Online Irregular Multivariate Time Series Forecasting via Uncertainty-Driven Dual-Expert Calibration

Researchers propose Under-Cali, a machine learning framework for forecasting irregular multivariate time series data in real-time online settings. The system uses uncertainty estimation and dual-expert calibration to maintain accuracy despite dynamic data distribution shifts, achieving improvements over existing methods with minimal computational overhead.

Under-Cali addresses a fundamental challenge in time series analysis: maintaining prediction accuracy when data arrives irregularly and patterns shift dynamically. Traditional forecasting methods rely on regular sampling and stable periodicity patterns, assumptions that fail in real-world applications like sensor networks, healthcare monitoring, and financial data feeds where sampling intervals vary and data quality fluctuates. The framework's innovation centers on uncertainty quantification as a control mechanism—by assessing confidence levels for each data batch, the system intelligently routes high-uncertainty samples to a calibration-focused expert while maintaining stable predictions through a reliable expert. This dual-track approach prevents catastrophic performance degradation during online deployment, a common problem when models trained offline encounter distribution shifts. The model-agnostic design allows adaptation without retraining the underlying forecasting engine, reducing computational demands critical for edge deployments and resource-constrained environments. For practitioners deploying time series systems in production, irregular sampling is endemic rather than exceptional—IoT networks drop packets, medical devices skip measurements, and market data gaps occur during trading halts. Under-Cali's ability to maintain forecast quality despite these challenges positions it as practically valuable across infrastructure monitoring, predictive maintenance, and anomaly detection use cases. The lightweight calibration module architecture suggests scalability advantages over full model retraining approaches. Academic contributions like this advance the maturity of online learning systems, though real-world adoption depends on integration with existing ML pipelines and validation across diverse industrial applications beyond the tested benchmarks.

• →Under-Cali uses uncertainty estimation as a core control mechanism to intelligently route and calibrate predictions in irregular time series forecasting
• →The dual-expert framework maintains a stable reliable expert while calibrating an unreliable expert with challenging samples, enabling efficient online adaptation
• →The system keeps the source model frozen and performs lightweight calibration only, reducing computational costs for production deployment
• →Framework demonstrates consistent performance improvements on irregular multivariate time series benchmarks despite dynamic distribution shifts
• →Model-agnostic design enables integration with existing forecasting architectures without architectural modifications