Predictive Model Plan
Model Used: Decision Tree Classifier
A tree-based model that segments customers based on conditions like high credit utilization, low credit
score, or frequent missed payments. It makes predictions by following a clear set of rule-based splits and
provides intuitive explanations for each classification.
Top 5 Input Features for Your Model
1. Credit Utilization – indicates how much credit the customer is using vs. limit.
2. Missed Payments – recent payment behavior directly linked to risk.
3. Debt to Income Ratio – shows financial strain or overleveraging.
4. Credit Score – traditional risk score used widely by lenders.
5. Income – supports ability to repay, especially when combined with loan balances.
1. Model Logic (Generated with Gen
AI):
GenAI-Generated Model Logic (Step-by-Step Process)
1. Input: Customer Data
python
CopyEdit
customer = {
"Income": 50000,
"Credit_Utilization": 0.85,
"Debt_to_Income_Ratio": 0.45,
"Missed_Payments": 2,
"Credit_Score": 580
}
2. Preprocessing Rules
o If any value is missing → fill with median value.
o Normalize Credit_Utilization if it exceeds 1.0.
o Convert categorical variables (e.g., "Employment Status") to numeric codes if
needed.
3. Decision Tree Logic (Pseudocode)
python
CopyEdit
if Credit_Utilization > 0.75:
if Credit_Score < 600:
if Missed_Payments >= 2:
return "Delinquent"
else:
return "Non-Delinquent"
else:
return "Non-Delinquent"
else:
if Debt_to_Income_Ratio > 0.4:
return "Delinquent"
else:
return "Non-Delinquent"
4. Model Output
python
CopyEdit
result = predict(customer) # → Output: 'Delinquent' or 'Non-Delinquent'
Simplified Python Code Snippet (Gen AI-Assisted)
python
CopyEdit
def predict(customer):
if customer["Credit_Utilization"] > 0.75:
if customer["Credit_Score"] < 600:
if customer["Missed_Payments"] >= 2:
return "Delinquent"
else:
return "Non-Delinquent"
else:
return "Non-Delinquent"
else:
if customer["Debt_to_Income_Ratio"] > 0.4:
return "Delinquent"
else:
return "Non-Delinquent"
# Example usage
sample = {
"Income": 50000,
"Credit_Utilization": 0.85,
"Debt_to_Income_Ratio": 0.45,
"Missed_Payments": 2,
"Credit_Score": 580
}
print(predict(sample)) # Output: Delinquent
What the Model Is Designed to Do
The model is designed to predict whether a customer is likely to become credit delinquent
(i.e., default or miss payments) based on key financial and behavioral variables like income,
credit utilization, debt ratio, credit score, and missed payments. This helps financial institutions
assess risk and make better lending decisions.
Sample Output of the Credit Delinquency Predictor App
When a user sends a valid POST request to the /predict endpoint with a JSON payload like:
{
"Income": 50000,
"Credit Utilization": 0.85,
"Debt to Income Ratio": 0.45,
"Missed Payments": 2,
"Credit Score": 580
}
The API processes the input through the trained decision tree model and returns:
{
"Prediction": "Delinquent"
}
This output means the model predicts that the customer is likely to be delinquent based on their
financial indicators. If the inputs had indicated lower credit utilization, higher credit score, or
fewer missed payments, the output might have been:
{
"Prediction": "Non-Delinquent"
}
This clear binary classification allows institutions to make risk-based decisions quickly and
effectively.
Why This Logic Works for Credit Risk?
1. Reflects Lender Behavior: The decision rules mirror how real-world lenders assess risk
—high credit utilization, low income, frequent missed payments, and poor credit scores
are well-established red flags in the credit industry.
2. Captures Nonlinear Relationships: Decision trees don't assume a straight-line (linear)
relationship between variables. This is useful because, in financial behavior, risk often
increases sharply after certain thresholds (e.g., credit utilization over 75%).
3. Handles Mixed Data Types: The model can seamlessly process both numeric features
(like income) and categorical ones (like employment status) without needing complex
transformations, making it ideal for financial datasets.
4. Interpretability = Compliance: Every prediction is traceable through a sequence of
clear decisions, which is critical for regulatory compliance (e.g., Fair Lending laws,
FCRA). Stakeholders can understand why a person was flagged as high risk.
5. Operational Simplicity: The model is lightweight, fast to run, and can be deployed
easily in digital loan systems or backend financial workflows—no advanced
infrastructure required.
2. Justification for Model Choice:
Model Selection Rationale
The decision tree model was selected due to its optimal balance between accuracy,
interpretability, and ease of implementation—three factors that are crucial in the financial
services industry. While more complex models like neural networks may offer marginally higher
accuracy, they lack the transparency and explain ability needed for regulatory compliance and
internal audit ability. Logistic regression, though highly interpretable, is limited in capturing
non-linear relationships common in financial behavior patterns, such as the interaction between
income level, credit utilization, and missed payments. Decision trees not only handle both
numerical and categorical data naturally, but also provide an intuitive, rule-based structure
that aligns with how risk managers and compliance officers think. This makes them particularly
relevant for Geldium, where accountability, fairness, and traceability are as important as
predictive performance.
Aspect Logistic Regression Decision Tree
Very high—coefficients High—decision paths can be
Interpretability
show direct impact visualized as rules
Good for linear Better for non-linear and
Performance
relationships complex patterns
Handling of Can handle missing data
Requires imputation
Missing Data during tree splitting
Extremely efficient for Reasonable for small to mid-
Scalability
large datasets sized datasets
Lightweight, easy to Also lightweight and intuitive
Deployment
integrate to deploy
Regulatory Easily explainable to Excellent for audit trails due
Transparency auditors and regulators to rule-based splits
High if not pruned or
Overfitting Risk Low with regularization
controlled
Use Case Fit Best for baseline, linear- Best for rule-based,
Aspect Logistic Regression Decision Tree
risk assessments interpretable credit decisions
3. Evaluation Strategy:
Key Metrics for Model Performance
1. Accuracy
o Measures the percentage of correct predictions overall.
o Useful as a general indicator but may be misleading in imbalanced datasets.
2. Precision
o The proportion of predicted "Delinquent" cases that are actually delinquent.
o Important to minimize false alarms and avoid denying credit to good customers.
3. Recall (Sensitivity)
o The proportion of actual delinquent customers correctly identified.
o Crucial to avoid missing risky customers who could default.
4. F1 Score
o Harmonic mean of precision and recall.
o Balances both false positives and false negatives, especially valuable when data is
imbalanced.
5. ROC-AUC (Area Under the Curve)
o Indicates the model's ability to distinguish between delinquent and non-delinquent
cases.
o AUC closer to 1.0 means better separation of classes; useful for comparing
models.
How to Interpret These Metrics?
High Recall with moderate Precision is acceptable when it's more important to flag
risky customers, even if it includes some false positives.
A balanced F1 score indicates the model isn’t skewed toward just one type of error.
AUC values above 0.80 indicate strong discriminatory power in most financial
applications.
Bias & Fairness Assessment in the Credit Risk Model
To determine whether the decision tree model exhibits bias or unfair treatment toward certain
customer groups, we evaluate prediction outcomes across demographic attributes such as
gender, age group, income level, and employment status.
1. Group-wise Prediction Analysis
We compare the proportion of customers predicted as “delinquent” across different
subgroups:
Age: Are younger or older applicants flagged more often?
Income Level: Does the model disproportionately flag low-income customers?
Employment Type: Are part-time or contract workers unfairly penalized?
Gender (if available): Are women or men being predicted as high-risk at different rates?
2. Fairness Metrics Used
Disparate Impact: We measure whether one group is flagged significantly more often
than another. A ratio below 0.8 (or 80%) often signals unfair treatment.
Equal Opportunity: Are all groups with actual delinquency being correctly flagged at
similar rates (true positive rate)?
False Positive Rate: Are some groups wrongly classified as delinquent more often?
3. Initial Observations
The model may flag lower-income individuals more frequently, potentially due to the
strong influence of credit utilization and debt-to-income ratio.
If categorical demographic variables (like Employment_Status or Location) are
proxying for protected classes (e.g., race, gender), the model may unintentionally embed
bias.
Younger age groups might be at higher risk not because of actual behavior, but due to
shorter credit histories—a systemic issue rather than a behavioral one.
4. Mitigation Strategies
Bias-Aware Preprocessing: Rebalance data using sampling or reweighting.
Post-processing: Apply fairness constraints or threshold adjustments after prediction.
Feature Review: Remove or limit the influence of features highly correlated with
protected attributes (e.g., zip code as proxy for socioeconomic status).
Conclusion
Preliminary analysis suggests that the decision tree model, while interpretable and efficient, may
reflect underlying data biases that disproportionately flag certain groups as high risk—
especially those with lower income, shorter credit histories, or irregular employment.
Ongoing fairness audits and potentially incorporating fairness-aware algorithms are
recommended to ensure ethical and compliant lending decisions.
Ethical Considerations
Transparency: Ensure decisions can be explained clearly to customers and regulators.
Fairness: Avoid using features that act as proxies for protected attributes (e.g., ZIP codes
indicating socioeconomic status).
Customer Impact: Use predictions as decision support—not automated denial—
especially in borderline cases.
Regulatory Compliance: Align with standards like the Fair Credit Reporting Act
(FCRA), Equal Credit Opportunity Act (ECOA), and Basel frameworks.
Evaluation Plan (Summary)
Accuracy: Measures overall correctness; useful but less reliable with class
imbalance.
F1 Score: Balances precision and recall; key for credit risk where both false
positives and false negatives matter.
ROC-AUC: Assesses model's ability to distinguish between risky and non-
risky customers; AUC > 0.8 preferred.
Fairness Checks: Use disparate impact ratio, equal opportunity difference,
and false positive rate gaps across demographics (e.g., income, age) to detect
bias.
Interpretation: Focus on minimizing missed high-risk cases (high recall)
while avoiding unfair denials (high precision), and ensure consistent
performance across groups.