FITS: Conditional Diffusion Model for Irregular Time Series Forecasting with Pseudo-future Exogenous Covariates

ICLR 2026 Conference Submission15952 Authors

Keywords: time series forecasting

TL;DR: Proposed a conditional diffusion model for IMTS forecasting that leverages pseudo-future exogenous covariates

Abstract: Irregular multivariate time series (IMTS) present unique challenges due to non-uniform intervals and different sampling rates. While existing methods struggle to capture both long-term dynamics and cross-channel dependencies under such irregularities, we tackle this by formulating time series forecasting as a conditional generation problem and introducing FITS, a conditional diffusion model for IMTS forecasting that leverages pseudo-future exogenous covariates. Our approach incorporates two key innovations. First, we propose a novel density-aware adaptive patching scheme that generates data-driven segments with dynamic boundaries determined by the information density. This scheme overcomes the limitations of traditional fixed-length or fixed-span segmentation in preserving continuous local semantics and modeling inter-time series correlations. Second, we develop a transformer-based prior knowledge extractor that captures forward-looking covariate dependencies via a novel cross-variate attention mechanism. The transformer structure is integrated into the conditional diffusion generative process as a unified framework, enabling precise distributional forecasting for IMTS. Extensive experiments on six datasets with four evaluation metrics validate the effectiveness of FITS.

Primary Area: learning on time series and dynamical systems

Submission Number: 15952