GARP-EFM: Improving Foundation Models with Revealed Preference Structure

TL;DR

This paper enhances foundation models' demand forecasting by integrating economic logic through fine-tuning on synthetic data generated from utility-maximizing agents, improving prediction accuracy.

Contribution

It introduces a method to incorporate economic rationality into foundation models using GARP-based synthetic data for better demand prediction.

Findings

Fine-tuning on GARP-consistent data improves forecast accuracy.

Economic theory can generate structured synthetic data for model enhancement.

The approach enhances predictions across multiple forecast horizons.

Abstract

Modern pretrained time-series foundation models can forecast without task-specific training, but they do not fully incorporate economic behavior. We show that teaching them basic economic logic improves how they predict demand using an experimental panel. We fine-tune Amazon Chronos-2, a transformer-based probabilistic time-series model, on synthetic data generated from utility-maximizing agents. We exploit Afriat's theorem, which guarantees that demand satisfies the Generalized Axiom of Revealed Preference (GARP) if and only if it can be generated by maximizing some utility function subject to a budget constraint. GARP is a simple condition to check that allows us to generate time series from a large class of utilities efficiently. The fine-tuned model serves as a rationality-constrained forecasting prior: it learns price-quantity relations from GARP-consistent synthetic histories and then uses those relations to predict the choices of real consumers. We find that fine-tuning on GARP-consistent synthetic data substantially improves prediction relative to zero-shot Chronos-2 at all forecast horizons we study. Our results show that economic theory can be used to generate structured synthetic data that improves foundation-model predictions when the theory implies observable patterns in the data.