Test train_auc scores


#1

Hello,
I am using XGBoost to build mortgage prepayment models trained at the loan level. Early stopping is based on AUC scores. The first 18 training rounds are 0.50 and 0.50 test and train, respectively. To get around this I set early stopping at 20. This solves my immediate problem of early stopping. The model trains well achieving an auc of 0.70 - which in this case is very good. To finish the code I think I need to get past the 0.50 and 0.5 auc in earlier rounds. Does anyone have a suggestion as to how achieve something like this?

Glenn


#2

Hi Glenn,

I am not sure I fully understand your post. But I think in may be helpful if you add 'base_score’: 0.70.

The base_score is set at 0.50 by default. The starting base score should not make a lot of difference if you have a fairly small eta. But if early stopping stops you at 20 rounds your eta is probably not small.

So maybe changing the base score to 0.70 could help. Maybe look at the eta too.

Sorry if I completely missed your question but maybe this will help some.

-Jim