Core: Transferable Long-Range Time Series Forecasting Enhanced by Covariates-Guided Representation

Abstract

In recent years, long-range time series forecasting has been actively studied and has shown promising results. However, since these methods mainly focus on predicting time series with a fixed dimension, they are inapplicable to the large-scale and ever-changing datasets that are common in real-world applications. Additionally, existing methods only take a window of the near past as input, which prevents the models from learning persistent historical patterns. To tackle these problems, we propose CoRe, a novel transferable long-term forecasting method enhanced by Covariates-guided Representation. By encoding the input series into a dense vector, CoRe is able to extract instance-wise global features. Specifically, the representation is learned by modeling the correlation between the target series and constructed auxiliary covariates, which is implemented by our proposed cross-dependency network. Comprehensive experiments on six real-world datasets show that CoRe achieves overall state-of-the-art results and can transfer to unseen data with stable performance.