In the literature you see small margins because many of the things that really matter in practice (data quantity, data quality, engineering) aren't generally of academic interest.