DynLMC: Dynamic Linear Coregionalization for Realistic Synthetic Multivariate Time Series

Annita Vapsi, Penghang Liu, Saheed Obitayo, Aakriti, Manoj Cherukumalli, Prathamesh Patil, Amit Varshney, Nicolas Marchesotti, Elizabeth Fons, Vamsi K. Potluru, Manuela Veloso

Presentation Attendance: No, we cannot present in-person

Keywords: Time Series, Synthetic Data Generation, Foundation Models for Time Series, Transfer Learning

TL;DR: We propose DynLMC, a synthetic data generator for multivariate time series that models dynamic, regime-switching correlations and cross-channel lags.

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.

Track: Research Track (max 4 pages)

Submission Number: 70