Give Me Some Credit: Predictive Credit Risk Modeling

Brya Patterson, et al., Spring 2025, University of Washington, MS Information Management (MSIM)

Brya K. Patterson

Executive Summary

  1. The Mission: With a team at the University of Washington - Masters of Information Management, I engineered a predictive credit risk framework to identify high-probability delinquency (90+ days) within a 2-year window for a portfolio of 250,000 borrowers.
  2. Impact: I Transformed the model’s defensive capabilities by increasing high-risk recall from 1% to 60%. This shift enables proactive debt restructuring and significantly reduces the cost of write-offs.
  3. Core Capabilities: Credit Risk Modeling, Cost-Sensitive Learning, Financial Resource Optimization, Portfolio Health.

Remediation & Recommendations

For this project, the goal was to move beyond simple binary approvals and build a predictive model that identifies borrowers likely to experience Serious Delinquency (90 days past due or worse) within a two-year window. Accurate predictions allow lenders to optimize their interest rates and mitigate the high costs associated with defaults.

  • The Remediation: Addressed the "Imbalanced Class" problem common in banking. Traditional models often ignore defaulters because they are a minority of the data; I re-engineered the model logic to "hunt" for these high-risk outliers, improving detection by 60x.
  • The Recommendation: De-emphasize "Revolving Utilization." My analysis showed that utilization spikes had a negligible impact on long-term default compared to income growth. I recommend a "Dynamic Income-to-Trend" metric for future credit limit increases.
  • The Future State: Proposed an Automated Intervention Pipeline. When the model identifies a high-probability delinquency, it should trigger an automated "Financial Health" outreach or loan restructuring offer before the first payment is missed.

    • Loss Mitigation Strategy

    Prioritized "Recall" over "Accuracy" to ensure the bank identifies 60% of potential defaults, directly protecting the institution's bottom line from high-cost write-offs.

    Evidence-Based Lending

    Proved that historical delinquency behavior is 100x more predictive than income levels, advocating for a shift in credit evaluation standards.

    Risk-Adjusted Logic

    Implemented cost-sensitive learning to balance the "Cost of Missed Defaults" vs. "Cost of False Alarms," aligning data science outputs with corporate risk appetite.

    View Full Case Study

    Overview

    This project re-evaluates the standard metrics used by financial institutions to assess borrower risk. Leveraging Python-based machine learning, our team analyzed historical data from 150,000 borrowers to identify the "pivot points" of financial default. We discovered that while income serves as a critical "risk shield"—where a $2,000 increase can reduce default probability by 97%—standard metrics like credit utilization had a surprisingly negligible impact on serious delinquency.

    By shifting focus from balance-based metrics to behavioral history, this model provides a framework for lenders to implement early-intervention protocols and more accurately tier risk across diverse demographic segments.

    Analytical Framework: Guiding Research Questions

    To ensure the model provided actionable business value, the analysis was structured around five core questions:

    1. Probability: What is the likelihood of a specific individual defaulting within the next two years?
    2. Predictors: Which financial indicators (e.g., age, debt ratio, or late payments) serve as the strongest predictors of default?
    3. Segmentation: Can we identify high-risk customer segments using unsupervised learning?
    4. Sensitivity: How do shifts in a borrower’s financial situation—such as a $2,000 income increase—impact their risk score?
    5. Behavioral Contrast: What are the defining demographic and financial differences between those who default and those who remain current?

    Key Insights

    • Behavior > Debt: Past payment history (specifically 90+ days late) is a 1,375% stronger predictor of default than current debt levels.
    • The Age Protective Factor: Risk steadily decreases with age, with a significant "stability drop-off" occurring after age 55.
    • Income Sensitivity: A $2,000 increase in monthly income was found to reduce default probability by up to 97% in high-risk simulations.
    • The Utilization Myth: Credit utilization and Debt Ratio had surprisingly low predictive power compared to behavioral history.

    Resources