African Credit Scoring Challenge
Helping Africa
$5 000 USD
Completed (~1 year ago)
1959 joined
1022 active
Start
Nov 29, 24
Close
Jan 12, 25
Reveal
Jan 13, 25
32nd Solution notebook
Notebooks · 14 Jan 2025, 16:32 · 4
Hello Zindians! 👋
I’m glad to share my solution to this Competition, which was actually my first competition in almost 2 years! it was a month fill of fun, failures, and lots of learning.
You can check out the full code here: https://www.kaggle.com/code/thelastsmilodon/zindi-s-africa-credit-challenge-solution . 🎉
For my solution I got 32nd place but could have gotten 16th if I had trusted my Cv more (a lesson well learned). My approach consisted of training four base models (XGBoost, CatBoost, LightGBM, and MLP ) and blending them using a stacking ensemble. No Post-processing or "magic" was used.
- What worked :
1. Feature Engineering : this was probably the most impactful step, the original features had very low correlation to the target but adding features like repayment_ratio = Total_Amount_to_Repay/Total_Amount had a high correlation (0.64) with the target. 2. Adding economic indicators data 3. hyperparameter optimisation using optuna 4. Stratified group (by customer_id) K-Fold CV 5. Feature Selection (inspired by https://www.kaggle.com/code/prashant111/comprehensive-guide-on-feature-selection/notebook)
- What didn't work :
1. target encoding
2. SMOTE
3. SVM and Tabnet and other models
4. post processing (changing the 0.5 threshold worsened score)
Congratulations for your good performance and thank you for sharing.
Thanks and congratulations to you too .
Would love to hear what worked for you too.
I did a lot of feature engineering ( mostly financial ratios, i even divided the debt rate by the duration ), i mostly worked with random forest classifier, probability threshold adjustement after predict_proba results to capture more true positives cases. I didn't have time to tune several models and then perform blending/stacking. This competition was tough for me but i'm still a beginner, i will continue to learn.
Congratulations to you! Unfortunately, the public/private leaderboard dilemma cost us dearly. Our biggest mistake was focusing too much on the public leaderboard, which resulted in a disappointing 106th place on the private leaderboard. What’s even more frustrating is that we had a submission capable of securing 31st place, but we overwrote it due to not saving our trials properly. This lack of organization ultimately cost us a much better rank.