Prof. Victor Chernozhukov | Adventures in Demand Analysis Using AI

Prof. Victor Chernozhukov | Adventures in Demand Analysis Using AI

by Victor Chernozhukov

Humanities, Social Sciences & Thought Economics & Finance KCEconomicsKCHEconometrics and economic statistics
🎙 Victor Chernozhukov 👥 8K 📅 January 28, 2026 ⏱ 59 min 👁 1K 📄 expert opinion 🧭 2026-08-15
Available in: English (current) Français

Keywords

demand elasticityAI embeddingstransformershedonic pricingcausal machine learning

Summary

Victor Chernozhukov presents his research on using AI and transformer-based embeddings for demand analysis, focusing on price elasticity estimation. He motivates the work with pedagogical goals and a tribute to JASA’s historical role in empirical demand analysis. The study uses public Amazon toy car data, constructing quantity and price signals from sales ranks and buy-box prices. They generate embeddings from text, images, and tabular data using transformer models (BERT, ViT, SAINT) and fine-tune them for prediction. Qualitative checks show that embeddings capture product similarity consistent with human perception, and clusters are meaningful. Quantitative results show that embeddings significantly improve out-of-sample prediction of price and quantity levels compared to traditional tabular features, but not for changes. The main finding is that embeddings act as strong effect modifiers for price elasticity but are weak confounders once past prices and visibility signals are controlled. The talk concludes with implications for hedonic pricing and inflation measurement, and mentions an accompanying online book.

158 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk provides valuable insights into the application of modern AI techniques to classical econometric problems. The argumentation is solid, grounded in a rigorous research paper and supported by empirical results. The speaker clearly explains the methodology, including the construction of embeddings, fine-tuning, and the causal inference framework. The findings are presented with appropriate caveats, such as the limitations of using sales rank as a proxy for quantity and the need for large datasets to fully exploit image information. The pedagogical motivation and the connection to historical literature add depth to the presentation.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is high, as the talk is based on a peer-reviewed paper and a working paper available on arXiv. The speaker cites relevant literature, including the work of Philip Wright and the hedonic modeling tradition. The title accurately reflects the content. The talk is part of a seminar at the Isaac Newton Institute, which adds credibility. The description provides links to the seminar page and the institute’s website, but no direct links to the paper or data are given in the description. The speaker mentions the paper is on arXiv and the data is public, but specific URLs are not provided in the video description.

215 words

Title / Content Match

The title accurately reflects the content, which is a presentation of research on demand analysis using AI methods.

Quality & Reliability

8/10

The talk is given by a leading econometrician (MIT professor) and is based on a peer-reviewed paper (Journal of Econometrics) and a working paper on arXiv. The methodology is clearly explained, and the data is publicly available. However, the talk is a seminar presentation, not a formal publication, and some details are simplified for a general audience.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

Concurring Sources

Contribution & Novelties

The talk presents a novel application of transformer-based embeddings to demand analysis, demonstrating their value in predicting prices and quantities and in estimating price elasticity. The finding that embeddings are strong effect modifiers but weak confounders is an important contribution to causal inference in economics. The use of publicly available data and the pedagogical approach make the methods accessible.

Pour aller plus loin :

  • Double Machine Learning — A framework for causal inference with high-dimensional controls, relevant to the residual-on-residual approach mentioned.
  • Hedonic regression — The method used to model prices as a function of characteristics, central to the talk.
  • Transformer (machine learning) — The architecture underlying the embeddings used in the study.

113 words

Radar Profile

The radar profile shows high scores in quantity of information, quality of information, and reliability, with a slightly lower technical level, indicating a well-balanced presentation that is rigorous yet accessible.

Reliability 8/10