Causal Inference with Time Series Foundation Models

Cyrus Illick, Saeyoung Rho, Vishal Misra

Keywords: Synthetic Control, Time Series Foundation Models

TL;DR: We study the use of Time Series Foundation Models for synthetic control and compare them against classical linear synthetic control methods.

Abstract: We study counterfactual estimation with time-series panel data where multiple units are observed and only one unit undergoes an intervention. Synthetic control (SC) addresses this problem by treating post-intervention outcomes of the treated unit as missing and imputing them using the rest of the panel. This can be viewed as a special forecasting problem in which future observations of untreated units are available. Motivated by this perspective, we explore the use of time-series foundation models (TSFMs) in the SC setting and compare with classical linear SC methods. The results show that linear models remain strong baselines, while TSFMs offer advantages in settings with stronger temporal trends, greater nonlinearity, or when the low-rank assumption fails. Finally, we discuss future research directions highlighting the need for better adaptation of TSFMs for SC.

Submission Number: 20