Reconstructing Interbank Networks for Counterfactual Policy Evaluation: A GNN-IPF Hybrid Approach with Application to Basel III Capital Buffers

This thesis develops a hybrid pipeline for reconstructing the Italian interbank network by combining a Graph Neural Network with Iterative Proportional Fitting, and applies it to evaluate the systemic risk implications of Basel III capital buffers. Using anonymized e-MID transaction data (2000--2009) and Orbis balance sheet data, I first demonstrate that exact bank de-anonymization is infeasible, but establish through permutation testing that hierarchical size-rank matching is sufficient for contagion analysis. The GNN, trained on daily interbank graphs enriched with macroeconomic variables, predicts edge existence with high precision, while IPF ensures consistency with known lending and borrowing aggregates. The hybrid method outperforms classical reconstruction benchmarks in reproducing the core-periphery structure of the real network. The pipeline is then deployed to generate counterfactual networks under pre- and post-Basel III conditions, with contagion simulated via DebtRank to quantify the impact of capital buffers on systemic risk propagation.