New CosDir loss function improves time series forecasting accuracy

Researchers have introduced CosDir, a novel loss function designed to improve time series forecasting by explicitly optimizing for the direction of change. Unlike traditional methods that focus on magnitude or shape, CosDir uses cosine similarity to align prediction and target difference vectors, providing a learning signal for small directional moves where other losses fail. An extension, CosDir-UW, adaptively learns the optimal mixing ratio for directional and magnitude terms, eliminating the need for manual tuning. Extensive experiments show CosDir consistently enhances directional accuracy while maintaining magnitude accuracy, outperforming existing loss functions. AI

IMPACT Enhances directional accuracy in time series forecasting, potentially improving decision-making in finance and risk management.

RANK_REASON Academic paper introducing a new methodology. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• CosDir
• CosDir-UW

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →