HaymiG/astu-community-financial-inclusion-hackathon

ML pipeline predicting bank account ownership across African countries using CatBoost, engineered features, and cross-validated ensemble averaging. Built for the Zindi Financial Inclusion challenge. # Bank Account Prediction — CatBoost Classifier A machine learning pipeline that predicts whether an individual has a bank account, built for the Zindi financial inclusion dataset. Uses CatBoost with 10-fold stratified cross-validation and feature engineering. --- ## Requirements ```bash pip install pandas numpy catboost scikit-learn ``` --- ## Dataset Expected at `/kaggle/input/datasets/haymig/zindi-dataset/`: | File | Description | |---|---| | `Train.csv` | Labeled training data with `bank_account` target column | | `Test.csv` | Unlabeled test data for submission | **Key columns used:** - `bank_account` — target: `Yes` / `No` - `cellphone_access`, `gender_of_respondent` — binary features - `country`, `location_type`, `relationship_with_head`, `marital_status`, `education_level`, `job_type` — categorical features - `household_size`, `age_of_respondent` — numeric features - `uniqueid` — row identifier (used in submission, dropped from features) --- ## Pipeline Overview ### 1. Encoding - Target `bank_account`: `Yes → 1`, `No → 0` - Binary columns (`cellphone_access`, `gender_of_respondent`): mapped to `1/0` ### 2. Feature Engineering | Feature | Formula | Rationale | |---|---|---| | `edu_job_combo` | `education_level + "_" + job_type` | Captures interaction between education and employment | | `household_pressure` | `household_size / (age_of_respondent + 1)` | Proxy for financial dependents relative to age | | `avg_age_country` | Mean age per country group | Country-level demographic baseline | | `age_gap` | `age_of_respondent − avg_age_country` | Individual deviation from country average age | > `avg_age_country` is used only to derive `age_gap` and is dropped before training. ### 3. Model — CatBoostClassifier | Parameter | Value | |---|---| | `iterations` | 2000 | | `learning_rate` | 0.02 | | `depth` | 7 | | `l2_leaf_reg` | 12 | | `bootstrap_type` | Bayesian | | `early_stopping_rounds` | 100 | Categorical columns are passed natively to CatBoost (no …