Exploring Industry-Distress Effects on Loan Recovery: A Double Machine Learning Approach for Quantiles

GAICIL Insights

Ask a banker how much is recovered when a loan goes bad, and an average will come back. This paper argues the average is the wrong answer.¹ Exploring Industry-Distress Effects on Loan Recovery: A Double Machine Learning Approach for Quantiles, by Hui-Ching Chuang and Jau-er Chen, measures how industry distress affects recovery rates across the whole distribution. Lenders care what they recover after default, and recovery depends heavily on economic conditions. The study combines unconditional quantile regression with lasso-based double selection, a machine learning technique for choosing among many potential control variables.

The data are substantial: 5334 debt and loan instruments in Moody's Default and Recovery Database from 1990 to 2017. Industry distress decreases recovery rates from 15.80% to 2.94% for the 15th to 55th percentile range, and slightly increases recovery rates in the lower and upper tails. The damage concentrates in the middle of the distribution. Distress hits the ordinary cases, leaving the extremes nearly untouched.

The method shapes the message. Unconditional quantile regression provides more interpretable and policy-relevant information on the predictive effect of the target variable than conditional quantile regression. Lasso-based double selection handles a broad set of macroeconomic and industry variables, selecting relevant controls while protecting inference on the variable of interest. Together they estimate distributional effects in a high-dimensional setting.

The regulatory payoff is explicit. The Basel Capital Accord requires banks to estimate loss given default during a downturn. The unconditional quantile estimates provide quantitative measurements of recovery rates in exactly such a downturn. Machine learning feeds straight into the measurement of risk for capital regulation, a far cry from chatbots.

That is a valuable, often overlooked use of artificial intelligence in economics: improving the reliability of inference. Popular discussion of AI centers on prediction. Here machine learning selects controls from many candidates, so the estimated effect of industry distress escapes distortion from omitted variables or arbitrary specification choices. Double machine learning techniques like this are making causal and predictive analysis more robust in finance. AI as a referee for the econometrician is unglamorous, and arguably very useful.

For banks and regulators, heterogeneity is the headline. Average effects can conceal large variation. Recovery rates in the middle of the distribution are most sensitive to industry distress, while the tails behave differently. Risk models capturing this heterogeneity can produce more accurate estimates of downturn losses, supporting appropriately calibrated capital requirements. A model tuned to the mean will misprice the middle of the book exactly when it matters.

Better measurement also serves financial stability. Accurate loss given default under stress helps banks hold appropriate capital, neither too little to absorb losses nor so much that lending is unnecessarily constrained. Tools improving these estimates support safety and the efficient allocation of credit simultaneously. Overcapitalization has a cost too, paid by borrowers who never get a loan.

Regulatory design can evolve with the tools. As AI methods improve, regulators can rely on more sophisticated, data-driven risk estimates. Frameworks such as the Basel Accord already ask banks to model downturn conditions. Allowing rigorous machine learning techniques, with appropriate validation and transparency, can make risk measurement more accurate.

The long sample, spanning several downturns between 1990 and 2017, strengthens the estimates. Capturing how recoveries behave across different distress episodes gives a sturdier basis for planning than any single crisis.

Interpretability is the quiet virtue. Regulators need estimates they can understand and scrutinize, and the choice of unconditional quantile regression reflects that need. Machine learning in regulated settings must balance flexibility with interpretability, and this paper models how to strike the balance.

At bottom, AI can strengthen the foundations of financial regulation. Combining machine learning variable selection with interpretable econometrics delivers policy-relevant estimates of how industry distress affects recoveries across the distribution, supporting a financial system that is resilient and able to fund growth.

Here is the uncomfortable test for supervisors: ask banks for the shape of downturn recoveries, with the average as one point on it. If a risk model cannot say which percentile of the book loses most in a downturn, it is not ready for one.