By the way this approach resolves all the issues revolving around regression fitting in log-log space. Is OLS regression better than quantile or Bayesian and so on. @TheRealPlanC

This method doesn't depend on regression at all. It simply starts with the assumption we follow a power law with unknown exponent.

Then we normalize the observed returns by log( (t+1)/t) that is the deterministic diminishing returns component.

These time independent returns then should have a symmetric distribution around n if we truly follow a power law.
Indeed we observe a symmetric distribution that is stable in time.

We can derive n from the distribution parameters.

It is the most robust way to find the power law and everything else is still useful but completely obsolete and less rigorous.