Probabilistic Multivariate Time-Series Forecasting using Deep Generative Models

At present, a new direction has emerged in the forecasting of time series using deep generative models, as evidenced by many works of recent years. For the forecasting problem, various approaches are used that have not been previously applied to time series, such as normalizing flows and generative-adversarial networks. On the other hand, neural network architectures for natural language processing tasks, such as recurrent neural networks and transformers, are currently being actively modified to solve time series forecasting problems. This article presents modern approaches to probabilistic time series modeling, as well as the most popular neural network architectures. The classification of models is given by the type of modeling and by the architecture of the neural network. In the course of the work, the results of measurements are presented, which showed that the best quality for various metrics is shown by such models as Gaussian TCN.