Time-o1: Time-Series Forecasting Needs Transformed Label Alignment
Hao Wang, Licheng Pan, Zhichao Chen, Xu Chen, Qingyang Dai, Lei Wang, Haoxuan Li, Zhouchen Lin
TL;DR
Time-o1 tackles two core issues in time-series forecasting objectives: autocorrelation bias and the explosion of multitask complexity with longer horizons. It transforms the label sequence into decorrelated components ranked by significance and trains models to align the most informative components, yielding debiased training and easier optimization. The approach is theoretically justified and implemented via a unified pipeline using SVD-based projection and a mixed objective L_{α,γ}. Empirically, Time-o1 consistently improves state-of-the-art forecasts across diverse datasets and backbone models, demonstrating strong practical utility and model-agnostic applicability. The work also provides extensive ablations and generalization studies, confirming the robustness and versatility of the proposed objective.
Abstract
Training time-series forecast models presents unique challenges in designing effective learning objectives. Existing methods predominantly utilize the temporal mean squared error, which faces two critical challenges: (1) label autocorrelation, which leads to bias from the label sequence likelihood; (2) excessive amount of tasks, which increases with the forecast horizon and complicates optimization. To address these challenges, we propose Time-o1, a transformation-augmented learning objective tailored for time-series forecasting. The central idea is to transform the label sequence into decorrelated components with discriminated significance. Models are then trained to align the most significant components, thereby effectively mitigating label autocorrelation and reducing task amount. Extensive experiments demonstrate that Time-o1 achieves state-of-the-art performance and is compatible with various forecast models. Code is available at https://github.com/Master-PLC/Time-o1.