High Risk, Poor Reward: Machine Learning Reveals the Obscured Distress Risk Puzzle = 머신러닝을 통해 재조명한 Distress Risk Puzzle

국문 초록 (Abstract)

머신러닝(고도화된 통계 모델)이 금융현상, 특히 기업의 부도 예측과 Distress Risk Puzzle에 미치는 영향은 무엇인가? 본 페이퍼는 미국 기업 데이터를 기반으로 전통적인 계량 모델(Logit)과 머신러닝(Random Forest, XGBoost)의 부도 예측력을 비교하며, 예측 성능의 개선이 공변량 확장(Variables Update)이 아니라 머신러닝 도입(Model Upgrade)에 의해 주도됨을 규명한다. 일부 선행연구에서 변수를 선택을 변경함으로써 Distress Risk Puzzle이 존재하지 않는다고 주장한 것과 달리, 본 연구는 고도화된 모델로 계량화한 부도확률을 바탕으로 자산가격 모형(CAPM 및 Fama–French 3, 4, 5요인 모형)을 이용한 포트폴리오 분석을 수행한 결과, 모든 모형에서 통계적으로 유의한 negative High-minus-Low default risk portfolio return을 확인한다. 또한 포트폴리오 수익률의 분포를 분석한 결과, 부도위험이 높은 기업일수록 하방 위험이 분포 전반에 걸쳐 확대되며, 이러한 하방 위험은 평균 수익률로 보상되지 않음을 보인다. 이와 같은 분포적 특성은 High-risk, poor-reward 패턴의 형성에 기여하며, 특히 규모가 작은 고부도위험 기업에서 두드러지게 나타난다.

다국어 초록 (Multilingual Abstract)

I examine how advances in measurement through machine learning affect inference from corporate default prediction and its implications for the distress risk puzzle. Using a U.S. panel dataset, I compare the predictive performance of a conventional econometric model, Logit, with machine learning methods, including Random Forest and XGBoost. Improvements in default prediction accuracy are driven primarily by upgrading the predictive model rather than by expanding covariates within a traditional specification. Portfolio analyses based on default probabilities estimated from upgraded models yield statistically significant negative High-minus-Low default risk returns across the CAPM and Fama-French factor models, indicating that the distress risk puzzle persists even when default risk is measured more precisely. In contrast, portfolios formed using logit-based default probabilities produce unstable and statistically insignificant return patterns. Distributional analysis further shows that firms with higher default risk experience substantially amplified downside risk that is spread broadly across negative return realizations and is not offset by higher average returns. This feature of returns underlies the high-risk, poor-reward phenomenon and is especially pronounced among smaller firms with elevated default risk.

목차 (Table of Contents)

• 1 Introduction, 3
• 1.1 Machine Learning in Default Prediction and Asset Pricing, 3
• 1.2 The Distress Puzzle (Default Risk Anomaly), 3
• 1.3 Conventional Methodologies for Calculating PDs (Probabilities of Default), 4
• 1.4 Machine Learning in Asset Pricing and Default Prediction, 4
• 1.5 Roadmap and Contributions, 5
• 2 Data and Sample, 5
• 2.1 Data Sources and Sample Construction, 5
• 2.2 Descriptive Statistics: Default Events and Sample, 6
• 2.3 The Obscured Distress Puzzle, 7
• 3 Model Specification and Methodology, 9
• 3.1 Logistic Model Specification, 9
• 3.2 Random Forest Specification, 10
• 3.3 Variable Selection and Multicollinearity Diagnostics, 10
• 3.4 Portfolio Year Definition and Construction, 11
• 4 Empirical Results, 13
• 4.1 Model Performance: Logit vs. Random Forest, 13
• 4.2 Model Upgrade versus Variable Update, 14
• 4.3 Machine Learning Reveals the Obscured Distress Puzzle, 15
• 4.4 High Risk, Poor Reward, 16
• 5 Conclusion, 20
• Appendix A Variable Definitions, 22
• Appendix B Benchmark Portfolios Based on Conventional Default Models, 24
• Appendix C XGBoost Robustness Test, 27
• Abstract in Korean, 29