mandfr19/data.org-Financial-Health-Prediction-Zindi

Solution for the Zindi Financial Health Index Prediction Challenge — Southern African SME classification using XGBoost, LightGBM, CatBoost and MLP ensemble # Financial Health Index (FHI) Prediction ## Zindi Competition — Southern African SME Financial Health Classification --- ## Results | Split | Score | |-------|-------| | Public Leaderboard | 0.8847 | | **Private Leaderboard** | **0.8860** ✅ | --- ## Overview This repository contains the solution for the **Financial Health Index Prediction Challenge**, a Zindi competition focused on predicting the financial well-being of small and medium-sized enterprises (SMEs) across four Southern African countries: Eswatini, Lesotho, Zimbabwe, and Malawi. The task is a **multiclass classification problem** — predicting whether a business has **Low**, **Medium**, or **High** financial health — evaluated using **Macro F1 Score**. --- ## Problem Statement Traditional measures like revenue or profit do not fully capture an SME's financial well-being. This competition introduces a holistic **Financial Health Index (FHI)** — a composite measure reflecting resilience, savings habits, and access to finance across four dimensions: - Savings and assets - Debt and repayment ability - Resilience to shocks - Access to credit and financial services Participants build machine learning models to predict FHI using socio-economic and business survey data. --- ## Repository Structure ``` ├── winning_solution.ipynb # ⭐ Main solution — best private LB (0.8860) ├── experiments.ipynb # Experimental notebook — additional approaches explored ├── README.md ├── VariableDefinitions.csv ├── requirements.txt ├── .gitignore └── outputs/ ├── submission_final.csv ├── experiment_log.json └── README.md ``` ### Solution Files | File | Description | Private LB | |------|-------------|------------| | `winning_solution.ipynb` | **Main solution** — clean preprocessing, feature engineering, 3-model ensemble (XGB + LGB + CAT), Optuna tuning, SMOTE, threshold optimization | **0.8860** ✅ | | `experiments.ipynb` | Extended experiments — pseudo-labeling, target encoding, MLP ensemble, additional f …