Visualising the Global Credit Risk Network

The network images presented in this website are computed from the forecast error variance decompositions of a set of vector autoregressive (VAR) models estimated on a rolling sample basis.

We employ a rolling sample of 250 trading days over the period January 2nd 2006 to July 27th 2015. There are 2,246 rolling samples in total.

For each rolling sample, we estimate a VAR model which approximates the dynamics of a vector of 76 endogenous variables (4 variables for each of 18 countries plus 4 global controls).

The rolling sample VAR models are estimated by LASSO on an equation-by-equation basis with 5-fold cross-validation in MATLAB.

For each rolling sample, we compute the 5-days-ahead generalised forecast error variance decomposition of the VAR model. This is then transformed by a simple normalisation to yield the adjacency matrix summarising the bilateral relations among the variables in the model.

When we plot graphs, the dates that are reported correspond to the last day of each rolling sample.

A rigorous technical discussion of our estimation framework can be found in the paper.