Frederik -

So this came from a simple linear regression, y = b_0 + (b_1)x + e, or just y = bx + e, correct? Well, guessing, first, was a variable transformed, say perhaps a log of another variable? That might be part of the reason for this. However, second, I wonder if you have one or more omitted variables of importance. As I recall, I have see data like this which was on a 3-D plot when you only looked at one independent variable, rather than both of them (in such a case). You might check some of the more extreme cases, if you can, and see if there is another important variable influencing these results. It would be a variable that only has large impact in some cases.

Also, once you figure out that, if you then have good predicted y-values, without substantial model or data issues, you may have a better chance to have heteroscedasticity, which is a feature, not a bug. See https://www.researchgate.net/project/OLS-Regression-Should-Not-Be-a-Default-for-WLS-Regression, and updates.

Cheers - Jim

What if you show us the standard plot of Actual versus Forecast?

By "uni-variate" if you mean not multivariate, so you just have one dependent variable and any number of independent variables, note that you want the best set of independent variables - which at the least means not too many and not too few. The model you are using and a definition of the variables might help here, if you are allowed to show this.