Credit risk prediction in an imbalanced social lending environment

Abstract: Credit risk prediction is an effective way of evaluating whether a potential borrower will repay a loan, particularly in peer-to-peer lending where class imbalance problems are prevalent. However, few credit risk prediction models for social lending consider imbalanced data and, further, the best resampling technique to use with imbalanced data is still controversial. In an attempt to address these problems, this paper presents an empirical comparison of various combinations of classifiers and resampling techniques within a novel risk assessment methodology that incorporates imbalanced data. The credit predictions from each combination are evaluated with a G-mean measure to avoid bias towards the majority class, which has not been considered in similar studies. The results reveal that combining random forest and random under-sampling may be an effective strategy for calculating the credit risk associated with loan applicants in social lending markets.

• The paper introduces a novel approach for credit risk prediction in P2P lending by addressing class imbalance using resampling methods.
• It demonstrates that a random forest with random under-sampling (RF-RUS) achieves superior G-mean performance by balancing sensitivity and specificity.
• The paper implies that rigorous feature engineering and model optimization can enhance real-world credit decision-making in peer-to-peer platforms.

Credit Risk Prediction in an Imbalanced Social Lending Environment

This research addresses credit risk prediction in peer-to-peer (P2P) lending, focusing on imbalanced data challenges typical in this domain. The study systematically explores and compares various resampling techniques in combination with machine learning classifiers to enhance prediction accuracy. This paper contributes a novel methodology for credit risk assessment, leveraging a current dataset from Lending Club to evaluate model performance using the G-mean metric, which accounts for the inherent bias towards majority classes in imbalanced datasets.

Introduction

P2P lending platforms have democratized financing by facilitating direct transactions between borrowers and lenders, often without traditional financial intermediaries. The absence of collateral and formal credit checks presents a unique challenge of evaluating borrower creditworthiness in a setting typically characterized by a high level of information asymmetry and imbalanced data. Traditional credit risk prediction models often fail in these environments, as they assume balanced distributions. Hence, this research embarks on an empirical study to address class imbalance issues in P2P lending through various resampling techniques combined with machine learning classifiers.

Resampling Methods and Classifiers

The study investigates three primary classes of resampling methods: under-sampling, over-sampling, and hybrid techniques.

Under-Sampling Techniques:

• Random Under-Sampling (RUS): Reduces the majority class by randomly removing instances.
• Instance Hardness Threshold (IHT): Selectively removes instances based on their classification margin.

Over-Sampling Techniques:

• Random Over-Sampling (ROS): Duplicates minority class instances.
• SMOTE: Utilizes k-nearest neighbors to synthetically generate new instances.
• ADASYN: Generates synthetic samples with a non-linear approach prioritizing harder-to-learn examples.

Hybrid Techniques:

• SMOTE + Tomek Links: Removes borderline instances to eliminate class overlap.
• SMOTE + Edited Nearest Neighbor (ENN): Applies edited nearest neighbor cleaning after SMOTE augmentation.

For classification, the study employs logistic regression, linear discriminant analysis, and random forest algorithms, assessing their performance using accuracy, AUC, sensitivity, specificity, and G-mean metrics.

Experimental Results

Using data from the Lending Club, inclusive of approximately 636,000 records, the research conducts extensive feature engineering and resampling to mitigate class imbalance effects, achieving more reliable credit scoring predictions. Feature engineering includes removing outliers, addressing data leakage, applying transformations, and computing correlations to optimize input data fidelity. The key findings indicate:

• RF-RUS achieves superior G-mean performance, effectively balancing specificity and sensitivity, making it the most reliable technique out of all combinations assessed.
• RF-ADASYN provides high sensitivity but suffers from specificity issues, highlighting a divergence in handling class imbalance.
• Hybrid methods like LR-SMOTE-Tomek demonstrate moderate performance improvements but fail to surpass the RF-RUS combination in G-mean metrics.

Implications and Future Work

This research implies that rigorous feature selection and the use of random under-sampling combined with a random forest classifier can significantly enhance credit risk predictions in imbalanced datasets typical in P2P lending. Practically, this framework can enable P2P platforms to make informed credit decisions, potentially reducing default rates and optimizing portfolio distributions.

Future research might explore the integration of support vector machines (SVM) with parameter optimization for enhanced classification. Continued publication and exploration of Lending Club's datasets can refine models against concept drift, which remains an essential consideration in volatile financial data streams.

Conclusion

Class imbalance remains a critical obstacle in credit risk prediction for P2P lending platforms. This study evidences that the strategic combination of resampling with advanced machine learning techniques can yield substantial improvements in predictive quality. The findings advocate for RF-RUS's adoption in real-world credit assessment applications, potentially transforming risk evaluation protocols across the peer-lending industry.