Due to the nature of how data is written onto HDDs and how each file system works. Getting an accurate ETC on massive file/directory transfers is like trying to seer a 5-day weather forecast.
And yet 5-day forecasts are exponentially better
than they used to be. "A modern 5-day forecast is as accurate as a 1-day forecast was in 1980, and useful forecasts now reach 9 to 10 days into the future."
When I took Atmospheric Physics classes in the early 80s, the department had a forecasting competition in which people could use any technique they wanted. The guy who won most consistently always forecast tomorrow to have exactly the same weather as today. He was wrong a lot, but he was still right more often than the people who were using rudimentary numerical models. Things have improved markedly since then. Beyond three days out was a complete crapshoot in 1980; today the forecast is right far more often than it's wrong -- and nobody was even talking about meaningful 10 day forecasts back then.
There are still failures, of course, and that's what people remember -- not the hundreds of times the forecasts got it right. And there are still difficult places and conditions, of course: the coastal Pacific Northwest (for example) has a complex local geography that creates many micro-climates, and in the winter it's right on the edge of the rain/snow line, so predicting snowfall amounts can be wildly variable (it doesn't help that the weather comes out of the north-east pacific, where non-satellite data-gathering is sparse). And doing small-scale prediction for things like tornado tracks is incredibly difficult as well, and yet they've made massive strides in that also. Hurricane prediction is good enough now that it has saved countless lives and billions of dollars (and that's enough to pay for the entire weather prediction apparatus).
Which is to say: if Microsoft wanted to throw a lot of money and modelling (and these days deep learning) at the "how long is a copy going to take" problem, they almost certainly could solve it for most use cases. But it's hardly worth it. And they would still get it wrong occasionally, which is the only thing people would notice or remember.