Koopa: Koopman Predictors for Non-stationary Time Series
Koopa (Koopman Predictors for Non-stationary Dynamics) · Also known as: Koopman Predictor, Koopman-based Time-Series Model, Koopa Forecaster, Koopman Tahmincisi
Koopa is a deep learning model for time-series forecasting introduced by Yong Liu, Chang Li, Jianmin Wang, and Mingsheng Long at NeurIPS 2023. It addresses the challenge of non-stationarity by disentangling time series into stationary and non-stationary components, then modeling the non-stationary dynamics using a learned approximation of the Koopman operator — a mathematical framework that lifts nonlinear systems into a linear space for tractable long-horizon prediction.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Koopa is well-suited for multivariate long-horizon time-series forecasting tasks where the data exhibits clear non-stationarity — shifting means, variances, or seasonal patterns over time. It assumes that dynamics in a lifted feature space can be reasonably approximated as linear, which holds for many real-world signals such as energy, weather, and financial series. It may underperform on series with highly irregular or chaotic dynamics where no approximately linear lifting exists. Alternatives include PatchTST (patch-based attention), DLinear (direct linear baseline), or state-space models when Koopman's linearity assumption is too restrictive.
Strengths & limitations
- Explicitly handles non-stationarity through a principled decomposition rather than relying on normalization heuristics
- Koopman-based linearization enables efficient multi-step forecasting without iterative error accumulation
- Theoretically grounded in dynamical systems theory, offering interpretable latent dynamics
- Computationally lighter than full attention-based transformers while remaining competitive on standard benchmarks
- The linear Koopman approximation may be insufficient for highly chaotic or irregular time series where no finite-dimensional linear lifting exists
- Performance depends on the quality of the trend-residual decomposition, which can fail when non-stationarity is abrupt or structural breaks occur
- The latent observable space dimension is a sensitive hyperparameter requiring careful tuning
- Less explored in low-data regimes compared to classical statistical models; may overfit on short series
Frequently asked
How does Koopa differ from Transformer-based forecasting models like Autoformer or FEDformer?
Transformer-based models rely on attention mechanisms to capture temporal dependencies but implicitly assume approximate stationarity. Koopa explicitly decomposes the series into stationary and non-stationary parts and models the non-stationary dynamics using a Koopman operator, providing a theoretically grounded mechanism for handling distribution shift without the quadratic complexity of full self-attention.
Is the Koopman operator actually learned, and how is it parameterized?
Yes. Koopa parameterizes the Koopman operator as a learnable matrix (or low-rank factorization) within the model. An MLP encoder lifts observations into the latent observable space, and the operator is trained end-to-end with the rest of the network using standard gradient descent on the forecasting loss.
Can Koopa handle multivariate time series?
Yes. The paper demonstrates Koopa on standard multivariate benchmarks. Each variate's dynamics can be modeled jointly in the lifted space, though the specific channel-mixing strategy and whether the Koopman operator is shared or per-channel are architectural choices that affect both capacity and generalization.
Sources
- Liu, Y., Li, C., Wang, J., & Long, M. (2023). Koopa: Learning non-stationary time series dynamics with Koopman predictors. NeurIPS. link ↗
How to cite this page
ScholarGate. (2026, June 2). Koopa (Koopman Predictors for Non-stationary Dynamics). ScholarGate. https://scholargate.app/en/deep-learning/koopa
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- DLinearDeep learning↔ compare
- Non-stationary TransformerDeep learning↔ compare
- State Space ModelEconometrics↔ compare