I have a question about the research project I’ve been working on. I am training a random forest to predict the Curie temperatures of magnetic materials and I ran into something interesting. When I plot the error between the actual temperatures and the predicted ones against the actual temperatures, it appears that there is a bias in the data where the model overestimates on materials with a low Curie temperature and underestimates on materials with a high Curie temperature. I assumed this bias came from my model which would have been and easy fix, but when I made the same plot with the predicted Curie temperatures instead of the real ones, the bias doesn’t show up and it appears like the model is functioning how it should. My research advisor and I concluded that even though the predictions are correlated with the real values and the real values are correlated with the bias, the predictions are not correlated with the bias. I just wanted to ask if this reasoning makes sense and see if any has any ideas about what is going on or a better explanation of what this means.