Probability of Default Prediction Model for a Retail Bank

The Client

The client is a retail bank managing a diverse portfolio of lending products. Accurate credit risk assessment is critical to their lending strategy and regulatory compliance. As regulatory requirements such as IFRS 9 increased in complexity, they needed more advanced models to estimate probability of default and expected credit losses.
 

The Challenge

Traditional credit scoring models rely heavily on static indicators such as credit history, often failing to capture the dynamic nature of borrower risk. This limits their ability to accurately predict defaults, especially under changing economic conditions.

The bank required a comprehensive framework to estimate both short-term (12-month point-in-time) and long-term (lifetime) probability of default. This included incorporating macroeconomic effects alongside customer-level behavioral and demographic data.

Additionally, the solution needed to support IFRS 9 Expected Credit Loss (ECL) calculations, requiring robust modeling of default probabilities over extended time horizons and across different product lines.

The challenge was further complicated by data fragmentation, the need for clustering similar customers into statistically meaningful groups, and ensuring model interpretability for regulatory use.

Without an advanced modeling approach, the bank risked inaccurate risk assessment, suboptimal pricing, and non-compliance with regulatory requirements.
 

What We Did

Approach

We developed a comprehensive credit risk modeling framework combining machine learning with probabilistic modeling techniques. The solution addressed both short- term risk prediction and long-term default dynamics.

During the discovery phase, we identified the need to segment customers into homogeneous risk pools to improve model stability and interpretability. Hierarchical clustering was used to group similar clients, enabling more robust estimation of default behavior.

A key decision was to separate the modeling into two layers: point-in-time PD models for near-term risk and Markov-based models for long-term default evolution, incorporating macroeconomic adjustments in both cases.
 

Methodology

• Data assessment:
  Collected and cleaned historical loan data, including defaults, customer demographics, and behavioural features
• Approach selection:
  Combined machine learning (e.g. logistic regression, decision trees) with clustering and Markov process modelling
• Model development:
  Built penalized LASSO logistic regression models for 12-month PiT PD and developed lifetime PD curves using Markov transition matrices
• Validation framework:
  Conducted out-of-sample testing to validate predictive accuracy and robustness across different portfolios
• Integration:
  Delivered model outputs as probability scores for integration into lending, pricing, and risk management workflows
• Compliance:
  Ensured alignment with IFRS 9 requirements for Expected Credit Loss modelling and regulatory transparency

The Outcome

Robust estimation of both short-term and lifetime probability of default

The solution enabled the bank to accurately assess default risk across multiple time horizons, supporting both operational decision-making and regulatory compliance.
 

Secondary outcomes:

• Improved accuracy of credit risk assessment at the individual and portfolio levels
• Long-term default forecasting over a 35-year horizon
• Enhanced support for IFRS 9 Expected Credit Loss calculations
• Scalable framework applicable across multiple lending products

Facing a similar challenge?

If you need to improve credit risk modelling and meet regulatory requirements, we can help build advanced AI-driven solutions tailored to your portfolio.

Book a Technical Consultation