A sparsity-controlled vector autoregressive model.

Area of Science:

• Statistics
• Time Series Analysis
• Econometrics

Background:

• Vector Autoregressive (VAR) models are standard for multivariate time series analysis.
• Standard VAR models often lack sparsity, hindering the identification of joint dependencies and causalities.
• Existing sparse VAR variants offer limited user control over sparsity dimensions.

Purpose of the Study:

• To propose a versatile sparsity-controlled VAR model.
• To enable users to control different dimensions of sparsity in VAR models.
• To facilitate visualization of potential causalities and improve prediction accuracy.

Main Methods:

• Developed a novel sparsity-controlled VAR model.
• Formulated model coefficient estimation as an optimization problem.
• Utilized standard numerical optimization routines for solving the problem.

Main Results:

• The proposed model allows users to control sparsity dimensions.
• Demonstrated superior performance over greedy and Lasso approaches.
• Achieved lower prediction errors and better sparsity on simulated and real-life data.

Conclusions:

• The developed sparsity-controlled VAR model offers enhanced flexibility and control.
• It provides a valuable tool for causal inference and prediction in multivariate time series.
• Outperforms existing methods in terms of prediction accuracy and sparsity control.