Discover and understand the inner workings of TimeMixer and apply it in your own forecasting project using Python
The field of time series forecasting keeps evolving at a rapid pace, with many models being proposed and claiming state-of-the-art performance.
Deep learning models are now common methods for time series forecasting, especially on large datasets with many features.
Although numerous models have been proposed in recent years, such as the iTransformer, SOFTS, and TimesNet, their performance often falls short in other benchmarks against models like NHITS, PatchTST and TSMixer.
In May 2024, a new model was proposed: TimeMixer. According to the original paper, TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting, this model uses mixing of features along with series decomposition in an MLP-based architecture to produce forecasts.
In this article, we first explore the inner workings of TimeMixer before running our own little benchmark in both short and long horizon forecasting tasks.
As always, make sure to read the original research article for more details.
Learn the latest time series…