A1580
Title: Deep learning for VAR modelling and forecasting
Authors: Xixi Li - The University of Manchester (United Kingdom) [presenting]
Jingsong Yuan - The University of Manchester (United Kingdom)
Abstract: Components of deep learning are incorporated into VAR (Vector Autoregressive) models for non-stationary data while maintaining the tractability and meaningfulness of the models. Three specific models are developed: (1).DeepVARwT: this is a time-invariant VAR model with a trend term generated from a long short-term memory (LSTM) network. (2).DeepTVAR: this is a VAR modelling with time-varying parameters generated from an LSTM network. It can be extended to integrated VAR with time-varying parameters. (3).DeepTVARwT: this more general model has a trend and a time-varying dependence structure, including DeepVARwT and DeepTVAR as special cases. For each model, deep learning methodology for maximum likelihood estimation of the parameters is employed. The causality condition on the VAR coefficients is also enforced to ensure the stability of each model and its interpretability as a prediction model. Simulation studies and real data applications demonstrate the proposed models' effectiveness and estimation methods.
