How to make distribution of predicted value similar to training example?

Running regression analysis, for weak relationships, the predicted value tends to cluster around the mean.

Is there any method to force the distribution of the predicted value similar to the training example?

I know it will sacrifice the MSE.