ZeroDiff: Zero-Shot Time Series Reconstruction via Informed-Prior Diffusion

2026-09-29

AI Summary

This study addresses the limitations of conventional mapping methods in target time series imputation, which often yield over-smoothed predictions and underestimate extreme values when observations are missing. To this end, we propose a zero-shot reconstruction framework that establishes an "information prior plus diffusion calibration" paradigm. Specifically, the method leverages exogenous variables to generate informative priors and employs diffusion models to calibrate reconstruction errors, effectively overcoming the challenge of capturing dynamic features in unsupervised settings. Experimental results demonstrate that the proposed framework significantly outperforms existing methods across multiple real-world datasets, achieving high-fidelity reconstruction at unobserved locations with strong generalization capabilities. The source code is publicly available.

Abstract

Time series modeling increasingly demands high-quality supervision, yet target observations remain scarce - exogenous inputs are broadly available, but target measurements are often unavailable due to cost, infrastructure, or accessibility constraints. Can models trained on observed locations reconstruct target time series where measurements have never been collected? We term this zero-shot time series reconstruction. A naive approach - directly mapping exogenous inputs to targets - can yield predictions at unobserved locations, but without target signals, such models fail to capture the intrinsic dynamics of the target variable, producing overly smooth outputs that underestimate extremes. This reveals systematic errors that call for explicit modeling and calibration. We propose ZeroDiff, which constructs an informed prior from exogenous variables alone, then learns to calibrate reconstruction errors through diffusion - training on observed locations and generalizing to unobserved ones. Experiments across diverse real-world datasets demonstrate significant improvements over existing approaches. Our code is available at https://github.com/YingdaFan/ZeroDiff-ICML2026.

Problem

• Zero-shot time series reconstruction
• Exogenous inputs
• Target scarcity
• Systematic errors
• Extreme underestimation

Innovation

• Zero-Shot Time Series Reconstruction
• Informed-Prior Diffusion
• Error Calibration
• Exogenous Variables