Netmetrics: Exploring Estimation Bias, Efficiency and Inference in Unknown, Partial Financial Networks

Abstract

Advanced analytics are an integral inferential component for knowledge discovery and dissemination. We employ network metrics on relational variables of interest to control for econometric estimation bias, inefficiency and inferential problems from non-spherical, network induced disturbances in sampled, unknown political networks. Our analysis follows a two-step procedure to establish a conceptual baseline and then integrate real-world data and questions pertinent for statistical inference. First, artificial networks are generated using two well-known graph models, Erdos-Renyi and Barabasi-Albert. Simulated, synthetic data is created across a range of graph sizes with multiple node-level network attributes and estimation bias at different sample sizes. This gives a picture of efficiency and consistency across network metrics, controlling for varied true-population sampling ratios. Second, we generate artificial networks across a broad range of graph model types compared to real networks from IMF’s Direction of Trade Statistics and Correlates of War dyadic trade data. We then examine the consistency of each chosen metric across a wider range of possible network-creation processes. This information is visualized against one known and one potentially unknown statistic, R2 and node sample size. Our unique approach provides analytical and estimative insights for opaque and dynamic real world financial networks.