I am using XGBoost (gradient boosting) to predict the value of a continuous dependent variable

The figure below shows a histogram of both the dependent variable data and the predicted data. (blue is the original dependent value and orange is the predicted).

enter image description here

Are there any methods/literature that talk about proper scaling or adjusting/mapping the output of a regression so that it can better relate to the original dependent variable?

For example the first histogram bar (from 0-0.25) for the dependent variable, what values of the predictions approximately correspond to that range?

For the sake of context the data is for credit risk scoring, for example 0 represents someone who has never made a payment and is in complete delinquency and 3 is someone with perfect credit. Then there are the individuals in between.

  • 1
    $\begingroup$ It seems like your question should be if it's possible to improve your predictions (from the accuracy standpoint) for such type of distribution of target variable, rather than doing something after you obtain model output. Do you agree? If so, I suggest you post your xgboost code (parameters) and also say if you've tried to perform any transformations on the target. $\endgroup$ – AlexK Apr 3 at 21:37
  • $\begingroup$ I second @AlexK, and on that note, there's a ton of tutorials on how to build credit scores using machine leanring, some references here. $\endgroup$ – Lucas Farias Apr 3 at 22:32

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.