If we go ahead and train our model on the sample data which has the wrong proportions it is likely that the model will be over-fitted to the training data and it is also likely that when we run the model against real-world or testing data that is in the right proportions it will underperform.