Task-aware joint imputation and forecasting for incomplete multivariate time series
Abstract
Incomplete multivariate time series are common in real-world applications due to sensor failures, communication interruptions, and data collection errors, which severely affect both historical analysis and future prediction. Existing studies usually handle imputation and forecasting as two separate tasks, where missing values are first recovered and the completed sequence is then used for forecasting. However, such a two-stage pipeline ignores the intrinsic relationship between the two tasks and may suffer from error accumulation. To address this issue, this paper proposes a Task-Aware Joint Imputation and Forecasting framework for incomplete multivariate time series. The proposed method first constructs a dense proxy input through normalization and coarse initialization, and then integrates coarse values, observation masks, and time-gap information to build informative input representations. A shared encoder is employed to jointly capture temporal dependencies and inter-variable correlations from incomplete historical observations, based on which an imputation head and a forecasting head are designed to recover missing historical values and predict future sequences, respectively. In addition, a task-consistency constraint is introduced to strengthen the interaction between the two tasks, allowing the forecasting objective to guide imputation-oriented representation learning. Experimental results on public datasets demonstrate that the proposed framework achieves competitive performance in both imputation and forecasting tasks, validating the effectiveness of joint task-aware learning.