CashFlow Credit Scoring

DSC 180B Capstone Project · UC San Diego · Prism Data

Evaluate credit risk using machine learning on transaction-level cash flows as a data-driven alternative to traditional scoring.

---

Introduction

Traditional credit scores rely on historical repayment data, often excluding those with limited formal credit history. This project develops a framework to assess creditworthiness via real-time behavioral signals such as income consistency, spending composition, and liquidity trends derived from bank transactions. Our results demonstrate that these cash-flow signals significantly enhance risk prediction for a more inclusive financial system.

Installation & Setup

1. Clone the Repository:

   git clone https://github.com/harisdsc/DSC180B.git
   cd DSC180B
   

2. Environment Setup:

   python3 -m venv env
   source env/bin/activate
   pip install -r requirements.txt
   

Data

The project processes hierarchical financial records:

• Consumer Data: Evaluation dates and delinquency outcomes (15,000 rows)

• Account Data: Types, balance dates, and amounts (24,466 rows)

• Transaction Data: Records with categories, amounts, and dates (6.4M rows)

• Category Map: Mapping for 50 transaction categories

Usage

1. Data Pipeline:

python3 src/pipeline.py

Output is cached at data/features.pqt

2. Model Training:

python3 train.py <model> [--tune]

• Models: log-reg, xgboost, catboost, lightgbm
• --tune: Runs 100 trials of Optuna hyperparameter optimization

Project Structure

• configs/: Optimized model hyperparameters
• data/: Local storage for datasets and features (git-ignored)
• models: Saved trained model binaries
• images/: Visualizations of model performance
• src/: Source code fro loading, engineerin, and evaluation
• train.py: Main script for model training and seletion

Results

The CatBoost model achieved the highest perfomance with an AUC-ROC of 0.8585.

• CatBoost: 0.8585 AUC

• XGBoost: 0.8539 AUC

• LightGBM: 0.8222 AUC

• Logistic Regression: 0.7532 AUC

---

Team: Ada Mo, Brighton Chan, Haris Saif, Kyle Choi Mentors: Kyle Nero & Daniel Matthew (Prism Data)