A Comparative Study on how Data Normalization Affects Zero-Shot Generalization in Time Series Foundation Models
A Comparative Study on how Data Normalization Affects Zero-Shot Generalization in Time Series Foundation Models
Authors: Ihab Ahmed, Denis Krompaß, Cheng Feng, Volker Tresp Date: 2026-04-21 Paper ID: openalex:2512.02833
Summary
This paper investigates the role of input normalization in time series foundation models (TSFMs) to address challenges posed by scale variation and non-stationarity. By evaluating four diverse TSFM architectures, the study determines that RevIN consistently achieves superior zero-shot performance and efficiency without requiring dataset-specific preprocessing. The findings highlight the critical interaction between normalization methods, model architecture, and training objectives in determining the effectiveness of zero-shot generalization.
Key Contributions
- Identifies RevIN as the most efficient normalization strategy for TSFMs, reducing zero-shot MASE by 89% against raw baselines.
- Demonstrates that optimal normalization for TSFMs is highly dependent on architectural choices (point-forecast vs. distribution vs. LLM-based) and loss functions.
- Provides an empirical analysis of how input normalization affects zero-shot generalization and accuracy-efficiency trade-offs in time series foundation models.
Open Questions & Future Work
Key Concepts
- revin: A reversible instance normalization technique designed to handle non-stationarity and scale variation in time series data.
Archivist Review
I approved ‘RevIN’ as a central concept because the paper establishes its efficacy in the context of TSFMs, shifting it from a standard preprocessing step to a key architectural consideration. I also added an open question regarding the sensitivity of foundation model architectures to different normalization techniques, as this addresses a specific research bottleneck in zero-shot time-series forecasting. I rejected no candidates because the user only provided one concept and no open questions.
Approved Concepts
- RevIN: The study provides empirical validation that RevIN, originally for traditional time-series models, is highly effective for improving zero-shot generalization in Foundation Models.
Approved Open Questions
- TSFM Normalization Architectural Sensitivity: Clarifying this architectural sensitivity is critical for designing future time-series foundation models that robustly handle distribution shifts.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2512.02833
- paper_id
- 2512.02833
- paper_source
- openalex
- domain
- time-series
- tags
- time-series-foundation-modelsnormalizationzero-shot-learningtime-series-forecastinggeneralization
- architectures
-
- datasets
-
- concept_slugs
- revin
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-04-24T05:50:29Z