Multivariate time series comprises more than one time-dependent variable. Each variable depends on its past values and possesses some dependency on other variables. It correlates between multiple time series variables to improve the accuracy of prediction. The primary goal of multivariate time series forecasting is to entrap the complex time series pattern and dependencies of the variable accurately. The most popular method for forecasting the multivariate time series is vector autoregression (VAR).
This method utilizes the relationship between multiple variables and defines the dynamic behavior of data with better forecasting results. Types of VAR models are reduced from VAR, recursive VAR, and structured VAR. Application areas of VAR models are economics, finance, and natural science. Some of the traditional deep learning models used for multivariate time series forecasting are Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent Units (GRU). Recent developments in the forecasting of multivariate time series data are multivariate time series forecasting with transformers, convolutional neural network-based multivariate time series, the factor-based framework for multivariate time series.