You can give a practical example for regression? I know this document but I failed to get any improvement using a lot of timestamps.
It\s not entirely clear to me what you mean by practical example. Is it about how it works or is it about what its used for?
Anyhow finding good (low phase noise) signal source is harder than the implementation itself.
This can be one very important factor why you cannot see improvement. Already 11 digits/s requires a good low phase noise SC cut OCXO (none of my TCXOs are good enough for this; OCXOs are OK).
I don't know what kind of reference and signal source of the author of that paper has used, as flicker free 12 digits/s requires clean signals that in my view are not easily obtainable on hobbyist level (I'm not willing to spend big amount of money on this). Let's not consider measuring own reference here as it can be misleading.
As for the implementation: Like described in the paper, huge amount of timestamps are needed for the measurement duration, let's say for 1 sec. If only one measurement is done, the time difference between the signal edge end the reference edge can be measured by the steps given by the TDC7200 that is ranging somewhere 54-57ps (actually the difference of 2 of such timestamps need to be used), influenced also by noise (that can kick some of the TDC measurement to near neighbor bin). That's the resolution of the measurement.
If time-stamping is continuously done and the signal and the reference is not phase coherent, there will be semi-random time differences between the signal edge and the reference edge. Those are still measured in discrete steps due to the TDC limitation but when statistics of those sufficiently distant (corresponding to measurement nominal duration) timestamps are done, it can resolve beyond the discrete step possible by the TDC.
Several commercial/branded frequency analyzers that have time interval resolution only somewhere between the 20-100ps range can do significantly better resolution due to statistics. The single timestamp resolution of FA2 is significantly worse, thus the statistics improved resolution also doesn't come close to the state of the art equipments. For hobbyist use however in my view it's still good value (IMHO actually beyond the need of most hobbyist and even TDC7200 based ones are pretty good without any statistics), especially that hobbyist typically don't have access to very good phase noise signal sources.
As for practical use it allows easy and quick analyses of most signal (or frequency reference) sources.