Dynamic Linear Coregionalization for Realistic Synthetic Multivariate Time Series

TL;DR

This paper introduces DynLMC, a dynamic model for generating synthetic multivariate time series with realistic, time-varying inter-channel correlations, improving foundation model training.

Contribution

The paper presents DynLMC, a novel dynamic linear model that captures regime-switching correlations and lag structures, enhancing synthetic data realism for time series modeling.

Findings

DynLMC produces synthetic data with correlation dynamics similar to real data.

Fine-tuning models on DynLMC data improves zero-shot forecasting across benchmarks.

Modeling dynamic correlations boosts transferability of foundation models.

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.