DynLMC: Dynamic Linear Coregionalization for Realistic Synthetic Multivariate Time Series

Researchers have introduced DynLMC, a Dynamic Linear Model of Coregionalization designed to generate realistic synthetic multivariate time series data. Current synthetic data generators often fail to capture realistic inter-channel dependencies because they assume static correlations. DynLMC addresses this limitation by incorporating time-varying, regime-switching correlations and cross-channel lag structures, producing synthetic data that closely mirrors the dynamic correlation patterns found in real-world datasets. The study demonstrates that fine-tuning three foundational models for time series (FMTS) on data generated by DynLMC leads to consistent zero-shot forecasting improvements across nine different benchmarks. These results highlight that modeling dynamic inter-channel correlations significantly enhances the transferability of foundation models. This research underscores the critical importance of data-centric pretraining strategies in advancing the performance of artificial intelligence systems dealing with complex temporal data, offering a new methodological approach for improving model robustness and accuracy in time series analysis tasks.