Learning from Lending in the Interbank Network

Abstract

Empirical analysis of a major overnight-funding network of European banks shows that, when liquidity disruptions occur in a part of the network, lending banks in other parts of the network broaden their cohorts of borrowers in the part of the network that is subject to the disruptions. Measures of this broadening are useful new statistics for the amount of information conveyed from one part of the network to another. In our setting, we call this broadening “counterparty sampling,” and present evidence that it improves the network’s stock of information about future interest rates. By comparing to linkages forecast by an LSTM deep learning model for counterparty linkages, we find that the extent of surprising new linkages predicts lower future rates. Our evidence supports the idea that interbank funding networks provide benefits of learning and information aggregation, and our measures suggest new ways of looking at sparse networks with stable structures but dynamically-changing linkages.