How a Leading Gold Loan Provider Eliminated Retention Guesswork with ML-Powered Churn Prediction

No Visibility Into Customer Churn Risk

The client had no system to identify which gold loan customers were likely to lapse at the end of their loan tenure. Every customer who closed a grant was treated identically by the marketing and call centre teams, regardless of their actual likelihood of returning. This meant that retention effort was dispersed indiscriminately across hundreds of thousands of accounts, with no prioritisation and no measurable impact.

Loan Approval Decisions Lacked Individual Intelligence

Loan-to-value ratios, a key lever for both customer satisfaction and portfolio risk management, were determined using blanket internal policies rather than individual repayment behaviour. Branch approval managers across the network had no data-backed framework to reward customers with strong repayment histories or to apply more conservative limits to higher-risk accounts. The result was a one-size-fits-all approach that neither served reliable customers well nor adequately managed risk.

Years of Valuable Transaction Data Sat Unused

Five years of granular transaction history across five interconnected SQL Server tables, including loan ticket records, gold article details, auction events, and customer master data, remained entirely unmodelled. The organisation was operationally data-rich but strategically insight-poor, with no pipeline to convert historical behaviour into forward-looking predictions.

Scale Made Manual Outreach Unsustainable

With an active customer base in the hundreds of thousands, the marketing and call centre teams faced an impossible task: reaching every customer with equal effort was both resource-intensive and ineffective. Without a risk-tiered prioritisation system, high-value retention conversations were lost in the noise of mass outreach, and the most at-risk customers received no differentiated attention.

01.

Churn Prediction Model

The team built an XGBoost-based machine learning model to predict the probability of customer churn for every active gold loan holder. A churn event was defined as any case where more than 90 days elapsed since the closure of a customer's most recent loan grant without a new loan being initiated, reflecting the business's natural customer lifecycle. The model was trained on five years of historical transaction data and engineered 54 features spanning loan tenure, repayment behaviour, arrear history, gold article characteristics, LTV ratios, and macro signals such as gold price movements. To handle the natural class imbalance between churned and retained customers, a scale weighting of 6.5 was applied during training. The model outputs a churn probability score between 0 and 1 for each target customer, which is automatically classified into three risk tiers: High Risk for probabilities at or above 0.70, Medium Risk for probabilities between 0.40 and 0.70, and Low Risk for probabilities below 0.40.

02.

Customer Scoring Model

A separate ML model was developed to score each customer on a 0 to 100 scale based on their full repayment history with the institution. The score was designed to serve as an objective, data-backed guideline for branch approval managers when determining the loan-to-value ratio to offer individual customers. Customers with strong repayment track records qualified for higher LTV ratios, up to 90% of gold collateral value, while customers with weaker histories were flagged for more conservative offers. This replaced the previous blanket policy approach with individual-level intelligence accessible to managers across all 50-plus branches.

03.

Tableau Reporting Dashboards

Two purpose-built Tableau dashboards were deployed on Tableau Server, connected live to the client's SQL Server database. The Customer Churn Overview dashboard surfaced the latest monthly churn probability scores and risk tier classifications for the marketing and call centre teams, enabling targeted outreach campaigns focused on the highest-risk segments. The Churned Customers view provided a longitudinal record of actual churn events alongside their predicted probabilities, supporting model performance tracking and campaign effectiveness analysis. Where Tableau viewer licences were insufficient to cover all 50-plus branch managers, a structured Excel export and distribution workflow was established to ensure universal access to customer scores across the branch network.

04.

Deployment and Automation

The churn model was operationalised as a Python script executed via a Windows Task Scheduler batch file on the client's dedicated virtual machine. The model runs automatically on the second of every month, updating the CustomerChurn output table in the VallibelCreditScore database with fresh predictions. The output table retains a rolling 24-month history of all model predictions, timestamped and tagged with a model run date to ensure full traceability and auditability for downstream reporting.

Targeted Retention Campaigns at Scale

The marketing and call centre teams gained a monthly, risk-tiered customer list that replaced indiscriminate mass outreach with focused, high-priority retention campaigns. By directing effort toward the High Risk segment, customers with a churn probability at or above 0.70, the client could concentrate its most valuable retention conversations on the accounts most likely to lapse, improving both campaign efficiency and customer engagement quality.

Data-Backed Loan Approval Decisions Across All Branches

Branch approval managers across the 50-plus branch network gained access to individual customer scores for the first time, providing an objective, standardised framework for loan-to-value decisions. Rather than applying a uniform policy regardless of customer history, managers could now reward strong repayment behaviour with more favourable LTV offers and apply appropriate caution to higher-risk accounts, improving both customer experience and portfolio risk management simultaneously.

Five Years of Historical Data Converted Into Live Intelligence

Transaction data spanning five years across five SQL Server tables, previously sitting unused, was fully operationalised into a monthly-refreshed ML pipeline. This transformed the client's data estate from a passive record of past activity into an active decision-support system, with predictions stored, versioned, and surfaced in Tableau dashboards for leadership, marketing, and branch operations.

Auditability and Model Performance Tracking

The output architecture was designed with long-term governance in mind. Every model run produces timestamped, versioned predictions retained for a rolling 24-month window, enabling the client to compare predictions against actual churn outcomes and track model accuracy over time. The ActualvsPrediction table further supports retrospective analysis, capturing both predicted probabilities and real churn events at the customer level.

Observable Revenue Impact Post-Deployment

Following deployment of the churn prediction model and targeted retention campaigns, the client reported measurable revenue improvement across the months after go-live. While concurrent gold price appreciation contributed to the revenue trend, the client directly attributed partial uplift to the model's ability to bring at-risk customers back into the loan portfolio, validating the system's effectiveness through campaign outcome tracking.