Khái niệm cốt lõi
Exploiting locally stationary lead-lag relationships between variates improves multivariate time series forecasting accuracy.
Tóm tắt
Introduction
Multivariate time series (MTS) forecasting is crucial in various domains like weather, traffic, and finance.
Traditional channel-dependent (CD) methods are being outperformed by channel-independent (CI) methods.
Locally Stationary Lead-Lag Relationships
Variates exhibit lead-lag relationships where some lagged variates follow leading indicators within a short period.
Exploiting this channel dependence can reduce forecasting difficulty by utilizing advance information.
LIFT Approach
LIFT method estimates leading indicators and their steps, aligns variates with leading indicators, and refines predictions using a Lead-aware Refiner.
Lightweight MTS Forecasting with LIFT
LightMTS, a lightweight method using LIFT, shows competitive performance compared to complex models.
Experiments
LIFT outperforms state-of-the-art CI and CD models on various datasets, improving average forecasting performance by 5.5%.
Ablation Study
Removing influence or difference terms in LightMTS leads to inferior performance, highlighting the importance of considering both lead-lag relationships.
Hyperparameter Study
Increasing the number of selected leading indicators generally improves performance but may introduce noise with excessively large values.
Thống kê
Extensive experiments on six real-world datasets demonstrate that LIFT improves the state-of-the-art methods by 5.5% in average forecasting performance.