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

TL;DR

Time-Series Foundation Models must generalize across domains with non-stationary data, and normalization plays a pivotal yet under-explored role. The authors benchmark six normalization methods across four TSFMs and multi-domain datasets, revealing that mean–std based normalization with RevIN at inference yields the best accuracy-efficiency trade-off. RevIN matches in-domain accuracy without dataset-level preprocessing while dramatically improving zero-shot MASE, and the effectiveness depends on the model's optimization objective. The findings offer practical guidelines for TSFM pretraining and generalization, and pave the way for TSFM-centric normalization strategies in other temporal tasks.

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).