MCMC Framework Preserves Temporal Dynamics in Time Series Generation
A new model-agnostic framework using Markov Chain Monte Carlo (MCMC) aims to preserve temporal dynamics in synthetic time series generation. Existing GAN-based methods focus on matching marginal distributions, causing distribution shift and temporal drift in multivariate sequences. The proposed approach mitigates these issues through sequential generation with MCMC correction. The work includes theoretical analysis of deviation accumulation in conditional generative models.
Key facts
- arXiv paper 2604.27182
- Proposes MCMC-based framework for time series generation
- Addresses distribution shift and temporal drift
- Model-agnostic approach
- Focuses on multivariate time series
- Theoretical analysis of deviation accumulation
- Targets regression-oriented forecasting tasks
- Improves fidelity of synthetic sequences
Entities
Institutions
- arXiv