TinevimboMusingadi/indabax_zimbabwe_hackathon

Hackathon: Machine Learning and Artificial Intelligence for Zimbabwe's Banking Sector and Financial Inclusion # IndabaX Zimbabwe 2026 — Loan Default Prediction End-to-end ML pipeline for the Deep Learning IndabaX Zimbabwe 2026 hackathon: *AI for Financial Inclusion — Predicting Loan Defaults in Zimbabwe's Banking Sector*. **Primary metric:** ROC-AUC | **Submission:** `ID,Target` (probabilities) ## Quick Start — Google Colab (recommended) 1. Upload `Train.csv`, `Test.csv`, `SampleSubmission.csv` to `Google Drive > My Drive > indabax_data/`. 2. Open `notebooks/01_train_predict_submit.ipynb` in Colab. 3. Set runtime to **GPU (T4)**. 4. Run all cells. A submission CSV will be downloaded automatically. ## Quick Start — Local ```bash git clone github.com cd indabax_zimbabwe_hackathon pip install -e ".[dev]" # Place Train.csv, Test.csv, SampleSubmission.csv into data/raw/ python -m src.pipelines.run_full --config configs/fast.yaml ``` ## Configs | Config | Models | Optuna trials | ~Runtime | Environment | |--------|--------|---------------|----------|-------------| | `fast.yaml` | LGBM, XGB | 0 (defaults) | ~5 min | CPU | | `colab_t4.yaml` | LGBM, XGB, CatBoost, TabNet, MLP | 50 | ~30 min | T4 GPU | | `full.yaml` | All models | 100 | ~90 min | GPU | ## Pipeline Phases 1. **Data loading** — schema validation, date parsing, missing-rate audit. 2. **Splitting** — 5-fold stratified CV, persisted to `data/splits/`. 3. **Feature engineering** — date deltas, ratios, interactions, cyclical features. Three encoding variants (OHE, ordinal, target+WOE). 4. **Training** — classical ML + gradient boosting + deep learning with fold-level OOF predictions. 5. **Tuning** — Optuna TPE per model (optional, controlled by config). 6. **Ensemble** — rank averaging, stacking, Optuna blend weights. 7. **Submission** — validates format against `SampleSubmission.csv`. ## Reproducibility - **Global seed:** 42 (threaded through Python, NumPy, PyTorch, sklearn). - **Fold IDs** persisted in `data/splits/folds.parquet`. - **Requirements file …