[Quantlib-dev] Adjoint Greeks

[Peter C. <pca...@gm...>]

Hello all,

happy new year.

I revisited Ferdinando's comments on adjoint greeks during our
December workshop and started to play around with that idea.

The approach I am trying to follow is to adapt the ql library code so
that automatic differentiation _tools_ can be used with it in a
transparent way. This is opposed to writing special adjoint engines by
_hand_ like e.g. advocated in Capriotti, Giles, Algorithmic
Differentiation: Adjoint Greeks Made Easy. The relatively small and
homogeneous code basis of ql seems to allow for this kind of more
fundamental approach.

I wrote a bit about my first steps in my blog

http://quantlib.wordpress.com/

and forked a new branch from Luigi's current master on github

https://github.com/pcaspers/quantlib/tree/adjoint

where I started to template'ize the library in order to allow for AD
tools to hook in. There are already first working examples (see the
blog) and I am starting to feel confident that the approach might work
as a whole, might be doable in a reasonable amount of time and is
worthwhile following.

About the feasibility: The library seems to consist of roughly 376k
lines of code currently (all hpp and cpp files under ql / ). From that
we can subtract "data" files

78862 ./math/randomnumbers/sobolrsg.cpp
21376 ./math/randomnumbers/primitivepolynomials.cpp
14495 ./math/randomnumbers/latticerules.cpp
10115 ./experimental/volatility/noarbsabrabsprobs.cpp

which leaves us with 251k lines. It seems that I have already reviewed
and adapted around 14k lines, which is 5% and which took me
approximately 60 hours. This gives an estimation of 130 person days
still left to do. For the whole (!) library where already parts will
make much sense and give interesting applications. E.g. excluding
experimental classes (90k) and the market model (25k) reduces the
estimate already to 65 person days to go.

I would be interested in your opinions on that, in particular
regarding the design choices to make (better now than later :-) ).

I'd also be grateful for people supporting the development by forking
the adjoint branch and sending pull requests with adapted code pieces.
My personal next steps would be
- rate deltas for Legs / Swap instruments
- rate vegas for vanilla interest rate options
- Hull White model

What do you think ?

Thank you
Peter