Nextbank: AI credit scoring engine. 97% prediction accuracy. 500 million applications processed.

For financial institutions across Southeast Asia - branchless mobile banking, business intelligence, loan origination, and credit scoring. Over 50 institutions use the platform. When Nextbank decided to replace the traditional credit scoring module in their SaaS offering, the ambition was specific: build an AI-powered engine that outperformed conventional scorecards on accuracy, transparency, and scale - and could be deployed across banks with different data environments and risk models.

Miquido has been Nextbank's technology partner since 2019. The credit scoring engine we built analyses over 600 data points per applicant, operates proactively to identify creditworthy borrowers before they apply, and has processed 500 million loan applications across 7 banks in Asia. Prediction accuracy sits at 97%.

01. Scorecards produced a score without an explanation

Conventional credit scoring models produce a three-digit number - and nothing else. No explanation of which factors drove the score, no visibility into the weighting logic, no way for a loan officer to challenge or contextualise the result. For a bank making lending decisions, that opacity creates compliance risk, allows biases to compound undetected, and makes it impossible to adapt the model when economic conditions change.

02. Limited predictive power across changing conditions

Traditional scorecards are maintained manually - teams of analysts updating models per client type and product. They operate within a narrow context: credit history as the primary signal, with limited ability to incorporate the broader economic variables that actually determine whether a borrower will repay. The result is a model that describes the past reasonably well but struggles to predict the future accurately - particularly for borrowers whose circumstances don't fit historical patterns.

03. A reactive system that missed creditworthy borrowers entirely

Conventional scoring only activates when a customer applies. Borrowers who are creditworthy but haven't considered applying - or who were previously declined by a model that couldn't account for their individual circumstances - are invisible to the system. That means revenue left on the table for the bank, and credit access denied to people who could manage it.

04. Deployment across multiple banks with different data environments

The engine had to work across 7 banks with different data structures, different risk appetites, and different regulatory environments. Building a model that was accurate and adaptable - without requiring a full rebuild for each institution - required architecture decisions that balanced standardisation with flexibility.

01. Failure defined before the model was built

The first decision was definitional: what counts as a loan failure? After analysing historical data and consulting with Nextbank's business team, the team agreed on a precise threshold - a loan instalment unpaid 90 days after the deadline. That definition became the target the model was trained to predict. Starting with a clear, agreed definition of failure is what makes the model's outputs meaningful to the banks using it.

02. Three machine learning models, selected for complementary strengths

The data science team evaluated three frameworks: linear regression as a baseline, LightGBM, and XGBoost. Linear regression provided the analytical transparency needed to measure each factor's contribution to the final score - critical for identifying model problems and explaining results to loan officers and regulators. LightGBM and XGBoost delivered the prediction accuracy the engine needed at production scale. The combination gave Nextbank both explainability and performance - the two things traditional scorecards couldn't provide simultaneously.

03. 600+ data points replacing the three-digit black box

The engine analyses over 600 variables per applicant: credit history nuances including number of active and closed loans, cumulative balances, and payment delay patterns; demographic attributes including age, gender, civil status, and income type; and loan-specific context including loan type, industry purpose, Margin of Safety group, and loan amount. That depth of analysis is what produces 97% prediction accuracy - and it runs automatically, without requiring teams of analysts to maintain per-client scorecard models.

04. Proactive borrower identification

The engine doesn't wait for an application. It operates proactively - identifying individuals who are likely creditworthy based on available data signals, before they initiate the lending process. For the banks using Nextbank's platform, that expands the pool of potential borrowers and creates revenue opportunities that a reactive system would never surface.

01. 97% loan repayment prediction accuracy

The engine predicts with 97% accuracy whether an applicant will repay a loan - significantly outperforming conventional scorecard models. That accuracy translates directly to lower default rates for the banks using the platform and better credit access for borrowers that traditional models would have declined without justification.

02. 500 million loan applications processed

The engine has processed over 500 million loan applications across its deployed banks. At that volume, the difference between 97% accuracy and the accuracy of a conventional scorecard is measured in millions of lending decisions made correctly rather than incorrectly.

03. Deployed across 7 banks in Asia

The scoring engine is live in 7 financial institutions across Southeast Asia - each with different data environments, risk models, and regulatory requirements. The architecture designed during the initial build has supported deployment across all of them without requiring a rebuild for each institution.

04. Manual scoring teams replaced by automated assessment

The process that previously required dedicated teams of analysts maintaining scorecards per client type and product now runs automatically. Loan officers receive a score, an explanation of the contributing factors, and a confidence level - without waiting for manual analysis. The speed and consistency of assessment improves the borrower experience and reduces operational cost for the institutions.

05. Singapore FinTech Awards Finalist 2019

The engine was recognised as a finalist at the Singapore FinTech Awards in 2019 - validation of the technical approach and its relevance to the financial services market Nextbank operates in.