Saying that your future estimates will improve based on your last ones, is like saying you have more chances of winning the lottery if you lost some games before.
Critics of estimation often use a sneaky trick called the motte and bailey fallacy. They mix up two different ideas: one that's pretty reasonable and easy to defend, like "we can't know _exactly_ what's going to happen in the future," with another that's way out there, like "the future is a total mystery, so trying to learn anything about it is useless because it's all just random noise."
"Critics of estimation" usually do this because managers don't actually know the concept of "estimation" and define your worth and paycheck on this concept. So it's just easier to simplify it to "just random".
Past experience can be used to drive good estimation. But it takes experience with the process of estimating to know when you "know" that your experience is relevant, and when it's not clear if the previous experience translates.
It doesn't matter how much experience you have with past estimates if one of the steps in the flow involves a scenario where you just can't estimate it. (Happens a lot when dealing with third parties)
I think for that scenario, there would have to be experience estimating based on the third party involved.
We have third parties building portions of our systems and for some of them I can estimate reasonably well their actual net duration because I've had a lot of experience with them, and they've been fairly consistent.
But there are others where there's a wide range, so I think it boils down to whether you can spot a pattern or not.