Problem
Research questions and friction points this paper is trying to address.
synthetic multivariate time series
dynamic correlations
inter-channel dependencies
time-varying correlations
foundation models for time series
Dynamic Linear Coregionalization for Realistic Synthetic Multivariate Time Series
🤖 AI Summary
This work addresses the limitation of existing time series synthesis methods, which typically assume static cross-channel correlations and thus fail to capture the dynamically evolving dependencies inherent in real-world multivariate data. To overcome this, we propose DynLMC, the first approach to integrate a dynamic linear model of coregionalization into synthetic data generation. DynLMC explicitly models complex temporal dynamics through time-varying correlations, a state-switching mechanism, and cross-channel lagged cointegration structures. Empirical results demonstrate that foundation models fine-tuned on data synthesized by DynLMC across nine benchmarks achieve substantially improved zero-shot forecasting performance, thereby validating that accurately modeling dynamic correlations is crucial for effective pretraining in time series representation learning.
📝 Abstract
Synthetic data is essential for training foundation models for time series (FMTS), but most generators assume static correlations, and are typically missing realistic inter-channel dependencies. We introduce DynLMC, a Dynamic Linear Model of Coregionalization, that incorporates time-varying, regime-switching correlations and cross-channel lag structures. Our approach produces synthetic multivariate time series with correlation dynamics that closely resemble real data. Fine-tuning three foundational models on DynLMC-generated data yields consistent zero-shot forecasting improvements across nine benchmarks. Our results demonstrate that modeling dynamic inter-channel correlations enhances FMTS transferability, highlighting the importance of data-centric pretraining.
Innovation
Methods, ideas, or system contributions that make the work stand out.
Dynamic Linear Coregionalization
Synthetic Multivariate Time Series
Time-varying Correlations
Cross-channel Lag
Foundation Models for Time Series
🔎 Similar Papers
No similar papers found.
