Paper reveals 'Forecast Collapse' in Time-Series Foundation Models
chaumian · x · 2026-08-17
A new paper identifies 'forecast collapse' where Time-Series Foundation Models produce nearly flat predictions with poor ranking when forecasting hourly stock returns. This phenomenon stems from low predictability and per-series objectives. The authors introduce CalibRank, an objective balancing calibration and ranking, which nearly triples cross-sectional correlation on the Finance1K dataset.
Related event: Research Reveals Forecast Collapse in Time-Series Foundation Models(2 posts)→
More from Research
- New MIT Paper: Intelligence Becomes Science via Closed Epistemic Loops — ProfBuehlerMIT · 2026-08-17
- Weaviate Podcast: Why "Retrieve More, Rerank" might be wrong — CShorten30 · 2026-08-17
- DeepSeek paper fixes Transformer flaw causing 3000x signal amplification — petrusenko_max · 2026-08-17
- Opinion: Not all knowledge needs to live in model weights — coallaoh · 2026-08-17
- Separating knowledge enables editing and addresses hallucinations — coallaoh · 2026-08-17
- PUREdrop Platform Automates Synthetic Cells to Screen Designed Proteins — bravo_abad · 2026-08-17