Credit scoring model for MD Finance
We built a credit scoring model for the Romanian market that predicts loan defaults with a GINI of around 40 and an AUC of around 70%, served through a REST API.
Published
- Industry
- Fintech, Lending
- Region
- Romania
- Services
- Data science, Machine learning, API development
- Duration
- 1 month
- GINI
- ~40
- AUC
- ~70%
- Project duration
- 1 month
Intro
MD Finance needed a credit scoring model for the Romanian market to predict loan defaults, improve lending decisions, reduce losses and support risk management.
We delivered it end to end as an outsourced project — from preprocessing to a containerized model behind a REST API.
Goals and challenges
- 01
Real-time latency
Scores had to be returned in real time.
- 02
Competitive accuracy
The model’s AUC had to compete with other vendors engaged by the client, against a target of GINI above 40 and AUC above 70%.
- 03
Easy deployment
The model had to be simple to deploy and integrate.
Solution
Data preprocessing
We handled missing values, filtered continuous features by missing-value thresholds and engineered new features.
Modelling
LightGBM for training, Bayesian optimization for hyperparameters, and Python for preprocessing, modelling and evaluation.
Deployment
Docker and Flask package the model as a container with a REST API for integration.
Performance evaluation
Predictions were grouped into ten bins by predicted likelihood of default, from the highest-risk clients in the first bin to the lowest-risk in the tenth.
A falling share of “bad” clients from bin to bin is the result that matters for lending: risk is concentrated where the model says it is, fewer good clients are misclassified as bad, and approval and limit decisions can follow the bins directly.
Team
- 1 Data Scientist
Results
- GINI around 40 and AUC around 70% — near the target, with room for refinement through more data or tuning.
- Clear risk separation across score bins supports lending decisions.
- The best AUC among the contractors engaged by the client.
- Delivered on schedule and within the agreed budget.
Postdata delivered a model with a GINI score of around 40 and an AUC of approximately 70%. The team consistently delivered top-quality work on schedule and within the agreed budget.
Project details
- Client
- MD Finance
- Industry
- Fintech, Lending
- Region
- Romania
- Services
- Data science, Machine learning, API development
- Technologies
- Python
- LightGBM
- Bayesian optimization
- Docker
- Flask
- REST
- Team
- 1 data scientist
- Duration
- 1 month
