Aug 24, 2017 · 1 min read
Quick Answer :
The GB got kind of an “artificial” memory through the manual features we created : The features were moving averages of the target with different “windows” and therefore we kept a “manual history” for the gradient boosting algorithm that it wouldn’t have been able to get otherwise.
PS : The AUC scores are really close (with the LSTM slightly better), that’s why we rely on other metrics such as the reliability score to evaluate the models.
