I’m staking on an ensemble of my best models, because building my own meta model has its advantages.
I’ve been calculating my ensemble as
ensemble = (prediction1 + prediction2) / 2
Which is simple, easy and wrong
I recognized it, because the result of the ensemble was always stronger influenced by one model then the other…
I learnt the hard way that different models output their predictions with different mean and standard deviation. The ensemble is then stronger influenced by the model, which has higher mean and/or standard deviation.
I believe the correct way of building an ensemble is
ensemble = MinMaxScaler.fit_transform( prediction1.rank() + prediction2.rank() )
This method should give models equal weight in the ensemble.
If you have a different or better method, please don’t hold it back!