data.org Financial Health Prediction Challenge
Helping Eswatini, Lesotho
and 2 other countries
and 2 other countries
- Eswatini
- Lesotho
- Zimbabwe
- Malawi
- Scroll to see more
$1 500 USD
Under code review
Prediction
Machine Learning
1686 joined
898 active
Start
Dec 12, 25
Close
Mar 15, 26
Reveal
Mar 16, 26
'/%3e%3cuse xlink:href='%23a' x='-108' transform='scale(-1 1)'/%3e%3c/g%3e%3cuse xlink:href='%23b' transform='rotate(72 54 60)'/%3e%3c/g%3e%3cuse xlink:href='%23b' transform='rotate(-72 54 60)'/%3e%3cuse xlink:href='%23c' transform='rotate(144 54 60)'/%3e%3c/g%3e%3cuse xlink:href='%23d' x='72'/%3e%3c/svg%3e){kind=small}
Share your private LB score and methods (34th on private LB)
19 Mar 2026, 11:47 · 5
Hey everyone, I would like to know what distinguished winning solutions, though I see no huge differences between contestants.
I ranked 34th on private LB, my method is simple: Train weighted ensemble of Extra trees and random forests with different seeds and slight variations in paramteres using 5 fold cross validation. Standard data preprocessing. And that's it.
I am 10th on private LB, I used MLP + multihead attention DAE. Feature engineering was important, I created some stuff like discipline score, insurance score, etc. I also pruned some features that were degrading performance.
wow didn't expect that to work so well, congrats on the 10th place !
I have a private LB solution 0.88427176 that would be in top 50 and a lot of from 0.883 to 0.884 solutions but unfortunately there was a system error that didnt allow me to change my selected submission. All these where with xgboost OOF stacking .Other solutions leaded my to overfitting.Maybe each country in the test dataset had very different high/medium/low levels percentages and due that xgboost try to find the best average solution ,learned patterns that were not generalized.
sorry to hear that, yeah the target has imbalanced splits, that's why I tried training my models on the most 2 frequent classes, if I recall it correctly, it was the low and medium if I am not wrong, all samples labeled 'high' were transformed to 'medium', but that didn't help
I placed 47th on the private leaderboard. My best-performing model was a weighted ensemble of Extra Trees (0.40), XGBoost (0.45), and Random Forest (0.15). This ensemble achieved a weighted F1 score of 0.8777 and a macro F1 score of 0.8190. On the competition leaderboard, it recorded a public F1 score of 0.902854951 and a private F1 score of 0.884097421.
The main challenge was the severe class imbalance, particularly in the “High” class, which represented only 4.89% of the dataset. As a result, performance on that class was lower, with an F1 score of 0.74, compared to 0.93 for “Low” and 0.79 for “Medium.”