Reversible Instance Normalization for Accurate Time-Series Forecasting Against Distribution Shift

Abstract

The statistical properties in a time-series often change, i.e., the data distributions differ over time. This temporal distribution change is one of the main challenges that prevent accurate time-series forecasting. To address this issue, we propose a simple but effective normalization method called reversible instance normalization (RevIN). To be specific, RevIN consists of two different steps, normalization and denormalization. The former normalizes the input to fix its distribution in terms of the mean and variance, while the latter returns the output to the original distribution. In addition, RevIN is model-agnostic, generally applicable to various time-series forecasting models with significant improvements in forecasting performance. As shown in Fig. 1, RevIN effectively enhances the performance of the baselines. Our extensive experimental results verify the general applicability and performance improvements on various real-world datasets.