A Machine Learning-Based Forecasting Model with Deep Learning Capabilities for Unbalanced Time Series Data

During the last several years, deep learning capabilities have surpassed those of more conventional models on a wide variety of machine learning-related tasks. Neural networks with deep layers have indeed been effectively used to time series forecasting issues, which are crucial to the field of data mining. Their ability to instantly understand the temporal connections contained in time series has made them a viable option. Yet, deciding on a deep neural network architecture and then optimizing its parameters calls for a high level of skill. Thus, it is essential to conduct comprehensive studies of the predictive capacities of all current architectural frameworks. The completion of this work is hampered by two primary factors: (1) a complete examination of the most recent research on depth learning's application to time series forecasting, and (2) an empirical inquiry analysing the effectiveness of the most commonly used architectures. Seven distinct deep learning models are evaluated and compared for their efficiency and accuracy. We present scores and distributions based on our analysis of the suggested models trained with various architectures and hyperparameter settings. The 50,000 time series in the databases are used to solve 12 unique forecasting issues. Using this data to train approximately 38,000 models, we present the largest deep learning research to date for forecasting time series. While LSTMs were shown to be the most significantly predictive, both LSTMs & CNNs showed promise in this research. Comparable performance is attained by CNNs, but with less variance in results as well as a more based on the best available overall.