One of my university professors advised "multiply your best-case-scenario estimate by Pi". It's still absolutely plucking numbers out of the air, but I've found it to be pretty reasonable over the last decade or so...