FlowState is a novel time-series modeling (TSFM) architecture proposed to address the problems of poor generalization performance, lack of sampling rate adaptability, and computational inefficiency inherent in existing TSFMs. It utilizes a state-space model (SSM)-based encoder and a function basis decoder to enable continuous-time modeling and dynamic time scaling. This enables generalization across all possible time resolutions and dynamically adjusting the prediction horizon. It also adapts internal dynamics to input scales without requiring training on all data for various sampling rates, reducing model size, data requirements, and improving efficiency. An efficient pre-training strategy enhances robustness and accelerates learning, achieving state-of-the-art performance on the GIFT-ZS and Chronos-ZS benchmarks.
