Beyond Static Positional Encoding: Bayesian Time-aware Attention for Non-Stationary Long-term Time Series Forecasting

arXiv:2610.19603 · cs.LG, cs.AI · Submitted 2026-10-14 · Read on arXiv

cs.LG, cs.AI

Submitted: 2026-10-14

Updated: 2026-10-14

Comments: 44 pages, 12 figures, including appendices

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Terminology

• Time Series Forecasting — 1× in this paper · explained in 4 episodes

Sources

• Recency Biased Causal Attention for Time-series Forecasting
• Positional Encoding in Transformer-Based Time Series Models: A Survey
• Foundation Models for Time Series: A Survey
• FredNormer: Frequency Domain Normalization for Non-stationary Time Series Forecasting
• Train Short, Test Long: Attention with Linear Biases Enables Input Length Extrapolation
• Self-Attention with Relative Position Representations
• Transformer Dissection: A Unified Understanding of Transformer's Attention via the Lens of Kernel
• TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting
• Transformers in Time Series: A Survey

Related papers

• Polynomial-Augmented Neural Networks (PANNs) with Weak Orthogonality Constraints for Enhanced Function and PDE Approximation
• AIRL-S: Unifying Reinforcement Learning and Search-Based Test-Time Scaling via Adversarial Inverse Reinforcement Learning
• Transformers as Bayesian In-Context Experimenters: Smoothness-Adaptive Efficient ATE Estimation
• Convergence issues in Relational Concept Analysis based on AOC-posets
• Beliefs Beyond Posteriors: The Predictive Cost of ELBO Binding
• Understanding Diffusion Models via Ratio-Based Function Approximation with SignReLU Networks