Koopa: Learning Non-stationary Time Series Dynamics with Koopman Predictors

Real-world time series are characterized by intrinsic non-stationarity that poses a principal challenge for deep forecasting models. While previous models suffer from complicated series variations induced by changing temporal distribution, we tackle non-stationary time series with modern Koopman theory that fundamentally considers the underlying time-variant dynamics. Inspired by Koopman theory of portraying complex dynamical systems, we disentangle time-variant and time-invariant components from intricate non-stationary series by Fourier Filter and design Koopman Predictor to advance respective dynamics forward. Technically, we propose Koopa as a novel Koopman forecaster composed of stackable blocks that learn hierarchical dynamics. Koopa seeks measurement functions for Koopman embedding and utilizes Koopman operators as linear portraits of implicit transition. To cope with time-variant dynamics that exhibits strong locality, Koopa calculates context-aware operators in the temporal neighborhood and is able to utilize incoming ground truth to scale up forecast horizon. Besides, by integrating Koopman Predictors into deep residual structure, we ravel out the binding reconstruction loss in previous Koopman forecasters and achieve end-to-end forecasting objective optimization. Compared with the state-of-the-art model, Koopa achieves competitive performance while saving 77.3% training time and 76.0% memory.

Paper

References (54)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer pXJ36/10 · confidence 5/52023-06-14

Summary

This paper tackles the task of non-stationary time series forecasting with Koopman operator. Specifically, the authors propose Koopa Blocks, which uses Fourier transform to disentangle the time series into time-variant and time-invariant components, and leverages Koopman operator to model their dynamics respectively. The method is evaluated on six real-world benchmark datasets.

Strengths

1. The paper is well-written and easy to follow. 2. This study represents a pioneering endeavor in integrating Fourier transform and Koopman theory, with the aim of disentangling time-variant and time-invariant components within the Fourier spectral domain. The idea is both innovative and thought-provoking, setting a precedent for future investigations into the research of deep Koopman methods. 3. The additional "Scaling Up Forecast Horizon" experiments demonstrate that Koopa is capable of accurately forecasting in situations involving length mismatches or long-term scenarios. In the realm of the Koopman method, these long-term prediction results are exceptional. 4. The proposed model is memory and time efficient compared to previous methods.

Weaknesses

1. The performance improvement of univariate forecasting is incremental. Also, the improvement of multivariate forecasting is also not so significant under most cases. 2. The essence of the deep Koopman method lies in the spectrum of the Koopman operator, because the eigenvalues determine the model's behavior during long-term evolution. However, the author did not provide any analysis or visualized results regarding the eigenvalues. This lead to the following issues: - Given the considerable forecasting length (up to 192 in this case), it is remarkable how the model manages to achieve accurate predictions without encountering issues such as explosion or decay during the forward pass. Is it true that all the eigenvalues of the system have a modulus of one, ensuring stable and non-divergent behavior throughout the prediction process? - Intuitively, it would be expected that the time-variant Koopman operator possesses more eigenvalues with larger phase angles compared to the time-invariant operator. I wonder whether this is true in this case. What does the spectrum for these two operator looks like?

Questions

1. The authors implement the DMD with torch.linalg.lstsq. Were there any concerns regarding gradient explosion during the autograd process through this layer? Did the authors employ any techniques to stabilize the training process?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

2 fair

Presentation

4 excellent

Contribution

3 good

Limitations

Yes.

Reviewer EGYL5/10 · confidence 4/52023-07-03

Summary

This work propose a new architecture inspired by Koopman theory for time-series forecasting task. To address the non-stationarity problem, this work follow the idea of disentangle local and global time-series representation and forecasting model respectively. Particularly, this work propose to use Fourier filter on long sequence to extract more robust and stable patterns for global model and Koopa Block to hierarchically disentangle local and global representation in a residual structure. The proposed method achieves competitive performance when comparing with recent baselines.

Strengths

1. The structure is well organized. Figures 2 and 3 provide a clear visual representation of how the proposed method works. 2. The experiments are comprehensive. The proposed method can beat the recent method PatchTST on the commonly used long sequence forecasting benchmark.

Weaknesses

1. Although the paper's title revolves around the Koopman theory, the proposed method is generic and not closely related to it. From my understanding, most of technique contributions lie in how to extract local and global representation and design forecasting model for non-stationary time-series. The work solely follow the the conclusion of Koopman theory and model the latent dynamic with linear transition model. 2. Technique novelty is limited. The general idea of using local and global presentation/model to tackle non-stationary time-series tasks has been used. This work builds upon these foundations and introduces several modifications by integrating some other methods used in recent deep time series forecasting, like residual block, Fourier filter. 3. Section 1 needs improvements. It is clear for me how Koopman operator address the non-linear dynamics, but section 1 does not clearly show how existing work of Koopman theory tackle non-stationary problems. Could the authors elaborate more on line 40-42? In line 52-54, could the authors explain the motivation of replacing reconstruction loss? In line 30-31, the authors introduce "Non-stationary ...... in different periods". However, the proposed method do not show the capability to detect period boundary. 4. The experiments are comprehensive, but it would be better to design experiments revolve around the topic and the proposed method. Ablation studies can successfully validate the effectiveness of the design of local and global model, but the experiment associated time and memory cost analysis might lose focus. The authors can also explore the effect of number of block to verify the capability of hierarchical disentanglement. 5. It would be better to discuss the related work of non-stationary time series forecasting, like regime switching state-space models, time-varying state-space models, etc.

Questions

how to set the hyperparameters D?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

3 good

Contribution

3 good

Limitations

N.A.

Reviewer a2Yj8/10 · confidence 2/52023-07-05

Summary

The authors address the issue that most real-world time-series are non-stationary. To tackle this problem they introduce Koopa, a globally learned (time-invariant) and localized (time-variant) linear Koopman operators, to exploit respective dynamics underlying different components. In practice, Koopa is composed of stackable Koopa Blocks, where each block learns operators hierarchically by taking the residual of previous block fitted dynamics as input. The authors evaluate the proposed model on variety of time-series data sets and compare their method with transformer and convolution based deep architectures. Koopa outperforms or performs on par with the existing methods, yet has a significantly lower computational costs. Moreover, the proposed method can be use in settings where the forecast horizon must be increased beyond the training set-up.

Strengths

Originality: The related work section in well written. It covers both relevant model types in the field, for (a) time series forecasting, TCN, RNN and transformers are mentioned, as well as the authors position the current work in its context; for (b) Koopman operators, the authors provide an overview of the related work and how their work differs. The work proposes a novel method leveraging Fourier analysis to disentangle time-series components, eDMD for estimating time-variant Koopman Predictor, and time-invariant Koopman Prediction as a learnable parameter. It differentiates itself from KAEs due to its block like structure, where the residual of the previous block is fed as input into the next block. Quality: The submission is techinncally sound. The authors build upon Koopman operator theory and Wold's Theorem. In addition, experimentally they validate whether their proposed sub-components of Koopa, Fourier Filter, truly disentangles time-variant from time-invariant series. As well as provide an ablation on Koopa structure. Clarity: The work is well written providing sufficient relevant background knowledge on Koopman operator theory also for reader who is unfamiliar with this theory. Significance: The obtained results are significant. The proposed method outperforms or performs on par with existing time-series forecasting methods. More importantly, the method is very efficient with respect to training time and memory costs compared to transformer or convolution based methods. Lastly, the authors showcase that their method is applicable for scaling up forecast horizon, where standard deep learning architectures fail. In all, based on the submission the author's suggest a very promising modelling technique, however, I am not an expert of Koopman operator theory (nor the works in this sub-domain).

Weaknesses

Clarity: When first mentioned, I would recommend using the complete name of the model type rather than only the acronym, for example, temporal convolutional network rather than just TCN, it makes it easier for the reader to read.

Questions

For time-variant KP how do you choose S, is the Time-variant KP sensitive to S?

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

The authors have not addressed limitation of their work. I would suggest the authors to include such section in the final version of their work.

Reviewer M3Xz5/10 · confidence 4/52023-07-06

Summary

This paper addresses non-stationary time series forecasting, in which case the temporal distribution of time series is changing over time. Koopman theory is introduced in this paper for its fundamental capacity on modeling dynamical systems and a Koopman forecaster called Koopa is proposed. Concretely, in this paper, time series is disentangled into time-variant and time-invariant components. For time-invariant component, Koopa introduces a learnable matrix as Koopman operator, and for time-variant component, Koopa leverages eDMD to find the best fitted matrix that advances forward the system. The proposed framework is technically sound but somewhat lacks novelty.

Strengths

1. Non-stationary time series forecasting is challenging and it is reasonable to apply Koopman theory to tackle this task. 2. Disentangling time series into variant and invariant components is intuitive. 3. The methods of constructing Koopman operators for both time-variant and -invariant components are straightforward.

Weaknesses

1. The proposed model lacks novelty. Both disentangling time series and applying Koopman theory are commonly used techniques in time series analysis. 2. The technical contribution of this paper is limited. The extraction of time variant and invariant components is too simple and lacks analysis. The application of Koopman theory is simply the construction of Koopman operators. 3. Hierarchically disentangling time series lacks intuition and motivation. What's the point of further disentangle variant component into invariant and variant components? 4. The experiment does not involve highly non-stationary datasets. 5. Koopman operators are highly correlated to measurement functions. In the branch of time variant component, although the model constructs diverse Koopman operators, the measurement, i.e., the encoder, is static. For more technical contributions, a deeper integration of Koopman's theory is needed.

Questions

See weaknesses.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

3 good

Contribution

1 poor

Limitations

Not applicable.

Reviewer 36zV7/10 · confidence 2/52023-07-06

Summary

The paper studied time series forecasting problem for weather forecasting, energy consumption, and financial assessment. To propose a model that generalizes on varying distribution, the authors proposed Koopa which is composed of modular Koopman predictors to describe and advance forward series dynamics. The motivation is Koopman-based methods are appropriate to learn non-stationary time series dynamics. The implementation is straightforward in the supplement material. The experimental results show the effective performance of the proposed methods compared with state-of-the-art methods.

Strengths

The paper proposed a novel method to solve the time series forecasting problem. The introduction of Koopa Block is interesting. The experimental results show that the model outperforms existing methods with more efficient memory usage.

Weaknesses

It is interesting to see more analysis about Koopa blocks. For example on Figure 5, authors showed the visualized blocks are different which corresponds to different curves. But it is clear that the curves of K1, K2, and K2 share the similarity that they goes down first, then goes up. It is interesting to see the correlation analysis if there is some. The experiments show the proposed method achieved better performance with efficient training time and memory usage. I am not sue if there is convergence analysis and what are the parameters, such as the number of epochs for training, batch size.

Questions

1. Is there any correlation between Koopa blocks if the curves share similarity as shown in Figure 5 ? 2. How is the convergence of the proposed method and compared methods? 3. What are the parameters, such as the number of epochs for training, batch size for experiments?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

There is no theoretical analysis but the main contribution is to propose a novel architecture.

Reviewer a2Yj2023-08-12

Thank you for the additional ablation experiment and for voicing the limitations of the presented work, I recommend the authors to include the limitations in the main paper rather than the supplementary material. I will keep my score as it is.

Authorsrebuttal2023-08-13

Thanks for your response

Thanks a lot again for your response and every effort on the review. We will include the limitations in the final version of the main paper.

Reviewer pXJ32023-08-16

Thanks for the response. While the concern regarding performance gain remains significant in the paper, I recognize that the proposed method presents an inspiring idea by combining Koopman theory and spectral methods. Moreover, its high interpretability adds value to the work, making it a valuable contribution to the Koopman learning community. I will raise my rating to weak accept.

Authorsrebuttal2023-08-17

Thanks for Your Response and Raising the Score

Many thanks to Reviewer pXJ3 for the insightful pre-rebuttal review and valuable feedback. Your detailed suggestions help us a lot in the rebuttal and paper revision!

Reviewer EGYL2023-08-16

- Could the authors clarify the connection between the proposed architecture design to the Koopman theory, how the proposed architecture can reduce approximation error of non-linear dynamic system? In theory, higher dimension dimension means lower approximation error w.r.t non-linear function. How to explain that the performance is insensitive to the choices of D? - Real-data is complicated, different datasets usually have different issues leading to performance bottlenecks. How to demonstrate that the performance improvement over the real-world datasets stems from Koopman theory via a better approximated non-linear function instead of deep architecture design (residual block, periodicity pattern extraction for this long-sequence forecasting benchmark, global and local disentanglement), whose effectiveness has been widely demonstrated in recent research work. - Can the authors elaborate more on my 3rd question in weakness?

Authorsrebuttal2023-08-17

Thanks for the Reviewer's Prompt Reply (Part 1)

Thanks a lot for your valuable prompt comments. We'd like to provide more elaborations on each point of your response. **1. How the proposed architecture reduces the approximation error of non-linear dynamics** Due to the complexity and non-divergent eigenvalues evolution of non-linear dynamics, it is always challenging to directly apply operator learning on non-stationary series as one dynamics, which motivates us to consider the following aspects of architecture. (a) Stackable blocks learning hierarchical dynamics To our best knowledge, we make Koopman-based forecasters "deep" for the first time. Each layer learns weak stationary process with well-natured operators for **stable training** (Please refer to $\underline{\text{Q2 of Reviewer pXJ3}}$ for our detailed analysis). Deeper layers learn the residual of previously fitted dynamics and aggregate layer-wise dynamics for the final forecast, which **enhances the model capacity** for learning complex non-linear dynamics. As shown in $\underline{\text{Table 1 of the global response}}$, the difficulty of dynamics approximation in one block can be larger than in multiple blocks. (b) Time-variant operators based on new disentanglement Inspired by Koopman theory portraying non-linear dynamics into sub-regions individually governed by local operators, we **associate dynamics sub-regions to varying windows** in time series and calculate time-variant operators by analytical method DMD. Though previous work also utilizes time-variant operators, they do not **explicitly design time-variant and invariant disentanglement** and **elaborate on respective dynamics modeling**. In our architectural ablation of $\underline{\text{Table 3 of the main text}}$, we validate the effectiveness of applying disentanglement, learning parameterized operator globally, and analytically calculating operators locally. **2. How to explain the performance is insensitive to the choices of D** Thanks a lot for your question with scientific rigor. As we clarified that stacking blocks could enhance model capacity, the sensitivity analysis of $D$ is conducted under a fixed number of Koopa blocks $B=3$, which can be large enough for dynamics modeling. To further address your concern, we further check the sensitivity of $D$ under varying block number $B$. | ETTh1 (MSE) | D=64 | D=128 | D=256 | D=512 | | ----------- | ----- | ----- | ----- | ----- | | B=1 | 0.400 | 0.396 | 0.393 | 0.390 | | B=2 | 0.393 | 0.392 | 0.389 | 0.388 | | B=3 | 0.388 | 0.387 | 0.386 | 0.386 | | B=4 | 0.385 | 0.385 | 0.385 | 0.384 | | Exchange (MSE) | D=64 | D=128 | D=256 | D=512 | | -------------- | ----- | ----- | ----- | ----- | | B=1 | 0.150 | 0.140 | 0.134 | 0.125 | | B=2 | 0.134 | 0.129 | 0.126 | 0.123 | | B=3 | 0.127 | 0.124 | 0.125 | 0.120 | | B=4 | 0.120 | 0.118 | 0.118 | 0.117 | It can be more obvious that a larger $D$ leads to lower error when $B$ is small. But stacking blocks as $B=3$ makes the performance insensitive to $D$. So we will include the findings and update hyperparameter sensitivity meticulously in the final version of our supplementary materials.

Authorsrebuttal2023-08-17

Thanks for the Reviewer's Prompt Reply (Part 2)

**3. How to demonstrate the performance improvement stems from Koopman theory via a better approximated non-linear function instead of deep architecture design.** To check how much the performance is improved by the Koopman theory-inspired building block. We compare Koopa with competitive baseline NBEATS[1] and its variants, which are all built on deep residual architecture with three available choices of the basic block. * Trend Block: MLP learning weighting on pre-defined polynomial basis $\{1, t, ..., t^p\}$. * Seasonal Block: MLP learning weighting on pre-defined Fourier series basis $\{cos(2\pi it), sin(2\pi it)\}$. * Generic Block: MLP learning point-wise weighting from lookback to forecast window. The basic blocks can be replaced by our Koopa Block, which is composed of Disentanglement (Fourier Filter) and Koopman Predictor (Enc+Dec+Operator). It is notable that NBEATS do not explicitly design disentanglement as well, so we leverage the well-acknowledged trend-seasonal disentanglement proposed by Autoformer[2]. We also trail on the periodicity selection of Fourier series in Seasonal Block based on the sample rate of datasets for a fair comparison. | Datasets (MSE) | ECL | ETTh2 | Exchange | ILI | Traffic | Weather | | - | - | - | - | - | - | - | | NBeats Generic Block | 0.190 | 0.246 | 0.050 | 3.302 | 0.620 | 0.141 | | NBeats Trend Block | 0.201 | 0.258 | 0.048 | 2.610 | 0.656 | 0.142 | | NBeats Seasonal Block | 0.181 | 0.238 | 0.056 | 3.359 | 0.613 | 0.135 | | NBeats Decomp + Seasonal + Trend Block | 0.198 | **0.224** | 0.043 | 2.552 | 0.698 | 0.155 | | Koopa Block (Ours) | **0.130** | 0.226 | **0.042** | **1.621** | **0.415** | **0.126** | The results demonstrate the significant performance gain brought by Koopman-based block (even with disentanglement) in ECL, ILI, and Traffic datasets. Notably, Koopman theory that employs varying localized operators is suitable to tackle distribution shift and the introduction of measurement function learning enhances the model capacity for non-linearity. It is also quite different from recent research works (global local decomp + respective modeling) that the localized part is still modeled by global-shared learnable parameters. **4. Elaborate more on 3rd question in weakness** (a) Existing works of Koopman theory tackling non-stationary problems. While Koopman theory has been widely incorporated with AutoEncoder for sequential modeling and forecasting [3, 4], **there are still few works that attend to the power of Koopman theory for non-stationary problems**. We sum up the pipeline as follows: - Introduce extra constraints to refine operator stability on non-stationary series, (e.g. PCL [5]). However, the operator can be intrinsically unstable learned from non-stationary data, since the temporal statistics and evolution regime can change greatly. So the predicted time series can be over-stationary. - Use local and global operators to deal with temporal distribution shifts. Since learning one operator for the whole dynamics is hard, KNF [6] as the most related work for non-stationary forecasting proposes to learn a unified global operator and utilize the self-attention map within the window as local operators. It directly **adds the global and local operator** as a time-variant transition, while we **add the components** given by respective predictors. It uses **classical AE structure** with reconstruction and forecasting loss and employs MLP for the coefficients on pre-defined measurement functions. (b) Elaborate more on lines 40-42. Take Duffing oscillator as shown on the left of $\underline{\text{Figure 1 of the main text}}$ for an example. It is hard to directly find one operator for portraying the dynamics but if we divide the dynamics into three localized sub-regions, we find several operators with simpler forms can be enough to describe. In that, for non-stationary time series with complicated time-variant $K_t$, **we discretize it into several window-wised operators** governing different periods (sub-regions). (c) Elaborate more on lines 52-54 and lines 30-31. As clarified in $\underline{\text{Q2 and Q3}}$, we hope our response can fulfilled the reviewer's expectations and would be very happy to answer any further detailed questions. [1] N-BEATS: Neural basis expansion analysis for interpretable time series forecasting. [2] Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting. [3] Deep learning for universal linear embeddings of nonlinear dynamics. [4] Learning deep neural network representations for koopman operators of nonlinear dynamical systems. [5] Forecasting sequential data using consistent koopman autoencoders. [6] Koopman neural forecaster for time series with temporal distribution shifts.

Authorsrebuttal2023-08-19

Looking forward to your reply.

We sincerely appreciate the time you dedicated to reviewing our paper. Given the limited timeframe for author-reviewer discussion, please kindly let us know if our response has addressed your concerns. Following your suggestions, we improve the paper in the following aspects: - We clarify how the proposed architecture reduces the approximation error of non-linear dynamics. - We further analyze the hyperparameter sensitivity of D. - We add comparisons with the deep architecture design and highlight the performance gain stemming from a better approximated non-linear dynamics with Koopman theory. - We elaborate on more on your 3rd question in weakness. Thanks again for your valuable review. Looking forward to your reply.

Authorsrebuttal2023-08-21

Request of Reviewer’s attention and feedback

Dear Reviewer EGYL, Thanks again for your valuable and constructive review, which has inspired us to improve our paper further substantially. Following your suggestions, we have clarified the connection between our architecture design to Koopman theory, further analyzed the hyperparameter sensitivity of $D$​, demonstrated the performance gain benefited from our Koopman-based blocks, and discussed all your mentioned weaknesses in every detail. We do our best to solve your concerns in the limited time and characters. We hope that this new version has addressed your concerns to your satisfaction. We eagerly await your reply and are happy to answer any further questions. We kindly remind you that the reviewer-author discussion phase will end by Aug 21st at 1 pm EDT, with just 4 hours left. After that, we may not have a chance to respond to your comments. Sincere thanks for your dedication! Authors

Reviewer M3Xz2023-08-18

Thanks for the efforts made on the response. The authors have clarified their motivations and emphasized the novelty. Additional experiment on Cryptos is included. Most of my concerns have been well addressed and I would like to raise my score to 5.

Authorsrebuttal2023-08-18

Thanks for Your Response and Raising the Score

We would like to thank Reviewer M3Xz for providing a detailed valuable pre-rebuttal review, which helps us a lot in the rebuttal and paper revision. Thanks again for your response and raising the score! In the final version, we will elaborate more on clarifing motivations and novelty, and include the additional experiments to the paper.

Reviewer 36zV2023-08-21

official comment from reviewer 36zV

Thanks for responses. The authors answer my questions. I will keep my score to support the paper.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC