TimePatch: Diffusion on Non-Stationary and Noisy Time Series Forecasting

Jihyeon Seong, Enver Menadjiev

Keywords: Diffusion, TimeSeires Forecasting, Noise Learning

Abstract: The success of Transformer-based systems in vision and language-related tasks led to the development of state-of-the-art approaches in time series. However, there are several fundamental problems with existing methods and their benchmarks. Although the real-world data is non-stationary, existing datasets are mostly presented with stationary data. This entails a focus shift of the models towards learning simple patterns. As a result, models struggle with predicting noisy data, such as stock data and Brownian Motion. In this work we present TimePatch Diffusion, which is a novel architecture, to deal with complex time-series forecasting tasks. It is composed of a DDPM encoder and a TPatch decoder. The former learns time-series noise, while the latter extract encoded features from captured representations.