Data prep and understanding it well is indeed the key. I actually use only one algo for data like this: XGBoost :D. I absolutely hate doing massive ensembles and stacking of 100 models which forces me to do better in feature engineering to get a better rank.