The journey of overcoming challenges in credit risk modeling

1.

16 Days of Despair,
Determination, and Success
The journey of overcoming
challenges in credit risk modeling

2.

Introduction to the Task
• Banks use credit risk management models to
evaluate the trustworthiness of clients in
fulfilling credit obligations. These models
automate the evaluation process, saving time
and enabling experts to audit the results when
necessary.
• The goal of this project: Create a model to
predict client default with a ROC-AUC score of
at least 75%.

3.

Data Overview
• We worked with 12 `.pq` files containing over
26 million records and a target `.csv` file with
3 million records. The dataset had 61 features,
all encoded or binary, adding complexity to
the task.

4.

Methods and Models
• Initial sampling of 100k IDs was not
representative. Moving to the full dataset
improved ROC-AUC. Key techniques:
• - Data type optimization to reduce memory
usage
• - Feature engineering using `Polars` for speed
• - GPU resources for faster computations
• Outcome: Reduced pipeline runtime from 5

5.

Results
• Model: CatBoostClassifier
• - ROC-AUC Score: 75.82%
• - Features: 183, after careful selection and
aggregation
• - Confusion Matrix: [Values not provided,
placeholder]
• The results exceed the project's requirements
and demonstrate the model's reliability.

6.

Future Improvements
• Potential enhancements:
• - Explore neural networks for complex feature
interactions
• - Fine-tune CatBoostClassifier parameters
further
• - Investigate advanced imbalance handling
techniques
• - Incorporate external data sources for richer
insights

7.

Acknowledgments and Contact
• Author: Alexey Zhuravlev
• Email: a.o.zhuravlev@gmail.com
• Thank you for your attention and support on
this journey of determination and success.