Reproduce “FreqFlow: Long-term forecasting using lightweight flow matching” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Seyed Mohamad Moghadas, Bruno Cornelis, Adrian Munteanu
FrèqFlow introduces a lightweight conditional flow matching model for multivariate time-series forecasting by operating primarily in the frequency domain. This approach achieves state-of-the-art accuracy, with an average 7% RMSE improvement over existing methods, while being over 15 times smaller in parameters and 3 times faster in inference than diffusion-based models.
Key insights
- Processing time-series data in the frequency domain allows for a more compact and efficient representation, where temporal dynamics are modeled as amplitude and phase shifts.
- Conditional flow matching enables efficient, single-pass deterministic sampling, circumventing the computational overhead of iterative diffusion models.
- A component-wise approach, utilizing frequency interpolation for trend/seasonality and flow matching for residual learning, significantly improves long-term forecasting accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.