An AI-driven loan brokerage platform: integrating socio-economic factors, consumer financial behaviour, and multi-criteria decision analysis for responsible lending.

The loan settlement process is a critical yet complex aspect of financial intermediation, involving multiple stages of decision-making that impact both lenders and borrowers. Traditional approaches to loan assessment rely heavily on manual review or simplified credit scoring mechanisms that primarily evaluate applicants based on income, liabilities, and historical repayment records. Although these methods are widely adopted, they are often time-consuming, error-prone, and limited in their ability to capture the nuanced socio-economic and behavioural realities of borrowers. As a result, eligible applicants may be excluded, high-risk applicants may be approved, and borrowers frequently unable to identify lenders that align with their needs and preferences. This thesis addresses these limitations by developing a comprehensive, data-driven, and transparent loan brokerage platform that integrates machine learning, knowledge graph reasoning, and multi-criteria decision analysis (MCDA) to streamline loan settlement in three interconnected phases: loan eligibility prediction, loan risk assessment, and suitable lender selection. The first phase of the research focuses on loan eligibility prediction, extending beyond conventional credit evaluation by incorporating five additional socio-economic factors: geographic region, bankruptcy history, gambling expenditure, Centrelink benefit reception, and court judgments. These attributes, combined with traditional applicant features such as income, expenses, credit score, loan amount, and loan term, are modelled using advanced machine learning algorithms including random forest, XGBoost, and AdaBoost. A knowledge graph is constructed to map causal dependencies among features, enhancing transparency and providing interpretable insights into how socio-economic conditions influence eligibility outcomes. Experimental validation using 5,000 real-world loan applications demonstrates that the inclusion of socio-economic variables significantly improves predictive accuracy while reducing bias against disadvantaged populations. The second phase addresses loan default prediction by analysing consumer financial behaviour derived from bank and credit card statements in combination with credit report features. Transactional attributes such as income stability; recurring obligations such as rent, utilities, subscription fees; and discretionary spending on entertainment, dining out, luxury spending and gambling are evaluated to provide a holistic view of repayment capacity. The integrated model achieves superior predictive performance compared to traditional credit score–based approaches, with SHAP (SHapley Additive Explanations) employed to interpret the contribution of individual features. This phase highlights the importance of behavioural and transactional insights in reducing default risk and strengthening responsible lending practices, particularly for younger or migrant borrowers lacking extensive credit histories. The third phase develops a suitable lender selection within the broker platform environment. A rule-based filtering mechanism is applied to screen applicants against lenders’ baseline criteria, after which multi-criteria decision analysis (MCDA) using the analytic network process (ANP) is employed to prioritise lenders based on applicant preferences. Four primary criteria such as interest rate, down payment requirement, response time, and customer satisfaction are modelled to capture trade-offs and interdependencies between preferences. This dual-sided approach ensures that applicants are not only matched with eligible lenders but also provided with options that align with their personal financial goals. The integration of these three phases into a unified brokerage platform represents a novel contribution to loan decision support systems. The framework advances responsible lending by improving fairness and transparency, enhancing operational efficiency through reduced reliance on manual assessments, and supporting borrower-centric outcomes by aligning lender recommendations with applicant priorities. The findings of this thesis hold significant implications for financial institutions, regulators, and technology-driven lending platforms, offering a scalable and practical pathway toward modernising loan settlement processes.

Grantor dc:publisher

UNSW, Sydney

Year dc:date

2026

Author dc:creator

• Kamruzzaman, Md

• Loan process optimization
• Loan Eligibility Prediction
• Loan Default Prediction
• Consumer financial behaviour
• Suitable lender selection
• ANP process
• anzsrc-for: 3502 Banking, finance and investment
• anzsrc-for: 4611 Machine learning

Statement dc:rights

• open access
• CC BY 4.0
• free_to_read

Licence

• https://purl.org/coar/access_right/c_abf2
• https://creativecommons.org/licenses/by/4.0/

Language dc:language

en

Identifier

https://doi.org/10.26190/unsworks/32035

OAI identifier oai:identifier

oai:unsworks.library.unsw.edu.au:1959.4/107047