Time Series Analysis and Modeling

The field of time series analysis and modeling is moving towards more robust and generalizable solutions, with a focus on learning from irregular and sparse data. Researchers are exploring new approaches to modeling continuous-time dynamics, such as using invertible neural flows and discretization schemes that preserve flatness. Additionally, there is a growing interest in time series foundation models and their applications in various domains, including hydrology and bearing health status classification. These models are being improved with techniques such as in-context learning and frequency-domain formulation, enabling more efficient and accurate forecasting. Noteworthy papers include FlowPath, which learns the geometry of the control path via an invertible neural flow, and FreqFlow, which leverages conditional flow matching in the frequency domain for deterministic multivariate time-series forecasting. Overall, the field is advancing towards more powerful and flexible models that can handle complex and heterogeneous time series data.

Sources

FlowPath: Learning Data-Driven Manifolds with Invertible Flows for Robust Irregularly-sampled Time Series Classification

Numerical Discretization Schemes that Preserve Flatness

Lightweight Time Series Data Valuation on Time Series Foundation Models via In-Context Finetuning

Leveraging Exogenous Signals for Hydrology Time Series Forecasting

Functional Mean Flow in Hilbert Space

Towards Stable and Structured Time Series Generation with Perturbation-Aware Flow Matching

Reservoir Computing via Multi-Scale Random Fourier Features for Forecasting Fast-Slow Dynamical Systems

On the Internal Semantics of Time-Series Foundation Models

TSFM in-context learning for time-series classification of bearing-health status

FreqFlow: Long-term forecasting using lightweight flow matching

Built with on top of