A deep exploration of TiDE, its implementation using Darts and an actual life use case comparison with DeepAR (a Transformer architecture)
As industries proceed to evolve, the importance of an accurate forecasting becomes a non-negotiable asset wether you’re employed in e-commerce, healthcare, retail and even in agriculture. The importance of with the ability to foresee what comes next and plan accordingly to beat future challenges is what could make you ahead of competition and thrive in an economy where margins are tight and the purchasers are more demanding than ever.
Transformer architectures have been the recent topic in AI for the past few years, specially on account of their success in Natural Language Processing (NLP) being some of the successful use cases the chatGPT that took the eye of everyone regardless should you were an AI enthusiastic or not. But NLP is just not the one subject where Transformers have been shown to outperform the state-of-the-art solutions, in Computer Vision as well with Stable Diffusion and its variants.
But can Transformers outperform state-of-the-art models in time series? Although many efforts have been done to develop Transformers for time series forecasting, it appears that evidently for long run horizons, easy linear models can outperform several Transformer based approaches.
In this text I explore TiDE, an easy deep learning model which is in a position to beat Transformer architectures in long run forecasting. I also provide a step-by-step implementation of TiDE to forecast weekly sales in a dataset from Walmart using Darts a forecasting library for Python. And at last, I compare the performance of TiDE and DeepAR in an actual life use case from my company.
As at all times, the code is obtainable on Github.
TiDE is a novel time-series encoder-decoder model that has shown to outperform state-of-the-art Transformer models in long-time horizon forecast [1]. It’s a multivariate time-series model that’s in a position to use static covariates (e.g. brand of a product) and dynamic covariates (e.g. price of a product) which will be…