There has been several papers improving the original Hagan's approximation formula (see this answer) to SABR model. At least, I know three below:
Is there any research comparing the accuracy of them? I assume that the error can be measured from the option value solved with finite difference method?
Also, is there finite difference implementation openly available?
