Nope, @Nimish_Khandelwal.
If the targets are labelled correctly we should not manipulate them by clipping or winsorizing them. But of course you can consider transforming the whole labels (e.g. with log scaling), training the model to predict your transformed label, and then revert the transformation to get to your actual label.
I understand the labels are scattering which cannot be explained by the model, but that is the reality then.
Can you show how the model residuals look like?
(See also this thread: True vs predicted values biased-intercept - #4 by Christian_Simonis)
In general points to consider are:
-
is the scattering of your label caused by the fact they they emerge from different distributions (e.g. let’s assume the insurance premium might be systematic different for a certain kind of characteristics) in that case you could consider to train a separate model for separate labels (e.g. insurance premium for car owners that just got their drivers license and one insurance premium for more experienced drivers)
-
try to incorporate domain knowledge into your features w/ feature engineering see also this thread: Time Series Linear Regression
-
dependent on your residual analysis consider other models than a linear model (which will only perform great if you can manage to model all the non-linearity in your features). Maybe Gaussian processes might be worth a look since they also allow to account for uncertainty / confidence and model non-linearity, see also this thread: Deep learning is a small part of ai - #6 by Christian_Simonis
Hope that helps, @Nimish_Khandelwal!
Best
Christian