Reassessing Deep Learning in Time Series: Predicting Time with Foresight

Machine learning relies heavily on time series forecasting, which has a wide range of uses, including although not limited to forecasting air quality, traffic patterns, and power use. Recent developments in deep training and matrix factorization models, however, have shown competitive performance and are emerging as alternatives. However, a fundamental disadvantage of these contemporary models over more established ones is their complexity. The Gradient Boosted Regression Tree, a well-known machine learning baseline, and popular deep learning models are compared in this work. We convert the time series forecast challenge into a window-based prediction issue, which is akin to (DNN) models. In addition, we design the GBRT model's input and output structure. In order to construct a single input instance to simulate a multi-output GBRT model, we concatenate the desired values with additional characteristics for each training window before flattening them. Using nine datasets, we undertake a thorough investigation and evaluate eight cutting-edge deep learning models recently presented at prestigious conferences. The findings show that a simple GBRT model performs much better than the state-of-the-art DNN models evaluated in this research thanks to the window-based input transformation.