Agenda

Rigers Behluli (Ca' Foscari University of Venice) - Heterogeneous g-priors for networks with dyadic covariance structures

Abstract:

The increasing availability of multivariate data calls for the use of matrix variate models for identifying hidden patterns within the data and for predicting the variables of interest. In this paper, we propose a novel regression model for sequences of matrix data with dyadic covariance structure and develop a Bayesian procedure for model selection. We show that it is possible to characterize the dyadic variance-covariance matrix analytically and perform inference and model selection using mixtures of g-priors adapted to accommodate the heterogeneity induced by dyadic covariance. We present some theoretical results and the algorithms used to sample low- and high-dimensional model spaces. We present simulation results and a real data comparison to an established method for dyadic covariance estimation.

Adriano Amati (Ca' Foscari University of Venice) - Graph Learning for Corporate Infiltration Risk

Abstract:

Learned representations have revolutionized the measurement of unstructured data—text and images—in economics. We extend this toolkit to unstructured relational data, demonstrating that Temporal Graph Networks (TGNs) can effectively encode node behavior in dynamic corporate networks. We apply this framework to a case study of Organized Crime Groups (OCGs) infiltration in Italian firms. Using a high-resolution dynamic ownership graph anchored by 6,200 judicially confiscated firms, we train a TGN on network dynamics (link creation, updates, and removal) to generate time-varying node embeddings. These embeddings capture each firm's and person's evolving structural position within the ownership network. We find that firms destined for future confiscation are already structurally similar (closer in the embedding space) to previously confiscated ones. In a forecasting task, the embeddings achieve near-perfect out-of-sample prediction of confiscation up to four years in advance. The next step in our research agenda is to deploy these embeddings in empirical applications, such as replicating and extending past findings on the economic impacts of OCG infiltration. This approach also addresses a key gap in the relevant literature by providing a novel, statistically rigorous method to identify firm exposure to OCGs, paving the way for a new approach for detecting criminal infiltration in the legal economy.

The seminar can be attended also remotely, connecting to ZOOM.

Link Zoom: bit.ly/insem-2425
ID riunione:  880 2639 9452
Passcode: InSem-2425