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

Taesung Kim, Jinhee Kim, Yunwon Tae, Cheonbok Park, Jang-Ho Choi, Jaegul Choo

Keywords: Time-series forecasting, Normalization, Distribution shift

Abstract: Statistical properties such as mean and variance often change over time in time series, i.e., time-series data suffer from a distribution shift problem. This change in temporal distribution is one of the main challenges that prevent accurate time-series forecasting. To address this issue, we propose a simple yet effective normalization method called reversible instance normalization (RevIN), a generally-applicable normalization-and-denormalization method with learnable affine transformation. The proposed method is symmetrically structured to remove and restore the statistical information of a time-series instance, leading to significant performance improvements in time-series forecasting, as shown in Fig. 1. We demonstrate the effectiveness of RevIN via extensive quantitative and qualitative analyses on various real-world datasets, addressing the distribution shift problem.

One-sentence Summary: We propose a simple yet effective normalization method, reversible instance normalization (RevIN), which solves the time-series forecasting task against the distribution shift problem.