RESEARCH: THE RESPONSIBLE USE OF AI
Machine Learning Explainability and Fairness
This empirical white paper is part of a broader research project on the explainability and fairness of machine learning in credit underwriting. The research is being conducted in collaboration with FinRegLab.
This study considers the capabilities, limitations and performance of proprietary and open-source tools to help lenders manage machine learning underwriting models as required by law. The report focuses on use of the tools in: (1) generating individualized disclosures that state why particular applicants were rejected or charged higher prices; and (2) analyzing what factors in the model drive disparities in model predictions among different demographic groups.
The evaluation analyzes model diagnostic tools from seven technology companies–Arthur AI, H2O.ai, Fiddler AI, Relational AI, Solas AI, Stratyfy, and Zest AI–as well as several open-source tools.
Unpacking the Black Box: Regulating Algorithmic Decisions
with Scott Nelson (Chicago Booth) and Jann Spiess (Stanford GSB)
We characterize optimal financial regulation in a world where lending decisions are made by complex algorithms and regulators are limited in the amount of information they can learn about the models these algorithms produce. We show that limiting lenders to algorithms that produce simple models is inefficient as long as the bias induced by misaligned lenders is small relative to the uncertainty about the true state of the world. Ex-post algorithmic audits can improve welfare but the gains depend on the design of the audit tools. Tools that focus on minimizing overall information loss, the focus of typical ‘explainer’ software tools, will generally be inefficient since they focus on explaining the average behavior of the model rather than sources of mis-prediction. Targeted tools that focus on the source of incentive misalignment, e.g. excess risk-taking or racial discrimination, can provide first best solutions. We provide empirical support for our theoretical findings using a large-scale credit bureau data set.
Presentations: NBER SI IT and Digitization 2021, Stanford SITE 2021, OCC, Philadelphia Fed New Perspectives on Consumer Behavior in Credit and Payments Markets, EEAMO, HBS, Microsoft Research, FTC, EC, OTC, Fed Board,
How Costly Is Noise? Data and Disparities in Consumer Credit
with Scott Nelson (Chicago Booth)
Submitted
We show that lenders face more uncertainty when assessing default risk of historically under-served groups in US credit markets and that this information disparity is a quantitatively important driver of inefficient and unequal credit market outcomes. We first document that widely used credit scores are statistically noisier indicators of default risk for historically under-served groups. This noise emerges primarily through the explanatory power of the underlying credit report data (e.g., thin credit files), not through issues with model fit (e.g., the inability to include protected class in the scoring model). Estimating a structural model of lending with heterogeneity in information, we quantify the gains from addressing these information disparities for the US mortgage market. We find that equalizing the precision of credit scores can reduce disparities in approval rates and in credit misallocation for disadvantaged groups by approximately half.
Presented at: NBER Economics of AI, NBER SI Household Finance, Stanford SITE Financial Regulation, UC Davis, St Louis Fed, Berkeley-Stanford Seminar, Federal Reserve Board, Stanford University, University of Chicago, NBER Summer Institute Corporate Finance, FIRS, NYU/NY Federal Reserve Financial Intermediation Conference, RedRock Conference
When Losses Turn Into Loans: The Cost of Weak Banks
Laura Blattner, Luísa Farinha (Banco de Portugal), and Francisco Rebelo (Boston College)
Conditional Accept at the American Economic Review
AQR Top Finance Graduate Award, BlackRock Applied Research Award (Winner),
ECB Lamfalussy Fellowship,
MFM dissertation grant from the Alfred P. Sloan Foundation, University of Chicago
We provide evidence that banks distort the composition of credit supply in order to comply with ratio-based capital requirements in times of economic distress. An unexpected intervention by the European Banking Authority provides a natural experiment to study how banks respond to falling below minimum required capital ratios during an economic downturn. We show that affected banks respond by cutting lending but also by reallocating credit to distressed firms with underreported loan losses. We develop a method to detect underreported losses using loan-level data. The credit reallocation leads to a reallocation of inputs across firms. We calculate that the resulting increase in input misallocation accounts for about 13% of the decline in productivity in Portugal in 2012.
Debt or Demand: Which Holds Investment Back? Evidence from an Investment Tax Credit
Laura Blattner, Luísa Farinha (Banco de Portugal), and Francisco Rebelo (Boston College)
We study how debt frictions and demand affect corporate investment using administrative data from a large temporary investment tax credit in Portugal. We obtain exogenous variation in demand for exporting firms from product-destination-level changes in foreign demand. We proxy debt frictions by an index of different debt-earnings ratios. We find that debt has a strong, non-linear effect on the likelihood that a firm invests in response to the tax credit. Firms in the lower two quartiles of our debt-earnings index have roughly equal predicted take-up probabilities. For firms in the third quartile predicted take-up drops by 50% while firms in the worst debt-earnings quartile have a predicted take-up rate close to zero. We show that the effect of demand is mediated by the size of a firm’s debt burden. While demand has a strong positive effect for the bottom debt quartiles, demand ceases to affect take-up in the highest debt quartile. These results highlight that the distribution of debt, rather than the absolute stock of debt, matters for understanding post-crisis investment dynamics.
Not All Shocks Are Created Equal: Assessing Heterogeneity in the Bank Lending Channel
Laura Blattner,
Luísa Farinha (Banco de Portugal), and Gil Nogueira (NYU Stern)
Lamfalussy Fellowship 2019
We provide evidence that the transmission of bank net worth shocks to lending varies significantly across key events in European sovereign debt markets. The onset of the sovereign debt crisis and the ECB's QE program lead to large effects on lending to firms and households. In contrast, the 2012 Greek PSI debt restructuring and the OMT announcement, both of which are positive net worth shocks, have moderate to no effects. Our results suggest an asymmetry in the transmission of bank net worth shocks: negative net worth shocks are transmitted one-to-one but positive shocks are only transmitted when banks realize the net worth gain by selling sovereign bonds. Our results imply that the ability of unconventional monetary policy to stimulate the real economy via a stealth recapitalization of the banking sector crucially depends on banks' incentive to realize trading gains. Our elasticity estimates also provide useful insights for the growing research on calibrating quantitative macroeconomic models using elasticities from micro data.
OLDER PAPERS
Sovereign Debt Composition in Advanced Economies: A Historical Perspective
Ali Abbas, Laura Blattner, Mark De Broeck, Asmaa A El Ganainy, Malin Hu
We examine how the composition of public debt, broken down by currency, maturity, holder profile and marketability, has responded to major debt accumulation and consolidation episodes during 1900-2011. Covering thirteen advanced economies, we focus on debt structure shifts that occurred around the two World Wars and global economic downturns, and the subsequent debt consolidations. Notwithstanding data gaps, we are able to recover some broad common patterns. Episodes of large debt accumulation—essentially, large increases in debt supply— were typically absorbed by increases in short-term, foreign currency-denominated, and banking-system-held debt. However, this pattern did not hold during the debt build-ups starting in the 1980s and 1990s, which were compositionally skewed toward long-term local-currency debt. We attribute this change to higher structural demand for sovereign paper, linked to capital account liberalization in advanced economies, the emergence of a large contractual saving sector, and innovative sovereign debt products. With regard to debt consolidations, we find support for the financial repression-cum-inflation channel for post World War II debt reductions. However, the scope for a repeat of this strategy appears limited unless financial liberalization and globalization were materially rolled back or the current globally agreed monetary policy regime built around price stability abandoned. Neither are significant favorable structural demand shifts, as witnessed in the 1980s and 1990s, likely.