Zero shot time series reconstruction improves prediction without target data
ZeroDiff: Zero-Shot Time Series Reconstruction via Informed-Prior Diffusion
Machine Learning
Summary
Sometimes, it's hard or expensive to collect certain measurements over time, but other related data (like external factors) is easier to get. The authors study how to predict missing measurements at new locations where no data was collected before. They show that just guessing from outside information tends to smooth out important details. To fix this, they build a model that first guesses with external data and then uses a special technique called diffusion to improve accuracy, learning from places where full data exists. Their method works better than previous ones on many real-world examples.
What this means in practice
- •For environmental monitoring teams: Generate accurate time series data at monitoring locations without direct measurements by calibrating predictions from external conditions.
- •For energy grid operators: Reconstruct missing sensor data across a power grid to improve system management when some sensors are offline or missing.
Authors
Yingda Fan, Dan Lu, Xiaowei Jia
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.