Discover the structure of TiDE and apply it in a forecasting mission utilizing Python
![Towards Data Science](https://miro.medium.com/v2/resize:fill:48:48/1*CJe3891yB1A1mzMdqemkdg.jpeg)
In our exploration of the newest advances within the subject of time sequence forecasting, we found N-HiTS, PatchTST, TimeGPT and likewise TSMixer.
Whereas many efforts have been deployed to use the Transformer structure for forecasting, it seems that it achieves a mediocre efficiency contemplating the computation necessities.
The truth is, easy linear fashions have been proven to outperform the complicated Transformer-based fashions on many benchmark datasets (see Zheng et al., 2022).
Motivated by that, in April 2023, researchers at Google proposed TiDE: a long-term forecasting mannequin with an encoder-decoder structure constructed with Multilayer Perceptrons (MLPs).
Of their paper Lengthy-term Forecasting with TiDE: Time-series Dense Encoder, the authors show that the mannequin achieves state-of-the-art outcomes on quite a few datasets when in comparison with different Transformer-based and MLP-based fashions, like PatchTST and N-HiTS respectively.
On this article, we first discover the structure and inside workings of TiDE. Then, we apply the mannequin in Python and use it in our personal small forecasting experiment.
For extra particulars on TiDE, be sure to learn the unique paper.
Be taught the newest time sequence evaluation methods with my free time sequence cheat sheet in Python! Get the implementation of statistical and deep studying methods, all in Python and TensorFlow!
Let’s get began!
TiDE stands for Time-series Dense Encoder. At its base, this mannequin implements the encoder-decoder idea with out the eye mechanism in Transformer-based fashions.
As a substitute, it depends on MLPs to realize sooner coaching and inference instances, whereas reaching good efficiency.
Throughout coaching, the mannequin will encode historic information together with covariates. Then, it’s going to decode the realized illustration together with recognized future covariates…