Hi everyone,

I've just open-sourced my complete machine learning pipeline for this challenge, currently scoring 0.8955 on the public leaderboard.

The approach focuses heavily on mathematical rigor, handling structural zeros, and contextualizing the MSME data across different regional economies. Here are a few highlights from the pipeline:

• Robust Feature Engineering: Custom composite metrics (financial access, informal access, and stress scores) and cross-interactions like stress_x_expense_ratio.
• Country-Specific Context: Financial features are log-transformed and scaled dynamically by the mean/std of each business's operating country, alongside country-percentile rankings. Segmented models are also spawned for countries with sufficient sample sizes.
• Calibrated Ensemble: A soft voting classifier combining heavily tuned LightGBM, XGBoost, and HistGradientBoosting models, balanced with SMOTE and calibrated via Isotonic Regression.

You can check out the full Google Colab notebook and preprocessing strategy here:

🔗 https://github.com/JONNY-ME/data.org-Financial-Health-Prediction

If you find the code or feature engineering strategies helpful for your own models, please consider dropping a ⭐ on the repo!

Good luck to everyone in the final stretch!