Dynamic Linear Coregionalization for Realistic Synthetic Multivariate Time Series
This addresses the need for realistic synthetic data to improve foundation models for time series forecasting, though it appears incremental as it builds on existing coregionalization methods with dynamic extensions.
The paper tackles the problem of generating synthetic multivariate time series with realistic inter-channel dependencies by introducing DynLMC, a dynamic linear model that incorporates time-varying correlations and lag structures. The result shows that fine-tuning foundation models on this synthetic data yields consistent zero-shot forecasting improvements across nine benchmarks.
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.