Economic Foundation Models

This project aims to create a first Science Foundation Model (EFM) that compresses massive amounts of granular, user-level data on economic behavior—specifically, clickstreams, potentially enriched with website screenshots—into a sequence model.

Inspired by large language models, an EFM is trained using self-supervision. In contrast to large language models, it carefully expresses heterogeneity among users.

Central to this effort is a sequence-modeling architecture, e.g., a transformer, trained to predict masked actions or screens in user histories together with a user-specific feature θᵢ. By continuously learning from the actual decisions people make in diverse contexts (online marketplaces, platforms, apps, etc.), the model captures both user-specific and market-level dynamics. The resulting policy π(action | screen, θᵢ) is paired with an empirical distribution of θᵢ to form an EFM “market.” This combination provides a powerful engine for downstream tasks, paralleling foundation models: By integrating over θᵢ, the EFM allows to compute market-level behaviors following a sequence of actions/screens, providing for market shares, demand elasticities, and a better understanding of different consumer types.  

When a new product enters a competitive environment, the EFM can integrate its learned distribution of user behaviors to generate robust estimates of market share or adoption rates. Because the model explicitly includes user identity and histories, it captures a level of personalization typically missing from standard industrial organization approaches. Moreover, by leveraging extensive micro-data with rich controls, we can address some endogeneity concerns systematically through deeper representation learning. Ultimately, the EFM framework offers a path toward more accurate, scalable, and responsive economic and investment decision-making tools, driven by truly heterogeneous user data and cutting-edge machine learning techniques.