[Paper Review] Improving a Credit Scoring Model by Incorporating Bank Statement Derived Features

Rory Bunker, Wenjun Zhang+1 more|arXiv (Cornell University)|Oct 30, 2016

Financial Distress and Bankruptcy Prediction25 references8 citations

TL;DR

This study enhances a credit scoring model for a New Zealand lender by integrating 5 novel features derived from 90-day bank statements—previously unused in application forms—into a Naive Bayes classifier. The combined model, using 11 existing application features plus 5 bank statement features, achieved the highest AUC (0.791), outperforming both baseline models, demonstrating that bank statement data significantly improves credit risk prediction, especially in highly imbalanced datasets.

ABSTRACT

In this paper, we investigate the extent to which features derived from bank statements provided by loan applicants, and which are not declared on an application form, can enhance a credit scoring model for a New Zealand lending company. Exploring the potential of such information to improve credit scoring models in this manner has not been studied previously. We construct a baseline model based solely on the existing scoring features obtained from the loan application form, and a second baseline model based solely on the new bank statement-derived features. A combined feature model is then created by augmenting the application form features with the new bank statement derived features. Our experimental results using ROC analysis show that a combined feature model performs better than both of the two baseline models, and show that a number of the bank statement-derived features have value in improving the credit scoring model. The target data set used for modelling was highly imbalanced, and Naive Bayes was found to be the best performing model, and outperformed a number of other classifiers commonly used in credit scoring, suggesting its potential for future use on highly imbalanced data sets.

Motivation & Objective

• To evaluate whether features derived from bank statements—unreported on loan application forms—can enhance credit scoring models.
• To compare the predictive performance of a model based solely on application form features versus one using only bank statement-derived features.
• To develop and evaluate a combined model that integrates both types of features to improve credit risk assessment.
• To assess the suitability of Naive Bayes for highly imbalanced credit scoring datasets.

Proposed method

• Constructed a baseline model using only the 11 existing application form features.
• Created a second baseline model using only 5 newly derived features from 90-day bank statement data.
• Combined both feature sets into a single 16-feature model for evaluation.
• Applied the Fayyad & Irani discretization method to convert continuous bank statement features into bins.
• Used Weight of Evidence (WOE) transformation for feature encoding to improve model interpretability and performance.
• Evaluated multiple classifiers using ROC AUC, with Naive Bayes selected as the best-performing model.

Experimental results

Research questions

• RQ1 Can bank statement-derived features that are not declared on loan application forms improve credit scoring model performance?
• RQ2 How does a model using only bank statement features compare to one using only application form features in terms of predictive power?
• RQ3 Does combining bank statement features with traditional application form features lead to a significant improvement in model performance?
• RQ4 Is Naive Bayes a viable and effective classifier for highly imbalanced credit scoring datasets?

Key findings

• The combined feature model, incorporating 11 application form features and 5 bank statement-derived features, achieved the highest AUC of 0.791, outperforming both baseline models.
• The baseline model using only application form features had a higher AUC than the model using only bank statement features, indicating the latter alone were insufficient for strong prediction.
• The Naive Bayes classifier outperformed other commonly used classifiers in credit scoring, particularly on highly imbalanced data.
• Five specific bank statement-derived features—such as holding a particular credit card type and not banking with a specific institution—showed significant Weight of Evidence (WOE) values, indicating predictive relevance.
• The model with a misclassification cost ratio of x=2 (where false positives are twice as costly as false negatives) achieved a true positive rate of 0.953 and correctly classified 7,059 out of 7,401 good applicants.
• The study suggests Naive Bayes merits further investigation in credit scoring, especially for imbalanced datasets, due to its strong performance in this context.