Researchers have introduced Timer-M1, a novel multivariate time series foundation model designed for zero-shot forecasting by learning temporal and relational primitives. This model utilizes a primitive-based data synthesis and pretraining pipeline, incorporating gated two-dimensional Transformer blocks for dynamic cross-variate attention. In parallel, a separate study proposes Scale-Invariant Training (ScaleIn) for time series foundation models, addressing issues of scale-contaminated training (ScaleCon) where series with larger scales disproportionately influence the model. ScaleIn aims to improve convergence and forecasting accuracy by ensuring gradients are not dominated by high-scale series, showing significant reductions in error across various benchmarks and architectures. AI
IMPACT These advancements in time series foundation models could lead to more robust and accurate forecasting across various domains, impacting fields like finance, climate science, and resource management.
RANK_REASON Two research papers published on arXiv detailing new methods for time series foundation models.
- GIFT-Eval
- Hugging Face
- MAE
- M-competitions
- Microsoft Security Essentials
- Quantile Loss
- Reversible Instance Normalization
- ScaleCon
- ScaleIn
- arXiv
- Timer-M1
- transformer
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →