How to choose a model from three approximately equal in ROC?

Question

Machine learning

Classification problem. Trained three models Random Forest, XGBoost, CatBoost. All three models show approximately equal ROC and Accuracy, but at the output I need probabilities, not a predicted class, and here problems arise. For example, on one of the model samples, the probabilities xgb: 0.38, fr: 0.22, catboost: 0.31 are given. As you can see, the spread is decent. Can these probabilities be averaged? What techniques are commonly used? As far as I understand, it is necessary to test the model on combat data and the one that gives the best result in terms of EV, for example, choose that model as the final one?

Answer 1

In general, now the trend is to combine models.
A panel of experts works better than each individual expert.
Get rid of this dogma.

Answer 2

Google stacking and bagging.