It seems that when moving, pathfinding first do horizontal and vertical moves (2 TU), then diagonal moves (3 TU). I recall old X-COM do the opposite, and ends up more acurate to guess which direction the player wants the actor to be facing on the end of the movement.
Doing X-COM way would be an improvement, but I sugest a further improvement.
Pathfinding could count how many diagonals (3 TU) and straigh (2 TU) movements the actor will make (between obstacles and after last obstacle of course) and do the kind of movement that appears most on that movement section by last, thus "guessing" better the direction player will want the actor to face on the end of the turn. In case of a tie, end up with diagonal one.
This will make actor look a lot smarter, because the kind of movement that occurs on most of the pathfnding is generaly the direction actor should be heading on the end of turn.