RESEARCH · RESEARCH · #2972
QUAL: Quantile alignment improves zero-shot time-series foundation models
An arXiv paper introduces QUAL, a label-free framework that aligns member-model quantiles relative to a model pool and then combines forecasts via Huber aggregation and rearrangement; on GIFT-Eval it reports average member MASE and CRPS reductions of 7.4% and 9.8%, and code is available on GitHub.
KEY POINTS
- An arXiv paper introduces QUAL, a label-free framework that aligns member-model quantiles relative to a model pool and then combines forecasts via Huber aggregation and rearrangement; on GIFT-Eval it reports average member MASE and CRPS reductions of 7.4% and 9.8%, and code is available on GitHub.
- It offers a simple, training-free way to reduce variability and improve zero-shot forecasts from TSFMs, enabling more robust ensemble deployment without target labels.
- QUAL: Quantile Alignment Improves Zero-Shot Time-Series Foundation Models
WHY IT MATTERS
It offers a simple, training-free way to reduce variability and improve zero-shot forecasts from TSFMs, enabling more robust ensemble deployment without target labels.