Learning to Generate Predictor for Long-Term Time Series Forecasting

Jun Seo, Jinseok Yang, Hyeokjun Choe, Soyeon Park, Seohui Bae, Wonbin Ahn, Dongwan Kang, Minjae Kim, Hyeseung Cho, Woohyung Lim

Primary Area: general machine learning (i.e., none of the above)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Time series forecasting, Learning to Generate

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract:

Long-term time series forecasting (LTSF) is a significant challenge in machine learning with numerous real-world applications. Although transformer architecture have shown promising performance in the LTSF task, recent research suggests that they are not suitable for time series forecasting due to their permutation invariant characteristic, and proposes a simple linear predictor which outperforms all existing transformer architectures. However, the linear predictor is inflexible and cannot reflect the characteristics of the time series for prediction due to its simple architecture. In this paper, we introduce a novel Learning to Generate Predictor (LGPred) framework, which generates a linear predictor adaptively to the given input time series by leveraging time series decomposition. LGPred obtains representations from the decomposed time series and generates a predictor suitable for the given time series from these representations.
Our extensive experiments demonstrate that LGPred achieves state-of-the-art performance for both multivariate and univariate forecasting tasks.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 4877