Which predicting models perform best?
In this paper Alonso and Carbó test the efficiency of using machine learning (ML) models (such as Lasso penalized logistic regression, Classification And Regression Tree, Random Forest, XGBoost and Deep Neural Networks) in predicting credit default as compared to traditional methods such as Logit. They find that:
- ML models perform better than the traditional Logit model, both in classification and calibration terms;
- While calibration is clearly a more difficult task than classification, XGBoost and Random Forest seem to provide the best results in both measures, despite not being the most algorithmically complex models;
- Their sensitivity analysis, which simulates how the results would change in case of different number of observations and features, demonstrates that statistically there exists a model advantage on top of an information advantage;
- The gains in terms of savings that an institution would achieve if they were to use ML models could amount to up to 17% of capital requirements.
Understanding the performance of machine learning models to predict credit default: a novel approach for supervisory evaluation
Authors: Andrés Alonso, José Manuel Carbó
From: Banco de España
Novel data set
Kriebel and Stitz use three deep learning architectures (convolutional neural networks, recurrent neural networks, convolutional recurrent, and average embeddings) to extract credit relevant information from user-generated text on Lending Club, one of the major platforms for peer-to-peer lending in the United States. Their paper contributes to existing research by providing more positive evidence on the value of using textual information in improving credit default prediction models.
Our findings indicate that the textual information extracted by deep learning is significant and valuable in predicting out-of-sample credit defaults. Average embedding networks outperform the other deep learning architectures. This is particularly interesting as these are less well-known compared to the other architectures. The deep learning approaches outperform approaches based on word frequencies and text characteristics in almost all cases. Among the text characteristics, spelling mistakes are found to be of importance.
Deep Learning: Credit Default Prediction from User-Generated Text
Authors: Johannes Kriebel, Lennart Stitz
From: University of Muenster
Which consumers have a high default risk: a behavioral approach
Annette Vissing-Jorgensen looks at account-level data from debit and credit cards in order to understand whether consumption patterns reveal any information about whether a given consumer is more likely to default on his/her credit. She uses the US Consumer Expenditure Survey (CEX) to study the link between spending patterns and consumer credit outcomes (paying positive finance/interest/late charges on consumer credit). She supplements this with an analysis of a large account-level data set from a Mexican retail chain that sells durables on credit and from which she can measure loan default. Her main finding is that spending patterns across goods and services are informative for consumer credit outcomes. Interestingly, spending on entertainment-related goods and services is related to a higher probability of paying positive finance/interest/late charges or of not repaying consumer credit in full (in Mexican data).
I conjecture that higher entertainment-related spending is an indicator of impatience and show in the US data that the same spending categories that predict paying positive finance/interest/late charges also predict smoking and lower education, outcomes typically associated with impatience.
Consumer Credit: Learning Your Customer’s Default Risk from What (S)he Buys
Author: Annette Vissing-Jorgensen
From: University of California Berkeley
