[Paper Review] On the Role of Reversible Instance Normalization

Gaspard Berthelier, Tahar Nabil+4 more|arXiv (Cornell University)|Mar 12, 2026

Forecasting Techniques and Applications0 citations

TL;DR

This paper critically analyzes Reversible Instance Normalization (RevIN) for time series forecasting, showing which components are essential and where RevIN falls short under distribution shifts. It provides ablations across benchmarks to guide enhancements for robustness and generalization.

ABSTRACT

Data normalization is a crucial component of deep learning models, yet its role in time series forecasting remains insufficiently understood. In this paper, we identify three central challenges for normalization in time series forecasting: temporal input distribution shift, spatial input distribution shift, and conditional output distribution shift. In this context, we revisit the widely used Reversible Instance Normalization (RevIN), by showing through ablation studies that several of its components are redundant or even detrimental. Based on these observations, we draw new perspectives to improve RevIN's robustness and generalization.

Motivation & Objective

• Identify the challenges normalization faces in time series forecasting (temporal, spatial, and conditional distribution shifts).
• Evaluate the components of Reversible Instance Normalization (RevIN) through controlled ablations.
• Determine which RevIN components are truly beneficial and under which conditions they help or hurt generalization.

Proposed method

• Review and categorize normalization methods used in neural forecasters for time series.
• Conduct extensive ablation studies of RevIN on standard forecasting benchmarks (Electricity, Solar, Traffic, Synthetic) using a consistent PatchTST backbone.
• Compare normalization strategies (standard, RevIN, RevIN without affine) under standard backpropagation and normalized backpropagation.
• Analyze the role of the affine layer (α, β) and the effect of training in normalized vs. denormalized spaces.
• Provide empirical and qualitative discussion on heterogeneity, conditional shifts, and limitations of RevIN.

(a) Temporal shift.

Experimental results

Research questions

• RQ1 Does RevIN mitigate temporal and spatial distribution shifts in time series forecasting?
• RQ2 Is the additional affine layer in RevIN (α, β) beneficial for handling conditional shifts?
• RQ3 Should training occur in the normalized space or the denormalized space for best generalization?
• RQ4 Under what data conditions might RevIN hurt performance due to heterogeneity or stationary bias?
• RQ5 What extensions or alternatives could address limitations of RevIN in conditional distribution shifts?

Key findings

• Instance normalization generally improves generalization for new dates and new users on average, indicating RevIN helps with temporal and spatial shifts.
• The affine layer (α, β) is not beneficial in practice and does not mitigate conditional shifts effectively.
• Training the model in the normalized space with backpropagation yields better models, even when evaluating with non-normalized MSE.
• RevIN may reduce heterogeneity between training and test distributions but does not fully address all forms of heterogeneity, and can hurt performance on some stationary datasets.
• Some extensions proposed in related work (e.g., normalization in different domains or slice-wise approaches) are not consistently superior across tasks.

Figure 2: Illustration of the RevIN process on three synthetic examples.