Predictive Model Plan – Geldium Delinquency Risk
1. Model Logic (Generated with GenAI)
Based on the dataset characteristics (imbalanced classes, mix of
numerical/categorical data) and the business need for intervention, the
recommended logic is a Weighted Logistic Regression Pipeline.
Model Workflow:
1. Data Ingestion: Load customer data including demographics (Age, Location),
financial status (Income, Loan_Balance), and behavior (Credit_Utilization,
Payment_History).
2. Preprocessing:
o Imputation: Fill missing Income with the median and Credit_Score with the
mean.
o Encoding: Convert Employment_Status to one-hot vectors and
Month_1...Month_6 to ordinal values (0=On-time, 1=Late, 2=Missed).
o Scaling: Normalize numerical features (e.g., Income, Loan_Balance) to a 0-1
range to prevent dominance by large numbers.
3. Feature Selection: Prioritize top predictors: Credit_Utilization,
Debt_to_Income_Ratio, Missed_Payments, and Month_1 (Recent Payment Status).
4. Model Execution:
o Apply Logistic Regression with class_weight='balanced' to heavily penalize
missing a delinquent customer.
o Calculate a probability score (0 to 1) for each customer.
5. Output Generation:
o Score: Probability of Default (e.g., 0.78).
o Flag: If Score > Threshold (0.45), classify as "At-Risk."
2. Justification for Model Choice
Selected Model: Logistic Regression with Class Weighting
I selected this model over complex options (like Neural Networks) or simple Decision
Trees for three specific reasons related to Geldium’s goals:
1. Explainability (Regulatory Requirement): Logistic Regression provides clear
coefficients (weights) for every feature. We can tell the Collections team exactly why
a customer was flagged (e.g., "High utilization increased their risk score by 30%").
This is crucial for compliance and "Fair Lending" regulations.
2. Handling Imbalanced Data: Our EDA showed that actual delinquencies are
rare. A standard Random Forest might ignore these rare cases to maximize overall
accuracy. Logistic Regression allows us to tune the "threshold" (lowering the bar to
flag risk) and apply class weights, ensuring we catch the minority of struggling
customers.
3. Actionable Probabilities: Unlike a Decision Tree which gives a hard "Yes/No,"
this model provides a probability score. This allows the Collections team to prioritize
their list—calling the 90% risk customers first, then the 70%, maximizing operational
efficiency.
3. Evaluation Strategy
To ensure the model is both effective and ethical, we will look beyond simple
"Accuracy."
Key Metrics:
• Recall (Sensitivity): Primary Metric. We want to minimize False Negatives
(missing a customer who is about to default). A high Recall score means we are
successfully catching most at-risk accounts.
• Precision: Secondary Metric. We monitor this to ensure we aren't flooding the
call center with false alarms (customers who are actually fine).
• AUC-ROC: To measure the model's overall ability to distinguish between
"Good" and "Bad" borrowers independent of the specific threshold.
Bias & Fairness Strategy:
• Demographic Parity Check: We will calculate the "Positive Prediction Rate"
across protected groups (e.g., Age brackets or Location as a proxy for
demographics). If the model flags 50% of young people but only 10% of older
people, we must investigate if Age is being used as a bias proxy rather than a true
risk factor.
• Feature Review: We will strictly exclude Customer_ID and carefully monitor
Location to ensure redlining (geographic discrimination) does not occur.