CAPTAIN: Conformal-Prediction-Based Multi-Source Time-Series Forecasting

TMLR Paper7556 Authors

Abstract: Uncertainty quantification is critical for real-world forecasting applications such as predictive maintenance, patient health monitoring, and environmental sensing, where decisions must account for confidence levels. Multi-source time-series forecasting introduces additional complexity due to inter-source interactions and temporal dependencies, which existing methods struggle to capture within a unified probabilistic framework, and most previous approaches also lack theoretical guarantees, leading to miscalibrated uncertainty estimates. We propose CAPTAIN (Conformal Prediction based multi-source Time-series forecasting), a two-stage framework that first employs Normal Inverse Gamma (NIG) distributions to model source-specific uncertainties and integrates a meta-source to capture inter-source interactions, then uses temporal copulas to model the evolution of joint uncertainties over time, ensuring robust and theoretically valid uncertainty coverage. Experiments on five diverse datasets (Synthetic, Shaoxing ECG, Air Quality, NGSIM Traffic, and ETTh1) demonstrate that CAPTAIN achieves valid coverage (>=90%) across all five benchmarks while other baselines achieve on 4 or fewer, confirming it is a better approach for multi-source uncertainty quantification over existing state-of-the-art baselines.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Taylor_W._Killian1

Submission Number: 7556