How to choose an AAD tool
Five questions decide whether an AAD library will hold in a bank's pricing and risk code, and each one you can check yourself before anything is signed. Speed matters too, but it depends on your code, so measure it there: we measure on your kernel; run any other tool on the same one.
Five questions to ask of any AAD tool
Source you can read
Ask: can our quants and model validation read every line that computes our sensitivities?
XAD: the full source and its history are public on GitHub, under AGPL-3.0. Nothing that computes a derivative is hidden in a binary.
XAD on GitHubEvaluate it inside your company, before any contract
Ask: can we try it on our own code, inside the bank, before any contract?
XAD: open source under the AGPL, so you can evaluate it and use it for research: pip install xad, or add it to a C++ build with CMake. No form, no licence key.
Public tests
Ask: are the tests that check the derivatives public, so we can run them ourselves?
XAD: its unit tests are public and stay public in the open-source repository. With a subscription, the XAD validation pack adds edge-case, stress and performance-regression suites and a validation report for your model-risk team.
The validation packSecond order by AAD
Ask: are gammas and cross gammas computed by AAD, or by bumping first-order sensitivities?
XAD: higher-order derivatives by AAD, nesting its forward and adjoint modes, in C++ and Python.
What you keep if the subscription ends
Ask: if we stop paying, what still runs, and who can maintain it?
XAD: the library stays open source, with its full source, history and tests, and the code we write for you is yours. What ends is the commercial licence, updates, support and the validation pack; what the AGPL requires is set out in its licence terms.
Procurement and continuitySpeed, measured on your code
A published timing measures someone else's workload. The figure that counts is the one on your pricer, your compiler and your hardware.
Our answer is the free speed-up estimate: we measure on your kernel; run any other tool on the same one. It runs inside your bank, with our engineer and your team, and you keep the report.