Rethinking the Power of Timestamps for Robust Time Series Forecasting: A Global-Local Fusion Perspective

Time series forecasting has played a pivotal role across various industries, including finance, transportation, energy, healthcare, and climate. Due to the abundant seasonal information they contain, timestamps possess the potential to offer robust global guidance for forecasting techniques. However, existing works primarily focus on local observations, with timestamps being treated merely as an optional supplement that remains underutilized. When data gathered from the real world is polluted, the absence of global information will damage the robust prediction capability of these algorithms. To address these problems, we propose a novel framework named GLAFF. Within this framework, the timestamps are modeled individually to capture the global dependencies. Working as a plugin, GLAFF adaptively adjusts the combined weights for global and local information, enabling seamless collaboration with any time series forecasting backbone. Extensive experiments conducted on nine real-world datasets demonstrate that GLAFF significantly enhances the average performance of widely used mainstream forecasting models by 12.5%, surpassing the previous state-of-the-art method by 5.5%.

Paper

Similar papers

Peer review

Reviewer N7JP7/10 · confidence 5/52024-06-30

Summary

This article focuses on enhancing time series forecasting capabilities using timestamps and introduces a plug-and-play module called GLAFF. Overall, GLAFF is designed to be simple and lightweight, significantly improving the predictive performance of existing time series forecasting algorithms such as ITransformer and DLinear.

Strengths

1. The article is well-structured, and the writing is clear and appropriate. 2. The motivation is explicit, and the method is straightforward and efficient. The core code is provided in the appendix for readers' convenience. 3. The experiments are relatively comprehensive, involving both large language models and task-specific general models.

Weaknesses

1. The article contains some typos, such as "damaged" in line 32 and "aimed" in line 52. 2. Formula 3 uses quantile for denormalization, but the reason for this choice is not explained. Why is quantile better than std for this purpose? 3. Observing Figure 3, I notice that while GLAFF improves the forecasting performance, it still fails to fully capture the traffic spikes. Can you explain why this is the case and suggest any other potential solutions to address this issue? 4. For the ablation experiment w/o Quantile, I would like to see the results of completely removing robust denormalization.

Questions

Please refer to the weaknesses section.

Rating

7

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors have discussed the limitations of the article in the appendix.

Reviewer Y7JY6/10 · confidence 3/52024-07-11

Summary

This paper introduces the GLAFF framework where time series models are adapted to also capture "global" information by using information content in the datetime parsed in components of habitual meaning to complement baseline models ("backbones"). The global information is represented through a Mapper, quantile-based Denormalizer, and Combiner, and is combined in an MLP head with the backbone output to form the prediction. While extracting the time units from datetimes is used in other NLP literatures, it has not been applied in time series forecasting models as far as I am aware.

Strengths

The work appears to be original and augments leading time series models, leading to performance improvements on a suite of time series prediction benchmarks as compared to the respective baseline models without GLAFF. The baseline models held previous SOTA performances in the last 5 years or so. Relevant alternatives (GPT2, Table 2) and relevant ablations are done for the best performing baseline iTransformer (Table 3). The results are presented clearly, with full tables of results in the appendix.

Weaknesses

The intuition paragraphs describing the motivation for the components could be made clearer, in particular with respect to "mitigating data drift"? What assumptions are made explicitly, and what is the precise formulation? The GLAFF framework could have be applied to a larger set of time series models, or at least applied to the best performing one for each dataset, since they are readily accessible and compared against: (e.g. https://github.com/thuml/Time-Series-Library). Hyperparameter searches were done on axes rather than a grid.

Questions

See first two limitations above.The

Rating

6

Confidence

3

Soundness

4

Presentation

3

Contribution

2

Limitations

The limitation section is in the appendix and describe computational cost. So the limitations section should be expanded upon, e.g. with respect to moving away from small benchmark datasets, irregular sampling, data drift settings, etc.

Authorsrebuttal2024-08-12

Dear Reviewer Y7JY, We greatly appreciate the time and effort you have invested in reviewing our paper and providing insightful feedback. As a gentle reminder, it has been more than 5 days since we submitted our rebuttal. As the discussion period is drawing to a close, we wish to ensure that our rebuttal has comprehensively addressed your concerns. We are keen to receive any further feedback you might have and are prepared to make additional clarifications or modifications as needed. Thank you once again for your valuable insights. We look forward to your final thoughts.

Reviewer Y7JY2024-08-12

The additional experiments further back the method and are appreciated. I remain with minor concerns about the exposition of the intuition/motivation which I believe could be further tightened, though I believe these could be addressed prior to a camera ready version.

Authorsrebuttal2024-08-13

Thank you for carefully reviewing our rebuttal and actively providing feedback. We will present a more tightened exposition of the motivation for this paper. > Time series forecasting is vital across various domains. However, existing models predominantly rely on local observations and inadequately utilize the extensive global information embedded in timestamps. This oversight reduces the robustness of these models, particularly when real-world data is noisy or contains anomalies. To address this issue, we propose GLAFF, an innovative framework that more comprehensively integrates timestamp information through late fusion (decision-level fusion), thereby enhancing the accuracy and robustness of time series forecasting backbones. Regarding late fusion, we provide the following explanation. > Early fusion (Informer) integrates modalities into a single representation at the input level and processes the fused representation through the model. Late fusion (GLAFF) allows each modality to run independently through its own model and fuses the outputs of each modality. Compared to early fusion, late fusion maximizes the processing effectiveness of each modality and is less susceptible to the noise of a single modality, resulting in greater robustness and reliability. We hope that these clarifications have addressed your concerns. Should you have any further concerns or questions, please do not hesitate to contact us.

Authorsrebuttal2024-08-13

Dear Reviewer Y7JY, We sincerely appreciate the time and effort you have devoted to reviewing our paper and offering valuable feedback. As the discussion period nears its conclusion, we wish to ensure that our rebuttal has thoroughly addressed your concerns. We are eager to receive any additional feedback you may have and are ready to provide further clarifications or make modifications as necessary. Lastly, we look forward to your final comments regarding the score.

Reviewer XBcx4/10 · confidence 4/52024-07-12

Summary

The paper introduces GLAFF, a novel framework that enhances time series forecasting by modeling timestamps to capture global dependencies and adaptively balancing global and local information, resulting in a significant improvement of 12.5% over existing methods in experiments across nine real-world datasets.

Strengths

1. The writing is clear and easy to understand. 2. Code is provided. 3. Comprehensive experiments consistently enhance performance.

Weaknesses

My concerns are as follows: 1. This paper focuses on utilizing timestamps but only discusses and compares some general time-series forecasting methods. Please discuss the differences and performance comparison with existing methods that focus on better utilizing timestamps, such as methods across temporal scales [1][2] and representation for each timestamp [3]. - [1] AutoCTS: Automated correlated time series forecasting - [2] METRO: a generic graph neural network framework for multivariate time series forecasting - [3] TS2Vec: Towards Universal Representation of Time Series 2. The method appears simple and lacks technical contributions, resembling a straightforward dual-pathway combination; it lacks theoretical backing, particularly in theoretical analysis of existing backbones' capabilities with timestamps, and only provides a case study for discussion. 3. I noticed that the adaptive weight has a minimal effect, and I am curious about how the weights in the Combiner are initialized, how they are tuned, and whether they are sensitive. 4. The usage has limitations, requiring high-precision timestamps. Please analyze the impact of timestamp granularity and noise on the results. 5. By simply increasing the lookback window, existing methods usually have better performance (better leverage periodicity). Please analyze the performance improvements under different lookback window lengths, especially longer lengths, to validate more realistic effectiveness.

Questions

Please check concerns

Rating

4

Confidence

4

Soundness

2

Presentation

3

Contribution

2

Limitations

The paper does not discuss limitations. Please elaborate more on when the method may perform poorly.

Authorsrebuttal2024-08-12

Dear Reviewer XBcx, We greatly appreciate the time and effort you have invested in reviewing our paper and providing insightful feedback. As a gentle reminder, it has been more than 5 days since we submitted our rebuttal. As the discussion period is drawing to a close, we wish to ensure that our rebuttal has comprehensively addressed your concerns. We are keen to receive any further feedback you might have and are prepared to make additional clarifications or modifications as needed. Thank you once again for your valuable insights. We look forward to your final thoughts.

Reviewer wbSq6/10 · confidence 4/52024-07-12

Summary

This paper proposes GLAFF which encodes the time stamps of time series and performs self attention across the encodings of the time dimensions, combining it with the output of a global time series forecaster via a learned weighting scheme. As GLAFF is a set of feature constructions, it’s generally additive in performance against any backbone architecture. The main requirement is that GLAFF seems to require knowledge of when the prediction should be produced for. More specifically, the time stamps themselves are used to generate de-medianed and de-quantized predictions. This is extra side information, and so it should generally tend to be helpful – some of the tested datasets have clear time information. For example, traffic tends to spike in the morning and afternoon hours, while electricity also has daily and monthly peaks and valleys. This is a natural encoding of seasonality and other similar regular events.

Strengths

Originality: - Most other papers in this literature tend to focus primarily on architecture, and the ones that do take into account some sort of feature information seem to not incorporate it that well. - Overall, I like the idea of separating the time encoding into what’s essentially its own network. Quality: - I appreciate the knockout studies of the varying parts, which is good experimental design. - I also appreciate the knockout study of iTransformer and would like to see more of these types of results. Clarity: - Overall, the paper is pretty well written. It’s pretty much clear what’s going on. Significance: - Improving time series forecasting is obviously a highly important problem, and improving the base model is a general purpose technique that should generally be quite useful.

Weaknesses

Quality: - My biggest concern in terms of usefulness (and unfortunately, this is somewhat a critique of the entire vein of literature here) is that the time stamps of the prediction window are trained on. This can be, in some sense, a pretty strong lookahead bias. In most deployment settings, we cannot pre-train on the time stamps that we’re going to use to generate the sequence because we do not know the absolute value for the time stamp during train time. o Thus, the method is limited to explicitly semi-regular time series where the forecast clock is known ahead of time. By non-regular forecasting, we can think of wanting to predict the inventory levels of a product after the next sixty sales, or to predict the price of a foreign currency after some amount of trading volume in it. Instead, in regular forecasting, at the start of the day, we may wish to predict the next day’s electricity demands (one of the benchmark datasets). Clarity: - I personally find $\tilde X$, $\hat X$, $\hat Y$ to be quite confusing in terms of notation. Perhaps $\hat T$, $\tilde S$, $\hat S$ could be used instead throughout to denote the fact that these values come from the encoded time stamps. o In line 197, I would encourage against setting the final prediction as $Y$, but rather as $\hat Y$. - I also find the methods section somewhat unclear, as it’s a bit tricky to parse out that the median / quantile stats are dependent on the input time series, X. Perhaps it would be better to not separate X and Y entirely, but rather point out that the time series is really cat(X, Y). - From the intro and contributions, it could probably be made clear that other papers _do_ consider time stamp information; however, it doesn’t seem to be helpful in knockout studies. This is more of a writing note than something major.

Questions

- In Eq 3, where do the median and quantiles of Y come from? Do they come from the global set of sequences? This is I think more for clarity at this point but want to confirm. - It seems that most of the improvement in exchange and weather, which are not periodic, is driven by improvement in the worst model (informer). In general, should we expect that there is less improvement on aperiodic data because the time embeddings are less helpful? - Often in time series prediction tasks, there is other side information, is this naturally best situated to being encoded solely inside the backbone architecture or should it be modeled alongside the time stamps? - Tables 1 and 2: I’d suggest using bar charts to make the presentation more engaging, and moving the tables to the appendix. - I find it fairly surprising that informer, timesNet, and iTransformer tend to have minimal dropoff when removing the timestamp pieces. Is this lack of dropoff consistent across datasets? - Could the authors include an average improvement across methods for a fixed horizon? It would be interesting to see if DLinear improves the most as a result of adding in the time step information.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

n/a

Reviewer NzBw7/10 · confidence 4/52024-07-12

Summary

The authors proposed a plugin to utilize global information from timestamps in time series forecasting tasks. The proposed plugin consists of three main components: attention-based timestamp mapper, robust denormalizer and adaptive combiner. The authors found that using the proposed plugin in combination with various backbone models helped achieve an average of 12.5% increase in forecasting prediction performance measured as mean square error and mean absolute error on 9 real-world datasets.

Strengths

• The paper is well structured and easy to follow. • The authors compared the proposed method with different timestamps treatments like summation, concatenation and omission. • Well-illustrated prediction showcases provide examples of the usefulness of the proposed plugin. • The experimental results are comprehensive and impressive. Extensive results on 9 real-world datasets in five domains confirm the superiority of the plugin models used. The authors presented the average percentage improvement of individual models and datasets, which is useful for reviewing the results. • The results with the proposed plugin are best for all datasets and backbone models used. • The prosed method requires only one core hyperparameter, the quantile q in the robust denormalizer, which is 0.75 by default. • The proposed method is flexible and can be used with any backbone model. • The authors conducted an ablation study, proving that all components are important. The authors explained how core components can be helpful data drift and concept drift mitigation. • The authors conducted computation time and memory usage study.

Weaknesses

• The overhead for lightweight models like DLinear is significant. • The authors did not provide average percentage increases in computation time and memory consumption for individual backbone models and datasets, which would be useful for quickly reviewing results. • Figure 2 suggests that statistics are calculated separately for historical and future mappings denormalization which is not true. • Lack of theoretical proof for the proposed plugin.

Questions

• Please refer to the weaknesses section. • Do the authors see a way to use the proposed method in datasets where timestamps are not available? • Do the authors plan to publish the exact same datasets used in the experiments? • It may be beneficial to see sample outputs from historical and future mappers.

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

As the authors mentioned, adding the plugin to backbone network cause additional computational time and memory usage.

Reviewer N7JP2024-08-08

The authors have effectively addressed my concerns. Based on these revisions and your thorough response, I have raised my score to 7. I recommend accepting this paper,

Authorsrebuttal2024-08-09

Thank you for thoroughly reviewing our rebuttal and deciding to raise the rating score. We appreciate your consideration and the time you have dedicated to evaluating our work.

Reviewer wbSq2024-08-11

Thanks for answering my questions, my opinion of the paper is somewhat improved. I trust the authors will cleanup the notation somewhat in the camera ready and update some of the plots. Overall, I think your approach (and you demonstrate this) does tend to account for time domain information better than existing methods. > As mainstream methods like iTransformer cannot address time series forecasting with missing values, we cannot expect a framework to solve all forecasting challenges. Indeed, my comment is really a critique of the existing literature and probably shouldn't be held too harshly against your work. However, I think it is a straightforward application from your approach (maybe you'd need to forecast when the next time stamp is), and one that could be extremely helpful in my different domain. The problems I suggested are also somewhat less periodic and likely more challenging than many of the benchmarks (hence increased practical utility). Thanks for providing more lookback windows as well in the updated experiments.

Authorsrebuttal2024-08-11

Thank you for meticulously reviewing our rebuttal. The non-regular forecasting scenario you describe is indeed more challenging, practically significant, and highly engaging. We will explore this further in the future. We greatly appreciate your consideration and time devoted to evaluating our work.

Authorsrebuttal2024-08-12

Dear Reviewer wbSq, We sincerely appreciate the time and effort you have devoted to reviewing our paper and offering valuable feedback. As the discussion period nears its conclusion, we wish to ensure that our rebuttal has thoroughly addressed your concerns. We are eager to receive any additional feedback you may have and are ready to provide further clarifications or make modifications as necessary. Lastly, we look forward to your final comments regarding the score.

Reviewer NzBw2024-08-11

I appreciate the authors' rebuttal which clarified my concerns and convinced me to raise the overall rating.

Authorsrebuttal2024-08-11

Thank you for thoroughly reviewing our rebuttal and deciding to raise the rating score. We appreciate your consideration and the time you have dedicated to evaluating our work.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC