Noise or Signal? Deconstructing Contradictions and An Adaptive Remedy for Reversible Normalization in Time Series Forecasting

TL;DR

This paper investigates the instability of reversible instance normalization in time series forecasting by revealing four fundamental contradictions between noise, past and future statistics, distribution fitness, and normalization scaling. It proposes a diagnostics-driven framework with two concrete solutions, R$^2$-IN+ and A_IN, and evaluates them against the standard RevIN and the naive R$^2$-IN on 11 real-world datasets using a DLinear backbone. The key finding is counterintuitive: the simple, robust R$^2$-IN often outperforms more sophisticated adaptive methods, and the attempted adaptive strategy A_IN can catastrophically fail due to flawed heuristics. The work emphasizes a move away from blind complexity toward data-driven diagnostics and robust baselines, offering practical guidelines and a cautious perspective on dynamic normalization in TSF.

Abstract

Reversible Instance Normalization (RevIN) is a key technique enabling simple linear models to achieve state-of-the-art performance in time series forecasting. While replacing its non-robust statistics with robust counterparts (termed R$^2$-IN) seems like a straightforward improvement, our findings reveal a far more complex reality. This paper deconstructs the perplexing performance of various normalization strategies by identifying four underlying theoretical contradictions. Our experiments provide two crucial findings: first, the standard RevIN catastrophically fails on datasets with extreme outliers, where its MSE surges by a staggering 683\%. Second, while the simple R$^2$-IN prevents this failure and unexpectedly emerges as the best overall performer, our adaptive model (A-IN), designed to test a diagnostics-driven heuristic, unexpectedly suffers a complete and systemic failure. This surprising outcome uncovers a critical, overlooked pitfall in time series analysis: the instability introduced by a simple or counter-intuitive heuristic can be more damaging than the statistical issues it aims to solve. The core contribution of this work is thus a new, cautionary paradigm for time series normalization: a shift from a blind search for complexity to a diagnostics-driven analysis that reveals not only the surprising power of simple baselines but also the perilous nature of naive adaptation.

Figures (6)

• Figure 1: Catastrophic failure of RevIN on a sample from the Electricity dataset. While the baseline DLinear model (without instance normalization) produces a reasonable forecast, the prediction from the RevIN-equipped model is severely distorted. This failure is caused by its statistical estimates being contaminated by extreme outliers present in the lookback window (shown in gray).
• Figure 2: Thw of Reversible Instance Normalization methods. Statistics (e.g., mean/std) are calculated from the input instance, used for normalization, and then re-applied for denormalization on the model's output.
• Figure 3: A visual illustration of the four core contradictions in instance normalization. (a) On time series with outliers, the mean is heavily skewed while the median remains robust. (b) For non-stationary series, statistics calculated from the past lookback window may not be a reliable proxy for the future. (c) For skewed distributions, the mean (center of gravity) and median (50th percentile) represent different notions of centrality. (d) On non-normal data, estimating the standard deviation using a fixed k-factor multiplier on MAD can lead to significant errors.
• Figure 4: The statically-configured adaptive mechanism of our proposed A-IN. It uses a pre-computed diagnostic metric (Change Point Risk) to select the most suitable normalization strategy for an entire dataset upfront.
• Figure 5: Average rank of normalization methods across all tested tasks. Lower is better. Counter-intuitively, the naive robust method, DLinear + R$^2$-IN, achieves the best overall performance, while the more sophisticated A-IN performs the worst, highlighting a strong "less is more" reality.