FlowState: Sampling Rate Invariant Time Series Forecasting

Lars Graf, Thomas Ortner, Stanisław Woźniak, Angeliki Pantazi

TL;DR

FlowState is a new "Time Series Foundation Model" (TSFM) that uses a State Space Model (SSM) encoder and a novel functional basis decoder to achieve state-of-the-art forecasting performance. Unlike previous models that struggle with varying temporal resolutions, FlowState is inherently sampling-rate invariant, allowing it to outperform models up to 192× its size while adapting to different time scales on the fly.

What Problem Does This Paper Address?

While Foundation Models (FMs) have revolutionized NLP, their success in time series forecasting has been limited. Existing Time Series Foundation Models (TSFMs) face three major hurdles:

  1. Rigidity: Most use linear decoders fixed to a specific target length, making them inflexible for different forecasting horizons.
  2. Sampling Rate Sensitivity: They often struggle to generalize across different temporal resolutions (e.g., moving from hourly to minutely data) unless they've seen those specific rates during training.
  3. Efficiency: Transformer-based architectures are computationally expensive and often rely on "patching" (grouping time steps), which can lose fine-grained information.

The authors aim to create a model that treats time as a continuous variable rather than a discrete sequence of tokens.

Methods

FlowState introduces an encoder-decoder architecture designed for continuous-time modeling:

  • SSM Encoder (S5): Instead of Transformers, the model uses a stack of S5 layers. This allows the model to process varying context lengths efficiently and maps the input "feature space" into a "coefficient space."
  • Functional Basis Decoder (FBD): This is the paper's standout innovation. Instead of predicting discrete future points, the FBD interprets the encoder's output as coefficients for a set of continuous basis functions (like Legendre polynomials). This creates a continuous forecast function that can be sampled at any interval ($\Delta$), allowing for dynamic adjustment of forecasting horizons and sampling rates without retraining.
  • Temporal Scaling ($s_\Delta$): The model adjusts its internal dynamics based on the seasonality of the data rather than raw frequency. This maps diverse datasets (e.g., hourly vs. weekly) into a common latent temporal scale.
  • Parallel Forecast Training: To improve efficiency, the authors developed a scheme that produces thousands of forecasts from a single sample by varying the context length in parallel.
  • Causal Normalization: A modified version of Reversible Instance Normalization (RevIN) that uses running statistics to prevent future data from leaking into the past during the parallel training process.

Datasets

The authors evaluated FlowState on two major zero-shot benchmarks:

  • GIFT-Eval: A comprehensive benchmark for time series foundation models. The authors used a subset called GIFT-ZS (79 tasks) to ensure no data overlap between training and testing.
  • Chronos-ZS: A benchmark derived from the Chronos model's evaluation suite.
  • Loop Seattle: Used specifically for "unseen sampling rate" experiments, where the model was tested on data subsampled at intervals ranging from 5 to 65 minutes.

Key Insights and Ideas

  • Continuous > Discrete: By using a functional basis decoder, the model treats forecasting as a function-fitting problem rather than a sequence-prediction problem. This makes the model "resolution-agnostic."
  • Seasonality as the Anchor: The authors found that scaling time based on seasonality (e.g., a 24-hour cycle) is more effective for generalization than scaling based on the raw sampling frequency.
  • Small is Mighty: FlowState achieves SOTA results with only 9.1M parameters. For context, it outperforms models like TimesFM (500M) and Moirai (311M), proving that architectural inductive biases (like SSMs and continuous decoders) can be more important than raw scale.
  • Zero-Shot Robustness: The model can adapt to a change in sampling rate during inference simply by adjusting a scaling parameter, a feat most discrete TSFMs cannot perform without performance degradation.

Key Results

  • SOTA Performance: FlowState(T)-9.1M achieved the top spot on both GIFT-ZS and Chronos-ZS benchmarks, surpassing the previous leader, TiRex.
  • Efficiency: The 2.6M parameter version of FlowState performs on par with models 10-15x its size.
  • Sampling Rate Invariance: In tests on the Loop Seattle dataset, FlowState maintained consistent accuracy across various sampling frequencies (5 min to 65 min), while baseline models showed significant performance drops at "uncommon" intervals that weren't well-represented in their training data.
  • Training Speed: The parallel forecasting scheme allows the model to learn from significantly more "context-target" pairs per training step, accelerating convergence.

Limitations and Future Work

  • Basis Function Selection: While Legendre polynomials were the default, the authors noted that different bases (like Fourier) might be better for highly periodic signals, suggesting a need for automated basis selection.
  • Multivariate Complexity: While the model handles multivariate data, the paper focuses heavily on the temporal aspect; further exploration into complex cross-channel correlations could be beneficial.
  • Full GIFT-Eval Analysis: The authors noted they haven't yet extensively analyzed the "full" GIFT-Eval benchmark (including tasks with potential data leakage) and see room for further improvement in general-purpose forecasting.

Summary updated

Abstract

Foundation models (FMs) have transformed natural language processing, but their success has not yet translated to time series forecasting. Existing time series foundation models (TSFMs), often based on transformer variants, struggle with generalization across varying context and target lengths, lack adaptability to different sampling rates, and are computationally inefficient. We introduce FlowState, a novel TSFM architecture that addresses these challenges through two key innovations: a state space model (SSM) based encoder and a functional basis decoder. This design enables continuous-time modeling and dynamic time-scale adjustment, allowing FlowState to inherently generalize across all possible temporal resolutions, and dynamically adjust the forecasting horizons. In contrast to other state-of-the-art TSFMs, which require training data across all possible sampling rates to memorize patterns at each scale, FlowState inherently adapts its internal dynamics to the input scale, enabling smaller models, reduced data requirements, and improved efficiency. We further propose an efficient pretraining strategy that improves robustness and accelerates training. Despite being the smallest model, FlowState outperforms all other models and is state-of-the-art for the GIFT-ZS and the Chronos-ZS benchmarks. Ablation studies confirm the effectiveness of its components, and we demonstrate its unique ability to adapt online to varying input sampling rates.
BibTeX
@inproceedings{graf2026flowstate,
  author = {Lars Graf and Thomas Ortner and Stanisław Woźniak and Angeliki Pantazi},
  title = {{FlowState: Sampling Rate Invariant Time Series Forecasting}},
  booktitle = {International Conference on Machine Learning},
  year = {2026},
  url = {https://openreview.net/forum?id=9yNGBueKgZ},
}
Figure