The tuning method that actually converges, rather than changing settings and waiting to be annoyed:
Leave the house and log the sensor for a day with the automation disabled. You want the raw state history, not the lights. Then look at when it reported presence in an empty house and correlate against what was happening — heating on, sun on the window, washing machine spinning, someone in the corridor.
Every false trigger has a cause and the log tells you which. Guessing does not.
The same log tells you whether a change worked. Change one thing, log another day, compare. Sounds slow, but it is a week and then it is done, whereas fiddling reactively goes on for months.
If your sensor exposes a raw distance or energy value rather than only a binary state, log that. It shows you exactly which distance band the phantom is at, which usually identifies the object immediately.