We are UET 杨蓉 who achieved 13th place in WiDS Global and 1st place in WiDS Hanoi. The below write-up demonstrates our performance in the WiDS competition.

• Removed all of the physical model forecast features, including the word "prate", "tmp2m".
• After that extract feature using CatBoost feature importance.

• We use K-Fold for train/validation strategy with K=5

• We train the dataset using CatBoost model with the below hyperparameters:

{ 
      'iterations': 24000,
      'learning_rate' : 0.1,
      'subsample' : 0.75, 
      'max_depth’ : 6,
      'use_best_model' : True, 
      'loss_function' : 'RMSE'
}

• The result in LB is 0.729 RMSE.

• Given that the train dataset time occurs during the El Nino phase, the temperature is above average. The test dataset time occurs during the La Nina phase, so the temperature is higher than usual. The predicted temperature will be higher, which is obvious.
• Accordingly, we adjust the temperature based on the report from Western Regional Climate Center. For some privacies, we must not share the coefficient of temperature modification.
• The final result in LB is 0.693 RMSE.