Ensuring Fair Comparisons in Time Series Forecasting: Addressing Quality Issues in Three Benchmark Datasets

Guillaume Habault, Atsunori Minamikawa

Keywords: Time Series; Dataset Quality; Fair Comparisons; Benchmark Datasets

TL;DR: Current Time Series Forecasting Datasets suffers from data inconsistencies, we propose cleaner version of three datasets in order to ensure fairer comparisons.

Abstract: Time series forecasting (TSF) is critical in numerous applications; however, unlike other AI domains where benchmark datasets are meticulously standardized, TSF datasets often suffer from data inconsistencies, missing values, and improper temporal splits. These issues have an impact on model performance and evaluation. This paper addresses these challenges by proposing inconsistency-free versions of three well-known TSF datasets. Our methodology involves identifying and correcting data inconsistencies using a combination of linear interpolation and context-aware imputation strategies. Additionally, we introduce a novel cycle-inclusive data splitting method, which respects the longest cycle in each dataset, ensuring that models are evaluated over meaningful temporal patterns. Through extensive testing of multiple transformer-based models, we demonstrate that our revised datasets and cycle-inclusive splitting lead to more accurate and interpretable forecasting results, as well as fairer comparison of TSF models. Finally, our findings highlight the need for proper dataset refinement and tailored data splitting strategies in TSF tasks, and pave the way for future work in the development of more robust forecasting benchmarks.

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.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

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: 8658