LST-Bench:A Benchmark for long sequence time-series forecasting Task
Primary Area: datasets and benchmarks
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, Deep Learning, Neural Networks, Data Mining
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: This paper introduces LST-Bench, a comprehensive benchmark designed for evaluating long sequence time-series forecasting(LSTF) models. This benchmark has been developed in response to recent advancements in deep learning methods in the field of LSTF tasks. LST-Bench includes Transformer-based, MLP-based, CNN-based, and RNN-based models, evaluating the performance of 11 major forecasting models on a set of commonly used 7 datasets and 7 new datasets that we have introduced. We conduct a thorough analysis of the experimental results, including the overall prediction performance of models and their generalization across different prediction lengths and datasets. Notably, we found that regardless of the model architecture, the phenomenon referred to as "Degeneracy" occurs when the model's predictions consistently maintain a low Mean Squared Error value but are characterized by repetitive and simplistic pattern generation, thus losing the meaningfulness of the predictions. Also, the model's optimal performance is very close to its performance after training for just one epoch. These two phenomenons emphasize the need for further investigation. Our LST-Bench will serve as a valuable resource for advancing research in the field of time series forecasting.
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: 9299
Loading