Member-only story
SOFTS: The Latest Innovation in Time Series Forecasting
Discover the architecture of SOFTS and the novel STAD module, and apply it in a forecasting project using Python.
In recent years, deep learning has been successfully applied for time series forecasting, where new architectures have incrementally set new standards for state-of-the-art performances.
It all started with N-BEATS in 2020, which was followed by NHITS in 2022. In 2023, PatchTST and TSMixer were proposed, and they still rank among the top forecasting models.
More recently, we discovered the iTransformer, which further pushed the performance of deep learning forecasting models.
Now, we introduce the Series-cOre Fused Time Series forecaster or SOFTS.
Proposed in April 2024 in the article SOFTS: Efficient Multivariate Time Series Forecasting with Series-Core Fusion, this model employs a centralized strategy for learning interactions across different series, resulting in state-of-the-art performances in multivariate forecasting tasks.
In this article, we explore the architecture of SOFTS in detail, and discover the novel STar Aggregate-Dispatch (STAD) module which is responsible for learning interactions between time series. Then, we apply…
