What we have to do now is gather signals we can draw from different models into one final prediction, so we can select from Middle East customers a list of probably defaulted customers to be sent to the internal audit team.

How would you seek this objective? Imagine you are chilling out with your friends at a pub when you suddenly start talking about your next vacation. You tell them that you are probably going to choose Austria as a destination, even if you don't know if your are going to like it.

Your friends, who really care about your well-being and love you, start sharing their opinions with you. One of them says you are not going to have a good time there because of the cold, one says you are not going to like it because of the humidity, and this goes on until all of your friends have shared their opinion with you. You now know that five of your friends think you are not going to have good time in Austria, while three think you are going to have a great vacation. What would you conclude?

The first possible idea could be to conclude that you are not going to love Austria because the majority of your friends think so. But then, you could probably start thinking that not all of your friends know you in the same way. There are some that really know your tastes, while others do not. Should you evaluate equally the opinion of both groups of friends, or give more relevance to the friends that know you better? Those reasonings are the foundations of ensemble learning techniques. Let's now have a more formal view of them.