Understanding the Performance of Machine Learning Models to Predict Credit Default: A Novel Approach for Supervisory Evaluation (184 KB)
Andrés Alonso and José Manuel Carbó
We study the economic impact for financial institutions of using machine learning (ML) models in credit default prediction. We do so by using a unique and anonymized database from a major Spanish bank. We first measure the statistical performance in terms of predictive power, both in classification and calibration, comparing models like Logit and Lasso, with more advanced ones like Trees (CART), Random Forest, XGBoost and Deep Learning. We find that ML models outperforms traditional ones, although more complex ML algorithms do not necessarily predict better. We then translate this into economic impact by estimating the savings in regulatory capital that an institution could achieve when using a ML model instead of a simpler one to compute the risk-weighted assets following the Internal Ratings Based (IRB) approach. Our benchmark results show that implementing XGBoost instead of Lasso could yield savings from 12.4% to 17% in capital requirements, depending on the type of underlying assets.
We quantify the aggregate, distributional, and welfare consequences of setting an optimal progressivity level in the Spanish personal income tax scheme. For that purpose, we use a heterogeneous households general equilibrium model featuring both life cycle and dynastic elements which is calibrated to replicate some aggregate and distributional characteristics of the Spanish economy. The findings suggest that increasing progressivity would be optimal, even though it would involve an efficiency loss. The optimal reform of the tax schedule would reduce wealth and income inequality at the cost of negative effects on capital, labor, and output. The evaluation of these theoretical results with tax microdata describes a current scenario where the income-top households typically face suboptimal effective average tax rates.
Urban air pollution and sick leaves: Evidence from social security data (169 KB)
Felix Holub, Laura Hospido and Ulrich J. Wagner
We estimate the impact of air pollution on the incidence of sick leaves in a representative panel of employees affiliated to the Spanish social security system. Using over 100 million worker-by-week observations from the period 2005-2014, we obtain that higher air pollution increases the incidence of sick leaves, controlling for weather, individual effects, and a wide range of time-by-location controls. The effect is stronger for workers with pre-existing medical conditions, and weaker for workers with low job security. Our estimates are instrumental for quantifying air pollution damages due to changes in labor supply. Improved air quality in urban Spain between 2005 and 2014 saved at least €503 million in foregone production thanks to 5.55 million fewer worker absence days.
Raising Markups to Survive: Small Spanish Firms during the Great Recession (159 KB)
Pilar Garcia-Perea, Aitor Lacuesta, Pau Roldan-Blanco
Markups and other measures of concentration have been on the rise in the United States in the last few decades, driven by the behavior of large and productive firms. We document that markups rose during the financial crisis in Spain, and that this increase was led by a behavioral response of small and unproductive firms: in response to a drop in sales, these firms were unable to increase their productive efficiency when average costs increased and, in order to escape a sharp decline in profit rates, they increase their markups to survive in the market.
Spanish non-financial corporations’ liquidity needs and solvency after the COVID-19 shock (185 KB)
Roberto Blanco, Sergio Mayordomo, Alvaro Menéndez and Maristela Mulino
The COVID-19 pandemic is exerting an unprecedented adverse impact on economic activity and, in particular, on firms’ income. This article presents the results of an exercise simulating Spanish non-financial corporations’ liquidity needs for the four quarters of this year. Liquidity needs, between April and December, might exceed €230 billion.
Keeping track of global trade in real time (146 KB)
Jaime Martínez-Martín and Elena Rusticelli
We build an innovative composite world trade-cycle index by means of a dynamic factor model for shortterm forecasts of world trade growth of both goods and (usually neglected) services. Trade indicators are selected using a multidimensional approach, including Bayesian model averaging techniques, dynamic correlations, and Granger non-causality tests in a linear vector autoregression framework. The dynamic factor model is extended to account for mixed frequencies, to deal with asynchronous data publication, and to include hard and survey data along with leading indicators. Nonlinearities are addressed with a Markov switching model. Pseudo-real-time empirical simulations suggest that: (i) the global trade index is a useful tool for tracking and forecasting world trade in real time; (ii) the model is able to infer global trade cycles very precisely and better than several competing alternatives; and (iii) global trade finance conditions seem to lead the trade cycle, a conclusion that is in line with the theoretical literature.