AFAIK, Size of raw data is not stored in any metadata/tsidx, so only option would be to query raw data. May be run for a smaller period to avoid very long running query.
your base search | eval size=len(_raw) | stats avg(size)
Yep. Event size was important to my system at one point so I set-up an accelerated data model using the same eval you have shown above. With the ADM it's easy to grab stats based on sourcetype, source, index and/or host. Once the need passed, I disabled the acceleration.