Performance enhancing techniques for deep learning models in time series forecasting

Time series forecasting uses deterministic algorithms to capture past temporal information or dependencies that can be used to predict future patterns. Studies have shown that traditional forecasting techniques are outperformed by deep learning models. Since then research work has been much focused on proposing different network models, little attention has been paid to improve the performance of existing models. In this paper, we compare the performance of several existing deep learning models used in both single and multiple time series forecasting tasks. We then propose two different approaches to improve the models' performance. Specifically, we present a fine-grained attention mechanism that achieves a much better performance for multi-step forecasting tasks. An ensemble technique is then proposed to further improve the performance of all the models.