A Comparative Study on How Data Normalization Affects Zero-Shot Generalization in Time Series Foundation Models

TL;DR

This paper systematically evaluates normalization methods for Time-Series Foundation Models, finding REVIN to significantly improve zero-shot generalization and efficiency across diverse architectures without dataset preprocessing.

Contribution

It provides the first comprehensive empirical analysis of normalization impacts on TSFMs, highlighting REVIN's effectiveness and dependencies on model design and training objectives.

Findings

REVIN reduces zero-shot MASE by 89% compared to unnormalized models.

REVIN achieves 44% better MASE than other normalization methods.

REVIN offers the best accuracy-efficiency trade-off without dataset preprocessing.

Abstract

We investigate input normalization methods for Time-Series Foundation Models (TSFMs). While normalization is well-studied in dataset-specific time-series models, it remains overlooked in TSFMs where generalization is critical. Time-series data, unlike text or images, exhibits significant scale variation across domains and channels, coupled with non-stationarity, can undermine TSFM performance regardless of architectural complexity. Through systematic evaluation across four architecturally diverse TSFMs, we empirically establish REVIN as the most efficient approach, reducing zero-shot MASE by 89\% relative to an un-normalized baseline and by 44\% versus other normalization methods, while matching the best in-domain accuracy (0.84 MASE) without any dataset-level preprocessing -- yielding the highest accuracy-efficiency trade-off. Yet its effect utilization depends on architectural design choices and optimization objective, particularly with respect to training loss scale sensitivity and model type (probabilistic, point-forecast, or LLM-based models).