Evaluating AI for Time- Series Forecasting

For instance, we may use time series forecasting, one of the most significant issue types in many machine learning scenarios, to estimate not just power consumption but also air quality or traffic patterns. Techniques like autoregressive integrated moving averages, rolling averages, and vector auto-regression are used in traditional forecasting models. On the other hand, more recent studies use deep learning methods and vector factorization to provide better results [24] – [26]. When compared to what the older methodologies give, the intricacy of these more advanced models is undoubtedly a drawback. In order to build a machine learning baseline, this research compares well-known deep learning models to the well-respected Gradient Booster Logistic Tree (GBRT) model. In deep neural network models, we consider time series forecasting as a window-based regression issue. To build the input/output structure for each training window of the GBRT model, we similarly flatten target values and external characteristics across all of the windows. This allows you to make a single input instance out of numerous outputs. Through an extensive comparison on nine datasets and eight recent popular examples of the state-of-the-art architectures, we show that the proposed window-based input perturbation method leads to significant performance improvements for a vanilla GBRT framework, with results even surpassing those obtained by modern deep learning models which were reported at top-level conferences.