 40 product categories.

👉 What do we do? We propose an approach to extract product information from unstructured text and image data that we then use as an input for a mixed logit demand model.

🚨 An updated version of "Demand Estimation with Text and Image Data" with Ilya Morozov and Stephan Seiler is now available online:

papers.ssrn.com/sol3/papers....

We also make the full code pipeline available here:

A reminder that to many folks in the world, you would be considered among the wealthiest

If you think the rich are too rich, you don't have to wait for new policies, consider giving to @give-directly.bsky.social

www.givedirectly.org

For those attending #ASSA2025, come to our session on ML/AI + unstructured data (e.g., text and images).

We'll have a couple of papers on the econometrics of it, and I'll talk about incorporating this type of data in demand estimation.

It's Sunday Jan 5 at 10:15 AM.

www.aeaweb.org/conference/2...

For those attending #ASSA2025, come to our session on ML/AI + unstructured data (e.g., text and images).

We'll have a couple of papers on the econometrics of it, and I'll talk about incorporating this type of data in demand estimation.

It's Sunday Jan 5 at 10:15 AM.

www.aeaweb.org/conference/2...

If you receive a ref request and can't finish in the allotted time, you have options:

- Decline the review.
- Ask editor for more time. Then turn in report close to your new proposed deadline.

Don't do:

- Ghost the request
- Agree to request then go months past the due date w/no communication

Please spread the word to anyone who might be interested!

The conference will bring together scholars across fields who use Machine Learning, NLP, and other tools to extract valuable insights from new types of data, including:
• unstructured data
• clickstream data
• data generated by AI.

For more information, visit: www.chicagobooth.edu/research/kil...

🚨📜CALL FOR PAPERS 📜🚨

*New Data for Consumer Insights Conference*

When: May 30-31, 2025
Where: Chicago Booth
Submission Deadline: Jan 6, 2025
Keynotes: @susanathey.bsky.social and Greg Lewis

Both methodological and empirical submissions are welcome.

Submit here: ndconference2025.hotcrp.com