Abstract:

Vector autoregressions (VARs) are the main method used by macroeconomists for modelling multivariate, economic time series data. The models are simple to interpret and fit but their assumptions of stationarity, linearity and normality are often too restrictive. Some of  these shortcomings can be addressed using time-varying VARs which allow the parameters of the VAR to change over time. However, these models allow for non-linearity, non-normality and non-stationarity using a single process which may be inappropriate. In this talk, I will describe a Bayesian nonparametric VAR and time-varying Bayesian nonparametric VAR which provide alternatives to TVP-VAR for flexibly modelling multivariate time series. The model will be illustrated and compared to competing models on US economic data.