A Machine learning approach in stress testing US bank holding companies (Job Market Paper)
In this paper, I analyze in two steps the role of machine learning techniques (ML), big macroeconomic, and bank balance sheet data in enhancing the top-down stress test. In the first step, I evaluate the relative contribution of ML and big data to improving the forecast accuracy of two key banking variables: net charge-off (NCO) and pre-provision net revenue (PPNR) over standard linear models extensively used in the literature.
I find that Random Forest and Adaptive Lasso are the best forecasts model for NCO and PPNR. This result highlights the importance of machine learning in building a more accurate model of NCO and PPNR, a cornerstone of the stress test exercise. In the second step, I construct a prediction of PPNR and NCO for selected bank holding companies (BHC) conditional on adverse stress scenarios using Adaptive Lasso and Random Forest.
Then, I use these predictions to simulate densities of Tier 1 common equity capital (T1CR). It is observed that the T1CR distribution has a heavier left tail than its linear counterpart for global systematically important banks. In addition, ML modeling better approximates T1CR densities in distressed conditions studying the case of the Great Recession. These findings suggest that linear models provide a too-optimistic picture of large bank vulnerabilities and may underestimate systemic risk. Therefore, machine learning better captures the complexity of macro-financial relationships and improves risk analysis in depressed economic conditions.
On bank credit and business cycle from a machine learning perspective, joint with Dalibor Stevanovic
In this paper, we propose to identify the credit shock by using machine learning approach in a data-rich setup. The analysis consists of two steps. First, we do a pseudo-out-of-sample forecasting exercise to approximate the bankers’ predictive model of the capital-to-asset ratio. Then, the aggregate credit shock is defined as the weighted average of bank-specific out-of-sample forecast errors. In the second step, we include the constructed shock in a VAR to assess its impact on the business cycle.
We find that a one standard deviation negative aggregate shock moves credit volume and credit prices in the opposite direction, corroborating that it identifies a bank credit supply shock. Furthermore, this shock has significant macroeconomic effects. It triggers a slowdown in GDP growth, inflation, and a substantial commercial and industrial credit volume drop. However, the decline in commercial and industrial credit is more pronounced and persistent than the slowdown in GDP growth.
Bank-level uncertainty and the business cycle (Work in progress)
I build a measure of bank-level uncertainty based on large bank-specific data and an extensive macroeconomic database. I compute the uncertainty as the mean square error of the forecasted return on asset (ROA). I build predictions of the ROA using different machine learning techniques. As Jurado et al. (2015), for macroeconomic variables, the measure reflects the difficulty of the banker in predicting ROA. Then, I compare this measure to Jurado’s macroeconomic uncertainty and find similar uncertainty episodes. Furthermore, in an 11- variable VAR, I establish that a shock in bank-level uncertainty has a significant but short-run effect on the financial sector and business cycle.