Hi Rui,

Thanks for the input!

That's an interesting take. Yes, I believe that, in the scenario you describe, taking the log might not be the best approach.

If there's a pattern in the way variance changes over time, you can use ARCH / GARCH approaches. These model the variance explicitly. Perhaps weighted regression could be a valid solution as well.

Hope this helps!

In my next article, I will explore this exact topic (how to deal with heteroskedasticity) in more detail.