ML-based risk scoring model reducing default rates by 44%
Key results
12-month default rate
8.7% → 4.9% (−44%)
Application decision time
4.2 days → 6 hours
Underwriter consistency
Gini 0.61 → 0.89
Early warning accuracy
73% of defaults flagged 90+ days early
Estimated year-1 net benefit
€3.2M (provisions + efficiency)
Problem
Medivest Capital specialized in SME lending in the healthcare sector. Their existing credit scoring methodology relied primarily on traditional financial ratios and manual underwriter assessments — a process that was slow (average decision time of 4.2 days), inconsistent (significant variation in approval rates across underwriters), and producing a default rate of 8.7% within 12 months of origination, significantly above the company's internal target of 5%.
The inconsistency issue was particularly damaging: analysis showed that the same application, reviewed by different underwriters, received approval decisions in 34% of cases with a difference of two or more rating grades. This created significant regulatory exposure and undermined the company's risk management framework.
A deteriorating macroeconomic environment was placing additional pressure on portfolio quality, making the need for more rigorous, data-driven credit assessment more urgent.
Approach
The project involved building a two-stage ML-based risk scoring system.
- Traditional financial features (revenue, EBITDA margin, leverage ratios, liquidity metrics)
- Behavioral features derived from banking transaction data (payment regularity, cash flow volatility, seasonal patterns)
- Sector-specific features related to healthcare industry dynamics (NHS reimbursement patterns, practice type, regulatory registration status)
- Bureau data features from multiple credit reference agencies
Particular care was taken to address class imbalance in the training data and to implement robust validation procedures including time-series cross-validation to prevent look-ahead bias.
Stage 2 — Behavioral Monitoring: A separate set of models for early warning of financial distress among existing borrowers, using monthly updated behavioral features to identify accounts requiring proactive management 3-6 months before formal default.
Both models were built with full explainability using SHAP values, allowing underwriters to understand the primary factors driving any individual score — critical for regulatory compliance and for building underwriter trust.
Solution & Delivery
The scoring system was integrated into Medivest's loan origination platform through an API, providing real-time credit scores with SHAP-based explanations for each application. The underwriter interface presents the model score alongside the key positive and negative factors, enabling informed credit decisions in a fraction of the previous time.
The behavioral monitoring system generates monthly risk alerts for the existing loan portfolio, flagging accounts showing early signs of distress and recommending specific management actions.