title	LoanGuard
emoji	🛡️
colorFrom	blue
colorTo	green
sdk	docker
app_port	7860
pinned	false
license	mit

LoanGuard

Loan application fraud detection on the LendingClub dataset (2007-2018, ~2.2M loans).

Live demo: https://huggingface.co/spaces/adwitiyashukla/loanguard

Stack

• Data: LendingClub accepted-loans CSV
• Labels: weak-supervision (first-payment default + anomaly rules)
• Features: behavioural, velocity, graph (networkx), WoE + smoothed target encoding
• Models: XGBoost, LightGBM, CatBoost, Isolation Forest, denoising autoencoder (PyTorch)
• Ensemble: logistic-regression stacker with isotonic calibration
• Serving: FastAPI + Streamlit dashboard, packaged as Docker
• Monitoring: Evidently drift report, Prometheus metrics

Results

Test metrics on a 50k-row time-split subset (0.37% positive rate).

Metric	XGBoost	LightGBM	CatBoost	Isolation Forest	Autoencoder	Ensemble
ROC-AUC	0.694	0.571	0.667	0.592	0.624	0.653
PR-AUC	0.014	0.007	0.010	0.005	0.005	0.006
KS	0.332	0.251	0.306	0.218	0.290	0.311
Brier	0.014	0.005	0.020	0.179	0.104	0.0038
Lift @ top 5%	3.57x	2.86x	2.86x	0.71x	0.71x	3.57x

Best threshold: 0.51. Expected net cost: $4.48 per applicant on a $50M origination book with the assumed cost matrix ($1,200 per missed fraud, $20 per false alarm).

Layout

src/
  data/         loader, weak-supervision labels, splitter, schema validation
  features/     behavioural, velocity, graph, WoE and target encoders
  models/       base class, XGBoost, LightGBM, CatBoost, Isolation Forest, autoencoder, ensemble
  training/     trainer, Optuna tuner
  evaluation/   metrics, cost sweep, SHAP explainer, fairness report
  api/          FastAPI schemas, service, main
  dashboard/    Streamlit risk console
  monitoring/   drift monitor
  utils/        config, logging, IO
scripts/        download_data, train
config/         config.yaml
artifacts/      trained models (Git LFS)

Run

git clone https://github.com/adwitiyashukla/loanguard.git
cd loanguard
python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt
python scripts/download_data.py
python scripts/train.py --config config/config.yaml
streamlit run app.py

Deploy

The repo doubles as a Docker Space on HuggingFace. Push to huggingface.co/spaces/adwitiyashukla/loanguard and the Space rebuilds automatically.