A0196
Title: A machine learning methodology for daily assessment of bank health, interconnectedness, and systemic risk
Authors: Shawn Mankad - Cornell University (United States)
Celso Brunetti - Bocconi University and Federal Reserve Board (United States) [presenting]
Jeffrey Harris - American University (United States)
Abstract: A novel methodology is proposed to estimate the portfolio composition of banks as a function of daily stock returns. Building on a model where individual bank balance sheets connect through common holdings, a constrained semi-non-negative matrix factorization problem is derived and solved, where the rows (corresponding to banks) of one latent matrix factor (representing asset holdings) are subject to probability constraints. Although banks report assets at low frequencies, estimating factorization over a rolling window allows for the derivation of daily estimates of bank portfolios. Estimates of asset holdings are validated by showing they closely match balance-sheet data reported in quarterly regulatory filings.
