Attached are a couple of examples that show my different ideas.
The first one (for the flat) focuses on horse characteristics as the key thing. Fewer than 5 runs suggests an unexposed horse that might particularly benefit from a trainer trying something new. Additional criteria are that the trainer has had a winner in the last 30 days (hopefully removing ones horribly out of form), CD winners (a standard control variable of mine) and odds.
The second one (for NH) focuses on trainers who have done well with new recruits. This has been discussed on other threads, but I looked back at 2011-2019 to test it in 2020, and also made a system with odd years and one with even years, and grouped the ones who did well in both. This was the short list.
Both are guilty of backfitting (which system isn't...) but i'd welcome people's thoughts, as I can see arguments for building trainer change systems around the horse data or around the trainer data.