My understanding is that 4 is a relatively sweet spot for allowing efficient consing/deconsing of the finger nodes. I became fascinated by fingertrees a few years ago and created several toy implementations. I recall that the size of the fingers did impact the performance of various operations though I no longer remember.

Sadly I do recall that the constant factors were prohibitively high which made finger trees rather unsuited as an every day list/array data structure for most applications. It would certainly simplify programming to have one sequential collection type that is adequately performant at everything.