IndabaX 2026: Loan Default
Deep Learning IndabaX Zimbabwe 2026 hackathon (AI for financial inclusion): predicting loan default in Zimbabwe's banking sector. Adversarial validation showed train and test were 91.5% separable, so the work focused on a shift-robust LightGBM / CatBoost / XGBoost ensemble tuned over 50+ cloud experiments on Modal.
- Python
- LightGBM
- CatBoost
- XGBoost
- Optuna
- PyTorch
- Modal
The problem
The Deep Learning IndabaX Zimbabwe 2026 hackathon challenge: predict which loans in Zimbabwe's banking sector will default, when the test data comes from a different time period than the training data, so ordinary validation scores are optimistic.
What was built
- Adversarial validation first: a classifier trained only to tell training rows from test rows reached AUC 0.915, revealing a strong temporal distribution shift before any model was trained.
- A shift-robust ensemble of LightGBM, CatBoost and XGBoost, with frozen 10-fold stratified splits shared across every experiment.
- 50+ experiments and 1,398 Optuna trials run as a cloud pipeline on Modal (A10G GPU and a 16-core CPU), for about $30 of compute in total.
- Reproducible locally in about 15 minutes, with no GPU required.
Result
Public leaderboard AUC 0.6840; private leaderboard AUC 0.6878.
Source: Competition leaderboard, as reported in the project README