Kisung Yang Assistant Professor, School of Finance, Soongsil University
In the default prediction problem, thecost from thefailure of forecasting defaults is much biggerthan that of forecastingnon-defaults. Thecost asymmetryis deeperin the corporatedefaultpredictionthan theretailas corporateloan portfolios arenot granular. However, thetwo typesof costs aretreatedequallyin generalas defaultpredictionmodels areusually estimatedto minimizepredictionerrorsor maximizestatistical performance. This practicemight not fulfill thegoal of risk managementto minimizeeconomiclosses. To mitigatethis issue, this study apply cost-sensitivelearningapproach to defaultprediction, which minimizeseconomiccosts insteadof statistical errors. Wedefineeconomic costs and testthemfor various levelsof thecost asymmetryby employingLogistic regression, XGBoost, and LightGBM. As a resultof empiricalexperimentswith Taiwaneseand Polish corporate default data, we first find that the proposed cost-sensitivemodels are superiorto thecost-insensitivecounterpartsin termsof economiccost, mostly regardless of thecost asymmetryscenarios. Secondly, neverthele,ssthedecreasesin thestatistical performanceare relativelysmall ? economic costs decrease24.6% at theexpenseof the decrease in AUC of 4.6% on average. This suggests that financial firms can adopt the proposed default prediction models without violating the regulatory requirement on model quality. Lastly, we find that the features of high prediction power in the cost-sensitive and insensitive models are different, which has an important implication for credit monitoring.
Chan Park
Seungyoo Jeon
Kisung Yang
In the default prediction problem, thecost from thefailure of forecasting defaults is much biggerthan that of forecastingnon-defaults. Thecost asymmetryis deeperin the corporatedefaultpredictionthan theretailas corporateloan portfolios arenot granular. However, thetwo typesof costs aretreatedequallyin generalas defaultpredictionmodels areusually estimatedto minimizepredictionerrorsor maximizestatistical performance. This practicemight not fulfill thegoal of risk managementto minimizeeconomiclosses. To mitigatethis issue, this study apply cost-sensitivelearningapproach to defaultprediction, which minimizeseconomiccosts insteadof statistical errors. Wedefineeconomic costs and testthemfor various levelsof thecost asymmetryby employingLogistic regression, XGBoost, and LightGBM. As a resultof empiricalexperimentswith Taiwaneseand Polish corporate default data, we first find that the proposed cost-sensitivemodels are superiorto thecost-insensitivecounterpartsin termsof economiccost, mostly regardless of thecost asymmetryscenarios. Secondly, neverthele,ssthedecreasesin thestatistical performanceare relativelysmall ? economic costs decrease24.6% at theexpenseof the decrease in AUC of 4.6% on average. This suggests that financial firms can adopt the proposed default prediction models without violating the regulatory requirement on model quality. Lastly, we find that the features of high prediction power in the cost-sensitive and insensitive models are different, which has an important implication for credit monitoring.
Cost sensitivity learning,Default prediction,Economic cost,Explainable AI
[ 07327 ] 67-8, Yeouinaru-ro, Yeongdeungpo-gu, Seoul, Korea TEL. +82-2-2003-9921 FAX. +82-2-2003-9979 E-mail. office@korfin.org COPYRIGHT(C) SINCE 1987 KOREAN FINANCE ASSOCIATION. All rights reserved.