Boost Credit Risk Model Performance -- 2
Budget: ₹12,500 – ₹37,500 INR
I have already deployed a full Streamlit application that predicts loan approvals in real time (live demo: https://lnkd.in/dtrRm-Jx, source: https://lnkd.in/dPYEfHgt). The pipeline currently includes Logistic Regression, K-Nearest Neighbors, and Naive Bayes models with standard scaling and the usual EDA-driven feature engineering.
What I want now is a measurable lift in overall model performance, with the F1-score as the guiding metric. Feel free to explore more advanced algorithms (e.g., Gradient Boosting, XGBoost, LightGBM, calibrated ensembles, or even a tuned version of my existing classifiers) as long as they integrate cleanly with the existing Python | Pandas | NumPy | Scikit-learn stack and can be surfaced through the current Streamlit front-end.
Key points you should address
• Re-examine preprocessing and feature selection only if it directly supports a higher F1-score; the interface and general UX can remain untouched.
• Provide well-commented, reproducible code and a concise notebook or markdown explaining your methodology, hyperparameter tuning strategy, and why the new model outperforms the baseline on unseen data.
• Update the Streamlit app so users can choose the improved model in real time, then redeploy (Heroku/Streamlit Cloud) or supply clear deployment instructions.
Acceptance criteria
1. End-to-end run on my dataset yields a materially higher F1-score than the current best model.
2. Code passes without errors in a fresh virtual environment using requirements.txt.
3. Updated app link (or PR) demonstrating the new model in production.
If this sounds like a challenge you enjoy, let’s get started.
What I want now is a measurable lift in overall model performance, with the F1-score as the guiding metric. Feel free to explore more advanced algorithms (e.g., Gradient Boosting, XGBoost, LightGBM, calibrated ensembles, or even a tuned version of my existing classifiers) as long as they integrate cleanly with the existing Python | Pandas | NumPy | Scikit-learn stack and can be surfaced through the current Streamlit front-end.
Key points you should address
• Re-examine preprocessing and feature selection only if it directly supports a higher F1-score; the interface and general UX can remain untouched.
• Provide well-commented, reproducible code and a concise notebook or markdown explaining your methodology, hyperparameter tuning strategy, and why the new model outperforms the baseline on unseen data.
• Update the Streamlit app so users can choose the improved model in real time, then redeploy (Heroku/Streamlit Cloud) or supply clear deployment instructions.
Acceptance criteria
1. End-to-end run on my dataset yields a materially higher F1-score than the current best model.
2. Code passes without errors in a fresh virtual environment using requirements.txt.
3. Updated app link (or PR) demonstrating the new model in production.
If this sounds like a challenge you enjoy, let’s get started.