Reversible Instance Normalization (RevIN)

• Reversible Instance Normalization (RevIN) is an instance-wise normalization technique that uses window-specific mean and variance to stabilize time series forecasting.
• It transforms inputs into a normalized space for prediction and then inversely maps the forecasted output to the original scale using computed statistics and optional affine parameters.
• While effectively mitigating temporal and spatial distribution shifts, RevIN has limitations in addressing higher-order statistics and seasonal non-stationarity.

Reversible Instance Normalization (RevIN) is an instance-wise normalize-then-denormalize wrapper for time-series forecasting: each input window is normalized with its own statistics, forecasting is performed in the normalized space, and the prediction is mapped back through an inverse transformation to the original scale. Later literature describes it both as a “symmetric, instance-wise normalize-then-denormalize wrapper with a learnable affine transform” and as a reversible normalization layer pair $(\mathcal N,\mathcal N^{-1})$ wrapped around a forecasting backbone $f_\theta$. The name is also referenced through the title “Reversible Instance Normalization for Deep Learning-based Time Series Forecasting” in subsequent work (Fu et al., 6 Oct 2025). Across long-term time-series forecasting, RevIN and its successors have become a de facto plug-in for handling non-stationarity and distribution shift, particularly when mean and variance drift across windows or regimes (Zhang et al., 11 May 2026, Baiga et al., 31 May 2026).

1. Canonical formulation

A common formalization writes RevIN as

$$\tilde{x}=\alpha\frac{x-\mu_x}{\sigma_x}+\beta, \qquad \hat{y}=\sigma_x\left(\frac{f_\theta(\tilde{x})-\beta}{\alpha}\right)+\mu_x,$$

with input-window statistics

$$\mu_x = \frac{1}{L}\sum_{i=1}^L x_i, \qquad \sigma_x^2 = \frac{1}{L-1}\sum_{i=1}^L (x_i-\mu_x)^2.$$

Here $x\in\mathbb{R}^L$ denotes an input window, $f_\theta$ is the forecasting model, and $\alpha,\beta\in\mathbb{R}$ are learnable affine parameters. In a more generic notation, a reversible normalization layer is expressed as

$$\hat{\mathbf y} = \mathcal N^{-1}\!\bigl(f_\theta(\mathcal N(\mathbf x));\,\mathbf s(\mathbf x)\bigr),$$

where RevIN uses $\mathbf s(\mathbf x)=(\mu,\sigma)$ and

$$\mathcal N(\mathbf x)=\frac{\mathbf x-\mu}{\sigma},$$

optionally followed by learnable affine parameters such as $f_\theta$0 (Berthelier et al., 12 Mar 2026, Zhang et al., 11 May 2026).

This formulation distinguishes RevIN from static dataset-level normalization. The statistics are computed from the current instance-window rather than from the training set as a whole, and the same stored statistics are later used in the inverse map. In some implementations this affine stage is deliberately kept minimal; for example, one formulation uses scalar $f_\theta$1 shared across all series and channels in order to preserve channel independence (Baiga et al., 31 May 2026).

2. Distribution shift as the motivating problem

Later analyses organize the motivation for RevIN around three distinct kinds of shift: temporal input distribution shift, spatial input distribution shift, and conditional output distribution shift. Temporal input distribution shift arises when training and test windows from the same series but different periods have different input distributions. Spatial input distribution shift arises when test-time series come from unseen users, sensors, or locations with different offsets or scales. Conditional output distribution shift is the case in which the future given the past varies across time or across entities even when the observed look-back windows appear similar after normalization (Berthelier et al., 12 Mar 2026).

Within this taxonomy, RevIN is principally aimed at input-side heterogeneity. By normalizing each window with its own $f_\theta$2, it reduces sensitivity to changing level and amplitude and can therefore mitigate temporal and spatial input shift. The conditional output case is more difficult, because normalization can remove context that is predictive of the horizon. This point recurs throughout later literature: the mechanism is effective when nuisance variation is concentrated in first- and second-order statistics, but it is not a general solution to all forms of train-test mismatch (Berthelier et al., 12 Mar 2026).

A closely related motivation appears in long-term forecasting settings where the statistics of the forecast window differ from those of the input window. TA-SparseMG, for example, argues that reversible instance normalization helps mitigate sample-level distribution discrepancies, but also argues that conventional RevIN restores outputs using fixed input-window statistics, which is limiting when the future window does not share the same mean and variance as the past (Liu et al., 26 Jun 2026).

3. Operational role in forecasting pipelines

Operationally, RevIN is usually a five-step procedure: compute instance statistics, normalize the input, apply an affine transform, forecast in normalized space, and then invert the affine transform and denormalize. This means that RevIN is not merely a preprocessing heuristic; it is an input-output wrapper around the predictive model. In TA-SparseMG, the normalization module is placed at the front end and the corresponding restoration step at the back end, so that the model is trained between these two reversible operations. In TimeGMM, the same pattern appears in a probabilistic setting: the input multivariate series is normalized before temporal encoding, and the predicted distribution parameters are mapped back before the final probabilistic objective is computed (Liu et al., 26 Jun 2026, Liu et al., 18 Jan 2026).

The training objective can also be placed either in the original space or in the normalized space. One analysis defines a normalized target

$f_\theta$3

and the normalized MSE

$f_\theta$4

Its empirical conclusion is that training in normalized space works better, even when evaluation is ultimately done in standard MSE. The same study also reports that the learnable affine parameters $f_\theta$5 are often redundant in practice, with RevIN and RevIN without $f_\theta$6 performing almost identically across the reported settings (Berthelier et al., 12 Mar 2026).

This suggests a useful distinction between the reversible standardization itself and the auxiliary affine correction. The former carries the central inductive bias of RevIN; the latter may or may not contribute, depending on the dataset and backbone.

4. Limitations and critical reassessments

A central criticism of RevIN is that its per-point transformation remains strictly affine. Later work states that RevIN and related methods such as Dish-TS, SAN, and FAN apply maps of the form

$f_\theta$7

No matter how $f_\theta$8 and $f_\theta$9 are computed, an affine map cannot change skewness or kurtosis; it only shifts and rescales the distribution. On this view, RevIN has a hard ceiling: heavy tails remain heavy and skewness remains uncorrected (Zhang et al., 11 May 2026).

A second criticism is that RevIN mainly removes and restores low-order temporal statistics such as mean and variance. Frequency Adaptive Normalization (FAN) argues that this makes RevIN suitable for trend-like shifts but “limited to expressing basic trends” and “incapable of handling seasonal patterns.” The criticism is not simply that RevIN is weak, but that seasonal non-stationarity may leave mean and variance nearly unchanged while the Fourier structure differs substantially. In that setting, time-domain statistical normalization can leave the dominant source of non-stationarity untouched (Ye et al., 2024).

A third criticism concerns denormalization across the forecast horizon. FreqLite argues that vanilla RevIN uses a single lookback statistic pair $$\tilde{x}=\alpha\frac{x-\mu_x}{\sigma_x}+\beta, \qquad \hat{y}=\sigma_x\left(\frac{f_\theta(\tilde{x})-\beta}{\alpha}\right)+\mu_x,$$0 to denormalize the entire forecast horizon, making it horizon-agnostic. Under non-stationarity, the lookback statistics may be suitable for the first future step but progressively less suitable for later steps, so the error induced by a fixed inverse can grow with horizon distance (Baiga et al., 31 May 2026).

A fourth line of criticism concerns robustness and dataset dependence. “On the Role of Reversible Instance Normalization” reports that instance normalization can help substantially on some datasets but can be detrimental on more stationary ones, especially Traffic, where it may even increase distribution distance and degrade performance. The same study argues that RevIN primarily stabilizes first and second moments, while the harder problem of conditional output distribution shift remains unresolved (Berthelier et al., 12 Mar 2026).

Outlier sensitivity has also become a prominent issue. A 2025 study reports that standard RevIN “catastrophically fails on datasets with extreme outliers,” with MSE surging by a “staggering 683\%,” while a robust variant termed $$\tilde{x}=\alpha\frac{x-\mu_x}{\sigma_x}+\beta, \qquad \hat{y}=\sigma_x\left(\frac{f_\theta(\tilde{x})-\beta}{\alpha}\right)+\mu_x,$$1-IN prevents this failure and “unexpectedly emerges as the best overall performer.” The same abstract reports that an adaptive model, A-IN, suffers a “complete and systemic failure,” and presents this as evidence that a simple or counter-intuitive heuristic can be more damaging than the statistical issue it is meant to solve (Fu et al., 6 Oct 2025).

5. Successors, adaptations, and generalizations

Several later methods retain the reversible normalization template but modify what is removed, how it is restored, or how the parameters are selected.

Variant	Key modification	Reported role
TA-RevIN	Drift-aware restoration statistics	Forecast-window distribution calibration
GRIN	GMM-adapted reversible normalization	Probabilistic denormalization of mixture parameters
A-RevIN	Gated horizon-adaptive inverse	Regime-adaptive correction under non-stationarity
FAN	Frequency-domain reversible normalization	Handles trend and seasonal patterns
NoRIN	Non-linear Johnson $$\tilde{x}=\alpha\frac{x-\mu_x}{\sigma_x}+\beta, \qquad \hat{y}=\sigma_x\left(\frac{f_\theta(\tilde{x})-\beta}{\alpha}\right)+\mu_x,$$2 transform	Reshapes tailedness and skewness

TA-RevIN keeps the standard normalization stage

$$\tilde{x}=\alpha\frac{x-\mu_x}{\sigma_x}+\beta, \qquad \hat{y}=\sigma_x\left(\frac{f_\theta(\tilde{x})-\beta}{\alpha}\right)+\mu_x,$$3

but augments restoration by splitting the input into preceding and succeeding segments, computing

$$\tilde{x}=\alpha\frac{x-\mu_x}{\sigma_x}+\beta, \qquad \hat{y}=\sigma_x\left(\frac{f_\theta(\tilde{x})-\beta}{\alpha}\right)+\mu_x,$$4

and estimating forecast-window recovery parameters as

$$\tilde{x}=\alpha\frac{x-\mu_x}{\sigma_x}+\beta, \qquad \hat{y}=\sigma_x\left(\frac{f_\theta(\tilde{x})-\beta}{\alpha}\right)+\mu_x,$$5

$$\tilde{x}=\alpha\frac{x-\mu_x}{\sigma_x}+\beta, \qquad \hat{y}=\sigma_x\left(\frac{f_\theta(\tilde{x})-\beta}{\alpha}\right)+\mu_x,$$6

The modification specifically targets the weakness of restoring with fixed input-window statistics when mean drift and volatility drift are present (Liu et al., 26 Jun 2026).

GRIN, introduced in TimeGMM, adapts RevIN to probabilistic forecasting with Gaussian mixture outputs. Its normalization stage is

$$\tilde{x}=\alpha\frac{x-\mu_x}{\sigma_x}+\beta, \qquad \hat{y}=\sigma_x\left(\frac{f_\theta(\tilde{x})-\beta}{\alpha}\right)+\mu_x,$$7

with instance-wise statistics computed along the temporal dimension for each variable. The paper’s key point is that reversibility must now be consistent not only for point predictions but also for GMM parameters $$\tilde{x}=\alpha\frac{x-\mu_x}{\sigma_x}+\beta, \qquad \hat{y}=\sigma_x\left(\frac{f_\theta(\tilde{x})-\beta}{\alpha}\right)+\mu_x,$$8, so that the mixture distribution in raw space remains calibrated after denormalization (Liu et al., 18 Jan 2026).

A-RevIN, proposed in FreqLite, retains RevIN as a special case but replaces the fixed inverse with a gated horizon-adaptive one. It introduces

$$\tilde{x}=\alpha\frac{x-\mu_x}{\sigma_x}+\beta, \qquad \hat{y}=\sigma_x\left(\frac{f_\theta(\tilde{x})-\beta}{\alpha}\right)+\mu_x,$$9

uses the drift feature

$$\mu_x = \frac{1}{L}\sum_{i=1}^L x_i, \qquad \sigma_x^2 = \frac{1}{L-1}\sum_{i=1}^L (x_i-\mu_x)^2.$$0

and defines

$$\mu_x = \frac{1}{L}\sum_{i=1}^L x_i, \qquad \sigma_x^2 = \frac{1}{L-1}\sum_{i=1}^L (x_i-\mu_x)^2.$$1

$$\mu_x = \frac{1}{L}\sum_{i=1}^L x_i, \qquad \sigma_x^2 = \frac{1}{L-1}\sum_{i=1}^L (x_i-\mu_x)^2.$$2

$$\mu_x = \frac{1}{L}\sum_{i=1}^L x_i, \qquad \sigma_x^2 = \frac{1}{L-1}\sum_{i=1}^L (x_i-\mu_x)^2.$$3

When $$\mu_x = \frac{1}{L}\sum_{i=1}^L x_i, \qquad \sigma_x^2 = \frac{1}{L-1}\sum_{i=1}^L (x_i-\mu_x)^2.$$4, the method reduces exactly to RevIN. The design claim is therefore a strict generalization rather than a replacement (Baiga et al., 31 May 2026).

FAN extends the reversible normalization idea into the frequency domain. Instead of removing $$\mu_x = \frac{1}{L}\sum_{i=1}^L x_i, \qquad \sigma_x^2 = \frac{1}{L-1}\sum_{i=1}^L (x_i-\mu_x)^2.$$5, it extracts an instance-specific non-stationary component through

$$\mu_x = \frac{1}{L}\sum_{i=1}^L x_i, \qquad \sigma_x^2 = \frac{1}{L-1}\sum_{i=1}^L (x_i-\mu_x)^2.$$6

then forms

$$\mu_x = \frac{1}{L}\sum_{i=1}^L x_i, \qquad \sigma_x^2 = \frac{1}{L-1}\sum_{i=1}^L (x_i-\mu_x)^2.$$7

The removed component is not simply added back; it is predicted forward by a pattern adaptation module before output reconstruction. In this sense FAN is a generalization of the reversible normalization paradigm but a replacement for RevIN’s specific mean-variance mechanism (Ye et al., 2024).

NoRIN pushes further by replacing the affine normalizer with a non-linear reversible transform based on the arcsinh-form Johnson $$\mu_x = \frac{1}{L}\sum_{i=1}^L x_i, \qquad \sigma_x^2 = \frac{1}{L-1}\sum_{i=1}^L (x_i-\mu_x)^2.$$8 family: $$\mu_x = \frac{1}{L}\sum_{i=1}^L x_i, \qquad \sigma_x^2 = \frac{1}{L-1}\sum_{i=1}^L (x_i-\mu_x)^2.$$9 The paper interprets $x\in\mathbb{R}^L$0 as controlling tailedness and compression strength and $x\in\mathbb{R}^L$1 as controlling skewness. Its central claim is that the linear RevIN regime is recovered only in the limit $x\in\mathbb{R}^L$2, and that naïvely learning $x\in\mathbb{R}^L$3 jointly with the backbone leads to a degeneration problem in which the parameters drift toward that near-linear corner (Zhang et al., 11 May 2026).

6. Empirical status and research significance

RevIN remains a strong practical baseline, but later empirical work consistently treats it as effective yet incomplete. “On the Role of Reversible Instance Normalization” reports that RevIN-style instance normalization gives a large average gain over no normalization on PatchTST, helps strongly on Electricity, helps on Solar in some settings but not uniformly, and can be worse than standard normalization on Traffic. The same study also reports that normalized backpropagation is the best-performing training strategy among the RevIN variants it considers (Berthelier et al., 12 Mar 2026).

Successor methods quantify where the baseline breaks. In TA-SparseMG, replacing TA-RevIN with original RevIN increases average MSE from 0.343 to 0.346, with consistent degradation on ETTh2, ETTh1, Traffic, and Weather. In TimeGMM, removing GRIN degrades CRPS from 0.2378 to 0.2743 on ETTm1, from 0.1409 to 0.1588 on ETTm2, from 0.1677 to 0.2036 on ETTh2, and from 0.0585 to 0.0611 on Weather. In FreqLite, A-RevIN is reported as neutral on stationary data but beneficial on non-stationary data, including “up to about 5\%” MSE reduction on ILI and monotonic benefit in a synthetic drift sweep as injected non-stationarity increases (Liu et al., 26 Jun 2026, Liu et al., 18 Jan 2026, Baiga et al., 31 May 2026).

Other extensions challenge the sufficiency of affine reversible normalization more directly. FAN reports 7.76\% ~ 37.90\% average improvements in MSE and argues that RevIN and related time-domain statistical methods are insufficient for evolving seasonal structure. NoRIN evaluates six representative backbones across five datasets and three horizons, for 90 configurations and 1,620 total runs, and reports that decoupled shape optimization significantly outperforms all linear-normalization baselines with paired Wilcoxon $x\in\mathbb{R}^L$4 (Ye et al., 2024, Zhang et al., 11 May 2026).

Taken together, this suggests that RevIN is best understood not as a universal normalization cure, but as a specific inductive bias: it removes instance-level level and scale variation, assumes that a reversible transform can stabilize learning, and delegates the remaining forecasting problem to the backbone. Later research has therefore moved in several directions at once—robust statistics, trend-aware restoration, probabilistic inverse mapping, horizon-adaptive denormalization, frequency-domain decomposition, and non-linear reversible transforms—while still preserving the core RevIN principle that normalization and denormalization should be explicitly coupled to the forecasting model.