Evolving Multi-Scale Normalization for Time Series Forecasting under Distribution Shifts

TL;DR

This paper introduces EvoMSN, a novel, model-agnostic framework that adaptively normalizes multi-scale distribution features to improve long-term time series forecasting under distribution shifts.

Contribution

It proposes a new evolving multi-scale normalization approach with a collaborative optimization strategy to better handle complex distribution dynamics in time series forecasting.

Findings

EvoMSN improves forecasting accuracy across five mainstream methods.

The framework outperforms existing normalization and online learning approaches.

EvoMSN effectively tracks shifting distributions in benchmark datasets.

Abstract

Complex distribution shifts are the main obstacle to achieving accurate long-term time series forecasting. Several efforts have been conducted to capture the distribution characteristics and propose adaptive normalization techniques to alleviate the influence of distribution shifts. However, these methods neglect the intricate distribution dynamics observed from various scales and the evolving functions of distribution dynamics and normalized mapping relationships. To this end, we propose a novel model-agnostic Evolving Multi-Scale Normalization (EvoMSN) framework to tackle the distribution shift problem. Flexible normalization and denormalization are proposed based on the multi-scale statistics prediction module and adaptive ensembling. An evolving optimization strategy is designed to update the forecasting model and statistics prediction module collaboratively to track the shifting distributions. We evaluate the effectiveness of EvoMSN in improving the performance of five mainstream forecasting methods on benchmark datasets and also show its superiority compared to existing advanced normalization and online learning approaches. The code is publicly available at https://github.com/qindalin/EvoMSN.