Is Fortran really that much quicker than Python+Numpy? I'm genuinely interested. I do large data set numerical stuff with Python and I find the performance is pretty good. As an example, I can z-score an array of 40 million data points in less than 2 seconds.
There is also decent support for linking in more high performance stuff, using Cython or Swig. I'm also not convinced that the execution speed benefits of a dated language like Fortran stacks up against the convenience of programming in Python. I would not want to manage the kind of complexity and variety I need to deal with in a language without modern OO support.
I'd rather it took a few weeks to write and a few days to run than a few months to write and a few hours to run. I guess this depends how many runs you expect to do.