Hi everyone,

I’d like to share a couple of important points that could improve the fairness of this competition and help reduce potential leakage:

1. **Cross-validation split method**

Since we are working with sequential **weather data**, shuffling in cross-validation introduces leakage across time.

Instead of:

```python

skf = StratifiedKFold(n_splits=5, shuffle=True, random_state=42)

```

we should use:

```python

skf = StratifiedKFold(n_splits=5, shuffle=False)

```

This respects the time order and avoids information from the future leaking into training folds.

2. **Evaluation metric**

The target is highly imbalanced (**≈88% NORAIN**). With this imbalance, **micro** and **weighted F1** tend to overstate performance because they are dominated by the majority class.

Using **macro F1** would provide a fairer assessment, as it balances the importance of all classes (including MEDIUMRAIN, HEAVYRAIN, and SMALLRAIN).

---

I hope the organizers will take these points into consideration to ensure more robust and fair evaluation.

Thanks!