Abstract:
We investigate the distribution of links in three large data-sets, one of these covering interbank loans in the electronic trading platform e-MID, the other two covering a large part of the loans of banks to non-financial companies in the Spanish and Japanese economies, respectively. In contrast to all previous literature, we do not assume homogeneity of the link distribution over time and across different categories of agents (banks, firms) but apply our hypothesized distributions as regression models. As it turns out, many of the tested sources of heterogeneity turn out to be significant regressors. For instance, we find pervasive time heterogeneity of link formation in all three data sets, and also heterogeneity for different categories of banks/firms that can be identified in the data. Across all networks, the Negative Binominal model always outperforms all alternative models confirming its good performance as a model of economic count data in many previous applications.