Tips on reducing regression bias?

Hi folks, just want to check if anyone has any pointer or encountered similar issues. So I’m trying to use keras to do regression on a dataset. And right now I have a pretty deep sequential model (batchnorm+conv1d+dense). However it always seems to have a strong bias (see example image below). I’m using ‘mean_absolute_percentage_error’ as the loss function. Wondering if there’s any tips to reduce the bias? Wondering if there’s any general tips in terms of the architecture engineering, or what issues I maybe overlooked. Thanks!