Are Time-Indexed Foundation Models the Future of Time Series Imputation?

TL;DR

This paper investigates the potential of time-indexed foundation models like TabPFN-TS and MoTM for zero-shot time series imputation, demonstrating their effectiveness across diverse datasets without retraining.

Contribution

It provides the first large-scale empirical evaluation of these models for zero-shot imputation, highlighting their practical utility and ability to incorporate covariates at inference.

Findings

Models perform well across 33 out-of-domain datasets.

Time-indexed foundation models enable zero-shot imputation without retraining.

Incorporating covariates improves imputation accuracy.

Abstract

Foundation models for time series imputation remain largely unexplored. Recently, two such models, TabPFN-TS and MoTM, have emerged. These models share a common philosophy that places them within the family of time-indexed foundation models. This paper presents the first large-scale empirical study of these models for zero-shot imputation, which enables missing value recovery without retraining across a wide range of scenarios. We conduct extensive univariate experiments across 33 out-of-domain datasets (approximately 1.3M imputation windows) and evaluate their ability to integrate covariates at inference time to improve accuracy without fine-tuning. Our results demonstrate that time-indexed foundation models are a powerful and practical step toward achieving general-purpose, zero-shot imputation for real-world time series.