New ORBIT training paradigm boosts time series foundation models

Researchers have introduced ORBIT, a novel training paradigm designed to enhance the performance of time series foundation models (TSFMs). ORBIT addresses limitations in current training methods by explicitly controlling the distribution of pre-training data, managing domain imbalance, context requirements, prediction horizons, and missing data. The paradigm integrates Bootstrap Multi-Level Sampling for dataset exposure and Omni-Range Incremental Training for varying context lengths and prediction horizons. Evaluations using the Falcon-2.0 model on GIFT-Eval and fev-bench benchmarks show significant improvements in zero-shot forecasting capabilities across various domains and frequencies. AI

IMPACT Introduces a new training methodology that could improve the accuracy and robustness of time series forecasting models.

RANK_REASON The cluster describes a new training methodology and model presented in an academic paper on arXiv. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• Bootstrap Multi-Level Sampling
• fev-bench
• GIFT-Eval
• Hugging Face
• Omni-Range Incremental Training
• ORBIT
• Rank-Guided Cross-Depth Alignment
• Transformer++

• paper
• model release

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