# How to avoid the alternating conditional expectations from transforming most of the response data to very close values?

I am applying alternating conditional expectations (ACE) in a forward stepwise manner, similarly to the authors of the original ACE paper.

My dataset has 103 predictor variables $$x_i$$, one response variable $$y$$, 255 samples. The distribution of response values is as follows:

I am trying to select those predictor variables that maximize $$R^2$$. However, ACE finds the transformations of the response variable something like this:

Even though $$R^2=0.83$$ here, the transformation of $$y$$ lumps most of the data near zero. So, if I zoom the figure on the right near zero, I get very small correlation.

How to avoid it? How to make ACE transform the response variable closer to linear?