How alirezamohamadian1998/default-prediction works
This repository implements an AMEX customer-default risk pipeline: validate monthly statements and labels, aggregate histories into customer-level features, train and select LightGBM models, then score new customers with an ensemble and saved threshold. Supporting branches tune hyperparameters and analyze model results and probability calibration. DuckDB-backed auditing and feature aggregation form the preparation layer. Relationships reflect the README and shown imports/calls; unsampled modules are included only for documented capabilities, with no inferred internal call chains.
Main entry points: prepare.py History preparation · predict.py Ensemble prediction · audit.py Dataset audit · train.py Baseline training
Components 17
Data preparation
- Dataset audit
src/amex_default/audit.py - DuckDB access
src/amex_default/database.py - Feature expressions
src/amex_default/features.py - History preparation
src/amex_default/prepare.py
Model development
- Baseline training
src/amex_default/train.py - Risk metrics
src/amex_default/metrics.py - Hyperparameter tuning
src/amex_default/tune.py - Candidate selection
src/amex_default/select_model.py
Prediction
- Ensemble prediction
src/amex_default/predict.py
Evaluation
- Model results analysis
analysis/model_results.py - Calibration evaluation
analysis/evaluate_calibration.py
Other components
- Statement data
- Default labels
- Pipeline operator
- Customer features
- Selected fold models
- Customer predictions
Connections 30
- Pipeline operator → Dataset audit: runs audit
README.md - Statement data → Dataset audit: validates
README.md - Default labels → Dataset audit: checks coverage
README.md - Dataset audit → DuckDB access: queries via
src/amex_default/audit.py - Pipeline operator → History preparation: runs preparation
README.md - History preparation → Dataset audit: audits inputs
src/amex_default/prepare.py - History preparation → Feature expressions: uses expressions
src/amex_default/prepare.py - History preparation → DuckDB access: connects through
src/amex_default/prepare.py - Statement data → History preparation: aggregates histories
README.md - History preparation → Customer features: writes features
README.md - Pipeline operator → Baseline training: runs training
README.md - Customer features → Baseline training: trains on
README.md - Default labels → Baseline training: trains against
README.md - Baseline training → Risk metrics: evaluates with
src/amex_default/train.py - Pipeline operator → Hyperparameter tuning: runs tuning
README.md - Customer features → Hyperparameter tuning: tunes on
README.md - Default labels → Hyperparameter tuning: tunes against
README.md - Hyperparameter tuning → Risk metrics: scores trials
src/amex_default/tune.py - Pipeline operator → Candidate selection: runs comparison
README.md - Candidate selection → Baseline training: reuses training
src/amex_default/select_model.py - Candidate selection → Selected fold models: retains winner
README.md - Pipeline operator → Ensemble prediction: runs scoring
README.md - Statement data → Ensemble prediction: provides statements
README.md - Ensemble prediction → History preparation: prepares features
src/amex_default/predict.py - Selected fold models → Ensemble prediction: supplies ensemble
README.md - Ensemble prediction → Customer predictions: writes probabilities
README.md - Pipeline operator → Model results analysis: runs analysis
README.md - Model results analysis → Selected fold models: compares artifacts
README.md - Pipeline operator → Calibration evaluation: runs evaluation
README.md - Customer predictions → Calibration evaluation: evaluates probabilities
README.md