How to deal with Errors (Regression) which are not following Normal distribution?

model-selection

#1

Hello Guys,
I am working on Model Selection problem, where there are two models which are predicting revenues for companies using 2 different formulas. So currently I have Actual revenue values and predicted revenue values. By using this data I have to do model selection. So first of all I have tried with various performance measures of Regression such as Explained Variance Score, Mean Absolute Error, Mean Squared Error , RMSE etc. these were giving weired values. So have removed outliers. It worked. But while checking distribution of errors I come to know that its not following Normal Distribution. So no use of applying above performance measures.

So now I am in confusion that how I should deal with this type of data. How to deal with Errors (Regression) which are not following Normal distribution? Is it okay if I will scale predicted and actual values to normal distribution? Please help… Thanks in advance

Vardhini Kundale


#2

Hi @Vardhini_Kundale,

You can scale the target variable to make its distribution normal. Just make sure while calculating the error or calculating the RMSE, scale the predicted values as well.


#3

@PulkitS Thanks for your Reply.

Okay. But will Scaling Targets ( Actual as well as Predicted) make errors to follow Normal Distribution?


#4

Hi @Vardhini_Kundale,

There is no clear-cut answer to this. You just have to try scaling the variables and see the results.


#5

Okay. thanks