At best it is of minimal value in these situations.

Take downvotes for example. A number I saw recently was 0.008% of votes being downvotes, essentially zero. Running a simulation over that data will tell you nothing about downvotes because (for practical purposes) no one uses them. Only after the cost structure attached to downvotes changes will, possibly, the usage of downvotes change, and nothing in the historical data will tell us how it will change or how much.

HF20 is a very different type of situation. Very little of HF20 was intended to or could reasonably be expected to change behavior via incentives on a widespread scale. What it did do is block certain actions (spamming mostly) which meant that it wouldn't even be possible to simulate in that way, because many of the previous recorded actions in the history would be blocked, resulting in a chain state from that point forward deviating from the historical state. Many subsequent actions in history would then become invalid, leading to further rejections and deviation.

One must use the right sorts of tools in any situation. Historical replay as you suggest is the right tool for some problems and the wrong one for others.