QUAL Framework Improves Time-Series Foundation Model Forecasts via Quantile Alignment
November 2, 2026
QUAL is a label-free framework that improves zero-shot time-series forecasting by aligning quantiles across a model pool using Huber aggregation. On the GIFT-Eval benchmark, it reduces average member MASE and CRPS by 7.4% and 9.8% respectively without requiring model modifications.
HOW THIS AFFECTS YOU
●
builderYou can improve forecasting accuracy in production by ensembling existing TSFMs with this label-free alignment method.
●
researcherThis provides a new way to evaluate and combine diverse time-series models without access to target labels.