TriSNAR : A Three-Layer Sparse Estimator for Large-Scale Network AutoRegressive Models

Understanding multi-market interactions and identifying leading markets in the global financial network is of interest to investors, regulators and policymakers. To discover the essential dynamic dependencies of digital currency exchanges, we propose TriSNAR, a three-layer sparse estimator for large-scale network autoregressive models, which imposes a structure on the lag-, network/group- and individual-level effects. We determine the asymptotic properties of the sparse estimator and investigate its finite-sample performance in extensive simulations. Numerical analysis shows that TriSNAR obtains a higher accuracy with less computational time per model contestant. We explore the applicability of TriSNAR on a network of 26 cryptocurrency exchanges with hourly pricing information. TriSNAR not only provides good out-of-sample prediction accuracy, but also exactly detects each leading exchange in North America, Europe and Asia