核心概念
Designing a novel loss function, TILDE-Q, enables shape-aware time-series forecasting by considering amplitude shifting, phase shifting, and uniform amplification invariances.
要約
Time-series forecasting is crucial across various domains. Existing models struggle with complex temporal patterns. TILDE-Q addresses this by introducing a shape-aware loss function that considers amplitude, phase distortions, and shapes of time-series sequences. Experimental results show TILDE-Q outperforms other metrics in real-world applications like electricity, traffic, and weather forecasting.
統計
MSE: 0.179 for Electricity dataset.
MAE: 0.412 for ETT dataset.
Improved accuracy by 8.85% with TILDE-Q compared to MSE in real-world applications.
引用
"In this paper, we examine the definition of shape and distortions crucial for shape-awareness in time-series forecasting."
"Our study enhances understanding of the impact of distortions on time-series forecasting problems."