Two Steps Forward and One Behind: Rethinking Time Series Forecasting with Deep Learning

The Transformer is a highly successful deep learning model that has revolutionised the world of artificial neural networks, first in natural language processing and later in computer vision. This model is based on the attention mechanism and is able to capture complex semantic relationships between a variety of patterns present in the input data. Precisely because of these characteristics, the Transformer has recently been exploited for time series forecasting problems, assuming its natural adaptability to the domain of continuous numerical series. Despite the acclaimed results in the literature, some works have raised doubts about the robustness of this approach. In this paper, we further investigate the effectiveness of Transformer-based models applied to the domain of time series forecasting, demonstrate their limitations, and propose a set of alternative models that are better performing and significantly less complex. In particular, we empirically show how simplifying this forecasting model almost always leads to an improvement, reaching the state of the art among Transformer-based architectures. We also propose shallow models without the attention mechanism, which compete with the overall state of the art in long time series forecasting, and demonstrate their ability to accurately predict extremely long windows. We show how it is always necessary to use a simple baseline to verify the effectiveness of one's models, and finally we conclude the paper with a reflection on recent research paths and the desire to follow trends and apply the latest model even where it may not be necessary.