I am working on a counterfeit medicine sales prediction regression model. As the relationship between target & response variables is non-linear I used tree based regressors random forests and XGB. The mse on validation data for both RF & XGB are similar but since XGB overfitted less I went ahead with it.
Then I calculated permutation feature importances & selected the important features.
The problem in the model is it is consistently under predicting high values of target variables as shown in the image.
Below is the distribution of target variable/sales.It is quite skewed but am not sure if that's the reason for bad predictions for high sales values. Can anybody please guide me how deal with this? Do I need to apply some multivariate outlier removal technique or is there any other solution to this.