A crucial point about macro-comparison is data. This sounds fairly obvious. But what is the use of comparing languages when primary data are not even well-described and compiled.
For example I've been comparing Hurrian, Hattic, Kassite and Elamite. They seem more related than is usually considered. But, there's a but. There exists no lexical compilation of these languages that would enable to compare these languages. This sounds to me as a necessary first step.
The issue is not just "reaching down" or the like, but reaching too big and reaching haze.