Car Finance 24/7 Project: Predicting Lender Approval Decisions and Optimising Broker Routing Using Machine Learning

Analysis of simulated car finance loan application data revealed significant differences in lender risk tolerance across three distinct lending profiles using logistic regression and gradient boosting models, highlighting opportunities to improve broker efficiency through intelligent, data-driven application routing. The project was developed using Python, with scikit-learn supporting the core modelling pipeline.

Predictive Modelling

To predict lender approval outcomes, two classification models were developed and evaluated across all three lenders.

Logistic Regression Baseline: Logistic regression was selected as the baseline model due to its interpretability and its ability to estimate approval probabilities directly. The dataset was split into training and testing sets using an 80/20 stratified split to maintain class balance. Model performance was evaluated using ROC-AUC, which measures the model’s ability to distinguish between approved and rejected applications. The logistic regression model achieved ROC-AUC scores of 0.954 for the prime lender, 0.895 for the mid-prime lender, and 0.889 for the stability-focused lender, demonstrating that borrower financial characteristics carry strong predictive signal for approval decisions.

Boosted Model: To improve on the baseline, a HistGradientBoostingClassifier was implemented. Boosting algorithms construct an ensemble of decision trees sequentially, enabling the model to capture nonlinear relationships and interactions between borrower characteristics that a linear model cannot represent. The boosted model achieved higher predictive performance across all three lenders, with ROC-AUC scores of 0.975 for the prime lender, 0.926 for the mid-prime lender, and 0.921 for the stability-focused lender. The consistent improvement over logistic regression suggests that lender approval decisions involve complex, nonlinear interactions between risk indicators.

Lender Decision Rules

By analysing approval probabilities across borrower risk profiles, clear decision rules characterising each lender’s behaviour were identified.

The prime lender applies the strictest criteria. Borrowers with debt-to-income ratios exceeding approximately 0.9 are consistently rejected, and applicants with two or more missed payments within the previous year receive no approvals, indicating a hard repayment reliability threshold with no tolerance for recent default behaviour.

The mid-prime lender demonstrates greater flexibility. Borrowers with moderate leverage and up to two missed payments may still receive approval, though applicants with extremely high debt levels or three or more missed payments are consistently rejected.

The stability-focused lender operates with the most flexible policy overall. While it accepts borrowers with higher leverage levels than either of the other lenders, it still enforces minimum repayment reliability requirements, rejecting borrowers with three or more missed payments regardless of other characteristics.

Broker Routing Decision Strategy

The predictive models developed in this study allow brokers to estimate the probability that each lender will approve a given loan application. Rather than submitting applications randomly or based on intuition, brokers can calculate approval probabilities for all three lenders simultaneously and route each application to the lender with the highest predicted likelihood of acceptance.

In practice, this produces a structured routing framework. Borrowers with strong financial profiles, low debt-to-income ratios, and clean repayment histories are best suited to the prime lender. Applicants with moderate leverage or utilisation can be routed to the mid-prime lender. Borrowers carrying higher debt loads but demonstrating stable repayment behaviour over time are best directed to the stability-focused lender.

This approach can increase first-time approval rates, reduce the volume of unnecessary credit checks that negatively affect borrower credit scores, and materially improve broker operational efficiency.

GDPR and Ethical Considerations

Although the dataset used in this study is synthetic, real-world deployment of such models would require careful compliance with the General Data Protection Regulation (GDPR). Loan approval models rely on personal financial data, which is considered sensitive under GDPR, and organisations using these systems must ensure that borrower data is processed lawfully, transparently, and only for legitimate purposes. Data minimisation principles require that only necessary borrower information be collected. Algorithmic decision systems in lending must also remain explainable and auditable, as financial institutions are legally required to provide meaningful explanations for automated decisions affecting individuals.

Limitations and Future Improvements

Several limitations should be acknowledged. The dataset is synthetic and may not fully represent real-world borrower behaviour or lender policies. Variables such as borrower savings, loan purpose, and asset ownership were not included in the analysis, which may limit the completeness of the risk picture available to the models. Future improvements could incorporate additional borrower attributes, further feature engineering, and hyperparameter tuning through cross-validation. Real-world deployment would also require continuous monitoring to ensure model predictions remain accurate as lending policies and borrower behaviour evolve over time.

Conclusion

This project demonstrates how machine learning can support broker decision-making in loan application routing. By analysing borrower financial characteristics and modelling lender approval behaviour, the study identifies the key risk indicators that drive lender decisions and reveals clear differences in risk tolerance across the three lenders examined.

Both logistic regression and boosted models achieved strong predictive performance, with the boosted model consistently outperforming the baseline across all lenders. These insights enable brokers to implement intelligent routing strategies that match borrowers with lenders whose risk preferences align with their financial profiles, increasing first-time approvals, reducing failed applications, and improving the overall efficiency of loan brokerage operations.

Contribution

I contributed to the exploratory data analysis, predictive modelling, and lender decision rule analysis components of the project, including the development and evaluation of the logistic regression and boosted classification models, and supported the design of the broker routing strategy.

Acknowledgement

This work was developed collaboratively with my MSc Data Analytics team members Ebrima Khan, Muhammad Qureshi and Shivangi Sinha. I would also like to thank our supervisor and the case study client team at Car Finance 24/7 for their guidance throughout the project.